Fix DealII type problems.
[official-gcc/Ramakrishna.git] / gcc / tree-ssa-pre.c
bloba3e7ad824b4c65e82eacd99d82c216e7860fda31
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 "tree-scalar-evolution.h"
49 #include "params.h"
50 #include "dbgcnt.h"
52 /* TODO:
54 1. Avail sets can be shared by making an avail_find_leader that
55 walks up the dominator tree and looks in those avail sets.
56 This might affect code optimality, it's unclear right now.
57 2. Strength reduction can be performed by anticipating expressions
58 we can repair later on.
59 3. We can do back-substitution or smarter value numbering to catch
60 commutative expressions split up over multiple statements.
63 /* For ease of terminology, "expression node" in the below refers to
64 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
65 represent the actual statement containing the expressions we care about,
66 and we cache the value number by putting it in the expression. */
68 /* Basic algorithm
70 First we walk the statements to generate the AVAIL sets, the
71 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
72 generation of values/expressions by a given block. We use them
73 when computing the ANTIC sets. The AVAIL sets consist of
74 SSA_NAME's that represent values, so we know what values are
75 available in what blocks. AVAIL is a forward dataflow problem. In
76 SSA, values are never killed, so we don't need a kill set, or a
77 fixpoint iteration, in order to calculate the AVAIL sets. In
78 traditional parlance, AVAIL sets tell us the downsafety of the
79 expressions/values.
81 Next, we generate the ANTIC sets. These sets represent the
82 anticipatable expressions. ANTIC is a backwards dataflow
83 problem. An expression is anticipatable in a given block if it could
84 be generated in that block. This means that if we had to perform
85 an insertion in that block, of the value of that expression, we
86 could. Calculating the ANTIC sets requires phi translation of
87 expressions, because the flow goes backwards through phis. We must
88 iterate to a fixpoint of the ANTIC sets, because we have a kill
89 set. Even in SSA form, values are not live over the entire
90 function, only from their definition point onwards. So we have to
91 remove values from the ANTIC set once we go past the definition
92 point of the leaders that make them up.
93 compute_antic/compute_antic_aux performs this computation.
95 Third, we perform insertions to make partially redundant
96 expressions fully redundant.
98 An expression is partially redundant (excluding partial
99 anticipation) if:
101 1. It is AVAIL in some, but not all, of the predecessors of a
102 given block.
103 2. It is ANTIC in all the predecessors.
105 In order to make it fully redundant, we insert the expression into
106 the predecessors where it is not available, but is ANTIC.
108 For the partial anticipation case, we only perform insertion if it
109 is partially anticipated in some block, and fully available in all
110 of the predecessors.
112 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
113 performs these steps.
115 Fourth, we eliminate fully redundant expressions.
116 This is a simple statement walk that replaces redundant
117 calculations with the now available values. */
119 /* Representations of value numbers:
121 Value numbers are represented by a representative SSA_NAME. We
122 will create fake SSA_NAME's in situations where we need a
123 representative but do not have one (because it is a complex
124 expression). In order to facilitate storing the value numbers in
125 bitmaps, and keep the number of wasted SSA_NAME's down, we also
126 associate a value_id with each value number, and create full blown
127 ssa_name's only where we actually need them (IE in operands of
128 existing expressions).
130 Theoretically you could replace all the value_id's with
131 SSA_NAME_VERSION, but this would allocate a large number of
132 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
133 It would also require an additional indirection at each point we
134 use the value id. */
136 /* Representation of expressions on value numbers:
138 Expressions consisting of value numbers are represented the same
139 way as our VN internally represents them, with an additional
140 "pre_expr" wrapping around them in order to facilitate storing all
141 of the expressions in the same sets. */
143 /* Representation of sets:
145 The dataflow sets do not need to be sorted in any particular order
146 for the majority of their lifetime, are simply represented as two
147 bitmaps, one that keeps track of values present in the set, and one
148 that keeps track of expressions present in the set.
150 When we need them in topological order, we produce it on demand by
151 transforming the bitmap into an array and sorting it into topo
152 order. */
154 /* Type of expression, used to know which member of the PRE_EXPR union
155 is valid. */
157 enum pre_expr_kind
159 NAME,
160 NARY,
161 REFERENCE,
162 CONSTANT
165 typedef union pre_expr_union_d
167 tree name;
168 tree constant;
169 vn_nary_op_t nary;
170 vn_reference_t reference;
171 } pre_expr_union;
173 typedef struct pre_expr_d
175 enum pre_expr_kind kind;
176 unsigned int id;
177 pre_expr_union u;
178 } *pre_expr;
180 #define PRE_EXPR_NAME(e) (e)->u.name
181 #define PRE_EXPR_NARY(e) (e)->u.nary
182 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
183 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
185 static int
186 pre_expr_eq (const void *p1, const void *p2)
188 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
189 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
191 if (e1->kind != e2->kind)
192 return false;
194 switch (e1->kind)
196 case CONSTANT:
197 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
198 PRE_EXPR_CONSTANT (e2));
199 case NAME:
200 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
201 case NARY:
202 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
203 case REFERENCE:
204 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
205 PRE_EXPR_REFERENCE (e2));
206 default:
207 abort();
211 static hashval_t
212 pre_expr_hash (const void *p1)
214 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
215 switch (e->kind)
217 case CONSTANT:
218 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
219 case NAME:
220 return iterative_hash_hashval_t (SSA_NAME_VERSION (PRE_EXPR_NAME (e)), 0);
221 case NARY:
222 return PRE_EXPR_NARY (e)->hashcode;
223 case REFERENCE:
224 return PRE_EXPR_REFERENCE (e)->hashcode;
225 default:
226 abort ();
231 /* Next global expression id number. */
232 static unsigned int next_expression_id;
234 /* Mapping from expression to id number we can use in bitmap sets. */
235 DEF_VEC_P (pre_expr);
236 DEF_VEC_ALLOC_P (pre_expr, heap);
237 static VEC(pre_expr, heap) *expressions;
238 static htab_t expression_to_id;
240 /* Allocate an expression id for EXPR. */
242 static inline unsigned int
243 alloc_expression_id (pre_expr expr)
245 void **slot;
246 /* Make sure we won't overflow. */
247 gcc_assert (next_expression_id + 1 > next_expression_id);
248 expr->id = next_expression_id++;
249 VEC_safe_push (pre_expr, heap, expressions, expr);
250 slot = htab_find_slot (expression_to_id, expr, INSERT);
251 gcc_assert (!*slot);
252 *slot = expr;
253 return next_expression_id - 1;
256 /* Return the expression id for tree EXPR. */
258 static inline unsigned int
259 get_expression_id (const pre_expr expr)
261 return expr->id;
264 static inline unsigned int
265 lookup_expression_id (const pre_expr expr)
267 void **slot;
269 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
270 if (!slot)
271 return 0;
272 return ((pre_expr)*slot)->id;
275 /* Return the existing expression id for EXPR, or create one if one
276 does not exist yet. */
278 static inline unsigned int
279 get_or_alloc_expression_id (pre_expr expr)
281 unsigned int id = lookup_expression_id (expr);
282 if (id == 0)
283 return alloc_expression_id (expr);
284 return expr->id = id;
287 /* Return the expression that has expression id ID */
289 static inline pre_expr
290 expression_for_id (unsigned int id)
292 return VEC_index (pre_expr, expressions, id);
295 /* Free the expression id field in all of our expressions,
296 and then destroy the expressions array. */
298 static void
299 clear_expression_ids (void)
301 VEC_free (pre_expr, heap, expressions);
304 static alloc_pool pre_expr_pool;
306 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
308 static pre_expr
309 get_or_alloc_expr_for_name (tree name)
311 pre_expr result = (pre_expr) pool_alloc (pre_expr_pool);
312 unsigned int result_id;
314 result->kind = NAME;
315 result->id = 0;
316 PRE_EXPR_NAME (result) = name;
317 result_id = lookup_expression_id (result);
318 if (result_id != 0)
320 pool_free (pre_expr_pool, result);
321 result = expression_for_id (result_id);
322 return result;
324 get_or_alloc_expression_id (result);
325 return result;
328 static bool in_fre = false;
330 /* An unordered bitmap set. One bitmap tracks values, the other,
331 expressions. */
332 typedef struct bitmap_set
334 bitmap expressions;
335 bitmap values;
336 } *bitmap_set_t;
338 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
339 EXECUTE_IF_SET_IN_BITMAP((set)->expressions, 0, (id), (bi))
341 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
342 EXECUTE_IF_SET_IN_BITMAP((set)->values, 0, (id), (bi))
344 /* Mapping from value id to expressions with that value_id. */
345 DEF_VEC_P (bitmap_set_t);
346 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
347 static VEC(bitmap_set_t, heap) *value_expressions;
349 /* Sets that we need to keep track of. */
350 typedef struct bb_bitmap_sets
352 /* The EXP_GEN set, which represents expressions/values generated in
353 a basic block. */
354 bitmap_set_t exp_gen;
356 /* The PHI_GEN set, which represents PHI results generated in a
357 basic block. */
358 bitmap_set_t phi_gen;
360 /* The TMP_GEN set, which represents results/temporaries generated
361 in a basic block. IE the LHS of an expression. */
362 bitmap_set_t tmp_gen;
364 /* The AVAIL_OUT set, which represents which values are available in
365 a given basic block. */
366 bitmap_set_t avail_out;
368 /* The ANTIC_IN set, which represents which values are anticipatable
369 in a given basic block. */
370 bitmap_set_t antic_in;
372 /* The PA_IN set, which represents which values are
373 partially anticipatable in a given basic block. */
374 bitmap_set_t pa_in;
376 /* The NEW_SETS set, which is used during insertion to augment the
377 AVAIL_OUT set of blocks with the new insertions performed during
378 the current iteration. */
379 bitmap_set_t new_sets;
381 /* A cache for value_dies_in_block_x. */
382 bitmap expr_dies;
384 /* True if we have visited this block during ANTIC calculation. */
385 unsigned int visited:1;
387 /* True we have deferred processing this block during ANTIC
388 calculation until its successor is processed. */
389 unsigned int deferred : 1;
390 } *bb_value_sets_t;
392 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
393 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
394 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
395 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
396 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
397 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
398 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
399 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
400 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
401 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
404 /* Basic block list in postorder. */
405 static int *postorder;
407 /* This structure is used to keep track of statistics on what
408 optimization PRE was able to perform. */
409 static struct
411 /* The number of RHS computations eliminated by PRE. */
412 int eliminations;
414 /* The number of new expressions/temporaries generated by PRE. */
415 int insertions;
417 /* The number of inserts found due to partial anticipation */
418 int pa_insert;
420 /* The number of new PHI nodes added by PRE. */
421 int phis;
423 /* The number of values found constant. */
424 int constified;
426 } pre_stats;
428 static bool do_partial_partial;
429 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
430 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
431 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
432 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
433 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
434 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
435 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr, bool);
436 static bitmap_set_t bitmap_set_new (void);
437 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
438 gimple, tree);
439 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
440 gimple);
441 static unsigned int get_expr_value_id (pre_expr);
443 /* We can add and remove elements and entries to and from sets
444 and hash tables, so we use alloc pools for them. */
446 static alloc_pool bitmap_set_pool;
447 static bitmap_obstack grand_bitmap_obstack;
449 /* To avoid adding 300 temporary variables when we only need one, we
450 only create one temporary variable, on demand, and build ssa names
451 off that. We do have to change the variable if the types don't
452 match the current variable's type. */
453 static tree pretemp;
454 static tree storetemp;
455 static tree prephitemp;
457 /* Set of blocks with statements that have had its EH information
458 cleaned up. */
459 static bitmap need_eh_cleanup;
461 /* Which expressions have been seen during a given phi translation. */
462 static bitmap seen_during_translate;
464 /* The phi_translate_table caches phi translations for a given
465 expression and predecessor. */
467 static htab_t phi_translate_table;
469 /* A three tuple {e, pred, v} used to cache phi translations in the
470 phi_translate_table. */
472 typedef struct expr_pred_trans_d
474 /* The expression. */
475 pre_expr e;
477 /* The predecessor block along which we translated the expression. */
478 basic_block pred;
480 /* The value that resulted from the translation. */
481 pre_expr v;
483 /* The hashcode for the expression, pred pair. This is cached for
484 speed reasons. */
485 hashval_t hashcode;
486 } *expr_pred_trans_t;
487 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
489 /* Return the hash value for a phi translation table entry. */
491 static hashval_t
492 expr_pred_trans_hash (const void *p)
494 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
495 return ve->hashcode;
498 /* Return true if two phi translation table entries are the same.
499 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
501 static int
502 expr_pred_trans_eq (const void *p1, const void *p2)
504 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
505 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
506 basic_block b1 = ve1->pred;
507 basic_block b2 = ve2->pred;
509 /* If they are not translations for the same basic block, they can't
510 be equal. */
511 if (b1 != b2)
512 return false;
513 return pre_expr_eq (ve1->e, ve2->e);
516 /* Search in the phi translation table for the translation of
517 expression E in basic block PRED.
518 Return the translated value, if found, NULL otherwise. */
520 static inline pre_expr
521 phi_trans_lookup (pre_expr e, basic_block pred)
523 void **slot;
524 struct expr_pred_trans_d ept;
526 ept.e = e;
527 ept.pred = pred;
528 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
529 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
530 NO_INSERT);
531 if (!slot)
532 return NULL;
533 else
534 return ((expr_pred_trans_t) *slot)->v;
538 /* Add the tuple mapping from {expression E, basic block PRED} to
539 value V, to the phi translation table. */
541 static inline void
542 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
544 void **slot;
545 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
546 new_pair->e = e;
547 new_pair->pred = pred;
548 new_pair->v = v;
549 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
550 pred->index);
552 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
553 new_pair->hashcode, INSERT);
554 if (*slot)
555 free (*slot);
556 *slot = (void *) new_pair;
560 /* Add expression E to the expression set of value id V. */
562 void
563 add_to_value (unsigned int v, pre_expr e)
565 bitmap_set_t set;
567 gcc_assert (get_expr_value_id (e) == v);
569 if (v >= VEC_length (bitmap_set_t, value_expressions))
571 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
572 v + 1);
575 set = VEC_index (bitmap_set_t, value_expressions, v);
576 if (!set)
578 set = bitmap_set_new ();
579 VEC_replace (bitmap_set_t, value_expressions, v, set);
582 bitmap_insert_into_set_1 (set, e, true);
585 /* Create a new bitmap set and return it. */
587 static bitmap_set_t
588 bitmap_set_new (void)
590 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
591 ret->expressions = BITMAP_ALLOC (&grand_bitmap_obstack);
592 ret->values = BITMAP_ALLOC (&grand_bitmap_obstack);
593 return ret;
596 /* Return the value id for a PRE expression EXPR. */
598 static unsigned int
599 get_expr_value_id (pre_expr expr)
601 switch (expr->kind)
603 case CONSTANT:
605 unsigned int id;
606 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
607 if (id == 0)
609 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
610 add_to_value (id, expr);
612 return id;
614 case NAME:
615 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
616 case NARY:
617 return PRE_EXPR_NARY (expr)->value_id;
618 case REFERENCE:
619 return PRE_EXPR_REFERENCE (expr)->value_id;
620 default:
621 gcc_unreachable ();
625 /* Remove an expression EXPR from a bitmapped set. */
627 static void
628 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
630 unsigned int val = get_expr_value_id (expr);
631 if (!value_id_constant_p (val))
633 bitmap_clear_bit (set->values, val);
634 bitmap_clear_bit (set->expressions, get_expression_id (expr));
638 static void
639 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
640 bool allow_constants)
642 unsigned int val = get_expr_value_id (expr);
643 if (allow_constants || !value_id_constant_p (val))
645 /* We specifically expect this and only this function to be able to
646 insert constants into a set. */
647 bitmap_set_bit (set->values, val);
648 bitmap_set_bit (set->expressions, get_or_alloc_expression_id (expr));
652 /* Insert an expression EXPR into a bitmapped set. */
654 static void
655 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
657 bitmap_insert_into_set_1 (set, expr, false);
660 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
662 static void
663 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
665 bitmap_copy (dest->expressions, orig->expressions);
666 bitmap_copy (dest->values, orig->values);
670 /* Free memory used up by SET. */
671 static void
672 bitmap_set_free (bitmap_set_t set)
674 BITMAP_FREE (set->expressions);
675 BITMAP_FREE (set->values);
679 /* Generate an topological-ordered array of bitmap set SET. */
681 static VEC(pre_expr, heap) *
682 sorted_array_from_bitmap_set (bitmap_set_t set)
684 unsigned int i, j;
685 bitmap_iterator bi, bj;
686 VEC(pre_expr, heap) *result = NULL;
688 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
690 /* The number of expressions having a given value is usually
691 relatively small. Thus, rather than making a vector of all
692 the expressions and sorting it by value-id, we walk the values
693 and check in the reverse mapping that tells us what expressions
694 have a given value, to filter those in our set. As a result,
695 the expressions are inserted in value-id order, which means
696 topological order.
698 If this is somehow a significant lose for some cases, we can
699 choose which set to walk based on the set size. */
700 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
701 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
703 if (bitmap_bit_p (set->expressions, j))
704 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
708 return result;
711 /* Perform bitmapped set operation DEST &= ORIG. */
713 static void
714 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
716 bitmap_iterator bi;
717 unsigned int i;
719 if (dest != orig)
721 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
723 bitmap_and_into (dest->values, orig->values);
724 bitmap_copy (temp, dest->expressions);
725 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
727 pre_expr expr = expression_for_id (i);
728 unsigned int value_id = get_expr_value_id (expr);
729 if (!bitmap_bit_p (dest->values, value_id))
730 bitmap_clear_bit (dest->expressions, i);
732 BITMAP_FREE (temp);
736 /* Subtract all values and expressions contained in ORIG from DEST. */
738 static bitmap_set_t
739 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
741 bitmap_set_t result = bitmap_set_new ();
742 bitmap_iterator bi;
743 unsigned int i;
745 bitmap_and_compl (result->expressions, dest->expressions,
746 orig->expressions);
748 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
750 pre_expr expr = expression_for_id (i);
751 unsigned int value_id = get_expr_value_id (expr);
752 bitmap_set_bit (result->values, value_id);
755 return result;
758 /* Subtract all the values in bitmap set B from bitmap set A. */
760 static void
761 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
763 unsigned int i;
764 bitmap_iterator bi;
765 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
767 bitmap_copy (temp, a->expressions);
768 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
770 pre_expr expr = expression_for_id (i);
771 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
772 bitmap_remove_from_set (a, expr);
774 BITMAP_FREE (temp);
778 /* Return true if bitmapped set SET contains the value VALUE_ID. */
780 static bool
781 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
783 if (value_id_constant_p (value_id))
784 return true;
786 if (!set || bitmap_empty_p (set->expressions))
787 return false;
789 return bitmap_bit_p (set->values, value_id);
792 static inline bool
793 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
795 return bitmap_bit_p (set->expressions, get_expression_id (expr));
798 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
800 static void
801 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
802 const pre_expr expr)
804 bitmap_set_t exprset;
805 unsigned int i;
806 bitmap_iterator bi;
808 if (value_id_constant_p (lookfor))
809 return;
811 if (!bitmap_set_contains_value (set, lookfor))
812 return;
814 /* The number of expressions having a given value is usually
815 significantly less than the total number of expressions in SET.
816 Thus, rather than check, for each expression in SET, whether it
817 has the value LOOKFOR, we walk the reverse mapping that tells us
818 what expressions have a given value, and see if any of those
819 expressions are in our set. For large testcases, this is about
820 5-10x faster than walking the bitmap. If this is somehow a
821 significant lose for some cases, we can choose which set to walk
822 based on the set size. */
823 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
824 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
826 if (bitmap_bit_p (set->expressions, i))
828 bitmap_clear_bit (set->expressions, i);
829 bitmap_set_bit (set->expressions, get_expression_id (expr));
830 return;
835 /* Return true if two bitmap sets are equal. */
837 static bool
838 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
840 return bitmap_equal_p (a->values, b->values);
843 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
844 and add it otherwise. */
846 static void
847 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
849 unsigned int val = get_expr_value_id (expr);
851 if (bitmap_set_contains_value (set, val))
852 bitmap_set_replace_value (set, val, expr);
853 else
854 bitmap_insert_into_set (set, expr);
857 /* Insert EXPR into SET if EXPR's value is not already present in
858 SET. */
860 static void
861 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
863 unsigned int val = get_expr_value_id (expr);
865 if (value_id_constant_p (val))
866 return;
868 if (!bitmap_set_contains_value (set, val))
869 bitmap_insert_into_set (set, expr);
872 /* Print out EXPR to outfile. */
874 static void
875 print_pre_expr (FILE *outfile, const pre_expr expr)
877 switch (expr->kind)
879 case CONSTANT:
880 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
881 break;
882 case NAME:
883 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
884 break;
885 case NARY:
887 unsigned int i;
888 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
889 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
890 for (i = 0; i < nary->length; i++)
892 print_generic_expr (outfile, nary->op[i], 0);
893 if (i != (unsigned) nary->length - 1)
894 fprintf (outfile, ",");
896 fprintf (outfile, "}");
898 break;
900 case REFERENCE:
902 vn_reference_op_t vro;
903 unsigned int i;
904 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
905 fprintf (outfile, "{");
906 for (i = 0;
907 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
908 i++)
910 bool closebrace = false;
911 if (vro->opcode != SSA_NAME
912 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
914 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
915 if (vro->op0)
917 fprintf (outfile, "<");
918 closebrace = true;
921 if (vro->op0)
923 print_generic_expr (outfile, vro->op0, 0);
924 if (vro->op1)
926 fprintf (outfile, ",");
927 print_generic_expr (outfile, vro->op1, 0);
929 if (vro->op2)
931 fprintf (outfile, ",");
932 print_generic_expr (outfile, vro->op2, 0);
935 if (closebrace)
936 fprintf (outfile, ">");
937 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
938 fprintf (outfile, ",");
940 fprintf (outfile, "}");
941 if (ref->vuse)
943 fprintf (outfile, "@");
944 print_generic_expr (outfile, ref->vuse, 0);
947 break;
950 void debug_pre_expr (pre_expr);
952 /* Like print_pre_expr but always prints to stderr. */
953 void
954 debug_pre_expr (pre_expr e)
956 print_pre_expr (stderr, e);
957 fprintf (stderr, "\n");
960 /* Print out SET to OUTFILE. */
962 static void
963 print_bitmap_set (FILE *outfile, bitmap_set_t set,
964 const char *setname, int blockindex)
966 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
967 if (set)
969 bool first = true;
970 unsigned i;
971 bitmap_iterator bi;
973 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
975 const pre_expr expr = expression_for_id (i);
977 if (!first)
978 fprintf (outfile, ", ");
979 first = false;
980 print_pre_expr (outfile, expr);
982 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
985 fprintf (outfile, " }\n");
988 void debug_bitmap_set (bitmap_set_t);
990 void
991 debug_bitmap_set (bitmap_set_t set)
993 print_bitmap_set (stderr, set, "debug", 0);
996 /* Print out the expressions that have VAL to OUTFILE. */
998 void
999 print_value_expressions (FILE *outfile, unsigned int val)
1001 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1002 if (set)
1004 char s[10];
1005 sprintf (s, "%04d", val);
1006 print_bitmap_set (outfile, set, s, 0);
1011 void
1012 debug_value_expressions (unsigned int val)
1014 print_value_expressions (stderr, val);
1017 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1018 represent it. */
1020 static pre_expr
1021 get_or_alloc_expr_for_constant (tree constant)
1023 unsigned int result_id;
1024 unsigned int value_id;
1025 pre_expr newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1026 newexpr->kind = CONSTANT;
1027 PRE_EXPR_CONSTANT (newexpr) = constant;
1028 result_id = lookup_expression_id (newexpr);
1029 if (result_id != 0)
1031 pool_free (pre_expr_pool, newexpr);
1032 newexpr = expression_for_id (result_id);
1033 return newexpr;
1035 value_id = get_or_alloc_constant_value_id (constant);
1036 get_or_alloc_expression_id (newexpr);
1037 add_to_value (value_id, newexpr);
1038 return newexpr;
1041 /* Given a value id V, find the actual tree representing the constant
1042 value if there is one, and return it. Return NULL if we can't find
1043 a constant. */
1045 static tree
1046 get_constant_for_value_id (unsigned int v)
1048 if (value_id_constant_p (v))
1050 unsigned int i;
1051 bitmap_iterator bi;
1052 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1054 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1056 pre_expr expr = expression_for_id (i);
1057 if (expr->kind == CONSTANT)
1058 return PRE_EXPR_CONSTANT (expr);
1061 return NULL;
1064 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1065 Currently only supports constants and SSA_NAMES. */
1066 static pre_expr
1067 get_or_alloc_expr_for (tree t)
1069 if (TREE_CODE (t) == SSA_NAME)
1070 return get_or_alloc_expr_for_name (t);
1071 else if (is_gimple_min_invariant (t))
1072 return get_or_alloc_expr_for_constant (t);
1073 else
1075 /* More complex expressions can result from SCCVN expression
1076 simplification that inserts values for them. As they all
1077 do not have VOPs the get handled by the nary ops struct. */
1078 vn_nary_op_t result;
1079 unsigned int result_id;
1080 vn_nary_op_lookup (t, &result);
1081 if (result != NULL)
1083 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1084 e->kind = NARY;
1085 PRE_EXPR_NARY (e) = result;
1086 result_id = lookup_expression_id (e);
1087 if (result_id != 0)
1089 pool_free (pre_expr_pool, e);
1090 e = expression_for_id (result_id);
1091 return e;
1093 alloc_expression_id (e);
1094 return e;
1097 return NULL;
1100 /* Return the folded version of T if T, when folded, is a gimple
1101 min_invariant. Otherwise, return T. */
1103 static pre_expr
1104 fully_constant_expression (pre_expr e)
1106 switch (e->kind)
1108 case CONSTANT:
1109 return e;
1110 case NARY:
1112 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1113 switch (TREE_CODE_CLASS (nary->opcode))
1115 case tcc_expression:
1116 if (nary->opcode == TRUTH_NOT_EXPR)
1117 goto do_unary;
1118 if (nary->opcode != TRUTH_AND_EXPR
1119 && nary->opcode != TRUTH_OR_EXPR
1120 && nary->opcode != TRUTH_XOR_EXPR)
1121 return e;
1122 /* Fallthrough. */
1123 case tcc_binary:
1124 case tcc_comparison:
1126 /* We have to go from trees to pre exprs to value ids to
1127 constants. */
1128 tree naryop0 = nary->op[0];
1129 tree naryop1 = nary->op[1];
1130 tree result;
1131 if (!is_gimple_min_invariant (naryop0))
1133 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1134 unsigned int vrep0 = get_expr_value_id (rep0);
1135 tree const0 = get_constant_for_value_id (vrep0);
1136 if (const0)
1137 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1139 if (!is_gimple_min_invariant (naryop1))
1141 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1142 unsigned int vrep1 = get_expr_value_id (rep1);
1143 tree const1 = get_constant_for_value_id (vrep1);
1144 if (const1)
1145 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1147 result = fold_binary (nary->opcode, nary->type,
1148 naryop0, naryop1);
1149 if (result && is_gimple_min_invariant (result))
1150 return get_or_alloc_expr_for_constant (result);
1151 /* We might have simplified the expression to a
1152 SSA_NAME for example from x_1 * 1. But we cannot
1153 insert a PHI for x_1 unconditionally as x_1 might
1154 not be available readily. */
1155 return e;
1157 case tcc_reference:
1158 if (nary->opcode != REALPART_EXPR
1159 && nary->opcode != IMAGPART_EXPR
1160 && nary->opcode != VIEW_CONVERT_EXPR)
1161 return e;
1162 /* Fallthrough. */
1163 case tcc_unary:
1164 do_unary:
1166 /* We have to go from trees to pre exprs to value ids to
1167 constants. */
1168 tree naryop0 = nary->op[0];
1169 tree const0, result;
1170 if (is_gimple_min_invariant (naryop0))
1171 const0 = naryop0;
1172 else
1174 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1175 unsigned int vrep0 = get_expr_value_id (rep0);
1176 const0 = get_constant_for_value_id (vrep0);
1178 result = NULL;
1179 if (const0)
1181 tree type1 = TREE_TYPE (nary->op[0]);
1182 const0 = fold_convert (type1, const0);
1183 result = fold_unary (nary->opcode, nary->type, const0);
1185 if (result && is_gimple_min_invariant (result))
1186 return get_or_alloc_expr_for_constant (result);
1187 return e;
1189 default:
1190 return e;
1193 case REFERENCE:
1195 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1196 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1197 vn_reference_op_t op;
1199 /* Try to simplify the translated expression if it is
1200 a call to a builtin function with at most two arguments. */
1201 op = VEC_index (vn_reference_op_s, operands, 0);
1202 if (op->opcode == CALL_EXPR
1203 && TREE_CODE (op->op0) == ADDR_EXPR
1204 && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL
1205 && DECL_BUILT_IN (TREE_OPERAND (op->op0, 0))
1206 && VEC_length (vn_reference_op_s, operands) >= 2
1207 && VEC_length (vn_reference_op_s, operands) <= 3)
1209 vn_reference_op_t arg0, arg1 = NULL;
1210 bool anyconst = false;
1211 arg0 = VEC_index (vn_reference_op_s, operands, 1);
1212 if (VEC_length (vn_reference_op_s, operands) > 2)
1213 arg1 = VEC_index (vn_reference_op_s, operands, 2);
1214 if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant
1215 || (arg0->opcode == ADDR_EXPR
1216 && is_gimple_min_invariant (arg0->op0)))
1217 anyconst = true;
1218 if (arg1
1219 && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant
1220 || (arg1->opcode == ADDR_EXPR
1221 && is_gimple_min_invariant (arg1->op0))))
1222 anyconst = true;
1223 if (anyconst)
1225 tree folded = build_call_expr (TREE_OPERAND (op->op0, 0),
1226 arg1 ? 2 : 1,
1227 arg0->op0,
1228 arg1 ? arg1->op0 : NULL);
1229 if (folded
1230 && TREE_CODE (folded) == NOP_EXPR)
1231 folded = TREE_OPERAND (folded, 0);
1232 if (folded
1233 && is_gimple_min_invariant (folded))
1234 return get_or_alloc_expr_for_constant (folded);
1237 return e;
1239 default:
1240 return e;
1242 return e;
1245 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1246 it has the value it would have in BLOCK. Set *SAME_VALID to true
1247 in case the new vuse doesn't change the value id of the OPERANDS. */
1249 static tree
1250 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1251 alias_set_type set, tree type, tree vuse,
1252 basic_block phiblock,
1253 basic_block block, bool *same_valid)
1255 gimple phi = SSA_NAME_DEF_STMT (vuse);
1256 ao_ref ref;
1257 edge e = NULL;
1258 bool use_oracle;
1260 *same_valid = true;
1262 if (gimple_bb (phi) != phiblock)
1263 return vuse;
1265 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1267 /* Use the alias-oracle to find either the PHI node in this block,
1268 the first VUSE used in this block that is equivalent to vuse or
1269 the first VUSE which definition in this block kills the value. */
1270 if (gimple_code (phi) == GIMPLE_PHI)
1271 e = find_edge (block, phiblock);
1272 else if (use_oracle)
1273 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1275 vuse = gimple_vuse (phi);
1276 phi = SSA_NAME_DEF_STMT (vuse);
1277 if (gimple_bb (phi) != phiblock)
1278 return vuse;
1279 if (gimple_code (phi) == GIMPLE_PHI)
1281 e = find_edge (block, phiblock);
1282 break;
1285 else
1286 return NULL_TREE;
1288 if (e)
1290 if (use_oracle)
1292 bitmap visited = NULL;
1293 /* Try to find a vuse that dominates this phi node by skipping
1294 non-clobbering statements. */
1295 vuse = get_continuation_for_phi (phi, &ref, &visited);
1296 if (visited)
1297 BITMAP_FREE (visited);
1299 else
1300 vuse = NULL_TREE;
1301 if (!vuse)
1303 /* If we didn't find any, the value ID can't stay the same,
1304 but return the translated vuse. */
1305 *same_valid = false;
1306 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1308 /* ??? We would like to return vuse here as this is the canonical
1309 upmost vdef that this reference is associated with. But during
1310 insertion of the references into the hash tables we only ever
1311 directly insert with their direct gimple_vuse, hence returning
1312 something else would make us not find the other expression. */
1313 return PHI_ARG_DEF (phi, e->dest_idx);
1316 return NULL_TREE;
1319 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1320 SET2. This is used to avoid making a set consisting of the union
1321 of PA_IN and ANTIC_IN during insert. */
1323 static inline pre_expr
1324 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1326 pre_expr result;
1328 result = bitmap_find_leader (set1, val, NULL);
1329 if (!result && set2)
1330 result = bitmap_find_leader (set2, val, NULL);
1331 return result;
1334 /* Get the tree type for our PRE expression e. */
1336 static tree
1337 get_expr_type (const pre_expr e)
1339 switch (e->kind)
1341 case NAME:
1342 return TREE_TYPE (PRE_EXPR_NAME (e));
1343 case CONSTANT:
1344 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1345 case REFERENCE:
1346 return PRE_EXPR_REFERENCE (e)->type;
1347 case NARY:
1348 return PRE_EXPR_NARY (e)->type;
1350 gcc_unreachable();
1353 /* Get a representative SSA_NAME for a given expression.
1354 Since all of our sub-expressions are treated as values, we require
1355 them to be SSA_NAME's for simplicity.
1356 Prior versions of GVNPRE used to use "value handles" here, so that
1357 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1358 either case, the operands are really values (IE we do not expect
1359 them to be usable without finding leaders). */
1361 static tree
1362 get_representative_for (const pre_expr e)
1364 tree exprtype;
1365 tree name;
1366 unsigned int value_id = get_expr_value_id (e);
1368 switch (e->kind)
1370 case NAME:
1371 return PRE_EXPR_NAME (e);
1372 case CONSTANT:
1373 return PRE_EXPR_CONSTANT (e);
1374 case NARY:
1375 case REFERENCE:
1377 /* Go through all of the expressions representing this value
1378 and pick out an SSA_NAME. */
1379 unsigned int i;
1380 bitmap_iterator bi;
1381 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1382 value_id);
1383 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1385 pre_expr rep = expression_for_id (i);
1386 if (rep->kind == NAME)
1387 return PRE_EXPR_NAME (rep);
1390 break;
1392 /* If we reached here we couldn't find an SSA_NAME. This can
1393 happen when we've discovered a value that has never appeared in
1394 the program as set to an SSA_NAME, most likely as the result of
1395 phi translation. */
1396 if (dump_file)
1398 fprintf (dump_file,
1399 "Could not find SSA_NAME representative for expression:");
1400 print_pre_expr (dump_file, e);
1401 fprintf (dump_file, "\n");
1404 exprtype = get_expr_type (e);
1406 /* Build and insert the assignment of the end result to the temporary
1407 that we will return. */
1408 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1410 pretemp = create_tmp_var (exprtype, "pretmp");
1411 get_var_ann (pretemp);
1414 name = make_ssa_name (pretemp, gimple_build_nop ());
1415 VN_INFO_GET (name)->value_id = value_id;
1416 if (e->kind == CONSTANT)
1417 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1418 else
1419 VN_INFO (name)->valnum = name;
1421 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1422 if (dump_file)
1424 fprintf (dump_file, "Created SSA_NAME representative ");
1425 print_generic_expr (dump_file, name, 0);
1426 fprintf (dump_file, " for expression:");
1427 print_pre_expr (dump_file, e);
1428 fprintf (dump_file, "\n");
1431 return name;
1437 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1438 the phis in PRED. SEEN is a bitmap saying which expression we have
1439 translated since we started translation of the toplevel expression.
1440 Return NULL if we can't find a leader for each part of the
1441 translated expression. */
1443 static pre_expr
1444 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1445 basic_block pred, basic_block phiblock, bitmap seen)
1447 pre_expr oldexpr = expr;
1448 pre_expr phitrans;
1450 if (!expr)
1451 return NULL;
1453 if (value_id_constant_p (get_expr_value_id (expr)))
1454 return expr;
1456 phitrans = phi_trans_lookup (expr, pred);
1457 if (phitrans)
1458 return phitrans;
1460 /* Prevent cycles when we have recursively dependent leaders. This
1461 can only happen when phi translating the maximal set. */
1462 if (seen)
1464 unsigned int expr_id = get_expression_id (expr);
1465 if (bitmap_bit_p (seen, expr_id))
1466 return NULL;
1467 bitmap_set_bit (seen, expr_id);
1470 switch (expr->kind)
1472 /* Constants contain no values that need translation. */
1473 case CONSTANT:
1474 return expr;
1476 case NARY:
1478 unsigned int i;
1479 bool changed = false;
1480 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1481 struct vn_nary_op_s newnary;
1482 /* The NARY structure is only guaranteed to have been
1483 allocated to the nary->length operands. */
1484 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1485 - sizeof (tree) * (4 - nary->length)));
1487 for (i = 0; i < newnary.length; i++)
1489 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1490 continue;
1491 else
1493 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1494 pre_expr leader = find_leader_in_sets (op_val_id, set1, set2);
1495 pre_expr result = phi_translate_1 (leader, set1, set2,
1496 pred, phiblock, seen);
1497 if (result && result != leader)
1499 tree name = get_representative_for (result);
1500 if (!name)
1501 return NULL;
1502 newnary.op[i] = name;
1504 else if (!result)
1505 return NULL;
1507 changed |= newnary.op[i] != nary->op[i];
1510 if (changed)
1512 pre_expr constant;
1514 tree result = vn_nary_op_lookup_pieces (newnary.length,
1515 newnary.opcode,
1516 newnary.type,
1517 newnary.op[0],
1518 newnary.op[1],
1519 newnary.op[2],
1520 newnary.op[3],
1521 &nary);
1522 unsigned int new_val_id;
1524 expr = (pre_expr) pool_alloc (pre_expr_pool);
1525 expr->kind = NARY;
1526 expr->id = 0;
1527 if (result && is_gimple_min_invariant (result))
1528 return get_or_alloc_expr_for_constant (result);
1531 if (nary)
1533 PRE_EXPR_NARY (expr) = nary;
1534 constant = fully_constant_expression (expr);
1535 if (constant != expr)
1536 return constant;
1538 new_val_id = nary->value_id;
1539 get_or_alloc_expression_id (expr);
1541 else
1543 new_val_id = get_next_value_id ();
1544 VEC_safe_grow_cleared (bitmap_set_t, heap,
1545 value_expressions,
1546 get_max_value_id() + 1);
1547 nary = vn_nary_op_insert_pieces (newnary.length,
1548 newnary.opcode,
1549 newnary.type,
1550 newnary.op[0],
1551 newnary.op[1],
1552 newnary.op[2],
1553 newnary.op[3],
1554 result, new_val_id);
1555 PRE_EXPR_NARY (expr) = nary;
1556 constant = fully_constant_expression (expr);
1557 if (constant != expr)
1558 return constant;
1559 get_or_alloc_expression_id (expr);
1561 add_to_value (new_val_id, expr);
1563 phi_trans_add (oldexpr, expr, pred);
1564 return expr;
1566 break;
1568 case REFERENCE:
1570 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1571 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1572 tree vuse = ref->vuse;
1573 tree newvuse = vuse;
1574 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1575 bool changed = false, same_valid = true;
1576 unsigned int i, j;
1577 vn_reference_op_t operand;
1578 vn_reference_t newref;
1580 for (i = 0, j = 0;
1581 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1583 pre_expr opresult;
1584 pre_expr leader;
1585 tree oldop0 = operand->op0;
1586 tree oldop1 = operand->op1;
1587 tree oldop2 = operand->op2;
1588 tree op0 = oldop0;
1589 tree op1 = oldop1;
1590 tree op2 = oldop2;
1591 tree type = operand->type;
1592 vn_reference_op_s newop = *operand;
1594 if (op0 && TREE_CODE (op0) == SSA_NAME)
1596 unsigned int op_val_id = VN_INFO (op0)->value_id;
1597 leader = find_leader_in_sets (op_val_id, set1, set2);
1598 opresult = phi_translate_1 (leader, set1, set2,
1599 pred, phiblock, seen);
1600 if (opresult && opresult != leader)
1602 tree name = get_representative_for (opresult);
1603 if (!name)
1604 break;
1605 op0 = name;
1607 else if (!opresult)
1608 break;
1610 changed |= op0 != oldop0;
1612 if (op1 && TREE_CODE (op1) == SSA_NAME)
1614 unsigned int op_val_id = VN_INFO (op1)->value_id;
1615 leader = find_leader_in_sets (op_val_id, set1, set2);
1616 opresult = phi_translate_1 (leader, set1, set2,
1617 pred, phiblock, seen);
1618 if (opresult && opresult != leader)
1620 tree name = get_representative_for (opresult);
1621 if (!name)
1622 break;
1623 op1 = name;
1625 else if (!opresult)
1626 break;
1628 /* We can't possibly insert these. */
1629 else if (op1 && !is_gimple_min_invariant (op1))
1630 break;
1631 changed |= op1 != oldop1;
1632 if (op2 && TREE_CODE (op2) == SSA_NAME)
1634 unsigned int op_val_id = VN_INFO (op2)->value_id;
1635 leader = find_leader_in_sets (op_val_id, set1, set2);
1636 opresult = phi_translate_1 (leader, set1, set2,
1637 pred, phiblock, seen);
1638 if (opresult && opresult != leader)
1640 tree name = get_representative_for (opresult);
1641 if (!name)
1642 break;
1643 op2 = name;
1645 else if (!opresult)
1646 break;
1648 /* We can't possibly insert these. */
1649 else if (op2 && !is_gimple_min_invariant (op2))
1650 break;
1651 changed |= op2 != oldop2;
1653 if (!newoperands)
1654 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1655 /* We may have changed from an SSA_NAME to a constant */
1656 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1657 newop.opcode = TREE_CODE (op0);
1658 newop.type = type;
1659 newop.op0 = op0;
1660 newop.op1 = op1;
1661 newop.op2 = op2;
1662 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1663 /* If it transforms from an SSA_NAME to an address, fold with
1664 a preceding indirect reference. */
1665 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1666 && VEC_index (vn_reference_op_s,
1667 newoperands, j - 1)->opcode == INDIRECT_REF)
1668 vn_reference_fold_indirect (&newoperands, &j);
1670 if (i != VEC_length (vn_reference_op_s, operands))
1672 if (newoperands)
1673 VEC_free (vn_reference_op_s, heap, newoperands);
1674 return NULL;
1677 if (vuse)
1679 newvuse = translate_vuse_through_block (newoperands,
1680 ref->set, ref->type,
1681 vuse, phiblock, pred,
1682 &same_valid);
1683 if (newvuse == NULL_TREE)
1685 VEC_free (vn_reference_op_s, heap, newoperands);
1686 return NULL;
1690 if (changed || newvuse != vuse)
1692 unsigned int new_val_id;
1693 pre_expr constant;
1695 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1696 ref->type,
1697 newoperands,
1698 &newref, true);
1699 if (newref)
1700 VEC_free (vn_reference_op_s, heap, newoperands);
1702 if (result && is_gimple_min_invariant (result))
1704 gcc_assert (!newoperands);
1705 return get_or_alloc_expr_for_constant (result);
1708 expr = (pre_expr) pool_alloc (pre_expr_pool);
1709 expr->kind = REFERENCE;
1710 expr->id = 0;
1712 if (newref)
1714 PRE_EXPR_REFERENCE (expr) = newref;
1715 constant = fully_constant_expression (expr);
1716 if (constant != expr)
1717 return constant;
1719 new_val_id = newref->value_id;
1720 get_or_alloc_expression_id (expr);
1722 else
1724 if (changed || !same_valid)
1726 new_val_id = get_next_value_id ();
1727 VEC_safe_grow_cleared (bitmap_set_t, heap,
1728 value_expressions,
1729 get_max_value_id() + 1);
1731 else
1732 new_val_id = ref->value_id;
1733 newref = vn_reference_insert_pieces (newvuse, ref->set,
1734 ref->type,
1735 newoperands,
1736 result, new_val_id);
1737 newoperands = NULL;
1738 PRE_EXPR_REFERENCE (expr) = newref;
1739 constant = fully_constant_expression (expr);
1740 if (constant != expr)
1741 return constant;
1742 get_or_alloc_expression_id (expr);
1744 add_to_value (new_val_id, expr);
1746 VEC_free (vn_reference_op_s, heap, newoperands);
1747 phi_trans_add (oldexpr, expr, pred);
1748 return expr;
1750 break;
1752 case NAME:
1754 gimple phi = NULL;
1755 edge e;
1756 gimple def_stmt;
1757 tree name = PRE_EXPR_NAME (expr);
1759 def_stmt = SSA_NAME_DEF_STMT (name);
1760 if (gimple_code (def_stmt) == GIMPLE_PHI
1761 && gimple_bb (def_stmt) == phiblock)
1762 phi = def_stmt;
1763 else
1764 return expr;
1766 e = find_edge (pred, gimple_bb (phi));
1767 if (e)
1769 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1770 pre_expr newexpr;
1772 if (TREE_CODE (def) == SSA_NAME)
1773 def = VN_INFO (def)->valnum;
1775 /* Handle constant. */
1776 if (is_gimple_min_invariant (def))
1777 return get_or_alloc_expr_for_constant (def);
1779 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1780 return NULL;
1782 newexpr = get_or_alloc_expr_for_name (def);
1783 return newexpr;
1786 return expr;
1788 default:
1789 gcc_unreachable ();
1793 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1794 the phis in PRED.
1795 Return NULL if we can't find a leader for each part of the
1796 translated expression. */
1798 static pre_expr
1799 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1800 basic_block pred, basic_block phiblock)
1802 bitmap_clear (seen_during_translate);
1803 return phi_translate_1 (expr, set1, set2, pred, phiblock,
1804 seen_during_translate);
1807 /* For each expression in SET, translate the values through phi nodes
1808 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1809 expressions in DEST. */
1811 static void
1812 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1813 basic_block phiblock)
1815 VEC (pre_expr, heap) *exprs;
1816 pre_expr expr;
1817 int i;
1819 if (!phi_nodes (phiblock))
1821 bitmap_set_copy (dest, set);
1822 return;
1825 exprs = sorted_array_from_bitmap_set (set);
1826 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1828 pre_expr translated;
1829 translated = phi_translate (expr, set, NULL, pred, phiblock);
1831 /* Don't add empty translations to the cache */
1832 if (translated)
1833 phi_trans_add (expr, translated, pred);
1835 if (translated != NULL)
1836 bitmap_value_insert_into_set (dest, translated);
1838 VEC_free (pre_expr, heap, exprs);
1841 /* Find the leader for a value (i.e., the name representing that
1842 value) in a given set, and return it. If STMT is non-NULL it
1843 makes sure the defining statement for the leader dominates it.
1844 Return NULL if no leader is found. */
1846 static pre_expr
1847 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1849 if (value_id_constant_p (val))
1851 unsigned int i;
1852 bitmap_iterator bi;
1853 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1855 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1857 pre_expr expr = expression_for_id (i);
1858 if (expr->kind == CONSTANT)
1859 return expr;
1862 if (bitmap_set_contains_value (set, val))
1864 /* Rather than walk the entire bitmap of expressions, and see
1865 whether any of them has the value we are looking for, we look
1866 at the reverse mapping, which tells us the set of expressions
1867 that have a given value (IE value->expressions with that
1868 value) and see if any of those expressions are in our set.
1869 The number of expressions per value is usually significantly
1870 less than the number of expressions in the set. In fact, for
1871 large testcases, doing it this way is roughly 5-10x faster
1872 than walking the bitmap.
1873 If this is somehow a significant lose for some cases, we can
1874 choose which set to walk based on which set is smaller. */
1875 unsigned int i;
1876 bitmap_iterator bi;
1877 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1879 EXECUTE_IF_AND_IN_BITMAP (exprset->expressions,
1880 set->expressions, 0, i, bi)
1882 pre_expr val = expression_for_id (i);
1883 /* At the point where stmt is not null, there should always
1884 be an SSA_NAME first in the list of expressions. */
1885 if (stmt)
1887 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1888 if (gimple_code (def_stmt) != GIMPLE_PHI
1889 && gimple_bb (def_stmt) == gimple_bb (stmt)
1890 && gimple_uid (def_stmt) >= gimple_uid (stmt))
1891 continue;
1893 return val;
1896 return NULL;
1899 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1900 BLOCK by seeing if it is not killed in the block. Note that we are
1901 only determining whether there is a store that kills it. Because
1902 of the order in which clean iterates over values, we are guaranteed
1903 that altered operands will have caused us to be eliminated from the
1904 ANTIC_IN set already. */
1906 static bool
1907 value_dies_in_block_x (pre_expr expr, basic_block block)
1909 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1910 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1911 gimple def;
1912 gimple_stmt_iterator gsi;
1913 unsigned id = get_expression_id (expr);
1914 bool res = false;
1915 ao_ref ref;
1917 if (!vuse)
1918 return false;
1920 /* Lookup a previously calculated result. */
1921 if (EXPR_DIES (block)
1922 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1923 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1925 /* A memory expression {e, VUSE} dies in the block if there is a
1926 statement that may clobber e. If, starting statement walk from the
1927 top of the basic block, a statement uses VUSE there can be no kill
1928 inbetween that use and the original statement that loaded {e, VUSE},
1929 so we can stop walking. */
1930 ref.base = NULL_TREE;
1931 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1933 tree def_vuse, def_vdef;
1934 def = gsi_stmt (gsi);
1935 def_vuse = gimple_vuse (def);
1936 def_vdef = gimple_vdef (def);
1938 /* Not a memory statement. */
1939 if (!def_vuse)
1940 continue;
1942 /* Not a may-def. */
1943 if (!def_vdef)
1945 /* A load with the same VUSE, we're done. */
1946 if (def_vuse == vuse)
1947 break;
1949 continue;
1952 /* Init ref only if we really need it. */
1953 if (ref.base == NULL_TREE
1954 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1955 refx->operands))
1957 res = true;
1958 break;
1960 /* If the statement may clobber expr, it dies. */
1961 if (stmt_may_clobber_ref_p_1 (def, &ref))
1963 res = true;
1964 break;
1968 /* Remember the result. */
1969 if (!EXPR_DIES (block))
1970 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1971 bitmap_set_bit (EXPR_DIES (block), id * 2);
1972 if (res)
1973 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1975 return res;
1979 #define union_contains_value(SET1, SET2, VAL) \
1980 (bitmap_set_contains_value ((SET1), (VAL)) \
1981 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
1983 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
1985 static bool
1986 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
1987 vn_reference_op_t vro)
1989 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
1991 struct pre_expr_d temp;
1992 temp.kind = NAME;
1993 temp.id = 0;
1994 PRE_EXPR_NAME (&temp) = vro->op0;
1995 temp.id = lookup_expression_id (&temp);
1996 if (temp.id == 0)
1997 return false;
1998 if (!union_contains_value (set1, set2,
1999 get_expr_value_id (&temp)))
2000 return false;
2002 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2004 struct pre_expr_d temp;
2005 temp.kind = NAME;
2006 temp.id = 0;
2007 PRE_EXPR_NAME (&temp) = vro->op1;
2008 temp.id = lookup_expression_id (&temp);
2009 if (temp.id == 0)
2010 return false;
2011 if (!union_contains_value (set1, set2,
2012 get_expr_value_id (&temp)))
2013 return false;
2016 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2018 struct pre_expr_d temp;
2019 temp.kind = NAME;
2020 temp.id = 0;
2021 PRE_EXPR_NAME (&temp) = vro->op2;
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;
2033 /* Determine if the expression EXPR is valid in SET1 U SET2.
2034 ONLY SET2 CAN BE NULL.
2035 This means that we have a leader for each part of the expression
2036 (if it consists of values), or the expression is an SSA_NAME.
2037 For loads/calls, we also see if the vuse is killed in this block. */
2039 static bool
2040 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2041 basic_block block)
2043 switch (expr->kind)
2045 case NAME:
2046 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2047 case NARY:
2049 unsigned int i;
2050 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2051 for (i = 0; i < nary->length; i++)
2053 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2055 struct pre_expr_d temp;
2056 temp.kind = NAME;
2057 temp.id = 0;
2058 PRE_EXPR_NAME (&temp) = nary->op[i];
2059 temp.id = lookup_expression_id (&temp);
2060 if (temp.id == 0)
2061 return false;
2062 if (!union_contains_value (set1, set2,
2063 get_expr_value_id (&temp)))
2064 return false;
2067 return true;
2069 break;
2070 case REFERENCE:
2072 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2073 vn_reference_op_t vro;
2074 unsigned int i;
2076 for (i = 0; VEC_iterate (vn_reference_op_s, ref->operands, i, vro); i++)
2078 if (!vro_valid_in_sets (set1, set2, vro))
2079 return false;
2081 if (ref->vuse)
2083 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2084 if (!gimple_nop_p (def_stmt)
2085 && gimple_bb (def_stmt) != block
2086 && !dominated_by_p (CDI_DOMINATORS,
2087 block, gimple_bb (def_stmt)))
2088 return false;
2090 return !value_dies_in_block_x (expr, block);
2092 default:
2093 gcc_unreachable ();
2097 /* Clean the set of expressions that are no longer valid in SET1 or
2098 SET2. This means expressions that are made up of values we have no
2099 leaders for in SET1 or SET2. This version is used for partial
2100 anticipation, which means it is not valid in either ANTIC_IN or
2101 PA_IN. */
2103 static void
2104 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2106 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2107 pre_expr expr;
2108 int i;
2110 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2112 if (!valid_in_sets (set1, set2, expr, block))
2113 bitmap_remove_from_set (set1, expr);
2115 VEC_free (pre_expr, heap, exprs);
2118 /* Clean the set of expressions that are no longer valid in SET. This
2119 means expressions that are made up of values we have no leaders for
2120 in SET. */
2122 static void
2123 clean (bitmap_set_t set, basic_block block)
2125 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2126 pre_expr expr;
2127 int i;
2129 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2131 if (!valid_in_sets (set, NULL, expr, block))
2132 bitmap_remove_from_set (set, expr);
2134 VEC_free (pre_expr, heap, exprs);
2137 static sbitmap has_abnormal_preds;
2139 /* List of blocks that may have changed during ANTIC computation and
2140 thus need to be iterated over. */
2142 static sbitmap changed_blocks;
2144 /* Decide whether to defer a block for a later iteration, or PHI
2145 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2146 should defer the block, and true if we processed it. */
2148 static bool
2149 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2150 basic_block block, basic_block phiblock)
2152 if (!BB_VISITED (phiblock))
2154 SET_BIT (changed_blocks, block->index);
2155 BB_VISITED (block) = 0;
2156 BB_DEFERRED (block) = 1;
2157 return false;
2159 else
2160 phi_translate_set (dest, source, block, phiblock);
2161 return true;
2164 /* Compute the ANTIC set for BLOCK.
2166 If succs(BLOCK) > 1 then
2167 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2168 else if succs(BLOCK) == 1 then
2169 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2171 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2174 static bool
2175 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2177 bool changed = false;
2178 bitmap_set_t S, old, ANTIC_OUT;
2179 bitmap_iterator bi;
2180 unsigned int bii;
2181 edge e;
2182 edge_iterator ei;
2184 old = ANTIC_OUT = S = NULL;
2185 BB_VISITED (block) = 1;
2187 /* If any edges from predecessors are abnormal, antic_in is empty,
2188 so do nothing. */
2189 if (block_has_abnormal_pred_edge)
2190 goto maybe_dump_sets;
2192 old = ANTIC_IN (block);
2193 ANTIC_OUT = bitmap_set_new ();
2195 /* If the block has no successors, ANTIC_OUT is empty. */
2196 if (EDGE_COUNT (block->succs) == 0)
2198 /* If we have one successor, we could have some phi nodes to
2199 translate through. */
2200 else if (single_succ_p (block))
2202 basic_block succ_bb = single_succ (block);
2204 /* We trade iterations of the dataflow equations for having to
2205 phi translate the maximal set, which is incredibly slow
2206 (since the maximal set often has 300+ members, even when you
2207 have a small number of blocks).
2208 Basically, we defer the computation of ANTIC for this block
2209 until we have processed it's successor, which will inevitably
2210 have a *much* smaller set of values to phi translate once
2211 clean has been run on it.
2212 The cost of doing this is that we technically perform more
2213 iterations, however, they are lower cost iterations.
2215 Timings for PRE on tramp3d-v4:
2216 without maximal set fix: 11 seconds
2217 with maximal set fix/without deferring: 26 seconds
2218 with maximal set fix/with deferring: 11 seconds
2221 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2222 block, succ_bb))
2224 changed = true;
2225 goto maybe_dump_sets;
2228 /* If we have multiple successors, we take the intersection of all of
2229 them. Note that in the case of loop exit phi nodes, we may have
2230 phis to translate through. */
2231 else
2233 VEC(basic_block, heap) * worklist;
2234 size_t i;
2235 basic_block bprime, first = NULL;
2237 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2238 FOR_EACH_EDGE (e, ei, block->succs)
2240 if (!first
2241 && BB_VISITED (e->dest))
2242 first = e->dest;
2243 else if (BB_VISITED (e->dest))
2244 VEC_quick_push (basic_block, worklist, e->dest);
2247 /* Of multiple successors we have to have visited one already. */
2248 if (!first)
2250 SET_BIT (changed_blocks, block->index);
2251 BB_VISITED (block) = 0;
2252 BB_DEFERRED (block) = 1;
2253 changed = true;
2254 VEC_free (basic_block, heap, worklist);
2255 goto maybe_dump_sets;
2258 if (phi_nodes (first))
2259 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2260 else
2261 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2263 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2265 if (phi_nodes (bprime))
2267 bitmap_set_t tmp = bitmap_set_new ();
2268 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2269 bitmap_set_and (ANTIC_OUT, tmp);
2270 bitmap_set_free (tmp);
2272 else
2273 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2275 VEC_free (basic_block, heap, worklist);
2278 /* Generate ANTIC_OUT - TMP_GEN. */
2279 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2281 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2282 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2283 TMP_GEN (block));
2285 /* Then union in the ANTIC_OUT - TMP_GEN values,
2286 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2287 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2288 bitmap_value_insert_into_set (ANTIC_IN (block),
2289 expression_for_id (bii));
2291 clean (ANTIC_IN (block), block);
2293 /* !old->expressions can happen when we deferred a block. */
2294 if (!old->expressions || !bitmap_set_equal (old, ANTIC_IN (block)))
2296 changed = true;
2297 SET_BIT (changed_blocks, block->index);
2298 FOR_EACH_EDGE (e, ei, block->preds)
2299 SET_BIT (changed_blocks, e->src->index);
2301 else
2302 RESET_BIT (changed_blocks, block->index);
2304 maybe_dump_sets:
2305 if (dump_file && (dump_flags & TDF_DETAILS))
2307 if (!BB_DEFERRED (block) || BB_VISITED (block))
2309 if (ANTIC_OUT)
2310 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2312 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2313 block->index);
2315 if (S)
2316 print_bitmap_set (dump_file, S, "S", block->index);
2318 else
2320 fprintf (dump_file,
2321 "Block %d was deferred for a future iteration.\n",
2322 block->index);
2325 if (old)
2326 bitmap_set_free (old);
2327 if (S)
2328 bitmap_set_free (S);
2329 if (ANTIC_OUT)
2330 bitmap_set_free (ANTIC_OUT);
2331 return changed;
2334 /* Compute PARTIAL_ANTIC for BLOCK.
2336 If succs(BLOCK) > 1 then
2337 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2338 in ANTIC_OUT for all succ(BLOCK)
2339 else if succs(BLOCK) == 1 then
2340 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2342 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2343 - ANTIC_IN[BLOCK])
2346 static bool
2347 compute_partial_antic_aux (basic_block block,
2348 bool block_has_abnormal_pred_edge)
2350 bool changed = false;
2351 bitmap_set_t old_PA_IN;
2352 bitmap_set_t PA_OUT;
2353 edge e;
2354 edge_iterator ei;
2355 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2357 old_PA_IN = PA_OUT = NULL;
2359 /* If any edges from predecessors are abnormal, antic_in is empty,
2360 so do nothing. */
2361 if (block_has_abnormal_pred_edge)
2362 goto maybe_dump_sets;
2364 /* If there are too many partially anticipatable values in the
2365 block, phi_translate_set can take an exponential time: stop
2366 before the translation starts. */
2367 if (max_pa
2368 && single_succ_p (block)
2369 && bitmap_count_bits (PA_IN (single_succ (block))->values) > max_pa)
2370 goto maybe_dump_sets;
2372 old_PA_IN = PA_IN (block);
2373 PA_OUT = bitmap_set_new ();
2375 /* If the block has no successors, ANTIC_OUT is empty. */
2376 if (EDGE_COUNT (block->succs) == 0)
2378 /* If we have one successor, we could have some phi nodes to
2379 translate through. Note that we can't phi translate across DFS
2380 back edges in partial antic, because it uses a union operation on
2381 the successors. For recurrences like IV's, we will end up
2382 generating a new value in the set on each go around (i + 3 (VH.1)
2383 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2384 else if (single_succ_p (block))
2386 basic_block succ = single_succ (block);
2387 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2388 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2390 /* If we have multiple successors, we take the union of all of
2391 them. */
2392 else
2394 VEC(basic_block, heap) * worklist;
2395 size_t i;
2396 basic_block bprime;
2398 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2399 FOR_EACH_EDGE (e, ei, block->succs)
2401 if (e->flags & EDGE_DFS_BACK)
2402 continue;
2403 VEC_quick_push (basic_block, worklist, e->dest);
2405 if (VEC_length (basic_block, worklist) > 0)
2407 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2409 unsigned int i;
2410 bitmap_iterator bi;
2412 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2413 bitmap_value_insert_into_set (PA_OUT,
2414 expression_for_id (i));
2415 if (phi_nodes (bprime))
2417 bitmap_set_t pa_in = bitmap_set_new ();
2418 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2419 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2420 bitmap_value_insert_into_set (PA_OUT,
2421 expression_for_id (i));
2422 bitmap_set_free (pa_in);
2424 else
2425 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2426 bitmap_value_insert_into_set (PA_OUT,
2427 expression_for_id (i));
2430 VEC_free (basic_block, heap, worklist);
2433 /* PA_IN starts with PA_OUT - TMP_GEN.
2434 Then we subtract things from ANTIC_IN. */
2435 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2437 /* For partial antic, we want to put back in the phi results, since
2438 we will properly avoid making them partially antic over backedges. */
2439 bitmap_ior_into (PA_IN (block)->values, PHI_GEN (block)->values);
2440 bitmap_ior_into (PA_IN (block)->expressions, PHI_GEN (block)->expressions);
2442 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2443 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2445 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2447 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2449 changed = true;
2450 SET_BIT (changed_blocks, block->index);
2451 FOR_EACH_EDGE (e, ei, block->preds)
2452 SET_BIT (changed_blocks, e->src->index);
2454 else
2455 RESET_BIT (changed_blocks, block->index);
2457 maybe_dump_sets:
2458 if (dump_file && (dump_flags & TDF_DETAILS))
2460 if (PA_OUT)
2461 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2463 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2465 if (old_PA_IN)
2466 bitmap_set_free (old_PA_IN);
2467 if (PA_OUT)
2468 bitmap_set_free (PA_OUT);
2469 return changed;
2472 /* Compute ANTIC and partial ANTIC sets. */
2474 static void
2475 compute_antic (void)
2477 bool changed = true;
2478 int num_iterations = 0;
2479 basic_block block;
2480 int i;
2482 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2483 We pre-build the map of blocks with incoming abnormal edges here. */
2484 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2485 sbitmap_zero (has_abnormal_preds);
2487 FOR_EACH_BB (block)
2489 edge_iterator ei;
2490 edge e;
2492 FOR_EACH_EDGE (e, ei, block->preds)
2494 e->flags &= ~EDGE_DFS_BACK;
2495 if (e->flags & EDGE_ABNORMAL)
2497 SET_BIT (has_abnormal_preds, block->index);
2498 break;
2502 BB_VISITED (block) = 0;
2503 BB_DEFERRED (block) = 0;
2504 /* While we are here, give empty ANTIC_IN sets to each block. */
2505 ANTIC_IN (block) = bitmap_set_new ();
2506 PA_IN (block) = bitmap_set_new ();
2509 /* At the exit block we anticipate nothing. */
2510 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2511 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2512 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2514 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2515 sbitmap_ones (changed_blocks);
2516 while (changed)
2518 if (dump_file && (dump_flags & TDF_DETAILS))
2519 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2520 num_iterations++;
2521 changed = false;
2522 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2524 if (TEST_BIT (changed_blocks, postorder[i]))
2526 basic_block block = BASIC_BLOCK (postorder[i]);
2527 changed |= compute_antic_aux (block,
2528 TEST_BIT (has_abnormal_preds,
2529 block->index));
2532 #ifdef ENABLE_CHECKING
2533 /* Theoretically possible, but *highly* unlikely. */
2534 gcc_assert (num_iterations < 500);
2535 #endif
2538 statistics_histogram_event (cfun, "compute_antic iterations",
2539 num_iterations);
2541 if (do_partial_partial)
2543 sbitmap_ones (changed_blocks);
2544 mark_dfs_back_edges ();
2545 num_iterations = 0;
2546 changed = true;
2547 while (changed)
2549 if (dump_file && (dump_flags & TDF_DETAILS))
2550 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2551 num_iterations++;
2552 changed = false;
2553 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2555 if (TEST_BIT (changed_blocks, postorder[i]))
2557 basic_block block = BASIC_BLOCK (postorder[i]);
2558 changed
2559 |= compute_partial_antic_aux (block,
2560 TEST_BIT (has_abnormal_preds,
2561 block->index));
2564 #ifdef ENABLE_CHECKING
2565 /* Theoretically possible, but *highly* unlikely. */
2566 gcc_assert (num_iterations < 500);
2567 #endif
2569 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2570 num_iterations);
2572 sbitmap_free (has_abnormal_preds);
2573 sbitmap_free (changed_blocks);
2576 /* Return true if we can value number the call in STMT. This is true
2577 if we have a pure or constant call. */
2579 static bool
2580 can_value_number_call (gimple stmt)
2582 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2583 return true;
2584 return false;
2587 /* Return true if OP is a tree which we can perform PRE on.
2588 This may not match the operations we can value number, but in
2589 a perfect world would. */
2591 static bool
2592 can_PRE_operation (tree op)
2594 return UNARY_CLASS_P (op)
2595 || BINARY_CLASS_P (op)
2596 || COMPARISON_CLASS_P (op)
2597 || TREE_CODE (op) == INDIRECT_REF
2598 || TREE_CODE (op) == COMPONENT_REF
2599 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2600 || TREE_CODE (op) == CALL_EXPR
2601 || TREE_CODE (op) == ARRAY_REF;
2605 /* Inserted expressions are placed onto this worklist, which is used
2606 for performing quick dead code elimination of insertions we made
2607 that didn't turn out to be necessary. */
2608 static VEC(gimple,heap) *inserted_exprs;
2609 static bitmap inserted_phi_names;
2611 /* Pool allocated fake store expressions are placed onto this
2612 worklist, which, after performing dead code elimination, is walked
2613 to see which expressions need to be put into GC'able memory */
2614 static VEC(gimple, heap) *need_creation;
2616 /* The actual worker for create_component_ref_by_pieces. */
2618 static tree
2619 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2620 unsigned int *operand, gimple_seq *stmts,
2621 gimple domstmt)
2623 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2624 *operand);
2625 tree genop;
2626 ++*operand;
2627 switch (currop->opcode)
2629 case CALL_EXPR:
2631 tree folded, sc = currop->op1;
2632 unsigned int nargs = 0;
2633 tree *args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2634 ref->operands) - 1);
2635 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2637 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2638 operand, stmts,
2639 domstmt);
2640 nargs++;
2642 folded = build_call_array (currop->type,
2643 TREE_CODE (currop->op0) == FUNCTION_DECL
2644 ? build_fold_addr_expr (currop->op0)
2645 : currop->op0,
2646 nargs, args);
2647 free (args);
2648 if (sc)
2650 pre_expr scexpr = get_or_alloc_expr_for (sc);
2651 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2652 if (!sc)
2653 return NULL_TREE;
2654 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2656 return folded;
2658 break;
2659 case TARGET_MEM_REF:
2661 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2662 *operand);
2663 pre_expr op0expr;
2664 tree genop0 = NULL_TREE;
2665 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2666 stmts, domstmt);
2667 if (!baseop)
2668 return NULL_TREE;
2669 if (currop->op0)
2671 op0expr = get_or_alloc_expr_for (currop->op0);
2672 genop0 = find_or_generate_expression (block, op0expr,
2673 stmts, domstmt);
2674 if (!genop0)
2675 return NULL_TREE;
2677 if (DECL_P (baseop))
2678 return build6 (TARGET_MEM_REF, currop->type,
2679 baseop, NULL_TREE,
2680 genop0, currop->op1, currop->op2,
2681 unshare_expr (nextop->op1));
2682 else
2683 return build6 (TARGET_MEM_REF, currop->type,
2684 NULL_TREE, baseop,
2685 genop0, currop->op1, currop->op2,
2686 unshare_expr (nextop->op1));
2688 break;
2689 case ADDR_EXPR:
2690 if (currop->op0)
2692 gcc_assert (is_gimple_min_invariant (currop->op0));
2693 return currop->op0;
2695 /* Fallthrough. */
2696 case REALPART_EXPR:
2697 case IMAGPART_EXPR:
2698 case VIEW_CONVERT_EXPR:
2700 tree folded;
2701 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2702 operand,
2703 stmts, domstmt);
2704 if (!genop0)
2705 return NULL_TREE;
2706 folded = fold_build1 (currop->opcode, currop->type,
2707 genop0);
2708 return folded;
2710 break;
2711 case ALIGN_INDIRECT_REF:
2712 case MISALIGNED_INDIRECT_REF:
2713 case INDIRECT_REF:
2715 tree folded;
2716 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2717 operand,
2718 stmts, domstmt);
2719 if (!genop1)
2720 return NULL_TREE;
2721 genop1 = fold_convert (build_pointer_type (currop->type),
2722 genop1);
2724 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2725 folded = fold_build2 (currop->opcode, currop->type,
2726 genop1, currop->op1);
2727 else
2728 folded = fold_build1 (currop->opcode, currop->type,
2729 genop1);
2730 return folded;
2732 break;
2733 case BIT_FIELD_REF:
2735 tree folded;
2736 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2737 stmts, domstmt);
2738 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2739 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2740 tree genop1;
2741 tree genop2;
2743 if (!genop0)
2744 return NULL_TREE;
2745 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2746 if (!genop1)
2747 return NULL_TREE;
2748 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2749 if (!genop2)
2750 return NULL_TREE;
2751 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2752 genop2);
2753 return folded;
2756 /* For array ref vn_reference_op's, operand 1 of the array ref
2757 is op0 of the reference op and operand 3 of the array ref is
2758 op1. */
2759 case ARRAY_RANGE_REF:
2760 case ARRAY_REF:
2762 tree genop0;
2763 tree genop1 = currop->op0;
2764 pre_expr op1expr;
2765 tree genop2 = currop->op1;
2766 pre_expr op2expr;
2767 tree genop3 = currop->op2;
2768 pre_expr op3expr;
2769 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2770 stmts, domstmt);
2771 if (!genop0)
2772 return NULL_TREE;
2773 op1expr = get_or_alloc_expr_for (genop1);
2774 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2775 if (!genop1)
2776 return NULL_TREE;
2777 if (genop2)
2779 op2expr = get_or_alloc_expr_for (genop2);
2780 genop2 = find_or_generate_expression (block, op2expr, stmts,
2781 domstmt);
2782 if (!genop2)
2783 return NULL_TREE;
2785 if (genop3)
2787 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2788 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2789 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2790 op3expr = get_or_alloc_expr_for (genop3);
2791 genop3 = find_or_generate_expression (block, op3expr, stmts,
2792 domstmt);
2793 if (!genop3)
2794 return NULL_TREE;
2796 return build4 (currop->opcode, currop->type, genop0, genop1,
2797 genop2, genop3);
2799 case COMPONENT_REF:
2801 tree op0;
2802 tree op1;
2803 tree genop2 = currop->op1;
2804 pre_expr op2expr;
2805 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2806 stmts, domstmt);
2807 if (!op0)
2808 return NULL_TREE;
2809 /* op1 should be a FIELD_DECL, which are represented by
2810 themselves. */
2811 op1 = currop->op0;
2812 if (genop2)
2814 op2expr = get_or_alloc_expr_for (genop2);
2815 genop2 = find_or_generate_expression (block, op2expr, stmts,
2816 domstmt);
2817 if (!genop2)
2818 return NULL_TREE;
2821 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2822 genop2);
2824 break;
2825 case SSA_NAME:
2827 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2828 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2829 return genop;
2831 case STRING_CST:
2832 case INTEGER_CST:
2833 case COMPLEX_CST:
2834 case VECTOR_CST:
2835 case REAL_CST:
2836 case CONSTRUCTOR:
2837 case VAR_DECL:
2838 case PARM_DECL:
2839 case CONST_DECL:
2840 case RESULT_DECL:
2841 case FUNCTION_DECL:
2842 return currop->op0;
2844 default:
2845 gcc_unreachable ();
2849 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2850 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2851 trying to rename aggregates into ssa form directly, which is a no no.
2853 Thus, this routine doesn't create temporaries, it just builds a
2854 single access expression for the array, calling
2855 find_or_generate_expression to build the innermost pieces.
2857 This function is a subroutine of create_expression_by_pieces, and
2858 should not be called on it's own unless you really know what you
2859 are doing. */
2861 static tree
2862 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2863 gimple_seq *stmts, gimple domstmt)
2865 unsigned int op = 0;
2866 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2869 /* Find a leader for an expression, or generate one using
2870 create_expression_by_pieces if it's ANTIC but
2871 complex.
2872 BLOCK is the basic_block we are looking for leaders in.
2873 EXPR is the expression to find a leader or generate for.
2874 STMTS is the statement list to put the inserted expressions on.
2875 Returns the SSA_NAME of the LHS of the generated expression or the
2876 leader.
2877 DOMSTMT if non-NULL is a statement that should be dominated by
2878 all uses in the generated expression. If DOMSTMT is non-NULL this
2879 routine can fail and return NULL_TREE. Otherwise it will assert
2880 on failure. */
2882 static tree
2883 find_or_generate_expression (basic_block block, pre_expr expr,
2884 gimple_seq *stmts, gimple domstmt)
2886 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2887 get_expr_value_id (expr), domstmt);
2888 tree genop = NULL;
2889 if (leader)
2891 if (leader->kind == NAME)
2892 genop = PRE_EXPR_NAME (leader);
2893 else if (leader->kind == CONSTANT)
2894 genop = PRE_EXPR_CONSTANT (leader);
2897 /* If it's still NULL, it must be a complex expression, so generate
2898 it recursively. Not so for FRE though. */
2899 if (genop == NULL
2900 && !in_fre)
2902 bitmap_set_t exprset;
2903 unsigned int lookfor = get_expr_value_id (expr);
2904 bool handled = false;
2905 bitmap_iterator bi;
2906 unsigned int i;
2908 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2909 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2911 pre_expr temp = expression_for_id (i);
2912 if (temp->kind != NAME)
2914 handled = true;
2915 genop = create_expression_by_pieces (block, temp, stmts,
2916 domstmt,
2917 get_expr_type (expr));
2918 break;
2921 if (!handled && domstmt)
2922 return NULL_TREE;
2924 gcc_assert (handled);
2926 return genop;
2929 #define NECESSARY GF_PLF_1
2931 /* Create an expression in pieces, so that we can handle very complex
2932 expressions that may be ANTIC, but not necessary GIMPLE.
2933 BLOCK is the basic block the expression will be inserted into,
2934 EXPR is the expression to insert (in value form)
2935 STMTS is a statement list to append the necessary insertions into.
2937 This function will die if we hit some value that shouldn't be
2938 ANTIC but is (IE there is no leader for it, or its components).
2939 This function may also generate expressions that are themselves
2940 partially or fully redundant. Those that are will be either made
2941 fully redundant during the next iteration of insert (for partially
2942 redundant ones), or eliminated by eliminate (for fully redundant
2943 ones).
2945 If DOMSTMT is non-NULL then we make sure that all uses in the
2946 expressions dominate that statement. In this case the function
2947 can return NULL_TREE to signal failure. */
2949 static tree
2950 create_expression_by_pieces (basic_block block, pre_expr expr,
2951 gimple_seq *stmts, gimple domstmt, tree type)
2953 tree temp, name;
2954 tree folded;
2955 gimple_seq forced_stmts = NULL;
2956 unsigned int value_id;
2957 gimple_stmt_iterator gsi;
2958 tree exprtype = type ? type : get_expr_type (expr);
2959 pre_expr nameexpr;
2960 gimple newstmt;
2962 switch (expr->kind)
2964 /* We may hit the NAME/CONSTANT case if we have to convert types
2965 that value numbering saw through. */
2966 case NAME:
2967 folded = PRE_EXPR_NAME (expr);
2968 break;
2969 case CONSTANT:
2970 folded = PRE_EXPR_CONSTANT (expr);
2971 break;
2972 case REFERENCE:
2974 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2975 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
2977 break;
2978 case NARY:
2980 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2981 switch (nary->length)
2983 case 2:
2985 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2986 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
2987 tree genop1 = find_or_generate_expression (block, op1,
2988 stmts, domstmt);
2989 tree genop2 = find_or_generate_expression (block, op2,
2990 stmts, domstmt);
2991 if (!genop1 || !genop2)
2992 return NULL_TREE;
2993 genop1 = fold_convert (TREE_TYPE (nary->op[0]),
2994 genop1);
2995 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
2996 may be a constant with the wrong type. */
2997 if (nary->opcode == POINTER_PLUS_EXPR)
2998 genop2 = fold_convert (sizetype, genop2);
2999 else
3000 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
3002 folded = fold_build2 (nary->opcode, nary->type,
3003 genop1, genop2);
3005 break;
3006 case 1:
3008 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3009 tree genop1 = find_or_generate_expression (block, op1,
3010 stmts, domstmt);
3011 if (!genop1)
3012 return NULL_TREE;
3013 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3015 folded = fold_build1 (nary->opcode, nary->type,
3016 genop1);
3018 break;
3019 default:
3020 return NULL_TREE;
3023 break;
3024 default:
3025 return NULL_TREE;
3028 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3029 folded = fold_convert (exprtype, folded);
3031 /* Force the generated expression to be a sequence of GIMPLE
3032 statements.
3033 We have to call unshare_expr because force_gimple_operand may
3034 modify the tree we pass to it. */
3035 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3036 false, NULL);
3038 /* If we have any intermediate expressions to the value sets, add them
3039 to the value sets and chain them in the instruction stream. */
3040 if (forced_stmts)
3042 gsi = gsi_start (forced_stmts);
3043 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3045 gimple stmt = gsi_stmt (gsi);
3046 tree forcedname = gimple_get_lhs (stmt);
3047 pre_expr nameexpr;
3049 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3050 if (TREE_CODE (forcedname) == SSA_NAME)
3052 VN_INFO_GET (forcedname)->valnum = forcedname;
3053 VN_INFO (forcedname)->value_id = get_next_value_id ();
3054 nameexpr = get_or_alloc_expr_for_name (forcedname);
3055 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3056 if (!in_fre)
3057 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3058 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3060 mark_symbols_for_renaming (stmt);
3062 gimple_seq_add_seq (stmts, forced_stmts);
3065 /* Build and insert the assignment of the end result to the temporary
3066 that we will return. */
3067 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3069 pretemp = create_tmp_var (exprtype, "pretmp");
3070 get_var_ann (pretemp);
3073 temp = pretemp;
3074 add_referenced_var (temp);
3076 if (TREE_CODE (exprtype) == COMPLEX_TYPE
3077 || TREE_CODE (exprtype) == VECTOR_TYPE)
3078 DECL_GIMPLE_REG_P (temp) = 1;
3080 newstmt = gimple_build_assign (temp, folded);
3081 name = make_ssa_name (temp, newstmt);
3082 gimple_assign_set_lhs (newstmt, name);
3083 gimple_set_plf (newstmt, NECESSARY, false);
3085 gimple_seq_add_stmt (stmts, newstmt);
3086 VEC_safe_push (gimple, heap, inserted_exprs, newstmt);
3088 /* All the symbols in NEWEXPR should be put into SSA form. */
3089 mark_symbols_for_renaming (newstmt);
3091 /* Add a value number to the temporary.
3092 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3093 we are creating the expression by pieces, and this particular piece of
3094 the expression may have been represented. There is no harm in replacing
3095 here. */
3096 VN_INFO_GET (name)->valnum = name;
3097 value_id = get_expr_value_id (expr);
3098 VN_INFO (name)->value_id = value_id;
3099 nameexpr = get_or_alloc_expr_for_name (name);
3100 add_to_value (value_id, nameexpr);
3101 if (!in_fre)
3102 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3103 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3105 pre_stats.insertions++;
3106 if (dump_file && (dump_flags & TDF_DETAILS))
3108 fprintf (dump_file, "Inserted ");
3109 print_gimple_stmt (dump_file, newstmt, 0, 0);
3110 fprintf (dump_file, " in predecessor %d\n", block->index);
3113 return name;
3117 /* Returns true if we want to inhibit the insertions of PHI nodes
3118 for the given EXPR for basic block BB (a member of a loop).
3119 We want to do this, when we fear that the induction variable we
3120 create might inhibit vectorization. */
3122 static bool
3123 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3125 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3126 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3127 vn_reference_op_t op;
3128 unsigned i;
3130 /* If we aren't going to vectorize we don't inhibit anything. */
3131 if (!flag_tree_vectorize)
3132 return false;
3134 /* Otherwise we inhibit the insertion when the address of the
3135 memory reference is a simple induction variable. In other
3136 cases the vectorizer won't do anything anyway (either it's
3137 loop invariant or a complicated expression). */
3138 for (i = 0; VEC_iterate (vn_reference_op_s, ops, i, op); ++i)
3140 switch (op->opcode)
3142 case ARRAY_REF:
3143 case ARRAY_RANGE_REF:
3144 if (TREE_CODE (op->op0) != SSA_NAME)
3145 break;
3146 /* Fallthru. */
3147 case SSA_NAME:
3149 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3150 affine_iv iv;
3151 /* Default defs are loop invariant. */
3152 if (!defbb)
3153 break;
3154 /* Defined outside this loop, also loop invariant. */
3155 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3156 break;
3157 /* If it's a simple induction variable inhibit insertion,
3158 the vectorizer might be interested in this one. */
3159 if (simple_iv (bb->loop_father, bb->loop_father,
3160 op->op0, &iv, true))
3161 return true;
3162 /* No simple IV, vectorizer can't do anything, hence no
3163 reason to inhibit the transformation for this operand. */
3164 break;
3166 default:
3167 break;
3170 return false;
3173 /* Insert the to-be-made-available values of expression EXPRNUM for each
3174 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3175 merge the result with a phi node, given the same value number as
3176 NODE. Return true if we have inserted new stuff. */
3178 static bool
3179 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3180 pre_expr *avail)
3182 pre_expr expr = expression_for_id (exprnum);
3183 pre_expr newphi;
3184 unsigned int val = get_expr_value_id (expr);
3185 edge pred;
3186 bool insertions = false;
3187 bool nophi = false;
3188 basic_block bprime;
3189 pre_expr eprime;
3190 edge_iterator ei;
3191 tree type = get_expr_type (expr);
3192 tree temp;
3193 gimple phi;
3195 if (dump_file && (dump_flags & TDF_DETAILS))
3197 fprintf (dump_file, "Found partial redundancy for expression ");
3198 print_pre_expr (dump_file, expr);
3199 fprintf (dump_file, " (%04d)\n", val);
3202 /* Make sure we aren't creating an induction variable. */
3203 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3205 bool firstinsideloop = false;
3206 bool secondinsideloop = false;
3207 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3208 EDGE_PRED (block, 0)->src);
3209 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3210 EDGE_PRED (block, 1)->src);
3211 /* Induction variables only have one edge inside the loop. */
3212 if ((firstinsideloop ^ secondinsideloop)
3213 && (expr->kind != REFERENCE
3214 || inhibit_phi_insertion (block, expr)))
3216 if (dump_file && (dump_flags & TDF_DETAILS))
3217 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3218 nophi = true;
3222 /* Make sure we are not inserting trapping expressions. */
3223 FOR_EACH_EDGE (pred, ei, block->preds)
3225 bprime = pred->src;
3226 eprime = avail[bprime->index];
3227 if (eprime->kind == NARY
3228 && vn_nary_may_trap (PRE_EXPR_NARY (eprime)))
3229 return false;
3232 /* Make the necessary insertions. */
3233 FOR_EACH_EDGE (pred, ei, block->preds)
3235 gimple_seq stmts = NULL;
3236 tree builtexpr;
3237 bprime = pred->src;
3238 eprime = avail[bprime->index];
3240 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3242 builtexpr = create_expression_by_pieces (bprime,
3243 eprime,
3244 &stmts, NULL,
3245 type);
3246 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3247 gsi_insert_seq_on_edge (pred, stmts);
3248 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3249 insertions = true;
3251 else if (eprime->kind == CONSTANT)
3253 /* Constants may not have the right type, fold_convert
3254 should give us back a constant with the right type.
3256 tree constant = PRE_EXPR_CONSTANT (eprime);
3257 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3259 tree builtexpr = fold_convert (type, constant);
3260 if (!is_gimple_min_invariant (builtexpr))
3262 tree forcedexpr = force_gimple_operand (builtexpr,
3263 &stmts, true,
3264 NULL);
3265 if (!is_gimple_min_invariant (forcedexpr))
3267 if (forcedexpr != builtexpr)
3269 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3270 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3272 if (stmts)
3274 gimple_stmt_iterator gsi;
3275 gsi = gsi_start (stmts);
3276 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3278 gimple stmt = gsi_stmt (gsi);
3279 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3280 gimple_set_plf (stmt, NECESSARY, false);
3282 gsi_insert_seq_on_edge (pred, stmts);
3284 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3289 else if (eprime->kind == NAME)
3291 /* We may have to do a conversion because our value
3292 numbering can look through types in certain cases, but
3293 our IL requires all operands of a phi node have the same
3294 type. */
3295 tree name = PRE_EXPR_NAME (eprime);
3296 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3298 tree builtexpr;
3299 tree forcedexpr;
3300 builtexpr = fold_convert (type, name);
3301 forcedexpr = force_gimple_operand (builtexpr,
3302 &stmts, true,
3303 NULL);
3305 if (forcedexpr != name)
3307 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3308 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3311 if (stmts)
3313 gimple_stmt_iterator gsi;
3314 gsi = gsi_start (stmts);
3315 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3317 gimple stmt = gsi_stmt (gsi);
3318 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3319 gimple_set_plf (stmt, NECESSARY, false);
3321 gsi_insert_seq_on_edge (pred, stmts);
3323 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3327 /* If we didn't want a phi node, and we made insertions, we still have
3328 inserted new stuff, and thus return true. If we didn't want a phi node,
3329 and didn't make insertions, we haven't added anything new, so return
3330 false. */
3331 if (nophi && insertions)
3332 return true;
3333 else if (nophi && !insertions)
3334 return false;
3336 /* Now build a phi for the new variable. */
3337 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3339 prephitemp = create_tmp_var (type, "prephitmp");
3340 get_var_ann (prephitemp);
3343 temp = prephitemp;
3344 add_referenced_var (temp);
3346 if (TREE_CODE (type) == COMPLEX_TYPE
3347 || TREE_CODE (type) == VECTOR_TYPE)
3348 DECL_GIMPLE_REG_P (temp) = 1;
3349 phi = create_phi_node (temp, block);
3351 gimple_set_plf (phi, NECESSARY, false);
3352 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3353 VN_INFO (gimple_phi_result (phi))->value_id = val;
3354 VEC_safe_push (gimple, heap, inserted_exprs, phi);
3355 bitmap_set_bit (inserted_phi_names,
3356 SSA_NAME_VERSION (gimple_phi_result (phi)));
3357 FOR_EACH_EDGE (pred, ei, block->preds)
3359 pre_expr ae = avail[pred->src->index];
3360 gcc_assert (get_expr_type (ae) == type
3361 || useless_type_conversion_p (type, get_expr_type (ae)));
3362 if (ae->kind == CONSTANT)
3363 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3364 else
3365 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3366 UNKNOWN_LOCATION);
3369 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3370 add_to_value (val, newphi);
3372 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3373 this insertion, since we test for the existence of this value in PHI_GEN
3374 before proceeding with the partial redundancy checks in insert_aux.
3376 The value may exist in AVAIL_OUT, in particular, it could be represented
3377 by the expression we are trying to eliminate, in which case we want the
3378 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3379 inserted there.
3381 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3382 this block, because if it did, it would have existed in our dominator's
3383 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3386 bitmap_insert_into_set (PHI_GEN (block), newphi);
3387 bitmap_value_replace_in_set (AVAIL_OUT (block),
3388 newphi);
3389 bitmap_insert_into_set (NEW_SETS (block),
3390 newphi);
3392 if (dump_file && (dump_flags & TDF_DETAILS))
3394 fprintf (dump_file, "Created phi ");
3395 print_gimple_stmt (dump_file, phi, 0, 0);
3396 fprintf (dump_file, " in block %d\n", block->index);
3398 pre_stats.phis++;
3399 return true;
3404 /* Perform insertion of partially redundant values.
3405 For BLOCK, do the following:
3406 1. Propagate the NEW_SETS of the dominator into the current block.
3407 If the block has multiple predecessors,
3408 2a. Iterate over the ANTIC expressions for the block to see if
3409 any of them are partially redundant.
3410 2b. If so, insert them into the necessary predecessors to make
3411 the expression fully redundant.
3412 2c. Insert a new PHI merging the values of the predecessors.
3413 2d. Insert the new PHI, and the new expressions, into the
3414 NEW_SETS set.
3415 3. Recursively call ourselves on the dominator children of BLOCK.
3417 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3418 do_regular_insertion and do_partial_insertion.
3422 static bool
3423 do_regular_insertion (basic_block block, basic_block dom)
3425 bool new_stuff = false;
3426 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3427 pre_expr expr;
3428 int i;
3430 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3432 if (expr->kind != NAME)
3434 pre_expr *avail;
3435 unsigned int val;
3436 bool by_some = false;
3437 bool cant_insert = false;
3438 bool all_same = true;
3439 pre_expr first_s = NULL;
3440 edge pred;
3441 basic_block bprime;
3442 pre_expr eprime = NULL;
3443 edge_iterator ei;
3444 pre_expr edoubleprime = NULL;
3445 bool do_insertion = false;
3447 val = get_expr_value_id (expr);
3448 if (bitmap_set_contains_value (PHI_GEN (block), val))
3449 continue;
3450 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3452 if (dump_file && (dump_flags & TDF_DETAILS))
3453 fprintf (dump_file, "Found fully redundant value\n");
3454 continue;
3457 avail = XCNEWVEC (pre_expr, last_basic_block);
3458 FOR_EACH_EDGE (pred, ei, block->preds)
3460 unsigned int vprime;
3462 /* We should never run insertion for the exit block
3463 and so not come across fake pred edges. */
3464 gcc_assert (!(pred->flags & EDGE_FAKE));
3465 bprime = pred->src;
3466 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3467 bprime, block);
3469 /* eprime will generally only be NULL if the
3470 value of the expression, translated
3471 through the PHI for this predecessor, is
3472 undefined. If that is the case, we can't
3473 make the expression fully redundant,
3474 because its value is undefined along a
3475 predecessor path. We can thus break out
3476 early because it doesn't matter what the
3477 rest of the results are. */
3478 if (eprime == NULL)
3480 cant_insert = true;
3481 break;
3484 eprime = fully_constant_expression (eprime);
3485 vprime = get_expr_value_id (eprime);
3486 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3487 vprime, NULL);
3488 if (edoubleprime == NULL)
3490 avail[bprime->index] = eprime;
3491 all_same = false;
3493 else
3495 avail[bprime->index] = edoubleprime;
3496 by_some = true;
3497 /* We want to perform insertions to remove a redundancy on
3498 a path in the CFG we want to optimize for speed. */
3499 if (optimize_edge_for_speed_p (pred))
3500 do_insertion = true;
3501 if (first_s == NULL)
3502 first_s = edoubleprime;
3503 else if (!pre_expr_eq (first_s, edoubleprime))
3504 all_same = false;
3507 /* If we can insert it, it's not the same value
3508 already existing along every predecessor, and
3509 it's defined by some predecessor, it is
3510 partially redundant. */
3511 if (!cant_insert && !all_same && by_some && do_insertion
3512 && dbg_cnt (treepre_insert))
3514 if (insert_into_preds_of_block (block, get_expression_id (expr),
3515 avail))
3516 new_stuff = true;
3518 /* If all edges produce the same value and that value is
3519 an invariant, then the PHI has the same value on all
3520 edges. Note this. */
3521 else if (!cant_insert && all_same && eprime
3522 && (edoubleprime->kind == CONSTANT
3523 || edoubleprime->kind == NAME)
3524 && !value_id_constant_p (val))
3526 unsigned int j;
3527 bitmap_iterator bi;
3528 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3529 value_expressions, val);
3531 unsigned int new_val = get_expr_value_id (edoubleprime);
3532 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3534 pre_expr expr = expression_for_id (j);
3536 if (expr->kind == NAME)
3538 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3539 /* Just reset the value id and valnum so it is
3540 the same as the constant we have discovered. */
3541 if (edoubleprime->kind == CONSTANT)
3543 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3544 pre_stats.constified++;
3546 else
3547 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3548 info->value_id = new_val;
3552 free (avail);
3556 VEC_free (pre_expr, heap, exprs);
3557 return new_stuff;
3561 /* Perform insertion for partially anticipatable expressions. There
3562 is only one case we will perform insertion for these. This case is
3563 if the expression is partially anticipatable, and fully available.
3564 In this case, we know that putting it earlier will enable us to
3565 remove the later computation. */
3568 static bool
3569 do_partial_partial_insertion (basic_block block, basic_block dom)
3571 bool new_stuff = false;
3572 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3573 pre_expr expr;
3574 int i;
3576 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3578 if (expr->kind != NAME)
3580 pre_expr *avail;
3581 unsigned int val;
3582 bool by_all = true;
3583 bool cant_insert = false;
3584 edge pred;
3585 basic_block bprime;
3586 pre_expr eprime = NULL;
3587 edge_iterator ei;
3589 val = get_expr_value_id (expr);
3590 if (bitmap_set_contains_value (PHI_GEN (block), val))
3591 continue;
3592 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3593 continue;
3595 avail = XCNEWVEC (pre_expr, last_basic_block);
3596 FOR_EACH_EDGE (pred, ei, block->preds)
3598 unsigned int vprime;
3599 pre_expr edoubleprime;
3601 /* We should never run insertion for the exit block
3602 and so not come across fake pred edges. */
3603 gcc_assert (!(pred->flags & EDGE_FAKE));
3604 bprime = pred->src;
3605 eprime = phi_translate (expr, ANTIC_IN (block),
3606 PA_IN (block),
3607 bprime, block);
3609 /* eprime will generally only be NULL if the
3610 value of the expression, translated
3611 through the PHI for this predecessor, is
3612 undefined. If that is the case, we can't
3613 make the expression fully redundant,
3614 because its value is undefined along a
3615 predecessor path. We can thus break out
3616 early because it doesn't matter what the
3617 rest of the results are. */
3618 if (eprime == NULL)
3620 cant_insert = true;
3621 break;
3624 eprime = fully_constant_expression (eprime);
3625 vprime = get_expr_value_id (eprime);
3626 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3627 vprime, NULL);
3628 if (edoubleprime == NULL)
3630 by_all = false;
3631 break;
3633 else
3634 avail[bprime->index] = edoubleprime;
3638 /* If we can insert it, it's not the same value
3639 already existing along every predecessor, and
3640 it's defined by some predecessor, it is
3641 partially redundant. */
3642 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3644 pre_stats.pa_insert++;
3645 if (insert_into_preds_of_block (block, get_expression_id (expr),
3646 avail))
3647 new_stuff = true;
3649 free (avail);
3653 VEC_free (pre_expr, heap, exprs);
3654 return new_stuff;
3657 static bool
3658 insert_aux (basic_block block)
3660 basic_block son;
3661 bool new_stuff = false;
3663 if (block)
3665 basic_block dom;
3666 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3667 if (dom)
3669 unsigned i;
3670 bitmap_iterator bi;
3671 bitmap_set_t newset = NEW_SETS (dom);
3672 if (newset)
3674 /* Note that we need to value_replace both NEW_SETS, and
3675 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3676 represented by some non-simple expression here that we want
3677 to replace it with. */
3678 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3680 pre_expr expr = expression_for_id (i);
3681 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3682 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3685 if (!single_pred_p (block))
3687 new_stuff |= do_regular_insertion (block, dom);
3688 if (do_partial_partial)
3689 new_stuff |= do_partial_partial_insertion (block, dom);
3693 for (son = first_dom_son (CDI_DOMINATORS, block);
3694 son;
3695 son = next_dom_son (CDI_DOMINATORS, son))
3697 new_stuff |= insert_aux (son);
3700 return new_stuff;
3703 /* Perform insertion of partially redundant values. */
3705 static void
3706 insert (void)
3708 bool new_stuff = true;
3709 basic_block bb;
3710 int num_iterations = 0;
3712 FOR_ALL_BB (bb)
3713 NEW_SETS (bb) = bitmap_set_new ();
3715 while (new_stuff)
3717 num_iterations++;
3718 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3720 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3724 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3726 static void
3727 add_to_exp_gen (basic_block block, tree op)
3729 if (!in_fre)
3731 pre_expr result;
3732 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3733 return;
3734 result = get_or_alloc_expr_for_name (op);
3735 bitmap_value_insert_into_set (EXP_GEN (block), result);
3739 /* Create value ids for PHI in BLOCK. */
3741 static void
3742 make_values_for_phi (gimple phi, basic_block block)
3744 tree result = gimple_phi_result (phi);
3746 /* We have no need for virtual phis, as they don't represent
3747 actual computations. */
3748 if (is_gimple_reg (result))
3750 pre_expr e = get_or_alloc_expr_for_name (result);
3751 add_to_value (get_expr_value_id (e), e);
3752 bitmap_insert_into_set (PHI_GEN (block), e);
3753 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3754 if (!in_fre)
3756 unsigned i;
3757 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3759 tree arg = gimple_phi_arg_def (phi, i);
3760 if (TREE_CODE (arg) == SSA_NAME)
3762 e = get_or_alloc_expr_for_name (arg);
3763 add_to_value (get_expr_value_id (e), e);
3770 /* Compute the AVAIL set for all basic blocks.
3772 This function performs value numbering of the statements in each basic
3773 block. The AVAIL sets are built from information we glean while doing
3774 this value numbering, since the AVAIL sets contain only one entry per
3775 value.
3777 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3778 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3780 static void
3781 compute_avail (void)
3784 basic_block block, son;
3785 basic_block *worklist;
3786 size_t sp = 0;
3787 unsigned i;
3789 /* We pretend that default definitions are defined in the entry block.
3790 This includes function arguments and the static chain decl. */
3791 for (i = 1; i < num_ssa_names; ++i)
3793 tree name = ssa_name (i);
3794 pre_expr e;
3795 if (!name
3796 || !SSA_NAME_IS_DEFAULT_DEF (name)
3797 || has_zero_uses (name)
3798 || !is_gimple_reg (name))
3799 continue;
3801 e = get_or_alloc_expr_for_name (name);
3802 add_to_value (get_expr_value_id (e), e);
3803 if (!in_fre)
3804 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3805 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3808 /* Allocate the worklist. */
3809 worklist = XNEWVEC (basic_block, n_basic_blocks);
3811 /* Seed the algorithm by putting the dominator children of the entry
3812 block on the worklist. */
3813 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3814 son;
3815 son = next_dom_son (CDI_DOMINATORS, son))
3816 worklist[sp++] = son;
3818 /* Loop until the worklist is empty. */
3819 while (sp)
3821 gimple_stmt_iterator gsi;
3822 gimple stmt;
3823 basic_block dom;
3824 unsigned int stmt_uid = 1;
3826 /* Pick a block from the worklist. */
3827 block = worklist[--sp];
3829 /* Initially, the set of available values in BLOCK is that of
3830 its immediate dominator. */
3831 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3832 if (dom)
3833 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3835 /* Generate values for PHI nodes. */
3836 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3837 make_values_for_phi (gsi_stmt (gsi), block);
3839 /* Now compute value numbers and populate value sets with all
3840 the expressions computed in BLOCK. */
3841 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3843 ssa_op_iter iter;
3844 tree op;
3846 stmt = gsi_stmt (gsi);
3847 gimple_set_uid (stmt, stmt_uid++);
3849 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3851 pre_expr e = get_or_alloc_expr_for_name (op);
3853 add_to_value (get_expr_value_id (e), e);
3854 if (!in_fre)
3855 bitmap_insert_into_set (TMP_GEN (block), e);
3856 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3859 if (gimple_has_volatile_ops (stmt)
3860 || stmt_could_throw_p (stmt))
3861 continue;
3863 switch (gimple_code (stmt))
3865 case GIMPLE_RETURN:
3866 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3867 add_to_exp_gen (block, op);
3868 continue;
3870 case GIMPLE_CALL:
3872 vn_reference_t ref;
3873 unsigned int i;
3874 vn_reference_op_t vro;
3875 pre_expr result = NULL;
3876 VEC(vn_reference_op_s, heap) *ops = NULL;
3878 if (!can_value_number_call (stmt))
3879 continue;
3881 copy_reference_ops_from_call (stmt, &ops);
3882 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3883 gimple_expr_type (stmt),
3884 ops, &ref, false);
3885 VEC_free (vn_reference_op_s, heap, ops);
3886 if (!ref)
3887 continue;
3889 for (i = 0; VEC_iterate (vn_reference_op_s,
3890 ref->operands, i,
3891 vro); i++)
3893 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3894 add_to_exp_gen (block, vro->op0);
3895 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3896 add_to_exp_gen (block, vro->op1);
3897 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3898 add_to_exp_gen (block, vro->op2);
3900 result = (pre_expr) pool_alloc (pre_expr_pool);
3901 result->kind = REFERENCE;
3902 result->id = 0;
3903 PRE_EXPR_REFERENCE (result) = ref;
3905 get_or_alloc_expression_id (result);
3906 add_to_value (get_expr_value_id (result), result);
3907 if (!in_fre)
3908 bitmap_value_insert_into_set (EXP_GEN (block), result);
3909 continue;
3912 case GIMPLE_ASSIGN:
3914 pre_expr result = NULL;
3915 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
3917 case tcc_unary:
3918 case tcc_binary:
3919 case tcc_comparison:
3921 vn_nary_op_t nary;
3922 unsigned int i;
3924 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
3925 gimple_assign_rhs_code (stmt),
3926 gimple_expr_type (stmt),
3927 gimple_assign_rhs1 (stmt),
3928 gimple_assign_rhs2 (stmt),
3929 NULL_TREE, NULL_TREE, &nary);
3931 if (!nary)
3932 continue;
3934 for (i = 0; i < nary->length; i++)
3935 if (TREE_CODE (nary->op[i]) == SSA_NAME)
3936 add_to_exp_gen (block, nary->op[i]);
3938 result = (pre_expr) pool_alloc (pre_expr_pool);
3939 result->kind = NARY;
3940 result->id = 0;
3941 PRE_EXPR_NARY (result) = nary;
3942 break;
3945 case tcc_declaration:
3946 case tcc_reference:
3948 vn_reference_t ref;
3949 unsigned int i;
3950 vn_reference_op_t vro;
3952 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3953 gimple_vuse (stmt),
3954 true, &ref);
3955 if (!ref)
3956 continue;
3958 for (i = 0; VEC_iterate (vn_reference_op_s,
3959 ref->operands, i,
3960 vro); i++)
3962 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3963 add_to_exp_gen (block, vro->op0);
3964 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3965 add_to_exp_gen (block, vro->op1);
3966 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3967 add_to_exp_gen (block, vro->op2);
3969 result = (pre_expr) pool_alloc (pre_expr_pool);
3970 result->kind = REFERENCE;
3971 result->id = 0;
3972 PRE_EXPR_REFERENCE (result) = ref;
3973 break;
3976 default:
3977 /* For any other statement that we don't
3978 recognize, simply add all referenced
3979 SSA_NAMEs to EXP_GEN. */
3980 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3981 add_to_exp_gen (block, op);
3982 continue;
3985 get_or_alloc_expression_id (result);
3986 add_to_value (get_expr_value_id (result), result);
3987 if (!in_fre)
3988 bitmap_value_insert_into_set (EXP_GEN (block), result);
3990 continue;
3992 default:
3993 break;
3997 /* Put the dominator children of BLOCK on the worklist of blocks
3998 to compute available sets for. */
3999 for (son = first_dom_son (CDI_DOMINATORS, block);
4000 son;
4001 son = next_dom_son (CDI_DOMINATORS, son))
4002 worklist[sp++] = son;
4005 free (worklist);
4008 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4009 than the available expressions for it. The insertion point is
4010 right before the first use in STMT. Returns the SSA_NAME that should
4011 be used for replacement. */
4013 static tree
4014 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4016 basic_block bb = gimple_bb (stmt);
4017 gimple_stmt_iterator gsi;
4018 gimple_seq stmts = NULL;
4019 tree expr;
4020 pre_expr e;
4022 /* First create a value expression from the expression we want
4023 to insert and associate it with the value handle for SSA_VN. */
4024 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4025 if (e == NULL)
4026 return NULL_TREE;
4028 /* Then use create_expression_by_pieces to generate a valid
4029 expression to insert at this point of the IL stream. */
4030 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4031 if (expr == NULL_TREE)
4032 return NULL_TREE;
4033 gsi = gsi_for_stmt (stmt);
4034 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4036 return expr;
4039 /* Eliminate fully redundant computations. */
4041 static unsigned int
4042 eliminate (void)
4044 VEC (gimple, heap) *to_remove = NULL;
4045 basic_block b;
4046 unsigned int todo = 0;
4047 gimple_stmt_iterator gsi;
4048 gimple stmt;
4049 unsigned i;
4051 FOR_EACH_BB (b)
4053 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4055 stmt = gsi_stmt (gsi);
4057 /* Lookup the RHS of the expression, see if we have an
4058 available computation for it. If so, replace the RHS with
4059 the available computation. */
4060 if (gimple_has_lhs (stmt)
4061 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4062 && !gimple_assign_ssa_name_copy_p (stmt)
4063 && (!gimple_assign_single_p (stmt)
4064 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4065 && !gimple_has_volatile_ops (stmt)
4066 && !has_zero_uses (gimple_get_lhs (stmt)))
4068 tree lhs = gimple_get_lhs (stmt);
4069 tree rhs = NULL_TREE;
4070 tree sprime = NULL;
4071 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4072 pre_expr sprimeexpr;
4074 if (gimple_assign_single_p (stmt))
4075 rhs = gimple_assign_rhs1 (stmt);
4077 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4078 get_expr_value_id (lhsexpr),
4079 NULL);
4081 if (sprimeexpr)
4083 if (sprimeexpr->kind == CONSTANT)
4084 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4085 else if (sprimeexpr->kind == NAME)
4086 sprime = PRE_EXPR_NAME (sprimeexpr);
4087 else
4088 gcc_unreachable ();
4091 /* If there is no existing leader but SCCVN knows this
4092 value is constant, use that constant. */
4093 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4095 sprime = VN_INFO (lhs)->valnum;
4096 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4097 TREE_TYPE (sprime)))
4098 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4100 if (dump_file && (dump_flags & TDF_DETAILS))
4102 fprintf (dump_file, "Replaced ");
4103 print_gimple_expr (dump_file, stmt, 0, 0);
4104 fprintf (dump_file, " with ");
4105 print_generic_expr (dump_file, sprime, 0);
4106 fprintf (dump_file, " in ");
4107 print_gimple_stmt (dump_file, stmt, 0, 0);
4109 pre_stats.eliminations++;
4110 propagate_tree_value_into_stmt (&gsi, sprime);
4111 stmt = gsi_stmt (gsi);
4112 update_stmt (stmt);
4113 continue;
4116 /* If there is no existing usable leader but SCCVN thinks
4117 it has an expression it wants to use as replacement,
4118 insert that. */
4119 if (!sprime || sprime == lhs)
4121 tree val = VN_INFO (lhs)->valnum;
4122 if (val != VN_TOP
4123 && TREE_CODE (val) == SSA_NAME
4124 && VN_INFO (val)->needs_insertion
4125 && can_PRE_operation (vn_get_expr_for (val)))
4126 sprime = do_SCCVN_insertion (stmt, val);
4128 if (sprime
4129 && sprime != lhs
4130 && (rhs == NULL_TREE
4131 || TREE_CODE (rhs) != SSA_NAME
4132 || may_propagate_copy (rhs, sprime)))
4134 gcc_assert (sprime != rhs);
4136 if (dump_file && (dump_flags & TDF_DETAILS))
4138 fprintf (dump_file, "Replaced ");
4139 print_gimple_expr (dump_file, stmt, 0, 0);
4140 fprintf (dump_file, " with ");
4141 print_generic_expr (dump_file, sprime, 0);
4142 fprintf (dump_file, " in ");
4143 print_gimple_stmt (dump_file, stmt, 0, 0);
4146 if (TREE_CODE (sprime) == SSA_NAME)
4147 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4148 NECESSARY, true);
4149 /* We need to make sure the new and old types actually match,
4150 which may require adding a simple cast, which fold_convert
4151 will do for us. */
4152 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4153 && !useless_type_conversion_p (gimple_expr_type (stmt),
4154 TREE_TYPE (sprime)))
4155 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4157 pre_stats.eliminations++;
4158 propagate_tree_value_into_stmt (&gsi, sprime);
4159 stmt = gsi_stmt (gsi);
4160 update_stmt (stmt);
4162 /* If we removed EH side effects from the statement, clean
4163 its EH information. */
4164 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4166 bitmap_set_bit (need_eh_cleanup,
4167 gimple_bb (stmt)->index);
4168 if (dump_file && (dump_flags & TDF_DETAILS))
4169 fprintf (dump_file, " Removed EH side effects.\n");
4173 /* If the statement is a scalar store, see if the expression
4174 has the same value number as its rhs. If so, the store is
4175 dead. */
4176 else if (gimple_assign_single_p (stmt)
4177 && !is_gimple_reg (gimple_assign_lhs (stmt))
4178 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4179 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4181 tree rhs = gimple_assign_rhs1 (stmt);
4182 tree val;
4183 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4184 gimple_vuse (stmt), true, NULL);
4185 if (TREE_CODE (rhs) == SSA_NAME)
4186 rhs = VN_INFO (rhs)->valnum;
4187 if (val
4188 && operand_equal_p (val, rhs, 0))
4190 if (dump_file && (dump_flags & TDF_DETAILS))
4192 fprintf (dump_file, "Deleted redundant store ");
4193 print_gimple_stmt (dump_file, stmt, 0, 0);
4196 /* Queue stmt for removal. */
4197 VEC_safe_push (gimple, heap, to_remove, stmt);
4200 /* Visit COND_EXPRs and fold the comparison with the
4201 available value-numbers. */
4202 else if (gimple_code (stmt) == GIMPLE_COND)
4204 tree op0 = gimple_cond_lhs (stmt);
4205 tree op1 = gimple_cond_rhs (stmt);
4206 tree result;
4208 if (TREE_CODE (op0) == SSA_NAME)
4209 op0 = VN_INFO (op0)->valnum;
4210 if (TREE_CODE (op1) == SSA_NAME)
4211 op1 = VN_INFO (op1)->valnum;
4212 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4213 op0, op1);
4214 if (result && TREE_CODE (result) == INTEGER_CST)
4216 if (integer_zerop (result))
4217 gimple_cond_make_false (stmt);
4218 else
4219 gimple_cond_make_true (stmt);
4220 update_stmt (stmt);
4221 todo = TODO_cleanup_cfg;
4224 /* Visit indirect calls and turn them into direct calls if
4225 possible. */
4226 if (gimple_code (stmt) == GIMPLE_CALL
4227 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4229 tree fn = VN_INFO (gimple_call_fn (stmt))->valnum;
4230 if (TREE_CODE (fn) == ADDR_EXPR
4231 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
4233 if (dump_file && (dump_flags & TDF_DETAILS))
4235 fprintf (dump_file, "Replacing call target with ");
4236 print_generic_expr (dump_file, fn, 0);
4237 fprintf (dump_file, " in ");
4238 print_gimple_stmt (dump_file, stmt, 0, 0);
4241 gimple_call_set_fn (stmt, fn);
4242 update_stmt (stmt);
4243 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4245 bitmap_set_bit (need_eh_cleanup,
4246 gimple_bb (stmt)->index);
4247 if (dump_file && (dump_flags & TDF_DETAILS))
4248 fprintf (dump_file, " Removed EH side effects.\n");
4251 /* Changing an indirect call to a direct call may
4252 have exposed different semantics. This may
4253 require an SSA update. */
4254 todo |= TODO_update_ssa_only_virtuals;
4259 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4261 gimple stmt, phi = gsi_stmt (gsi);
4262 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4263 pre_expr sprimeexpr, resexpr;
4264 gimple_stmt_iterator gsi2;
4266 /* We want to perform redundant PHI elimination. