bitmap.h (bitmap_ior_and_into): New.
[official-gcc.git] / gcc / tree-ssa-pre.c
blob11640a8895e0a783fe6165e10272a8191b8eb4d9
1 /* SSA-PRE for trees.
2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
3 Free Software Foundation, Inc.
4 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
5 <stevenb@suse.de>
7 This file is part of GCC.
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.
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.
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/>. */
23 #include "config.h"
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "ggc.h"
28 #include "tree.h"
29 #include "basic-block.h"
30 #include "diagnostic.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 "real.h"
40 #include "alloc-pool.h"
41 #include "obstack.h"
42 #include "tree-pass.h"
43 #include "flags.h"
44 #include "bitmap.h"
45 #include "langhooks.h"
46 #include "cfgloop.h"
47 #include "tree-ssa-sccvn.h"
48 #include "params.h"
49 #include "dbgcnt.h"
51 /* TODO:
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.
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. */
67 /* Basic algorithm
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.
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.
94 Third, we perform insertions to make partially redundant
95 expressions fully redundant.
97 An expression is partially redundant (excluding partial
98 anticipation) if:
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.
104 In order to make it fully redundant, we insert the expression into
105 the predecessors where it is not available, but is ANTIC.
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.
111 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
112 performs these steps.
114 Fourth, we eliminate fully redundant expressions.
115 This is a simple statement walk that replaces redundant
116 calculations with the now available values. */
118 /* Representations of value numbers:
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).
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. */
135 /* Representation of expressions on value numbers:
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. */
142 /* Representation of sets:
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.
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. */
153 /* Type of expression, used to know which member of the PRE_EXPR union
154 is valid. */
156 enum pre_expr_kind
158 NAME,
159 NARY,
160 REFERENCE,
161 CONSTANT
164 typedef union pre_expr_union_d
166 tree name;
167 tree constant;
168 vn_nary_op_t nary;
169 vn_reference_t reference;
170 } pre_expr_union;
172 typedef struct pre_expr_d
174 enum pre_expr_kind kind;
175 unsigned int id;
176 pre_expr_union u;
177 } *pre_expr;
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
184 static int
185 pre_expr_eq (const void *p1, const void *p2)
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;
190 if (e1->kind != e2->kind)
191 return false;
193 switch (e1->kind)
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 abort();
210 static hashval_t
211 pre_expr_hash (const void *p1)
213 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
214 switch (e->kind)
216 case CONSTANT:
217 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
218 case NAME:
219 return iterative_hash_hashval_t (SSA_NAME_VERSION (PRE_EXPR_NAME (e)), 0);
220 case NARY:
221 return PRE_EXPR_NARY (e)->hashcode;
222 case REFERENCE:
223 return PRE_EXPR_REFERENCE (e)->hashcode;
224 default:
225 abort ();
230 /* Next global expression id number. */
231 static unsigned int next_expression_id;
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;
239 /* Allocate an expression id for EXPR. */
241 static inline unsigned int
242 alloc_expression_id (pre_expr expr)
244 void **slot;
245 /* Make sure we won't overflow. */
246 gcc_assert (next_expression_id + 1 > next_expression_id);
247 expr->id = next_expression_id++;
248 VEC_safe_push (pre_expr, heap, expressions, expr);
249 slot = htab_find_slot (expression_to_id, expr, INSERT);
250 gcc_assert (!*slot);
251 *slot = expr;
252 return next_expression_id - 1;
255 /* Return the expression id for tree EXPR. */
257 static inline unsigned int
258 get_expression_id (const pre_expr expr)
260 return expr->id;
263 static inline unsigned int
264 lookup_expression_id (const pre_expr expr)
266 void **slot;
268 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
269 if (!slot)
270 return 0;
271 return ((pre_expr)*slot)->id;
274 /* Return the existing expression id for EXPR, or create one if one
275 does not exist yet. */
277 static inline unsigned int
278 get_or_alloc_expression_id (pre_expr expr)
280 unsigned int id = lookup_expression_id (expr);
281 if (id == 0)
282 return alloc_expression_id (expr);
283 return expr->id = id;
286 /* Return the expression that has expression id ID */
288 static inline pre_expr
289 expression_for_id (unsigned int id)
291 return VEC_index (pre_expr, expressions, id);
294 /* Free the expression id field in all of our expressions,
295 and then destroy the expressions array. */
297 static void
298 clear_expression_ids (void)
300 VEC_free (pre_expr, heap, expressions);
303 static alloc_pool pre_expr_pool;
305 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
307 static pre_expr
308 get_or_alloc_expr_for_name (tree name)
310 pre_expr result = (pre_expr) pool_alloc (pre_expr_pool);
311 unsigned int result_id;
313 result->kind = NAME;
314 result->id = 0;
315 PRE_EXPR_NAME (result) = name;
316 result_id = lookup_expression_id (result);
317 if (result_id != 0)
319 pool_free (pre_expr_pool, result);
320 result = expression_for_id (result_id);
321 return result;
323 get_or_alloc_expression_id (result);
324 return result;
327 static bool in_fre = false;
329 /* An unordered bitmap set. One bitmap tracks values, the other,
330 expressions. */
331 typedef struct bitmap_set
333 bitmap expressions;
334 bitmap values;
335 } *bitmap_set_t;
337 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
338 EXECUTE_IF_SET_IN_BITMAP((set)->expressions, 0, (id), (bi))
340 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
341 EXECUTE_IF_SET_IN_BITMAP((set)->values, 0, (id), (bi))
343 /* Mapping from value id to expressions with that value_id. */
344 DEF_VEC_P (bitmap_set_t);
345 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
346 static VEC(bitmap_set_t, heap) *value_expressions;
348 /* Sets that we need to keep track of. */
349 typedef struct bb_bitmap_sets
351 /* The EXP_GEN set, which represents expressions/values generated in
352 a basic block. */
353 bitmap_set_t exp_gen;
355 /* The PHI_GEN set, which represents PHI results generated in a
356 basic block. */
357 bitmap_set_t phi_gen;
359 /* The TMP_GEN set, which represents results/temporaries generated
360 in a basic block. IE the LHS of an expression. */
361 bitmap_set_t tmp_gen;
363 /* The AVAIL_OUT set, which represents which values are available in
364 a given basic block. */
365 bitmap_set_t avail_out;
367 /* The ANTIC_IN set, which represents which values are anticipatable
368 in a given basic block. */
369 bitmap_set_t antic_in;
371 /* The PA_IN set, which represents which values are
372 partially anticipatable in a given basic block. */
373 bitmap_set_t pa_in;
375 /* The NEW_SETS set, which is used during insertion to augment the
376 AVAIL_OUT set of blocks with the new insertions performed during
377 the current iteration. */
378 bitmap_set_t new_sets;
380 /* A cache for value_dies_in_block_x. */
381 bitmap expr_dies;
383 /* True if we have visited this block during ANTIC calculation. */
384 unsigned int visited:1;
386 /* True we have deferred processing this block during ANTIC
387 calculation until its successor is processed. */
388 unsigned int deferred : 1;
389 } *bb_value_sets_t;
391 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
392 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
393 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
394 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
395 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
396 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
397 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
398 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
399 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
400 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
403 /* Maximal set of values, used to initialize the ANTIC problem, which
404 is an intersection problem. */
405 static bitmap_set_t maximal_set;
407 /* Basic block list in postorder. */
408 static int *postorder;
410 /* This structure is used to keep track of statistics on what
411 optimization PRE was able to perform. */
412 static struct
414 /* The number of RHS computations eliminated by PRE. */
415 int eliminations;
417 /* The number of new expressions/temporaries generated by PRE. */
418 int insertions;
420 /* The number of inserts found due to partial anticipation */
421 int pa_insert;
423 /* The number of new PHI nodes added by PRE. */
424 int phis;
426 /* The number of values found constant. */
427 int constified;
429 } pre_stats;
431 static bool do_partial_partial;
432 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
433 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
434 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
435 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
436 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
437 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
438 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr, bool);
439 static bitmap_set_t bitmap_set_new (void);
440 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
441 gimple, tree);
442 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
443 gimple);
444 static unsigned int get_expr_value_id (pre_expr);
446 /* We can add and remove elements and entries to and from sets
447 and hash tables, so we use alloc pools for them. */
449 static alloc_pool bitmap_set_pool;
450 static bitmap_obstack grand_bitmap_obstack;
452 /* To avoid adding 300 temporary variables when we only need one, we
453 only create one temporary variable, on demand, and build ssa names
454 off that. We do have to change the variable if the types don't
455 match the current variable's type. */
456 static tree pretemp;
457 static tree storetemp;
458 static tree prephitemp;
460 /* Set of blocks with statements that have had its EH information
461 cleaned up. */
462 static bitmap need_eh_cleanup;
464 /* Which expressions have been seen during a given phi translation. */
465 static bitmap seen_during_translate;
467 /* The phi_translate_table caches phi translations for a given
468 expression and predecessor. */
470 static htab_t phi_translate_table;
472 /* A three tuple {e, pred, v} used to cache phi translations in the
473 phi_translate_table. */
475 typedef struct expr_pred_trans_d
477 /* The expression. */
478 pre_expr e;
480 /* The predecessor block along which we translated the expression. */
481 basic_block pred;
483 /* The value that resulted from the translation. */
484 pre_expr v;
486 /* The hashcode for the expression, pred pair. This is cached for
487 speed reasons. */
488 hashval_t hashcode;
489 } *expr_pred_trans_t;
490 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
492 /* Return the hash value for a phi translation table entry. */
494 static hashval_t
495 expr_pred_trans_hash (const void *p)
497 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
498 return ve->hashcode;
501 /* Return true if two phi translation table entries are the same.
502 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
504 static int
505 expr_pred_trans_eq (const void *p1, const void *p2)
507 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
508 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
509 basic_block b1 = ve1->pred;
510 basic_block b2 = ve2->pred;
512 /* If they are not translations for the same basic block, they can't
513 be equal. */
514 if (b1 != b2)
515 return false;
516 return pre_expr_eq (ve1->e, ve2->e);
519 /* Search in the phi translation table for the translation of
520 expression E in basic block PRED.
521 Return the translated value, if found, NULL otherwise. */
523 static inline pre_expr
524 phi_trans_lookup (pre_expr e, basic_block pred)
526 void **slot;
527 struct expr_pred_trans_d ept;
529 ept.e = e;
530 ept.pred = pred;
531 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
532 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
533 NO_INSERT);
534 if (!slot)
535 return NULL;
536 else
537 return ((expr_pred_trans_t) *slot)->v;
541 /* Add the tuple mapping from {expression E, basic block PRED} to
542 value V, to the phi translation table. */
544 static inline void
545 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
547 void **slot;
548 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
549 new_pair->e = e;
550 new_pair->pred = pred;
551 new_pair->v = v;
552 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
553 pred->index);
555 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
556 new_pair->hashcode, INSERT);
557 if (*slot)
558 free (*slot);
559 *slot = (void *) new_pair;
563 /* Add expression E to the expression set of value id V. */
565 void
566 add_to_value (unsigned int v, pre_expr e)
568 bitmap_set_t set;
570 gcc_assert (get_expr_value_id (e) == v);
572 if (v >= VEC_length (bitmap_set_t, value_expressions))
574 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
575 v + 1);
578 set = VEC_index (bitmap_set_t, value_expressions, v);
579 if (!set)
581 set = bitmap_set_new ();
582 VEC_replace (bitmap_set_t, value_expressions, v, set);
585 bitmap_insert_into_set_1 (set, e, true);
588 /* Create a new bitmap set and return it. */
590 static bitmap_set_t
591 bitmap_set_new (void)
593 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
594 ret->expressions = BITMAP_ALLOC (&grand_bitmap_obstack);
595 ret->values = BITMAP_ALLOC (&grand_bitmap_obstack);
596 return ret;
599 /* Return the value id for a PRE expression EXPR. */
601 static unsigned int
602 get_expr_value_id (pre_expr expr)
604 switch (expr->kind)
606 case CONSTANT:
608 unsigned int id;
609 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
610 if (id == 0)
612 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
613 add_to_value (id, expr);
615 return id;
617 case NAME:
618 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
619 case NARY:
620 return PRE_EXPR_NARY (expr)->value_id;
621 case REFERENCE:
622 return PRE_EXPR_REFERENCE (expr)->value_id;
623 default:
624 gcc_unreachable ();
628 /* Remove an expression EXPR from a bitmapped set. */
630 static void
631 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
633 unsigned int val = get_expr_value_id (expr);
634 if (!value_id_constant_p (val))
636 bitmap_clear_bit (set->values, val);
637 bitmap_clear_bit (set->expressions, get_expression_id (expr));
641 static void
642 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
643 bool allow_constants)
645 unsigned int val = get_expr_value_id (expr);
646 if (allow_constants || !value_id_constant_p (val))
648 /* We specifically expect this and only this function to be able to
649 insert constants into a set. */
650 bitmap_set_bit (set->values, val);
651 bitmap_set_bit (set->expressions, get_or_alloc_expression_id (expr));
655 /* Insert an expression EXPR into a bitmapped set. */
657 static void
658 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
660 bitmap_insert_into_set_1 (set, expr, false);
663 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
665 static void
666 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
668 bitmap_copy (dest->expressions, orig->expressions);
669 bitmap_copy (dest->values, orig->values);
673 /* Free memory used up by SET. */
674 static void
675 bitmap_set_free (bitmap_set_t set)
677 BITMAP_FREE (set->expressions);
678 BITMAP_FREE (set->values);
682 /* Generate an topological-ordered array of bitmap set SET. */
684 static VEC(pre_expr, heap) *
685 sorted_array_from_bitmap_set (bitmap_set_t set)
687 unsigned int i, j;
688 bitmap_iterator bi, bj;
689 VEC(pre_expr, heap) *result = NULL;
691 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
693 /* The number of expressions having a given value is usually
694 relatively small. Thus, rather than making a vector of all
695 the expressions and sorting it by value-id, we walk the values
696 and check in the reverse mapping that tells us what expressions
697 have a given value, to filter those in our set. As a result,
698 the expressions are inserted in value-id order, which means
699 topological order.
701 If this is somehow a significant lose for some cases, we can
702 choose which set to walk based on the set size. */
703 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
704 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
706 if (bitmap_bit_p (set->expressions, j))
707 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
711 return result;
714 /* Perform bitmapped set operation DEST &= ORIG. */
716 static void
717 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
719 bitmap_iterator bi;
720 unsigned int i;
722 if (dest != orig)
724 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
726 bitmap_and_into (dest->values, orig->values);
727 bitmap_copy (temp, dest->expressions);
728 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
730 pre_expr expr = expression_for_id (i);
731 unsigned int value_id = get_expr_value_id (expr);
732 if (!bitmap_bit_p (dest->values, value_id))
733 bitmap_clear_bit (dest->expressions, i);
735 BITMAP_FREE (temp);
739 /* Subtract all values and expressions contained in ORIG from DEST. */
741 static bitmap_set_t
742 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
744 bitmap_set_t result = bitmap_set_new ();
745 bitmap_iterator bi;
746 unsigned int i;
748 bitmap_and_compl (result->expressions, dest->expressions,
749 orig->expressions);
751 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
753 pre_expr expr = expression_for_id (i);
754 unsigned int value_id = get_expr_value_id (expr);
755 bitmap_set_bit (result->values, value_id);
758 return result;
761 /* Subtract all the values in bitmap set B from bitmap set A. */
763 static void
764 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
766 unsigned int i;
767 bitmap_iterator bi;
768 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
770 bitmap_copy (temp, a->expressions);
771 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
773 pre_expr expr = expression_for_id (i);
774 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
775 bitmap_remove_from_set (a, expr);
777 BITMAP_FREE (temp);
781 /* Return true if bitmapped set SET contains the value VALUE_ID. */
783 static bool
784 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
786 if (value_id_constant_p (value_id))
787 return true;
789 if (!set || bitmap_empty_p (set->expressions))
790 return false;
792 return bitmap_bit_p (set->values, value_id);
795 static inline bool
796 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
798 return bitmap_bit_p (set->expressions, get_expression_id (expr));
801 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
803 static void
804 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
805 const pre_expr expr)
807 bitmap_set_t exprset;
808 unsigned int i;
809 bitmap_iterator bi;
811 if (value_id_constant_p (lookfor))
812 return;
814 if (!bitmap_set_contains_value (set, lookfor))
815 return;
817 /* The number of expressions having a given value is usually
818 significantly less than the total number of expressions in SET.
819 Thus, rather than check, for each expression in SET, whether it
820 has the value LOOKFOR, we walk the reverse mapping that tells us
821 what expressions have a given value, and see if any of those
822 expressions are in our set. For large testcases, this is about
823 5-10x faster than walking the bitmap. If this is somehow a
824 significant lose for some cases, we can choose which set to walk
825 based on the set size. */
826 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
827 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
829 if (bitmap_bit_p (set->expressions, i))
831 bitmap_clear_bit (set->expressions, i);
832 bitmap_set_bit (set->expressions, get_expression_id (expr));
833 return;
838 /* Return true if two bitmap sets are equal. */
840 static bool
841 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
843 return bitmap_equal_p (a->values, b->values);
846 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
847 and add it otherwise. */
849 static void
850 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
852 unsigned int val = get_expr_value_id (expr);
854 if (bitmap_set_contains_value (set, val))
855 bitmap_set_replace_value (set, val, expr);
856 else
857 bitmap_insert_into_set (set, expr);
860 /* Insert EXPR into SET if EXPR's value is not already present in
861 SET. */
863 static void
864 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
866 unsigned int val = get_expr_value_id (expr);
868 if (value_id_constant_p (val))
869 return;
871 if (!bitmap_set_contains_value (set, val))
872 bitmap_insert_into_set (set, expr);
875 /* Print out EXPR to outfile. */
877 static void
878 print_pre_expr (FILE *outfile, const pre_expr expr)
880 switch (expr->kind)
882 case CONSTANT:
883 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
884 break;
885 case NAME:
886 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
887 break;
888 case NARY:
890 unsigned int i;
891 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
892 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
893 for (i = 0; i < nary->length; i++)
895 print_generic_expr (outfile, nary->op[i], 0);
896 if (i != (unsigned) nary->length - 1)
897 fprintf (outfile, ",");
899 fprintf (outfile, "}");
901 break;
903 case REFERENCE:
905 vn_reference_op_t vro;
906 unsigned int i;
907 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
908 fprintf (outfile, "{");
909 for (i = 0;
910 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
911 i++)
913 bool closebrace = false;
914 if (vro->opcode != SSA_NAME
915 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
917 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
918 if (vro->op0)
920 fprintf (outfile, "<");
921 closebrace = true;
924 if (vro->op0)
926 print_generic_expr (outfile, vro->op0, 0);
927 if (vro->op1)
929 fprintf (outfile, ",");
930 print_generic_expr (outfile, vro->op1, 0);
932 if (vro->op2)
934 fprintf (outfile, ",");
935 print_generic_expr (outfile, vro->op2, 0);
938 if (closebrace)
939 fprintf (outfile, ">");
940 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
941 fprintf (outfile, ",");
943 fprintf (outfile, "}");
944 if (ref->vuse)
946 fprintf (outfile, "@");
947 print_generic_expr (outfile, ref->vuse, 0);
950 break;
953 void debug_pre_expr (pre_expr);
955 /* Like print_pre_expr but always prints to stderr. */
956 void
957 debug_pre_expr (pre_expr e)
959 print_pre_expr (stderr, e);
960 fprintf (stderr, "\n");
963 /* Print out SET to OUTFILE. */
965 static void
966 print_bitmap_set (FILE *outfile, bitmap_set_t set,
967 const char *setname, int blockindex)
969 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
970 if (set)
972 bool first = true;
973 unsigned i;
974 bitmap_iterator bi;
976 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
978 const pre_expr expr = expression_for_id (i);
980 if (!first)
981 fprintf (outfile, ", ");
982 first = false;
983 print_pre_expr (outfile, expr);
985 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
988 fprintf (outfile, " }\n");
991 void debug_bitmap_set (bitmap_set_t);
993 void
994 debug_bitmap_set (bitmap_set_t set)
996 print_bitmap_set (stderr, set, "debug", 0);
999 /* Print out the expressions that have VAL to OUTFILE. */
1001 void
1002 print_value_expressions (FILE *outfile, unsigned int val)
1004 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1005 if (set)
1007 char s[10];
1008 sprintf (s, "%04d", val);
1009 print_bitmap_set (outfile, set, s, 0);
1014 void
1015 debug_value_expressions (unsigned int val)
1017 print_value_expressions (stderr, val);
1020 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1021 represent it. */
1023 static pre_expr
1024 get_or_alloc_expr_for_constant (tree constant)
1026 unsigned int result_id;
1027 unsigned int value_id;
1028 pre_expr newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1029 newexpr->kind = CONSTANT;
1030 PRE_EXPR_CONSTANT (newexpr) = constant;
1031 result_id = lookup_expression_id (newexpr);
1032 if (result_id != 0)
1034 pool_free (pre_expr_pool, newexpr);
1035 newexpr = expression_for_id (result_id);
1036 return newexpr;
1038 value_id = get_or_alloc_constant_value_id (constant);
1039 get_or_alloc_expression_id (newexpr);
1040 add_to_value (value_id, newexpr);
1041 return newexpr;
1044 /* Given a value id V, find the actual tree representing the constant
1045 value if there is one, and return it. Return NULL if we can't find
1046 a constant. */
1048 static tree
1049 get_constant_for_value_id (unsigned int v)
1051 if (value_id_constant_p (v))
1053 unsigned int i;
1054 bitmap_iterator bi;
1055 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1057 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1059 pre_expr expr = expression_for_id (i);
1060 if (expr->kind == CONSTANT)
1061 return PRE_EXPR_CONSTANT (expr);
1064 return NULL;
1067 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1068 Currently only supports constants and SSA_NAMES. */
1069 static pre_expr
1070 get_or_alloc_expr_for (tree t)
1072 if (TREE_CODE (t) == SSA_NAME)
1073 return get_or_alloc_expr_for_name (t);
1074 else if (is_gimple_min_invariant (t)
1075 || TREE_CODE (t) == EXC_PTR_EXPR
1076 || TREE_CODE (t) == FILTER_EXPR)
1077 return get_or_alloc_expr_for_constant (t);
1078 else
1080 /* More complex expressions can result from SCCVN expression
1081 simplification that inserts values for them. As they all
1082 do not have VOPs the get handled by the nary ops struct. */
1083 vn_nary_op_t result;
1084 unsigned int result_id;
1085 vn_nary_op_lookup (t, &result);
1086 if (result != NULL)
1088 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1089 e->kind = NARY;
1090 PRE_EXPR_NARY (e) = result;
1091 result_id = lookup_expression_id (e);
1092 if (result_id != 0)
1094 pool_free (pre_expr_pool, e);
1095 e = expression_for_id (result_id);
1096 return e;
1098 alloc_expression_id (e);
1099 return e;
1102 return NULL;
1105 /* Return the folded version of T if T, when folded, is a gimple
1106 min_invariant. Otherwise, return T. */
1108 static pre_expr
1109 fully_constant_expression (pre_expr e)
1111 switch (e->kind)
1113 case CONSTANT:
1114 return e;
1115 case NARY:
1117 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1118 switch (TREE_CODE_CLASS (nary->opcode))
1120 case tcc_expression:
1121 if (nary->opcode == TRUTH_NOT_EXPR)
1122 goto do_unary;
1123 if (nary->opcode != TRUTH_AND_EXPR
1124 && nary->opcode != TRUTH_OR_EXPR
1125 && nary->opcode != TRUTH_XOR_EXPR)
1126 return e;
1127 /* Fallthrough. */
1128 case tcc_binary:
1129 case tcc_comparison:
1131 /* We have to go from trees to pre exprs to value ids to
1132 constants. */
1133 tree naryop0 = nary->op[0];
1134 tree naryop1 = nary->op[1];
1135 tree result;
1136 if (!is_gimple_min_invariant (naryop0))
1138 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1139 unsigned int vrep0 = get_expr_value_id (rep0);
1140 tree const0 = get_constant_for_value_id (vrep0);
1141 if (const0)
1142 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1144 if (!is_gimple_min_invariant (naryop1))
1146 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1147 unsigned int vrep1 = get_expr_value_id (rep1);
1148 tree const1 = get_constant_for_value_id (vrep1);
1149 if (const1)
1150 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1152 result = fold_binary (nary->opcode, nary->type,
1153 naryop0, naryop1);
1154 if (result && is_gimple_min_invariant (result))
1155 return get_or_alloc_expr_for_constant (result);
1156 /* We might have simplified the expression to a
1157 SSA_NAME for example from x_1 * 1. But we cannot
1158 insert a PHI for x_1 unconditionally as x_1 might
1159 not be available readily. */
1160 return e;
1162 case tcc_reference:
1163 if (nary->opcode != REALPART_EXPR
1164 && nary->opcode != IMAGPART_EXPR
1165 && nary->opcode != VIEW_CONVERT_EXPR)
1166 return e;
1167 /* Fallthrough. */
1168 case tcc_unary:
1169 do_unary:
1171 /* We have to go from trees to pre exprs to value ids to
1172 constants. */
1173 tree naryop0 = nary->op[0];
1174 tree const0, result;
1175 if (is_gimple_min_invariant (naryop0))
1176 const0 = naryop0;
1177 else
1179 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1180 unsigned int vrep0 = get_expr_value_id (rep0);
1181 const0 = get_constant_for_value_id (vrep0);
1183 result = NULL;
1184 if (const0)
1186 tree type1 = TREE_TYPE (nary->op[0]);
1187 const0 = fold_convert (type1, const0);
1188 result = fold_unary (nary->opcode, nary->type, const0);
1190 if (result && is_gimple_min_invariant (result))
1191 return get_or_alloc_expr_for_constant (result);
1192 return e;
1194 default:
1195 return e;
1198 case REFERENCE:
1200 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1201 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1202 vn_reference_op_t op;
1204 /* Try to simplify the translated expression if it is
1205 a call to a builtin function with at most two arguments. */
1206 op = VEC_index (vn_reference_op_s, operands, 0);
1207 if (op->opcode == CALL_EXPR
1208 && TREE_CODE (op->op0) == ADDR_EXPR
1209 && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL
1210 && DECL_BUILT_IN (TREE_OPERAND (op->op0, 0))
1211 && VEC_length (vn_reference_op_s, operands) >= 2
1212 && VEC_length (vn_reference_op_s, operands) <= 3)
1214 vn_reference_op_t arg0, arg1 = NULL;
1215 bool anyconst = false;
1216 arg0 = VEC_index (vn_reference_op_s, operands, 1);
1217 if (VEC_length (vn_reference_op_s, operands) > 2)
1218 arg1 = VEC_index (vn_reference_op_s, operands, 2);
1219 if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant
1220 || (arg0->opcode == ADDR_EXPR
1221 && is_gimple_min_invariant (arg0->op0)))
1222 anyconst = true;
1223 if (arg1
1224 && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant
1225 || (arg1->opcode == ADDR_EXPR
1226 && is_gimple_min_invariant (arg1->op0))))
1227 anyconst = true;
1228 if (anyconst)
1230 tree folded = build_call_expr (TREE_OPERAND (op->op0, 0),
1231 arg1 ? 2 : 1,
1232 arg0->op0,
1233 arg1 ? arg1->op0 : NULL);
1234 if (folded
1235 && TREE_CODE (folded) == NOP_EXPR)
1236 folded = TREE_OPERAND (folded, 0);
1237 if (folded
1238 && is_gimple_min_invariant (folded))
1239 return get_or_alloc_expr_for_constant (folded);
1242 return e;
1244 default:
1245 return e;
1247 return e;
1250 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1251 it has the value it would have in BLOCK. */
1253 static tree
1254 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1255 alias_set_type set, tree type, tree vuse,
1256 basic_block phiblock,
1257 basic_block block)
1259 gimple phi = SSA_NAME_DEF_STMT (vuse);
1260 ao_ref ref;
1262 if (gimple_bb (phi) != phiblock)
1263 return vuse;
1265 if (gimple_code (phi) == GIMPLE_PHI)
1267 edge e = find_edge (block, phiblock);
1268 return PHI_ARG_DEF (phi, e->dest_idx);
1271 if (!ao_ref_init_from_vn_reference (&ref, set, type, operands))
1272 return NULL_TREE;
1274 /* Use the alias-oracle to find either the PHI node in this block,
1275 the first VUSE used in this block that is equivalent to vuse or
1276 the first VUSE which definition in this block kills the value. */
1277 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1279 vuse = gimple_vuse (phi);
1280 phi = SSA_NAME_DEF_STMT (vuse);
1281 if (gimple_bb (phi) != phiblock)
1282 return vuse;
1283 if (gimple_code (phi) == GIMPLE_PHI)
1285 edge e = find_edge (block, phiblock);
1286 return PHI_ARG_DEF (phi, e->dest_idx);
1290 return NULL_TREE;
1293 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1294 SET2. This is used to avoid making a set consisting of the union
1295 of PA_IN and ANTIC_IN during insert. */
1297 static inline pre_expr
1298 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1300 pre_expr result;
1302 result = bitmap_find_leader (set1, val, NULL);
1303 if (!result && set2)
1304 result = bitmap_find_leader (set2, val, NULL);
1305 return result;
1308 /* Get the tree type for our PRE expression e. */
1310 static tree
1311 get_expr_type (const pre_expr e)
1313 switch (e->kind)
1315 case NAME:
1316 return TREE_TYPE (PRE_EXPR_NAME (e));
1317 case CONSTANT:
1318 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1319 case REFERENCE:
1320 return PRE_EXPR_REFERENCE (e)->type;
1321 case NARY:
1322 return PRE_EXPR_NARY (e)->type;
1324 gcc_unreachable();
1327 /* Get a representative SSA_NAME for a given expression.
1328 Since all of our sub-expressions are treated as values, we require
1329 them to be SSA_NAME's for simplicity.
1330 Prior versions of GVNPRE used to use "value handles" here, so that
1331 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1332 either case, the operands are really values (IE we do not expect
1333 them to be usable without finding leaders). */
1335 static tree
1336 get_representative_for (const pre_expr e)
1338 tree exprtype;
1339 tree name;
1340 unsigned int value_id = get_expr_value_id (e);
1342 switch (e->kind)
1344 case NAME:
1345 return PRE_EXPR_NAME (e);
1346 case CONSTANT:
1347 return PRE_EXPR_CONSTANT (e);
1348 case NARY:
1349 case REFERENCE:
1351 /* Go through all of the expressions representing this value
1352 and pick out an SSA_NAME. */
1353 unsigned int i;
1354 bitmap_iterator bi;
1355 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1356 value_id);
1357 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1359 pre_expr rep = expression_for_id (i);
1360 if (rep->kind == NAME)
1361 return PRE_EXPR_NAME (rep);
1364 break;
1366 /* If we reached here we couldn't find an SSA_NAME. This can
1367 happen when we've discovered a value that has never appeared in
1368 the program as set to an SSA_NAME, most likely as the result of
1369 phi translation. */
1370 if (dump_file)
1372 fprintf (dump_file,
1373 "Could not find SSA_NAME representative for expression:");
1374 print_pre_expr (dump_file, e);
1375 fprintf (dump_file, "\n");
1378 exprtype = get_expr_type (e);
1380 /* Build and insert the assignment of the end result to the temporary
1381 that we will return. */
1382 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1384 pretemp = create_tmp_var (exprtype, "pretmp");
1385 get_var_ann (pretemp);
1388 name = make_ssa_name (pretemp, gimple_build_nop ());
1389 VN_INFO_GET (name)->value_id = value_id;
1390 if (e->kind == CONSTANT)
1391 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1392 else
1393 VN_INFO (name)->valnum = name;
1395 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1396 if (dump_file)
1398 fprintf (dump_file, "Created SSA_NAME representative ");
1399 print_generic_expr (dump_file, name, 0);
1400 fprintf (dump_file, " for expression:");
1401 print_pre_expr (dump_file, e);
1402 fprintf (dump_file, "\n");
1405 return name;
1411 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1412 the phis in PRED. SEEN is a bitmap saying which expression we have
1413 translated since we started translation of the toplevel expression.
1414 Return NULL if we can't find a leader for each part of the
1415 translated expression. */
1417 static pre_expr
1418 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1419 basic_block pred, basic_block phiblock, bitmap seen)
1421 pre_expr oldexpr = expr;
1422 pre_expr phitrans;
1424 if (!expr)
1425 return NULL;
1427 if (value_id_constant_p (get_expr_value_id (expr)))
1428 return expr;
1430 phitrans = phi_trans_lookup (expr, pred);
1431 if (phitrans)
1432 return phitrans;
1434 /* Prevent cycles when we have recursively dependent leaders. This
1435 can only happen when phi translating the maximal set. */
1436 if (seen)
1438 unsigned int expr_id = get_expression_id (expr);
1439 if (bitmap_bit_p (seen, expr_id))
1440 return NULL;
1441 bitmap_set_bit (seen, expr_id);
1444 switch (expr->kind)
1446 /* Constants contain no values that need translation. */
1447 case CONSTANT:
1448 return expr;
1450 case NARY:
1452 unsigned int i;
1453 bool changed = false;
1454 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1455 struct vn_nary_op_s newnary;
1456 /* The NARY structure is only guaranteed to have been
1457 allocated to the nary->length operands. */
1458 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1459 - sizeof (tree) * (4 - nary->length)));
1461 for (i = 0; i < newnary.length; i++)
1463 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1464 continue;
1465 else
1467 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1468 pre_expr leader = find_leader_in_sets (op_val_id, set1, set2);
1469 pre_expr result = phi_translate_1 (leader, set1, set2,
1470 pred, phiblock, seen);
1471 if (result && result != leader)
1473 tree name = get_representative_for (result);
1474 if (!name)
1475 return NULL;
1476 newnary.op[i] = name;
1478 else if (!result)
1479 return NULL;
1481 changed |= newnary.op[i] != nary->op[i];
1484 if (changed)
1486 pre_expr constant;
1488 tree result = vn_nary_op_lookup_pieces (newnary.length,
1489 newnary.opcode,
1490 newnary.type,
1491 newnary.op[0],
1492 newnary.op[1],
1493 newnary.op[2],
1494 newnary.op[3],
1495 &nary);
1496 unsigned int new_val_id;
1498 expr = (pre_expr) pool_alloc (pre_expr_pool);
1499 expr->kind = NARY;
1500 expr->id = 0;
1501 if (result && is_gimple_min_invariant (result))
1502 return get_or_alloc_expr_for_constant (result);
1505 if (nary)
1507 PRE_EXPR_NARY (expr) = nary;
1508 constant = fully_constant_expression (expr);
1509 if (constant != expr)
1510 return constant;
1512 new_val_id = nary->value_id;
1513 get_or_alloc_expression_id (expr);
1515 else
1517 new_val_id = get_next_value_id ();
1518 VEC_safe_grow_cleared (bitmap_set_t, heap,
1519 value_expressions,
1520 get_max_value_id() + 1);
1521 nary = vn_nary_op_insert_pieces (newnary.length,
1522 newnary.opcode,
1523 newnary.type,
1524 newnary.op[0],
1525 newnary.op[1],
1526 newnary.op[2],
1527 newnary.op[3],
1528 result, new_val_id);
1529 PRE_EXPR_NARY (expr) = nary;
1530 constant = fully_constant_expression (expr);
1531 if (constant != expr)
1532 return constant;
1533 get_or_alloc_expression_id (expr);
1535 add_to_value (new_val_id, expr);
1537 phi_trans_add (oldexpr, expr, pred);
1538 return expr;
1540 break;
1542 case REFERENCE:
1544 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1545 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1546 tree vuse = ref->vuse;
1547 tree newvuse = vuse;
1548 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1549 bool changed = false;
1550 unsigned int i, j;
1551 vn_reference_op_t operand;
1552 vn_reference_t newref;
1554 for (i = 0, j = 0;
1555 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1557 pre_expr opresult;
1558 pre_expr leader;
1559 tree oldop0 = operand->op0;
1560 tree oldop1 = operand->op1;
1561 tree oldop2 = operand->op2;
1562 tree op0 = oldop0;
1563 tree op1 = oldop1;
1564 tree op2 = oldop2;
1565 tree type = operand->type;
1566 vn_reference_op_s newop = *operand;
1568 if (op0 && TREE_CODE (op0) == SSA_NAME)
1570 unsigned int op_val_id = VN_INFO (op0)->value_id;
1571 leader = find_leader_in_sets (op_val_id, set1, set2);
1572 opresult = phi_translate_1 (leader, set1, set2,
1573 pred, phiblock, seen);
1574 if (opresult && opresult != leader)
1576 tree name = get_representative_for (opresult);
1577 if (!name)
1578 break;
1579 op0 = name;
1581 else if (!opresult)
1582 break;
1584 changed |= op0 != oldop0;
1586 if (op1 && TREE_CODE (op1) == SSA_NAME)
1588 unsigned int op_val_id = VN_INFO (op1)->value_id;
1589 leader = find_leader_in_sets (op_val_id, set1, set2);
1590 opresult = phi_translate_1 (leader, set1, set2,
1591 pred, phiblock, seen);
1592 if (opresult && opresult != leader)
1594 tree name = get_representative_for (opresult);
1595 if (!name)
1596 break;
1597 op1 = name;
1599 else if (!opresult)
1600 break;
1602 changed |= op1 != oldop1;
1603 if (op2 && TREE_CODE (op2) == SSA_NAME)
1605 unsigned int op_val_id = VN_INFO (op2)->value_id;
1606 leader = find_leader_in_sets (op_val_id, set1, set2);
1607 opresult = phi_translate_1 (leader, set1, set2,
1608 pred, phiblock, seen);
1609 if (opresult && opresult != leader)
1611 tree name = get_representative_for (opresult);
1612 if (!name)
1613 break;
1614 op2 = name;
1616 else if (!opresult)
1617 break;
1619 changed |= op2 != oldop2;
1621 if (!newoperands)
1622 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1623 /* We may have changed from an SSA_NAME to a constant */
1624 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1625 newop.opcode = TREE_CODE (op0);
1626 newop.type = type;
1627 newop.op0 = op0;
1628 newop.op1 = op1;
1629 newop.op2 = op2;
1630 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1631 /* If it transforms from an SSA_NAME to an address, fold with
1632 a preceding indirect reference. */
1633 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1634 && VEC_index (vn_reference_op_s,
1635 newoperands, j - 1)->opcode == INDIRECT_REF)
1636 vn_reference_fold_indirect (&newoperands, &j);
1638 if (i != VEC_length (vn_reference_op_s, operands))
1640 if (newoperands)
1641 VEC_free (vn_reference_op_s, heap, newoperands);
1642 return NULL;
1645 if (vuse)
1647 newvuse = translate_vuse_through_block (newoperands,
1648 ref->set, ref->type,
1649 vuse, phiblock, pred);
1650 if (newvuse == NULL_TREE)
1652 VEC_free (vn_reference_op_s, heap, newoperands);
1653 return NULL;
1656 changed |= newvuse != vuse;
1658 if (changed)
1660 unsigned int new_val_id;
1661 pre_expr constant;
1663 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1664 ref->type,
1665 newoperands,
1666 &newref, true);
1667 if (newref)
1668 VEC_free (vn_reference_op_s, heap, newoperands);
1670 if (result && is_gimple_min_invariant (result))
1672 gcc_assert (!newoperands);
1673 return get_or_alloc_expr_for_constant (result);
1676 expr = (pre_expr) pool_alloc (pre_expr_pool);
1677 expr->kind = REFERENCE;
1678 expr->id = 0;
1680 if (newref)
1682 PRE_EXPR_REFERENCE (expr) = newref;
1683 constant = fully_constant_expression (expr);
1684 if (constant != expr)
1685 return constant;
1687 new_val_id = newref->value_id;
1688 get_or_alloc_expression_id (expr);
1690 else
1692 new_val_id = get_next_value_id ();
1693 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
1694 get_max_value_id() + 1);
1695 newref = vn_reference_insert_pieces (newvuse, ref->set,
1696 ref->type,
1697 newoperands,
1698 result, new_val_id);
1699 newoperands = NULL;
1700 PRE_EXPR_REFERENCE (expr) = newref;
1701 constant = fully_constant_expression (expr);
1702 if (constant != expr)
1703 return constant;
1704 get_or_alloc_expression_id (expr);
1706 add_to_value (new_val_id, expr);
1708 VEC_free (vn_reference_op_s, heap, newoperands);
1709 phi_trans_add (oldexpr, expr, pred);
1710 return expr;
1712 break;
1714 case NAME:
1716 gimple phi = NULL;
1717 edge e;
1718 gimple def_stmt;
1719 tree name = PRE_EXPR_NAME (expr);
1721 def_stmt = SSA_NAME_DEF_STMT (name);
1722 if (gimple_code (def_stmt) == GIMPLE_PHI
1723 && gimple_bb (def_stmt) == phiblock)
1724 phi = def_stmt;
1725 else
1726 return expr;
1728 e = find_edge (pred, gimple_bb (phi));
1729 if (e)
1731 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1732 pre_expr newexpr;
1734 if (TREE_CODE (def) == SSA_NAME)
1735 def = VN_INFO (def)->valnum;
1737 /* Handle constant. */
1738 if (is_gimple_min_invariant (def))
1739 return get_or_alloc_expr_for_constant (def);
1741 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1742 return NULL;
1744 newexpr = get_or_alloc_expr_for_name (def);
1745 return newexpr;
1748 return expr;
1750 default:
1751 gcc_unreachable ();
1755 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1756 the phis in PRED.
1757 Return NULL if we can't find a leader for each part of the
1758 translated expression. */
1760 static pre_expr
1761 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1762 basic_block pred, basic_block phiblock)
1764 bitmap_clear (seen_during_translate);
1765 return phi_translate_1 (expr, set1, set2, pred, phiblock,
1766 seen_during_translate);
1769 /* For each expression in SET, translate the values through phi nodes
1770 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1771 expressions in DEST. */
1773 static void
1774 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1775 basic_block phiblock)
1777 VEC (pre_expr, heap) *exprs;
1778 pre_expr expr;
1779 int i;
1781 if (!phi_nodes (phiblock))
1783 bitmap_set_copy (dest, set);
1784 return;
1787 exprs = sorted_array_from_bitmap_set (set);
1788 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1790 pre_expr translated;
1791 translated = phi_translate (expr, set, NULL, pred, phiblock);
1793 /* Don't add empty translations to the cache */
1794 if (translated)
1795 phi_trans_add (expr, translated, pred);
1797 if (translated != NULL)
1798 bitmap_value_insert_into_set (dest, translated);
1800 VEC_free (pre_expr, heap, exprs);
1803 /* Find the leader for a value (i.e., the name representing that
1804 value) in a given set, and return it. If STMT is non-NULL it
1805 makes sure the defining statement for the leader dominates it.
1806 Return NULL if no leader is found. */
1808 static pre_expr
1809 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1811 if (value_id_constant_p (val))
1813 unsigned int i;
1814 bitmap_iterator bi;
1815 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1817 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1819 pre_expr expr = expression_for_id (i);
1820 if (expr->kind == CONSTANT)
1821 return expr;
1824 if (bitmap_set_contains_value (set, val))
1826 /* Rather than walk the entire bitmap of expressions, and see
1827 whether any of them has the value we are looking for, we look
1828 at the reverse mapping, which tells us the set of expressions
1829 that have a given value (IE value->expressions with that
1830 value) and see if any of those expressions are in our set.
1831 The number of expressions per value is usually significantly
1832 less than the number of expressions in the set. In fact, for
1833 large testcases, doing it this way is roughly 5-10x faster
1834 than walking the bitmap.
1835 If this is somehow a significant lose for some cases, we can
1836 choose which set to walk based on which set is smaller. */
1837 unsigned int i;
1838 bitmap_iterator bi;
1839 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1841 EXECUTE_IF_AND_IN_BITMAP (exprset->expressions,
1842 set->expressions, 0, i, bi)
1844 pre_expr val = expression_for_id (i);
1845 /* At the point where stmt is not null, there should always
1846 be an SSA_NAME first in the list of expressions. */
1847 if (stmt)
1849 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1850 if (gimple_code (def_stmt) != GIMPLE_PHI
1851 && gimple_bb (def_stmt) == gimple_bb (stmt)
1852 && gimple_uid (def_stmt) >= gimple_uid (stmt))
1853 continue;
1855 return val;
1858 return NULL;
1861 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1862 BLOCK by seeing if it is not killed in the block. Note that we are
1863 only determining whether there is a store that kills it. Because
1864 of the order in which clean iterates over values, we are guaranteed
1865 that altered operands will have caused us to be eliminated from the
1866 ANTIC_IN set already. */
1868 static bool
1869 value_dies_in_block_x (pre_expr expr, basic_block block)
1871 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1872 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1873 gimple def;
1874 gimple_stmt_iterator gsi;
1875 unsigned id = get_expression_id (expr);
1876 bool res = false;
1877 ao_ref ref;
1879 if (!vuse)
1880 return false;
1882 /* Lookup a previously calculated result. */
1883 if (EXPR_DIES (block)
1884 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1885 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1887 /* A memory expression {e, VUSE} dies in the block if there is a
1888 statement that may clobber e. If, starting statement walk from the
1889 top of the basic block, a statement uses VUSE there can be no kill
1890 inbetween that use and the original statement that loaded {e, VUSE},
1891 so we can stop walking. */
1892 ref.base = NULL_TREE;
1893 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1895 tree def_vuse, def_vdef;
1896 def = gsi_stmt (gsi);
1897 def_vuse = gimple_vuse (def);
1898 def_vdef = gimple_vdef (def);
1900 /* Not a memory statement. */
1901 if (!def_vuse)
1902 continue;
1904 /* Not a may-def. */
1905 if (!def_vdef)
1907 /* A load with the same VUSE, we're done. */
1908 if (def_vuse == vuse)
1909 break;
1911 continue;
1914 /* Init ref only if we really need it. */
1915 if (ref.base == NULL_TREE
1916 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1917 refx->operands))
1919 res = true;
1920 break;
1922 /* If the statement may clobber expr, it dies. */
1923 if (stmt_may_clobber_ref_p_1 (def, &ref))
1925 res = true;
1926 break;
1930 /* Remember the result. */
1931 if (!EXPR_DIES (block))
1932 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1933 bitmap_set_bit (EXPR_DIES (block), id * 2);
1934 if (res)
1935 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1937 return res;
1941 #define union_contains_value(SET1, SET2, VAL) \
1942 (bitmap_set_contains_value ((SET1), (VAL)) \
1943 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
1945 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
1947 static bool
1948 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
1949 vn_reference_op_t vro)
1951 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
1953 struct pre_expr_d temp;
1954 temp.kind = NAME;
1955 temp.id = 0;
1956 PRE_EXPR_NAME (&temp) = vro->op0;
1957 temp.id = lookup_expression_id (&temp);
1958 if (temp.id == 0)
1959 return false;
1960 if (!union_contains_value (set1, set2,
1961 get_expr_value_id (&temp)))
1962 return false;
1964 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
1966 struct pre_expr_d temp;
1967 temp.kind = NAME;
1968 temp.id = 0;
1969 PRE_EXPR_NAME (&temp) = vro->op1;
1970 temp.id = lookup_expression_id (&temp);
1971 if (temp.id == 0)
1972 return false;
1973 if (!union_contains_value (set1, set2,
1974 get_expr_value_id (&temp)))
1975 return false;
1978 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
1980 struct pre_expr_d temp;
1981 temp.kind = NAME;
1982 temp.id = 0;
1983 PRE_EXPR_NAME (&temp) = vro->op2;
1984 temp.id = lookup_expression_id (&temp);
1985 if (temp.id == 0)
1986 return false;
1987 if (!union_contains_value (set1, set2,
1988 get_expr_value_id (&temp)))
1989 return false;
1992 return true;
1995 /* Determine if the expression EXPR is valid in SET1 U SET2.
1996 ONLY SET2 CAN BE NULL.
1997 This means that we have a leader for each part of the expression
1998 (if it consists of values), or the expression is an SSA_NAME.
1999 For loads/calls, we also see if the vuse is killed in this block.
2002 static bool
2003 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2004 basic_block block)
2006 switch (expr->kind)
2008 case NAME:
2009 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2010 case NARY:
2012 unsigned int i;
2013 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2014 for (i = 0; i < nary->length; i++)
2016 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2018 struct pre_expr_d temp;
2019 temp.kind = NAME;
2020 temp.id = 0;
2021 PRE_EXPR_NAME (&temp) = nary->op[i];
2022 temp.id = lookup_expression_id (&temp);
2023 if (temp.id == 0)
2024 return false;
2025 if (!union_contains_value (set1, set2,
2026 get_expr_value_id (&temp)))
2027 return false;
2030 return true;
2032 break;
2033 case REFERENCE:
2035 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2036 vn_reference_op_t vro;
2037 unsigned int i;
2039 for (i = 0; VEC_iterate (vn_reference_op_s, ref->operands, i, vro); i++)
2041 if (!vro_valid_in_sets (set1, set2, vro))
2042 return false;
2044 if (ref->vuse)
2046 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2047 if (!gimple_nop_p (def_stmt)
2048 && gimple_bb (def_stmt) != block
2049 && !dominated_by_p (CDI_DOMINATORS,
2050 block, gimple_bb (def_stmt)))
2051 return false;
2053 return !value_dies_in_block_x (expr, block);
2055 default:
2056 gcc_unreachable ();
2060 /* Clean the set of expressions that are no longer valid in SET1 or
2061 SET2. This means expressions that are made up of values we have no
2062 leaders for in SET1 or SET2. This version is used for partial
2063 anticipation, which means it is not valid in either ANTIC_IN or
2064 PA_IN. */
2066 static void
2067 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2069 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2070 pre_expr expr;
2071 int i;
2073 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2075 if (!valid_in_sets (set1, set2, expr, block))
2076 bitmap_remove_from_set (set1, expr);
2078 VEC_free (pre_expr, heap, exprs);
2081 /* Clean the set of expressions that are no longer valid in SET. This
2082 means expressions that are made up of values we have no leaders for
2083 in SET. */
2085 static void
2086 clean (bitmap_set_t set, basic_block block)
2088 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2089 pre_expr expr;
2090 int i;
2092 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2094 if (!valid_in_sets (set, NULL, expr, block))
2095 bitmap_remove_from_set (set, expr);
2097 VEC_free (pre_expr, heap, exprs);
2100 static sbitmap has_abnormal_preds;
2102 /* List of blocks that may have changed during ANTIC computation and
2103 thus need to be iterated over. */
2105 static sbitmap changed_blocks;
2107 /* Decide whether to defer a block for a later iteration, or PHI
2108 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2109 should defer the block, and true if we processed it. */
2111 static bool
2112 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2113 basic_block block, basic_block phiblock)
2115 if (!BB_VISITED (phiblock))
2117 SET_BIT (changed_blocks, block->index);
2118 BB_VISITED (block) = 0;
2119 BB_DEFERRED (block) = 1;
2120 return false;
2122 else
2123 phi_translate_set (dest, source, block, phiblock);
2124 return true;
2127 /* Compute the ANTIC set for BLOCK.
2129 If succs(BLOCK) > 1 then
2130 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2131 else if succs(BLOCK) == 1 then
2132 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2134 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2137 static bool
2138 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2140 bool changed = false;
2141 bitmap_set_t S, old, ANTIC_OUT;
2142 bitmap_iterator bi;
2143 unsigned int bii;
2144 edge e;
2145 edge_iterator ei;
2147 old = ANTIC_OUT = S = NULL;
2148 BB_VISITED (block) = 1;
2150 /* If any edges from predecessors are abnormal, antic_in is empty,
2151 so do nothing. */
2152 if (block_has_abnormal_pred_edge)
2153 goto maybe_dump_sets;
2155 old = ANTIC_IN (block);
2156 ANTIC_OUT = bitmap_set_new ();
2158 /* If the block has no successors, ANTIC_OUT is empty. */
2159 if (EDGE_COUNT (block->succs) == 0)
2161 /* If we have one successor, we could have some phi nodes to
2162 translate through. */
2163 else if (single_succ_p (block))
2165 basic_block succ_bb = single_succ (block);
2167 /* We trade iterations of the dataflow equations for having to
2168 phi translate the maximal set, which is incredibly slow
2169 (since the maximal set often has 300+ members, even when you
2170 have a small number of blocks).
2171 Basically, we defer the computation of ANTIC for this block
2172 until we have processed it's successor, which will inevitably
2173 have a *much* smaller set of values to phi translate once
2174 clean has been run on it.
2175 The cost of doing this is that we technically perform more
2176 iterations, however, they are lower cost iterations.
2178 Timings for PRE on tramp3d-v4:
2179 without maximal set fix: 11 seconds
2180 with maximal set fix/without deferring: 26 seconds
2181 with maximal set fix/with deferring: 11 seconds
2184 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2185 block, succ_bb))
2187 changed = true;
2188 goto maybe_dump_sets;
2191 /* If we have multiple successors, we take the intersection of all of
2192 them. Note that in the case of loop exit phi nodes, we may have
2193 phis to translate through. */
2194 else
2196 VEC(basic_block, heap) * worklist;
2197 size_t i;
2198 basic_block bprime, first;
2200 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2201 FOR_EACH_EDGE (e, ei, block->succs)
2202 VEC_quick_push (basic_block, worklist, e->dest);
2203 first = VEC_index (basic_block, worklist, 0);
2205 if (phi_nodes (first))
2207 bitmap_set_t from = ANTIC_IN (first);
2209 if (!BB_VISITED (first))
2210 from = maximal_set;
2211 phi_translate_set (ANTIC_OUT, from, block, first);
2213 else
2215 if (!BB_VISITED (first))
2216 bitmap_set_copy (ANTIC_OUT, maximal_set);
2217 else
2218 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2221 for (i = 1; VEC_iterate (basic_block, worklist, i, bprime); i++)
2223 if (phi_nodes (bprime))
2225 bitmap_set_t tmp = bitmap_set_new ();
2226 bitmap_set_t from = ANTIC_IN (bprime);
2228 if (!BB_VISITED (bprime))
2229 from = maximal_set;
2230 phi_translate_set (tmp, from, block, bprime);
2231 bitmap_set_and (ANTIC_OUT, tmp);
2232 bitmap_set_free (tmp);
2234 else
2236 if (!BB_VISITED (bprime))
2237 bitmap_set_and (ANTIC_OUT, maximal_set);
2238 else
2239 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2242 VEC_free (basic_block, heap, worklist);
2245 /* Generate ANTIC_OUT - TMP_GEN. */
2246 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2248 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2249 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2250 TMP_GEN (block));
2252 /* Then union in the ANTIC_OUT - TMP_GEN values,
2253 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2254 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2255 bitmap_value_insert_into_set (ANTIC_IN (block),
2256 expression_for_id (bii));
2258 clean (ANTIC_IN (block), block);
2260 /* !old->expressions can happen when we deferred a block. */
2261 if (!old->expressions || !bitmap_set_equal (old, ANTIC_IN (block)))
2263 changed = true;
2264 SET_BIT (changed_blocks, block->index);
2265 FOR_EACH_EDGE (e, ei, block->preds)
2266 SET_BIT (changed_blocks, e->src->index);
2268 else
2269 RESET_BIT (changed_blocks, block->index);
2271 maybe_dump_sets:
2272 if (dump_file && (dump_flags & TDF_DETAILS))
2274 if (!BB_DEFERRED (block) || BB_VISITED (block))
2276 if (ANTIC_OUT)
2277 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2279 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2280 block->index);
2282 if (S)
2283 print_bitmap_set (dump_file, S, "S", block->index);
2285 else
2287 fprintf (dump_file,
2288 "Block %d was deferred for a future iteration.\n",
2289 block->index);
2292 if (old)
2293 bitmap_set_free (old);
2294 if (S)
2295 bitmap_set_free (S);
2296 if (ANTIC_OUT)
2297 bitmap_set_free (ANTIC_OUT);
2298 return changed;
2301 /* Compute PARTIAL_ANTIC for BLOCK.
2303 If succs(BLOCK) > 1 then
2304 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2305 in ANTIC_OUT for all succ(BLOCK)
2306 else if succs(BLOCK) == 1 then
2307 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2309 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2310 - ANTIC_IN[BLOCK])
2313 static bool
2314 compute_partial_antic_aux (basic_block block,
2315 bool block_has_abnormal_pred_edge)
2317 bool changed = false;
2318 bitmap_set_t old_PA_IN;
2319 bitmap_set_t PA_OUT;
2320 edge e;
2321 edge_iterator ei;
2322 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2324 old_PA_IN = PA_OUT = NULL;
2326 /* If any edges from predecessors are abnormal, antic_in is empty,
2327 so do nothing. */
2328 if (block_has_abnormal_pred_edge)
2329 goto maybe_dump_sets;
2331 /* If there are too many partially anticipatable values in the
2332 block, phi_translate_set can take an exponential time: stop
2333 before the translation starts. */
2334 if (max_pa
2335 && single_succ_p (block)
2336 && bitmap_count_bits (PA_IN (single_succ (block))->values) > max_pa)
2337 goto maybe_dump_sets;
2339 old_PA_IN = PA_IN (block);
2340 PA_OUT = bitmap_set_new ();
2342 /* If the block has no successors, ANTIC_OUT is empty. */
2343 if (EDGE_COUNT (block->succs) == 0)
2345 /* If we have one successor, we could have some phi nodes to
2346 translate through. Note that we can't phi translate across DFS
2347 back edges in partial antic, because it uses a union operation on
2348 the successors. For recurrences like IV's, we will end up
2349 generating a new value in the set on each go around (i + 3 (VH.1)
2350 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2351 else if (single_succ_p (block))
2353 basic_block succ = single_succ (block);
2354 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2355 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2357 /* If we have multiple successors, we take the union of all of
2358 them. */
2359 else
2361 VEC(basic_block, heap) * worklist;
2362 size_t i;
2363 basic_block bprime;
2365 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2366 FOR_EACH_EDGE (e, ei, block->succs)
2368 if (e->flags & EDGE_DFS_BACK)
2369 continue;
2370 VEC_quick_push (basic_block, worklist, e->dest);
2372 if (VEC_length (basic_block, worklist) > 0)
2374 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2376 unsigned int i;
2377 bitmap_iterator bi;
2379 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2380 bitmap_value_insert_into_set (PA_OUT,
2381 expression_for_id (i));
2382 if (phi_nodes (bprime))
2384 bitmap_set_t pa_in = bitmap_set_new ();
2385 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2386 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2387 bitmap_value_insert_into_set (PA_OUT,
2388 expression_for_id (i));
2389 bitmap_set_free (pa_in);
2391 else
2392 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2393 bitmap_value_insert_into_set (PA_OUT,
2394 expression_for_id (i));
2397 VEC_free (basic_block, heap, worklist);
2400 /* PA_IN starts with PA_OUT - TMP_GEN.
2401 Then we subtract things from ANTIC_IN. */
2402 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2404 /* For partial antic, we want to put back in the phi results, since
2405 we will properly avoid making them partially antic over backedges. */
2406 bitmap_ior_into (PA_IN (block)->values, PHI_GEN (block)->values);
2407 bitmap_ior_into (PA_IN (block)->expressions, PHI_GEN (block)->expressions);
2409 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2410 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2412 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2414 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2416 changed = true;
2417 SET_BIT (changed_blocks, block->index);
2418 FOR_EACH_EDGE (e, ei, block->preds)
2419 SET_BIT (changed_blocks, e->src->index);
2421 else
2422 RESET_BIT (changed_blocks, block->index);
2424 maybe_dump_sets:
2425 if (dump_file && (dump_flags & TDF_DETAILS))
2427 if (PA_OUT)
2428 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2430 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2432 if (old_PA_IN)
2433 bitmap_set_free (old_PA_IN);
2434 if (PA_OUT)
2435 bitmap_set_free (PA_OUT);
2436 return changed;
2439 /* Compute ANTIC and partial ANTIC sets. */
2441 static void
2442 compute_antic (void)
2444 bool changed = true;
2445 int num_iterations = 0;
2446 basic_block block;
2447 int i;
2449 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2450 We pre-build the map of blocks with incoming abnormal edges here. */
2451 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2452 sbitmap_zero (has_abnormal_preds);
2454 FOR_EACH_BB (block)
2456 edge_iterator ei;
2457 edge e;
2459 FOR_EACH_EDGE (e, ei, block->preds)
2461 e->flags &= ~EDGE_DFS_BACK;
2462 if (e->flags & EDGE_ABNORMAL)
2464 SET_BIT (has_abnormal_preds, block->index);
2465 break;
2469 BB_VISITED (block) = 0;
2470 BB_DEFERRED (block) = 0;
2471 /* While we are here, give empty ANTIC_IN sets to each block. */
2472 ANTIC_IN (block) = bitmap_set_new ();
2473 PA_IN (block) = bitmap_set_new ();
2476 /* At the exit block we anticipate nothing. */
2477 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2478 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2479 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2481 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2482 sbitmap_ones (changed_blocks);
2483 while (changed)
2485 if (dump_file && (dump_flags & TDF_DETAILS))
2486 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2487 num_iterations++;
2488 changed = false;
2489 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2491 if (TEST_BIT (changed_blocks, postorder[i]))
2493 basic_block block = BASIC_BLOCK (postorder[i]);
2494 changed |= compute_antic_aux (block,
2495 TEST_BIT (has_abnormal_preds,
2496 block->index));
2499 #ifdef ENABLE_CHECKING
2500 /* Theoretically possible, but *highly* unlikely. */
2501 gcc_assert (num_iterations < 500);
2502 #endif
2505 statistics_histogram_event (cfun, "compute_antic iterations",
2506 num_iterations);
2508 if (do_partial_partial)
2510 sbitmap_ones (changed_blocks);
2511 mark_dfs_back_edges ();
2512 num_iterations = 0;
2513 changed = true;
2514 while (changed)
2516 if (dump_file && (dump_flags & TDF_DETAILS))
2517 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2518 num_iterations++;
2519 changed = false;
2520 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2522 if (TEST_BIT (changed_blocks, postorder[i]))
2524 basic_block block = BASIC_BLOCK (postorder[i]);
2525 changed
2526 |= compute_partial_antic_aux (block,
2527 TEST_BIT (has_abnormal_preds,
2528 block->index));
2531 #ifdef ENABLE_CHECKING
2532 /* Theoretically possible, but *highly* unlikely. */
2533 gcc_assert (num_iterations < 500);
2534 #endif
2536 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2537 num_iterations);
2539 sbitmap_free (has_abnormal_preds);
2540 sbitmap_free (changed_blocks);
2543 /* Return true if we can value number the call in STMT. This is true
2544 if we have a pure or constant call. */
2546 static bool
2547 can_value_number_call (gimple stmt)
2549 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2550 return true;
2551 return false;
2554 /* Return true if OP is an exception handler related operation, such as
2555 FILTER_EXPR or EXC_PTR_EXPR. */
2557 static bool
2558 is_exception_related (gimple stmt)
2560 return (is_gimple_assign (stmt)
2561 && (gimple_assign_rhs_code (stmt) == FILTER_EXPR
2562 || gimple_assign_rhs_code (stmt) == EXC_PTR_EXPR));
2565 /* Return true if OP is a tree which we can perform PRE on.
2566 This may not match the operations we can value number, but in
2567 a perfect world would. */
2569 static bool
2570 can_PRE_operation (tree op)
2572 return UNARY_CLASS_P (op)
2573 || BINARY_CLASS_P (op)
2574 || COMPARISON_CLASS_P (op)
2575 || TREE_CODE (op) == INDIRECT_REF
2576 || TREE_CODE (op) == COMPONENT_REF
2577 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2578 || TREE_CODE (op) == CALL_EXPR
2579 || TREE_CODE (op) == ARRAY_REF;
2583 /* Inserted expressions are placed onto this worklist, which is used
2584 for performing quick dead code elimination of insertions we made
2585 that didn't turn out to be necessary. */
2586 static VEC(gimple,heap) *inserted_exprs;
2587 static bitmap inserted_phi_names;
2589 /* Pool allocated fake store expressions are placed onto this
2590 worklist, which, after performing dead code elimination, is walked
2591 to see which expressions need to be put into GC'able memory */
2592 static VEC(gimple, heap) *need_creation;
2594 /* The actual worker for create_component_ref_by_pieces. */
2596 static tree
2597 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2598 unsigned int *operand, gimple_seq *stmts,
2599 gimple domstmt)
2601 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2602 *operand);
2603 tree genop;
2604 ++*operand;
2605 switch (currop->opcode)
2607 case CALL_EXPR:
2609 tree folded, sc = currop->op1;
2610 unsigned int nargs = 0;
2611 tree *args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2612 ref->operands) - 1);
2613 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2615 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2616 operand, stmts,
2617 domstmt);
2618 nargs++;
2620 folded = build_call_array (currop->type,
2621 TREE_CODE (currop->op0) == FUNCTION_DECL
2622 ? build_fold_addr_expr (currop->op0)
2623 : currop->op0,
2624 nargs, args);
2625 free (args);
2626 if (sc)
2628 pre_expr scexpr = get_or_alloc_expr_for (sc);
2629 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2630 if (!sc)
2631 return NULL_TREE;
2632 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2634 return folded;
2636 break;
2637 case TARGET_MEM_REF:
2639 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2640 *operand);
2641 pre_expr op0expr;
2642 tree genop0 = NULL_TREE;
2643 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2644 stmts, domstmt);
2645 if (!baseop)
2646 return NULL_TREE;
2647 if (currop->op0)
2649 op0expr = get_or_alloc_expr_for (currop->op0);
2650 genop0 = find_or_generate_expression (block, op0expr,
2651 stmts, domstmt);
2652 if (!genop0)
2653 return NULL_TREE;
2655 if (DECL_P (baseop))
2656 return build6 (TARGET_MEM_REF, currop->type,
2657 baseop, NULL_TREE,
2658 genop0, currop->op1, currop->op2,
2659 unshare_expr (nextop->op1));
2660 else
2661 return build6 (TARGET_MEM_REF, currop->type,
2662 NULL_TREE, baseop,
2663 genop0, currop->op1, currop->op2,
2664 unshare_expr (nextop->op1));
2666 break;
2667 case ADDR_EXPR:
2668 if (currop->op0)
2670 gcc_assert (is_gimple_min_invariant (currop->op0));
2671 return currop->op0;
2673 /* Fallthrough. */
2674 case REALPART_EXPR:
2675 case IMAGPART_EXPR:
2676 case VIEW_CONVERT_EXPR:
2678 tree folded;
2679 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2680 operand,
2681 stmts, domstmt);
2682 if (!genop0)
2683 return NULL_TREE;
2684 folded = fold_build1 (currop->opcode, currop->type,
2685 genop0);
2686 return folded;
2688 break;
2689 case ALIGN_INDIRECT_REF:
2690 case MISALIGNED_INDIRECT_REF:
2691 case INDIRECT_REF:
2693 tree folded;
2694 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2695 operand,
2696 stmts, domstmt);
2697 if (!genop1)
2698 return NULL_TREE;
2699 genop1 = fold_convert (build_pointer_type (currop->type),
2700 genop1);
2702 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2703 folded = fold_build2 (currop->opcode, currop->type,
2704 genop1, currop->op1);
2705 else
2706 folded = fold_build1 (currop->opcode, currop->type,
2707 genop1);
2708 return folded;
2710 break;
2711 case BIT_FIELD_REF:
2713 tree folded;
2714 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2715 stmts, domstmt);
2716 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2717 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2718 tree genop1;
2719 tree genop2;
2721 if (!genop0)
2722 return NULL_TREE;
2723 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2724 if (!genop1)
2725 return NULL_TREE;
2726 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2727 if (!genop2)
2728 return NULL_TREE;
2729 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2730 genop2);
2731 return folded;
2734 /* For array ref vn_reference_op's, operand 1 of the array ref
2735 is op0 of the reference op and operand 3 of the array ref is
2736 op1. */
2737 case ARRAY_RANGE_REF:
2738 case ARRAY_REF:
2740 tree genop0;
2741 tree genop1 = currop->op0;
2742 pre_expr op1expr;
2743 tree genop2 = currop->op1;
2744 pre_expr op2expr;
2745 tree genop3;
2746 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2747 stmts, domstmt);
2748 if (!genop0)
2749 return NULL_TREE;
2750 op1expr = get_or_alloc_expr_for (genop1);
2751 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2752 if (!genop1)
2753 return NULL_TREE;
2754 if (genop2)
2756 op2expr = get_or_alloc_expr_for (genop2);
2757 genop2 = find_or_generate_expression (block, op2expr, stmts,
2758 domstmt);
2759 if (!genop2)
2760 return NULL_TREE;
2763 genop3 = currop->op2;
2764 return build4 (currop->opcode, currop->type, genop0, genop1,
2765 genop2, genop3);
2767 case COMPONENT_REF:
2769 tree op0;
2770 tree op1;
2771 tree genop2 = currop->op1;
2772 pre_expr op2expr;
2773 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2774 stmts, domstmt);
2775 if (!op0)
2776 return NULL_TREE;
2777 /* op1 should be a FIELD_DECL, which are represented by
2778 themselves. */
2779 op1 = currop->op0;
2780 if (genop2)
2782 op2expr = get_or_alloc_expr_for (genop2);
2783 genop2 = find_or_generate_expression (block, op2expr, stmts,
2784 domstmt);
2785 if (!genop2)
2786 return NULL_TREE;
2789 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2790 genop2);
2792 break;
2793 case SSA_NAME:
2795 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2796 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2797 return genop;
2799 case STRING_CST:
2800 case INTEGER_CST:
2801 case COMPLEX_CST:
2802 case VECTOR_CST:
2803 case REAL_CST:
2804 case CONSTRUCTOR:
2805 case VAR_DECL:
2806 case PARM_DECL:
2807 case CONST_DECL:
2808 case RESULT_DECL:
2809 case FUNCTION_DECL:
2810 return currop->op0;
2812 default:
2813 gcc_unreachable ();
2817 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2818 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2819 trying to rename aggregates into ssa form directly, which is a no no.
2821 Thus, this routine doesn't create temporaries, it just builds a
2822 single access expression for the array, calling
2823 find_or_generate_expression to build the innermost pieces.
2825 This function is a subroutine of create_expression_by_pieces, and
2826 should not be called on it's own unless you really know what you
2827 are doing. */
2829 static tree
2830 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2831 gimple_seq *stmts, gimple domstmt)
2833 unsigned int op = 0;
2834 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2837 /* Find a leader for an expression, or generate one using
2838 create_expression_by_pieces if it's ANTIC but
2839 complex.
2840 BLOCK is the basic_block we are looking for leaders in.
2841 EXPR is the expression to find a leader or generate for.
2842 STMTS is the statement list to put the inserted expressions on.
2843 Returns the SSA_NAME of the LHS of the generated expression or the
2844 leader.
2845 DOMSTMT if non-NULL is a statement that should be dominated by
2846 all uses in the generated expression. If DOMSTMT is non-NULL this
2847 routine can fail and return NULL_TREE. Otherwise it will assert
2848 on failure. */
2850 static tree
2851 find_or_generate_expression (basic_block block, pre_expr expr,
2852 gimple_seq *stmts, gimple domstmt)
2854 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2855 get_expr_value_id (expr), domstmt);
2856 tree genop = NULL;
2857 if (leader)
2859 if (leader->kind == NAME)
2860 genop = PRE_EXPR_NAME (leader);
2861 else if (leader->kind == CONSTANT)
2862 genop = PRE_EXPR_CONSTANT (leader);
2865 /* If it's still NULL, it must be a complex expression, so generate
2866 it recursively. Not so for FRE though. */
2867 if (genop == NULL
2868 && !in_fre)
2870 bitmap_set_t exprset;
2871 unsigned int lookfor = get_expr_value_id (expr);
2872 bool handled = false;
2873 bitmap_iterator bi;
2874 unsigned int i;
2876 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2877 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2879 pre_expr temp = expression_for_id (i);
2880 if (temp->kind != NAME)
2882 handled = true;
2883 genop = create_expression_by_pieces (block, temp, stmts,
2884 domstmt,
2885 get_expr_type (expr));
2886 break;
2889 if (!handled && domstmt)
2890 return NULL_TREE;
2892 gcc_assert (handled);
2894 return genop;
2897 #define NECESSARY GF_PLF_1
2899 /* Create an expression in pieces, so that we can handle very complex
2900 expressions that may be ANTIC, but not necessary GIMPLE.
2901 BLOCK is the basic block the expression will be inserted into,
2902 EXPR is the expression to insert (in value form)
2903 STMTS is a statement list to append the necessary insertions into.
2905 This function will die if we hit some value that shouldn't be
2906 ANTIC but is (IE there is no leader for it, or its components).
2907 This function may also generate expressions that are themselves
2908 partially or fully redundant. Those that are will be either made
2909 fully redundant during the next iteration of insert (for partially
2910 redundant ones), or eliminated by eliminate (for fully redundant
2911 ones).
2913 If DOMSTMT is non-NULL then we make sure that all uses in the
2914 expressions dominate that statement. In this case the function
2915 can return NULL_TREE to signal failure. */
2917 static tree
2918 create_expression_by_pieces (basic_block block, pre_expr expr,
2919 gimple_seq *stmts, gimple domstmt, tree type)
2921 tree temp, name;
2922 tree folded;
2923 gimple_seq forced_stmts = NULL;
2924 unsigned int value_id;
2925 gimple_stmt_iterator gsi;
2926 tree exprtype = type ? type : get_expr_type (expr);
2927 pre_expr nameexpr;
2928 gimple newstmt;
2930 switch (expr->kind)
2932 /* We may hit the NAME/CONSTANT case if we have to convert types
2933 that value numbering saw through. */
2934 case NAME:
2935 folded = PRE_EXPR_NAME (expr);
2936 break;
2937 case CONSTANT:
2938 folded = PRE_EXPR_CONSTANT (expr);
2939 break;
2940 case REFERENCE:
2942 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2943 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
2945 break;
2946 case NARY:
2948 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2949 switch (nary->length)
2951 case 2:
2953 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2954 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
2955 tree genop1 = find_or_generate_expression (block, op1,
2956 stmts, domstmt);
2957 tree genop2 = find_or_generate_expression (block, op2,
2958 stmts, domstmt);
2959 if (!genop1 || !genop2)
2960 return NULL_TREE;
2961 genop1 = fold_convert (TREE_TYPE (nary->op[0]),
2962 genop1);
2963 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
2964 may be a constant with the wrong type. */
2965 if (nary->opcode == POINTER_PLUS_EXPR)
2966 genop2 = fold_convert (sizetype, genop2);
2967 else
2968 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
2970 folded = fold_build2 (nary->opcode, nary->type,
2971 genop1, genop2);
2973 break;
2974 case 1:
2976 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2977 tree genop1 = find_or_generate_expression (block, op1,
2978 stmts, domstmt);
2979 if (!genop1)
2980 return NULL_TREE;
2981 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
2983 folded = fold_build1 (nary->opcode, nary->type,
2984 genop1);
2986 break;
2987 default:
2988 return NULL_TREE;
2991 break;
2992 default:
2993 return NULL_TREE;
2996 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2997 folded = fold_convert (exprtype, folded);
2999 /* Force the generated expression to be a sequence of GIMPLE
3000 statements.
3001 We have to call unshare_expr because force_gimple_operand may
3002 modify the tree we pass to it. */
3003 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3004 false, NULL);
3006 /* If we have any intermediate expressions to the value sets, add them
3007 to the value sets and chain them in the instruction stream. */
3008 if (forced_stmts)
3010 gsi = gsi_start (forced_stmts);
3011 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3013 gimple stmt = gsi_stmt (gsi);
3014 tree forcedname = gimple_get_lhs (stmt);
3015 pre_expr nameexpr;
3017 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3018 if (TREE_CODE (forcedname) == SSA_NAME)
3020 VN_INFO_GET (forcedname)->valnum = forcedname;
3021 VN_INFO (forcedname)->value_id = get_next_value_id ();
3022 nameexpr = get_or_alloc_expr_for_name (forcedname);
3023 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3024 if (!in_fre)
3025 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3026 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3028 mark_symbols_for_renaming (stmt);
3030 gimple_seq_add_seq (stmts, forced_stmts);
3033 /* Build and insert the assignment of the end result to the temporary
3034 that we will return. */
3035 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3037 pretemp = create_tmp_var (exprtype, "pretmp");
3038 get_var_ann (pretemp);
3041 temp = pretemp;
3042 add_referenced_var (temp);
3044 if (TREE_CODE (exprtype) == COMPLEX_TYPE
3045 || TREE_CODE (exprtype) == VECTOR_TYPE)
3046 DECL_GIMPLE_REG_P (temp) = 1;
3048 newstmt = gimple_build_assign (temp, folded);
3049 name = make_ssa_name (temp, newstmt);
3050 gimple_assign_set_lhs (newstmt, name);
3051 gimple_set_plf (newstmt, NECESSARY, false);
3053 gimple_seq_add_stmt (stmts, newstmt);
3054 VEC_safe_push (gimple, heap, inserted_exprs, newstmt);
3056 /* All the symbols in NEWEXPR should be put into SSA form. */
3057 mark_symbols_for_renaming (newstmt);
3059 /* Add a value number to the temporary.
3060 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3061 we are creating the expression by pieces, and this particular piece of
3062 the expression may have been represented. There is no harm in replacing
3063 here. */
3064 VN_INFO_GET (name)->valnum = name;
3065 value_id = get_expr_value_id (expr);
3066 VN_INFO (name)->value_id = value_id;
3067 nameexpr = get_or_alloc_expr_for_name (name);
3068 add_to_value (value_id, nameexpr);
3069 if (!in_fre)
3070 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3071 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3073 pre_stats.insertions++;
3074 if (dump_file && (dump_flags & TDF_DETAILS))
3076 fprintf (dump_file, "Inserted ");
3077 print_gimple_stmt (dump_file, newstmt, 0, 0);
3078 fprintf (dump_file, " in predecessor %d\n", block->index);
3081 return name;
3085 /* Insert the to-be-made-available values of expression EXPRNUM for each
3086 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3087 merge the result with a phi node, given the same value number as
3088 NODE. Return true if we have inserted new stuff. */
3090 static bool
3091 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3092 pre_expr *avail)
3094 pre_expr expr = expression_for_id (exprnum);
3095 pre_expr newphi;
3096 unsigned int val = get_expr_value_id (expr);
3097 edge pred;
3098 bool insertions = false;
3099 bool nophi = false;
3100 basic_block bprime;
3101 pre_expr eprime;
3102 edge_iterator ei;
3103 tree type = get_expr_type (expr);
3104 tree temp;
3105 gimple phi;
3107 if (dump_file && (dump_flags & TDF_DETAILS))
3109 fprintf (dump_file, "Found partial redundancy for expression ");
3110 print_pre_expr (dump_file, expr);
3111 fprintf (dump_file, " (%04d)\n", val);
3114 /* Make sure we aren't creating an induction variable. */
3115 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2
3116 && expr->kind != REFERENCE)
3118 bool firstinsideloop = false;
3119 bool secondinsideloop = false;
3120 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3121 EDGE_PRED (block, 0)->src);
3122 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3123 EDGE_PRED (block, 1)->src);
3124 /* Induction variables only have one edge inside the loop. */
3125 if (firstinsideloop ^ secondinsideloop)
3127 if (dump_file && (dump_flags & TDF_DETAILS))
3128 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3129 nophi = true;
3133 /* Make sure we are not inserting trapping expressions. */
3134 FOR_EACH_EDGE (pred, ei, block->preds)
3136 bprime = pred->src;
3137 eprime = avail[bprime->index];
3138 if (eprime->kind == NARY
3139 && vn_nary_may_trap (PRE_EXPR_NARY (eprime)))
3140 return false;
3143 /* Make the necessary insertions. */
3144 FOR_EACH_EDGE (pred, ei, block->preds)
3146 gimple_seq stmts = NULL;
3147 tree builtexpr;
3148 bprime = pred->src;
3149 eprime = avail[bprime->index];
3151 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3153 builtexpr = create_expression_by_pieces (bprime,
3154 eprime,
3155 &stmts, NULL,
3156 type);
3157 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3158 gsi_insert_seq_on_edge (pred, stmts);
3159 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3160 insertions = true;
3162 else if (eprime->kind == CONSTANT)
3164 /* Constants may not have the right type, fold_convert
3165 should give us back a constant with the right type.
3167 tree constant = PRE_EXPR_CONSTANT (eprime);
3168 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3170 tree builtexpr = fold_convert (type, constant);
3171 if (!is_gimple_min_invariant (builtexpr))
3173 tree forcedexpr = force_gimple_operand (builtexpr,
3174 &stmts, true,
3175 NULL);
3176 if (!is_gimple_min_invariant (forcedexpr))
3178 if (forcedexpr != builtexpr)
3180 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3181 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3183 if (stmts)
3185 gimple_stmt_iterator gsi;
3186 gsi = gsi_start (stmts);
3187 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3189 gimple stmt = gsi_stmt (gsi);
3190 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3191 gimple_set_plf (stmt, NECESSARY, false);
3193 gsi_insert_seq_on_edge (pred, stmts);
3195 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3200 else if (eprime->kind == NAME)
3202 /* We may have to do a conversion because our value
3203 numbering can look through types in certain cases, but
3204 our IL requires all operands of a phi node have the same
3205 type. */
3206 tree name = PRE_EXPR_NAME (eprime);
3207 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3209 tree builtexpr;
3210 tree forcedexpr;
3211 builtexpr = fold_convert (type, name);
3212 forcedexpr = force_gimple_operand (builtexpr,
3213 &stmts, true,
3214 NULL);
3216 if (forcedexpr != name)
3218 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3219 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3222 if (stmts)
3224 gimple_stmt_iterator gsi;
3225 gsi = gsi_start (stmts);
3226 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3228 gimple stmt = gsi_stmt (gsi);
3229 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3230 gimple_set_plf (stmt, NECESSARY, false);
3232 gsi_insert_seq_on_edge (pred, stmts);
3234 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3238 /* If we didn't want a phi node, and we made insertions, we still have
3239 inserted new stuff, and thus return true. If we didn't want a phi node,
3240 and didn't make insertions, we haven't added anything new, so return
3241 false. */
3242 if (nophi && insertions)
3243 return true;
3244 else if (nophi && !insertions)
3245 return false;
3247 /* Now build a phi for the new variable. */
3248 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3250 prephitemp = create_tmp_var (type, "prephitmp");
3251 get_var_ann (prephitemp);
3254 temp = prephitemp;
3255 add_referenced_var (temp);
3257 if (TREE_CODE (type) == COMPLEX_TYPE
3258 || TREE_CODE (type) == VECTOR_TYPE)
3259 DECL_GIMPLE_REG_P (temp) = 1;
3260 phi = create_phi_node (temp, block);
3262 gimple_set_plf (phi, NECESSARY, false);
3263 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3264 VN_INFO (gimple_phi_result (phi))->value_id = val;
3265 VEC_safe_push (gimple, heap, inserted_exprs, phi);
3266 bitmap_set_bit (inserted_phi_names,
3267 SSA_NAME_VERSION (gimple_phi_result (phi)));
3268 FOR_EACH_EDGE (pred, ei, block->preds)
3270 pre_expr ae = avail[pred->src->index];
3271 gcc_assert (get_expr_type (ae) == type
3272 || useless_type_conversion_p (type, get_expr_type (ae)));
3273 if (ae->kind == CONSTANT)
3274 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred);
3275 else
3276 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred);
3279 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3280 add_to_value (val, newphi);
3282 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3283 this insertion, since we test for the existence of this value in PHI_GEN
3284 before proceeding with the partial redundancy checks in insert_aux.
3286 The value may exist in AVAIL_OUT, in particular, it could be represented
3287 by the expression we are trying to eliminate, in which case we want the
3288 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3289 inserted there.
3291 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3292 this block, because if it did, it would have existed in our dominator's
3293 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3296 bitmap_insert_into_set (PHI_GEN (block), newphi);
3297 bitmap_value_replace_in_set (AVAIL_OUT (block),
3298 newphi);
3299 bitmap_insert_into_set (NEW_SETS (block),
3300 newphi);
3302 if (dump_file && (dump_flags & TDF_DETAILS))
3304 fprintf (dump_file, "Created phi ");
3305 print_gimple_stmt (dump_file, phi, 0, 0);
3306 fprintf (dump_file, " in block %d\n", block->index);
3308 pre_stats.phis++;
3309 return true;
3314 /* Perform insertion of partially redundant values.
3315 For BLOCK, do the following:
3316 1. Propagate the NEW_SETS of the dominator into the current block.
3317 If the block has multiple predecessors,
3318 2a. Iterate over the ANTIC expressions for the block to see if
3319 any of them are partially redundant.
3320 2b. If so, insert them into the necessary predecessors to make
3321 the expression fully redundant.
3322 2c. Insert a new PHI merging the values of the predecessors.
3323 2d. Insert the new PHI, and the new expressions, into the
3324 NEW_SETS set.
3325 3. Recursively call ourselves on the dominator children of BLOCK.
3327 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3328 do_regular_insertion and do_partial_insertion.
3332 static bool
3333 do_regular_insertion (basic_block block, basic_block dom)
3335 bool new_stuff = false;
3336 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3337 pre_expr expr;
3338 int i;
3340 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3342 if (expr->kind != NAME)
3344 pre_expr *avail;
3345 unsigned int val;
3346 bool by_some = false;
3347 bool cant_insert = false;
3348 bool all_same = true;
3349 pre_expr first_s = NULL;
3350 edge pred;
3351 basic_block bprime;
3352 pre_expr eprime = NULL;
3353 edge_iterator ei;
3354 pre_expr edoubleprime = NULL;
3356 val = get_expr_value_id (expr);
3357 if (bitmap_set_contains_value (PHI_GEN (block), val))
3358 continue;
3359 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3361 if (dump_file && (dump_flags & TDF_DETAILS))
3362 fprintf (dump_file, "Found fully redundant value\n");
3363 continue;
3366 avail = XCNEWVEC (pre_expr, last_basic_block);
3367 FOR_EACH_EDGE (pred, ei, block->preds)
3369 unsigned int vprime;
3371 /* We should never run insertion for the exit block
3372 and so not come across fake pred edges. */
3373 gcc_assert (!(pred->flags & EDGE_FAKE));
3374 bprime = pred->src;
3375 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3376 bprime, block);
3378 /* eprime will generally only be NULL if the
3379 value of the expression, translated
3380 through the PHI for this predecessor, is
3381 undefined. If that is the case, we can't
3382 make the expression fully redundant,
3383 because its value is undefined along a
3384 predecessor path. We can thus break out
3385 early because it doesn't matter what the
3386 rest of the results are. */
3387 if (eprime == NULL)
3389 cant_insert = true;
3390 break;
3393 eprime = fully_constant_expression (eprime);
3394 vprime = get_expr_value_id (eprime);
3395 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3396 vprime, NULL);
3397 if (edoubleprime == NULL)
3399 avail[bprime->index] = eprime;
3400 all_same = false;
3402 else
3404 avail[bprime->index] = edoubleprime;
3405 by_some = true;
3406 if (first_s == NULL)
3407 first_s = edoubleprime;
3408 else if (!pre_expr_eq (first_s, edoubleprime))
3409 all_same = false;
3412 /* If we can insert it, it's not the same value
3413 already existing along every predecessor, and
3414 it's defined by some predecessor, it is
3415 partially redundant. */
3416 if (!cant_insert && !all_same && by_some && dbg_cnt (treepre_insert))
3418 if (insert_into_preds_of_block (block, get_expression_id (expr),
3419 avail))
3420 new_stuff = true;
3422 /* If all edges produce the same value and that value is
3423 an invariant, then the PHI has the same value on all
3424 edges. Note this. */
3425 else if (!cant_insert && all_same && eprime
3426 && (edoubleprime->kind == CONSTANT
3427 || edoubleprime->kind == NAME)
3428 && !value_id_constant_p (val))
3430 unsigned int j;
3431 bitmap_iterator bi;
3432 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3433 value_expressions, val);
3435 unsigned int new_val = get_expr_value_id (edoubleprime);
3436 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3438 pre_expr expr = expression_for_id (j);
3440 if (expr->kind == NAME)
3442 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3443 /* Just reset the value id and valnum so it is
3444 the same as the constant we have discovered. */
3445 if (edoubleprime->kind == CONSTANT)
3447 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3448 pre_stats.constified++;
3450 else
3451 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3452 info->value_id = new_val;
3456 free (avail);
3460 VEC_free (pre_expr, heap, exprs);
3461 return new_stuff;
3465 /* Perform insertion for partially anticipatable expressions. There
3466 is only one case we will perform insertion for these. This case is
3467 if the expression is partially anticipatable, and fully available.
3468 In this case, we know that putting it earlier will enable us to
3469 remove the later computation. */
3472 static bool
3473 do_partial_partial_insertion (basic_block block, basic_block dom)
3475 bool new_stuff = false;
3476 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3477 pre_expr expr;
3478 int i;
3480 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3482 if (expr->kind != NAME)
3484 pre_expr *avail;
3485 unsigned int val;
3486 bool by_all = true;
3487 bool cant_insert = false;
3488 edge pred;
3489 basic_block bprime;
3490 pre_expr eprime = NULL;
3491 edge_iterator ei;
3493 val = get_expr_value_id (expr);
3494 if (bitmap_set_contains_value (PHI_GEN (block), val))
3495 continue;
3496 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3497 continue;
3499 avail = XCNEWVEC (pre_expr, last_basic_block);
3500 FOR_EACH_EDGE (pred, ei, block->preds)
3502 unsigned int vprime;
3503 pre_expr edoubleprime;
3505 /* We should never run insertion for the exit block
3506 and so not come across fake pred edges. */
3507 gcc_assert (!(pred->flags & EDGE_FAKE));
3508 bprime = pred->src;
3509 eprime = phi_translate (expr, ANTIC_IN (block),
3510 PA_IN (block),
3511 bprime, block);
3513 /* eprime will generally only be NULL if the
3514 value of the expression, translated
3515 through the PHI for this predecessor, is
3516 undefined. If that is the case, we can't
3517 make the expression fully redundant,
3518 because its value is undefined along a
3519 predecessor path. We can thus break out
3520 early because it doesn't matter what the
3521 rest of the results are. */
3522 if (eprime == NULL)
3524 cant_insert = true;
3525 break;
3528 eprime = fully_constant_expression (eprime);
3529 vprime = get_expr_value_id (eprime);
3530 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3531 vprime, NULL);
3532 if (edoubleprime == NULL)
3534 by_all = false;
3535 break;
3537 else
3538 avail[bprime->index] = edoubleprime;
3542 /* If we can insert it, it's not the same value
3543 already existing along every predecessor, and
3544 it's defined by some predecessor, it is
3545 partially redundant. */
3546 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3548 pre_stats.pa_insert++;
3549 if (insert_into_preds_of_block (block, get_expression_id (expr),
3550 avail))
3551 new_stuff = true;
3553 free (avail);
3557 VEC_free (pre_expr, heap, exprs);
3558 return new_stuff;
3561 static bool
3562 insert_aux (basic_block block)
3564 basic_block son;
3565 bool new_stuff = false;
3567 if (block)
3569 basic_block dom;
3570 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3571 if (dom)
3573 unsigned i;
3574 bitmap_iterator bi;
3575 bitmap_set_t newset = NEW_SETS (dom);
3576 if (newset)
3578 /* Note that we need to value_replace both NEW_SETS, and
3579 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3580 represented by some non-simple expression here that we want
3581 to replace it with. */
3582 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3584 pre_expr expr = expression_for_id (i);
3585 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3586 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3589 if (!single_pred_p (block))
3591 new_stuff |= do_regular_insertion (block, dom);
3592 if (do_partial_partial)
3593 new_stuff |= do_partial_partial_insertion (block, dom);
3597 for (son = first_dom_son (CDI_DOMINATORS, block);
3598 son;
3599 son = next_dom_son (CDI_DOMINATORS, son))
3601 new_stuff |= insert_aux (son);
3604 return new_stuff;
3607 /* Perform insertion of partially redundant values. */
3609 static void
3610 insert (void)
3612 bool new_stuff = true;
3613 basic_block bb;
3614 int num_iterations = 0;
3616 FOR_ALL_BB (bb)
3617 NEW_SETS (bb) = bitmap_set_new ();
3619 while (new_stuff)
3621 num_iterations++;
3622 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3624 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3628 /* Add OP to EXP_GEN (block), and possibly to the maximal set if it is
3629 not defined by a phi node.
3630 PHI nodes can't go in the maximal sets because they are not in
3631 TMP_GEN, so it is possible to get into non-monotonic situations
3632 during ANTIC calculation, because it will *add* bits. */
3634 static void
3635 add_to_exp_gen (basic_block block, tree op)
3637 if (!in_fre)
3639 pre_expr result;
3640 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3641 return;
3642 result = get_or_alloc_expr_for_name (op);
3643 bitmap_value_insert_into_set (EXP_GEN (block), result);
3644 if (TREE_CODE (op) != SSA_NAME
3645 || gimple_code (SSA_NAME_DEF_STMT (op)) != GIMPLE_PHI)
3646 bitmap_value_insert_into_set (maximal_set, result);
3650 /* Create value ids for PHI in BLOCK. */
3652 static void
3653 make_values_for_phi (gimple phi, basic_block block)
3655 tree result = gimple_phi_result (phi);
3657 /* We have no need for virtual phis, as they don't represent
3658 actual computations. */
3659 if (is_gimple_reg (result))
3661 pre_expr e = get_or_alloc_expr_for_name (result);
3662 add_to_value (get_expr_value_id (e), e);
3663 bitmap_insert_into_set (PHI_GEN (block), e);
3664 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3668 /* Compute the AVAIL set for all basic blocks.
3670 This function performs value numbering of the statements in each basic
3671 block. The AVAIL sets are built from information we glean while doing
3672 this value numbering, since the AVAIL sets contain only one entry per
3673 value.
3675 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3676 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3678 static void
3679 compute_avail (void)
3682 basic_block block, son;
3683 basic_block *worklist;
3684 size_t sp = 0;
3685 unsigned i;
3687 /* We pretend that default definitions are defined in the entry block.
3688 This includes function arguments and the static chain decl. */
3689 for (i = 1; i < num_ssa_names; ++i)
3691 tree name = ssa_name (i);
3692 pre_expr e;
3693 if (!name
3694 || !SSA_NAME_IS_DEFAULT_DEF (name)
3695 || has_zero_uses (name)
3696 || !is_gimple_reg (name))
3697 continue;
3699 e = get_or_alloc_expr_for_name (name);
3700 add_to_value (get_expr_value_id (e), e);
3701 if (!in_fre)
3703 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3704 bitmap_value_insert_into_set (maximal_set, e);
3706 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3709 /* Allocate the worklist. */
3710 worklist = XNEWVEC (basic_block, n_basic_blocks);
3712 /* Seed the algorithm by putting the dominator children of the entry
3713 block on the worklist. */
3714 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3715 son;
3716 son = next_dom_son (CDI_DOMINATORS, son))
3717 worklist[sp++] = son;
3719 /* Loop until the worklist is empty. */
3720 while (sp)
3722 gimple_stmt_iterator gsi;
3723 gimple stmt;
3724 basic_block dom;
3725 unsigned int stmt_uid = 1;
3727 /* Pick a block from the worklist. */
3728 block = worklist[--sp];
3730 /* Initially, the set of available values in BLOCK is that of
3731 its immediate dominator. */
3732 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3733 if (dom)
3734 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3736 /* Generate values for PHI nodes. */
3737 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3738 make_values_for_phi (gsi_stmt (gsi), block);
3740 /* Now compute value numbers and populate value sets with all
3741 the expressions computed in BLOCK. */
3742 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3744 ssa_op_iter iter;
3745 tree op;
3747 stmt = gsi_stmt (gsi);
3748 gimple_set_uid (stmt, stmt_uid++);
3750 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3752 pre_expr e = get_or_alloc_expr_for_name (op);
3754 add_to_value (get_expr_value_id (e), e);
3755 if (!in_fre)
3756 bitmap_insert_into_set (TMP_GEN (block), e);
3757 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3760 if (gimple_has_volatile_ops (stmt)
3761 || stmt_could_throw_p (stmt))
3762 continue;
3764 switch (gimple_code (stmt))
3766 case GIMPLE_RETURN:
3767 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3768 add_to_exp_gen (block, op);
3769 continue;
3771 case GIMPLE_CALL:
3773 vn_reference_t ref;
3774 unsigned int i;
3775 vn_reference_op_t vro;
3776 pre_expr result = NULL;
3777 VEC(vn_reference_op_s, heap) *ops = NULL;
3779 if (!can_value_number_call (stmt))
3780 continue;
3782 copy_reference_ops_from_call (stmt, &ops);
3783 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3784 gimple_expr_type (stmt),
3785 ops, &ref, false);
3786 VEC_free (vn_reference_op_s, heap, ops);
3787 if (!ref)
3788 continue;
3790 for (i = 0; VEC_iterate (vn_reference_op_s,
3791 ref->operands, i,
3792 vro); i++)
3794 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3795 add_to_exp_gen (block, vro->op0);
3796 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3797 add_to_exp_gen (block, vro->op1);
3798 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3799 add_to_exp_gen (block, vro->op2);
3801 result = (pre_expr) pool_alloc (pre_expr_pool);
3802 result->kind = REFERENCE;
3803 result->id = 0;
3804 PRE_EXPR_REFERENCE (result) = ref;
3806 get_or_alloc_expression_id (result);
3807 add_to_value (get_expr_value_id (result), result);
3808 if (!in_fre)
3810 bitmap_value_insert_into_set (EXP_GEN (block),
3811 result);
3812 bitmap_value_insert_into_set (maximal_set, result);
3814 continue;
3817 case GIMPLE_ASSIGN:
3819 pre_expr result = NULL;
3820 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
3822 case tcc_unary:
3823 if (is_exception_related (stmt))
3824 continue;
3825 case tcc_binary:
3826 case tcc_comparison:
3828 vn_nary_op_t nary;
3829 unsigned int i;
3831 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
3832 gimple_assign_rhs_code (stmt),
3833 gimple_expr_type (stmt),
3834 gimple_assign_rhs1 (stmt),
3835 gimple_assign_rhs2 (stmt),
3836 NULL_TREE, NULL_TREE, &nary);
3838 if (!nary)
3839 continue;
3841 for (i = 0; i < nary->length; i++)
3842 if (TREE_CODE (nary->op[i]) == SSA_NAME)
3843 add_to_exp_gen (block, nary->op[i]);
3845 result = (pre_expr) pool_alloc (pre_expr_pool);
3846 result->kind = NARY;
3847 result->id = 0;
3848 PRE_EXPR_NARY (result) = nary;
3849 break;
3852 case tcc_declaration:
3853 case tcc_reference:
3855 vn_reference_t ref;
3856 unsigned int i;
3857 vn_reference_op_t vro;
3859 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3860 gimple_vuse (stmt),
3861 false, &ref);
3862 if (!ref)
3863 continue;
3865 for (i = 0; VEC_iterate (vn_reference_op_s,
3866 ref->operands, i,
3867 vro); i++)
3869 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3870 add_to_exp_gen (block, vro->op0);
3871 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3872 add_to_exp_gen (block, vro->op1);
3873 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3874 add_to_exp_gen (block, vro->op2);
3876 result = (pre_expr) pool_alloc (pre_expr_pool);
3877 result->kind = REFERENCE;
3878 result->id = 0;
3879 PRE_EXPR_REFERENCE (result) = ref;
3880 break;
3883 default:
3884 /* For any other statement that we don't
3885 recognize, simply add all referenced
3886 SSA_NAMEs to EXP_GEN. */
3887 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3888 add_to_exp_gen (block, op);
3889 continue;
3892 get_or_alloc_expression_id (result);
3893 add_to_value (get_expr_value_id (result), result);
3894 if (!in_fre)
3896 bitmap_value_insert_into_set (EXP_GEN (block), result);
3897 bitmap_value_insert_into_set (maximal_set, result);
3900 continue;
3902 default:
3903 break;
3907 /* Put the dominator children of BLOCK on the worklist of blocks
3908 to compute available sets for. */
3909 for (son = first_dom_son (CDI_DOMINATORS, block);
3910 son;
3911 son = next_dom_son (CDI_DOMINATORS, son))
3912 worklist[sp++] = son;
3915 free (worklist);
3918 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
3919 than the available expressions for it. The insertion point is
3920 right before the first use in STMT. Returns the SSA_NAME that should
3921 be used for replacement. */
3923 static tree
3924 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
3926 basic_block bb = gimple_bb (stmt);
3927 gimple_stmt_iterator gsi;
3928 gimple_seq stmts = NULL;
3929 tree expr;
3930 pre_expr e;
3932 /* First create a value expression from the expression we want
3933 to insert and associate it with the value handle for SSA_VN. */
3934 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
3935 if (e == NULL)
3936 return NULL_TREE;
3938 /* Then use create_expression_by_pieces to generate a valid
3939 expression to insert at this point of the IL stream. */
3940 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
3941 if (expr == NULL_TREE)
3942 return NULL_TREE;
3943 gsi = gsi_for_stmt (stmt);
3944 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
3946 return expr;
3949 /* Eliminate fully redundant computations. */
3951 static unsigned int
3952 eliminate (void)
3954 VEC (gimple, heap) *to_remove = NULL;
3955 basic_block b;
3956 unsigned int todo = 0;
3957 gimple_stmt_iterator gsi;
3958 gimple stmt;
3959 unsigned i;
3961 FOR_EACH_BB (b)
3963 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
3965 stmt = gsi_stmt (gsi);
3967 /* Lookup the RHS of the expression, see if we have an
3968 available computation for it. If so, replace the RHS with
3969 the available computation. */
3970 if (gimple_has_lhs (stmt)
3971 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
3972 && !gimple_assign_ssa_name_copy_p (stmt)
3973 && (!gimple_assign_single_p (stmt)
3974 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
3975 && !gimple_has_volatile_ops (stmt)
3976 && !has_zero_uses (gimple_get_lhs (stmt)))
3978 tree lhs = gimple_get_lhs (stmt);
3979 tree rhs = NULL_TREE;
3980 tree sprime = NULL;
3981 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
3982 pre_expr sprimeexpr;
3984 if (gimple_assign_single_p (stmt))
3985 rhs = gimple_assign_rhs1 (stmt);
3987 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
3988 get_expr_value_id (lhsexpr),
3989 NULL);
3991 if (sprimeexpr)
3993 if (sprimeexpr->kind == CONSTANT)
3994 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
3995 else if (sprimeexpr->kind == NAME)
3996 sprime = PRE_EXPR_NAME (sprimeexpr);
3997 else
3998 gcc_unreachable ();
4001 /* If there is no existing leader but SCCVN knows this
4002 value is constant, use that constant. */
4003 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4005 sprime = VN_INFO (lhs)->valnum;
4006 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4007 TREE_TYPE (sprime)))
4008 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4010 if (dump_file && (dump_flags & TDF_DETAILS))
4012 fprintf (dump_file, "Replaced ");
4013 print_gimple_expr (dump_file, stmt, 0, 0);
4014 fprintf (dump_file, " with ");
4015 print_generic_expr (dump_file, sprime, 0);
4016 fprintf (dump_file, " in ");
4017 print_gimple_stmt (dump_file, stmt, 0, 0);
4019 pre_stats.eliminations++;
4020 propagate_tree_value_into_stmt (&gsi, sprime);
4021 stmt = gsi_stmt (gsi);
4022 update_stmt (stmt);
4023 continue;
4026 /* If there is no existing usable leader but SCCVN thinks
4027 it has an expression it wants to use as replacement,
4028 insert that. */
4029 if (!sprime || sprime == lhs)
4031 tree val = VN_INFO (lhs)->valnum;
4032 if (val != VN_TOP
4033 && TREE_CODE (val) == SSA_NAME
4034 && VN_INFO (val)->needs_insertion
4035 && can_PRE_operation (vn_get_expr_for (val)))
4036 sprime = do_SCCVN_insertion (stmt, val);
4038 if (sprime
4039 && sprime != lhs
4040 && (rhs == NULL_TREE
4041 || TREE_CODE (rhs) != SSA_NAME
4042 || may_propagate_copy (rhs, sprime)))
4044 gcc_assert (sprime != rhs);
4046 if (dump_file && (dump_flags & TDF_DETAILS))
4048 fprintf (dump_file, "Replaced ");
4049 print_gimple_expr (dump_file, stmt, 0, 0);
4050 fprintf (dump_file, " with ");
4051 print_generic_expr (dump_file, sprime, 0);
4052 fprintf (dump_file, " in ");
4053 print_gimple_stmt (dump_file, stmt, 0, 0);
4056 if (TREE_CODE (sprime) == SSA_NAME)
4057 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4058 NECESSARY, true);
4059 /* We need to make sure the new and old types actually match,
4060 which may require adding a simple cast, which fold_convert
4061 will do for us. */
4062 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4063 && !useless_type_conversion_p (gimple_expr_type (stmt),
4064 TREE_TYPE (sprime)))
4065 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4067 pre_stats.eliminations++;
4068 propagate_tree_value_into_stmt (&gsi, sprime);
4069 stmt = gsi_stmt (gsi);
4070 update_stmt (stmt);
4072 /* If we removed EH side effects from the statement, clean
4073 its EH information. */
4074 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4076 bitmap_set_bit (need_eh_cleanup,
4077 gimple_bb (stmt)->index);
4078 if (dump_file && (dump_flags & TDF_DETAILS))
4079 fprintf (dump_file, " Removed EH side effects.\n");
4083 /* If the statement is a scalar store, see if the expression
4084 has the same value number as its rhs. If so, the store is
4085 dead. */
4086 else if (gimple_assign_single_p (stmt)
4087 && !is_gimple_reg (gimple_assign_lhs (stmt))
4088 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4089 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4091 tree rhs = gimple_assign_rhs1 (stmt);
4092 tree val;
4093 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4094 gimple_vuse (stmt), true, NULL);
4095 if (TREE_CODE (rhs) == SSA_NAME)
4096 rhs = VN_INFO (rhs)->valnum;
4097 if (val
4098 && operand_equal_p (val, rhs, 0))
4100 if (dump_file && (dump_flags & TDF_DETAILS))
4102 fprintf (dump_file, "Deleted redundant store ");
4103 print_gimple_stmt (dump_file, stmt, 0, 0);
4106 /* Queue stmt for removal. */
4107 VEC_safe_push (gimple, heap, to_remove, stmt);
4110 /* Visit COND_EXPRs and fold the comparison with the
4111 available value-numbers. */
4112 else if (gimple_code (stmt) == GIMPLE_COND)
4114 tree op0 = gimple_cond_lhs (stmt);
4115 tree op1 = gimple_cond_rhs (stmt);
4116 tree result;
4118 if (TREE_CODE (op0) == SSA_NAME)
4119 op0 = VN_INFO (op0)->valnum;
4120 if (TREE_CODE (op1) == SSA_NAME)
4121 op1 = VN_INFO (op1)->valnum;
4122 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4123 op0, op1);
4124 if (result && TREE_CODE (result) == INTEGER_CST)
4126 if (integer_zerop (result))
4127 gimple_cond_make_false (stmt);
4128 else
4129 gimple_cond_make_true (stmt);
4130 update_stmt (stmt);
4131 todo = TODO_cleanup_cfg;
4134 /* Visit indirect calls and turn them into direct calls if
4135 possible. */
4136 if (gimple_code (stmt) == GIMPLE_CALL
4137 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4139 tree fn = VN_INFO (gimple_call_fn (stmt))->valnum;
4140 if (TREE_CODE (fn) == ADDR_EXPR
4141 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
4143 if (dump_file && (dump_flags & TDF_DETAILS))
4145 fprintf (dump_file, "Replacing call target with ");
4146 print_generic_expr (dump_file, fn, 0);
4147 fprintf (dump_file, " in ");
4148 print_gimple_stmt (dump_file, stmt, 0, 0);
4151 gimple_call_set_fn (stmt, fn);
4152 update_stmt (stmt);
4153 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4154 gimple_purge_dead_eh_edges (b);
4156 /* Changing an indirect call to a direct call may
4157 have exposed different semantics. This may
4158 require an SSA update. */
4159 todo |= TODO_update_ssa_only_virtuals;
4164 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4166 gimple stmt, phi = gsi_stmt (gsi);
4167 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4168 pre_expr sprimeexpr, resexpr;
4169 gimple_stmt_iterator gsi2;
4171 /* We want to perform redundant PHI elimination. Do so by
4172 replacing the PHI with a single copy if possible.
4173 Do not touch inserted, single-argument or virtual PHIs. */
4174 if (gimple_phi_num_args (phi) == 1
4175 || !is_gimple_reg (res)
4176 || bitmap_bit_p (inserted_phi_names, SSA_NAME_VERSION (res)))
4178 gsi_next (&gsi);
4179 continue;
4182 resexpr = get_or_alloc_expr_for_name (res);
4183 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4184 get_expr_value_id (resexpr), NULL);
4185 if (sprimeexpr)
4187 if (sprimeexpr->kind == CONSTANT)
4188 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4189 else if (sprimeexpr->kind == NAME)
4190 sprime = PRE_EXPR_NAME (sprimeexpr);
4191 else
4192 gcc_unreachable ();
4194 if (!sprimeexpr
4195 || sprime == res)
4197 gsi_next (&gsi);
4198 continue;
4201 if (dump_file && (dump_flags & TDF_DETAILS))
4203 fprintf (dump_file, "Replaced redundant PHI node defining ");
4204 print_generic_expr (dump_file, res, 0);
4205 fprintf (dump_file, " with ");
4206 print_generic_expr (dump_file, sprime, 0);
4207 fprintf (dump_file, "\n");
4210 remove_phi_node (&gsi, false);
4212 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4213 sprime = fold_convert (TREE_TYPE (res), sprime);
4214 stmt = gimple_build_assign (res, sprime);
4215 SSA_NAME_DEF_STMT (res) = stmt;
4216 if (TREE_CODE (sprime) == SSA_NAME)
4217 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4218 NECESSARY, true);
4219 gsi2 = gsi_after_labels (b);
4220 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4221 /* Queue the copy for eventual removal. */
4222 VEC_safe_push (gimple, heap, to_remove, stmt);
4223 pre_stats.eliminations++;
4227 /* We cannot remove stmts during BB walk, especially not release SSA
4228 names there as this confuses the VN machinery. The stmts ending
4229 up in to_remove are either stores or simple copies. */
4230 for (i = 0; VEC_iterate (gimple, to_remove, i, stmt); ++i)
4232 tree lhs = gimple_assign_lhs (stmt);
4233 use_operand_p use_p;
4234 gimple use_stmt;
4236 /* If there is a single use only, propagate the equivalency
4237 instead of keeping the copy. */
4238 if (TREE_CODE (lhs) == SSA_NAME
4239 && single_imm_use (lhs, &use_p, &use_stmt)
4240 && may_propagate_copy (USE_FROM_PTR (use_p),
4241 gimple_assign_rhs1 (stmt)))
4243 SET_USE (use_p, gimple_assign_rhs1 (stmt));
4244 update_stmt (use_stmt);
4247 /* If this is a store or a now unused copy, remove it. */
4248 if (TREE_CODE (lhs) != SSA_NAME
4249 || has_zero_uses (lhs))
4251 gsi = gsi_for_stmt (stmt);
4252 unlink_stmt_vdef (stmt);
4253 gsi_remove (&gsi, true);
4254 release_defs (stmt);
4257 VEC_free (gimple, heap, to_remove);
4259 return todo;
4262 /* Borrow a bit of tree-ssa-dce.c for the moment.
4263 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4264 this may be a bit faster, and we may want critical edges kept split. */
4266 /* If OP's defining statement has not already been determined to be necessary,
4267 mark that statement necessary. Return the stmt, if it is newly
4268 necessary. */
4270 static inline gimple
4271 mark_operand_necessary (tree op)
4273 gimple stmt;
4275 gcc_assert (op);
4277 if (TREE_CODE (op) != SSA_NAME)
4278 return NULL;
4280 stmt = SSA_NAME_DEF_STMT (op);
4281 gcc_assert (stmt);
4283 if (gimple_plf (stmt, NECESSARY)
4284 || gimple_nop_p (stmt))
4285 return NULL;
4287 gimple_set_plf (stmt, NECESSARY, true);
4288 return stmt;
4291 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4292 to insert PHI nodes sometimes, and because value numbering of casts isn't
4293 perfect, we sometimes end up inserting dead code. This simple DCE-like
4294 pass removes any insertions we made that weren't actually used. */
4296 static void
4297 remove_dead_inserted_code (void)
4299 VEC(gimple,heap) *worklist = NULL;
4300 int i;
4301 gimple t;
4303 worklist = VEC_alloc (gimple, heap, VEC_length (gimple, inserted_exprs));
4304 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4306 if (gimple_plf (t, NECESSARY))
4307 VEC_quick_push (gimple, worklist, t);
4309 while (VEC_length (gimple, worklist) > 0)
4311 t = VEC_pop (gimple, worklist);
4313 /* PHI nodes are somewhat special in that each PHI alternative has
4314 data and control dependencies. All the statements feeding the
4315 PHI node's arguments are always necessary. */
4316 if (gimple_code (t) == GIMPLE_PHI)
4318 unsigned k;
4320 VEC_reserve (gimple, heap, worklist, gimple_phi_num_args (t));
4321 for (k = 0; k < gimple_phi_num_args (t); k++)
4323 tree arg = PHI_ARG_DEF (t, k);
4324 if (TREE_CODE (arg) == SSA_NAME)
4326 gimple n = mark_operand_necessary (arg);
4327 if (n)
4328 VEC_quick_push (gimple, worklist, n);
4332 else
4334 /* Propagate through the operands. Examine all the USE, VUSE and
4335 VDEF operands in this statement. Mark all the statements
4336 which feed this statement's uses as necessary. */
4337 ssa_op_iter iter;
4338 tree use;
4340 /* The operands of VDEF expressions are also needed as they
4341 represent potential definitions that may reach this
4342 statement (VDEF operands allow us to follow def-def
4343 links). */
4345 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4347 gimple n = mark_operand_necessary (use);
4348 if (n)
4349 VEC_safe_push (gimple, heap, worklist, n);
4354 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4356 if (!gimple_plf (t, NECESSARY))
4358 gimple_stmt_iterator gsi;
4360 if (dump_file && (dump_flags & TDF_DETAILS))
4362 fprintf (dump_file, "Removing unnecessary insertion:");
4363 print_gimple_stmt (dump_file, t, 0, 0);
4366 gsi = gsi_for_stmt (t);
4367 if (gimple_code (t) == GIMPLE_PHI)
4368 remove_phi_node (&gsi, true);
4369 else
4370 gsi_remove (&gsi, true);
4371 release_defs (t);
4374 VEC_free (gimple, heap, worklist);
4377 /* Initialize data structures used by PRE. */
4379 static void
4380 init_pre (bool do_fre)
4382 basic_block bb;
4384 next_expression_id = 1;
4385 expressions = NULL;
4386 VEC_safe_push (pre_expr, heap, expressions, NULL);
4387 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4388 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4389 get_max_value_id() + 1);
4391 in_fre = do_fre;
4393 inserted_exprs = NULL;
4394 need_creation = NULL;
4395 pretemp = NULL_TREE;
4396 storetemp = NULL_TREE;
4397 prephitemp = NULL_TREE;
4399 connect_infinite_loops_to_exit ();
4400 memset (&pre_stats, 0, sizeof (pre_stats));
4403 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4404 post_order_compute (postorder, false, false);
4406 FOR_ALL_BB (bb)
4407 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4409 calculate_dominance_info (CDI_POST_DOMINATORS);
4410 calculate_dominance_info (CDI_DOMINATORS);
4412 bitmap_obstack_initialize (&grand_bitmap_obstack);
4413 inserted_phi_names = BITMAP_ALLOC (&grand_bitmap_obstack);
4414 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4415 expr_pred_trans_eq, free);
4416 expression_to_id = htab_create (num_ssa_names * 3,
4417 pre_expr_hash,
4418 pre_expr_eq, NULL);
4419 seen_during_translate = BITMAP_ALLOC (&grand_bitmap_obstack);
4420 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4421 sizeof (struct bitmap_set), 30);
4422 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4423 sizeof (struct pre_expr_d), 30);
4424 FOR_ALL_BB (bb)
4426 EXP_GEN (bb) = bitmap_set_new ();
4427 PHI_GEN (bb) = bitmap_set_new ();
4428 TMP_GEN (bb) = bitmap_set_new ();
4429 AVAIL_OUT (bb) = bitmap_set_new ();
4431 maximal_set = in_fre ? NULL : bitmap_set_new ();
4433 need_eh_cleanup = BITMAP_ALLOC (NULL);
4437 /* Deallocate data structures used by PRE. */
4439 static void
4440 fini_pre (bool do_fre)
4442 basic_block bb;
4444 free (postorder);
4445 VEC_free (bitmap_set_t, heap, value_expressions);
4446 VEC_free (gimple, heap, inserted_exprs);
4447 VEC_free (gimple, heap, need_creation);
4448 bitmap_obstack_release (&grand_bitmap_obstack);
4449 free_alloc_pool (bitmap_set_pool);
4450 free_alloc_pool (pre_expr_pool);
4451 htab_delete (phi_translate_table);
4452 htab_delete (expression_to_id);
4454 FOR_ALL_BB (bb)
4456 free (bb->aux);
4457 bb->aux = NULL;
4460 free_dominance_info (CDI_POST_DOMINATORS);
4462 if (!bitmap_empty_p (need_eh_cleanup))
4464 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4465 cleanup_tree_cfg ();
4468 BITMAP_FREE (need_eh_cleanup);
4470 if (!do_fre)
4471 loop_optimizer_finalize ();
4474 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4475 only wants to do full redundancy elimination. */
4477 static unsigned int
4478 execute_pre (bool do_fre ATTRIBUTE_UNUSED)
4480 unsigned int todo = 0;
4482 do_partial_partial = optimize > 2;
4484 /* This has to happen before SCCVN runs because
4485 loop_optimizer_init may create new phis, etc. */
4486 if (!do_fre)
4487 loop_optimizer_init (LOOPS_NORMAL);
4489 if (!run_scc_vn (do_fre))
4491 if (!do_fre)
4493 remove_dead_inserted_code ();
4494 loop_optimizer_finalize ();
4497 return 0;
4499 init_pre (do_fre);
4502 /* Collect and value number expressions computed in each basic block. */
4503 compute_avail ();
4505 if (dump_file && (dump_flags & TDF_DETAILS))
4507 basic_block bb;
4509 FOR_ALL_BB (bb)
4511 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4512 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen",
4513 bb->index);
4514 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out",
4515 bb->index);
4519 /* Insert can get quite slow on an incredibly large number of basic
4520 blocks due to some quadratic behavior. Until this behavior is
4521 fixed, don't run it when he have an incredibly large number of
4522 bb's. If we aren't going to run insert, there is no point in
4523 computing ANTIC, either, even though it's plenty fast. */
4524 if (!do_fre && n_basic_blocks < 4000)
4526 compute_antic ();
4527 insert ();
4530 /* Remove all the redundant expressions. */
4531 todo |= eliminate ();
4533 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4534 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4535 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4536 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4537 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4539 /* Make sure to remove fake edges before committing our inserts.
4540 This makes sure we don't end up with extra critical edges that
4541 we would need to split. */
4542 remove_fake_exit_edges ();
4543 gsi_commit_edge_inserts ();
4545 clear_expression_ids ();
4546 free_scc_vn ();
4547 if (!do_fre)
4548 remove_dead_inserted_code ();
4550 fini_pre (do_fre);
4552 return todo;
4555 /* Gate and execute functions for PRE. */
4557 static unsigned int
4558 do_pre (void)
4560 return execute_pre (false);
4563 static bool
4564 gate_pre (void)
4566 /* PRE tends to generate bigger code. */
4567 return flag_tree_pre != 0 && optimize_function_for_speed_p (cfun);
4570 struct gimple_opt_pass pass_pre =
4573 GIMPLE_PASS,
4574 "pre", /* name */
4575 gate_pre, /* gate */
4576 do_pre, /* execute */
4577 NULL, /* sub */
4578 NULL, /* next */
4579 0, /* static_pass_number */
4580 TV_TREE_PRE, /* tv_id */
4581 PROP_no_crit_edges | PROP_cfg
4582 | PROP_ssa, /* properties_required */
4583 0, /* properties_provided */
4584 0, /* properties_destroyed */
4585 TODO_rebuild_alias, /* todo_flags_start */
4586 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4587 | TODO_verify_ssa /* todo_flags_finish */
4592 /* Gate and execute functions for FRE. */
4594 static unsigned int
4595 execute_fre (void)
4597 return execute_pre (true);
4600 static bool
4601 gate_fre (void)
4603 return flag_tree_fre != 0;
4606 struct gimple_opt_pass pass_fre =
4609 GIMPLE_PASS,
4610 "fre", /* name */
4611 gate_fre, /* gate */
4612 execute_fre, /* execute */
4613 NULL, /* sub */
4614 NULL, /* next */
4615 0, /* static_pass_number */
4616 TV_TREE_FRE, /* tv_id */
4617 PROP_cfg | PROP_ssa, /* properties_required */
4618 0, /* properties_provided */
4619 0, /* properties_destroyed */
4620 0, /* todo_flags_start */
4621 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */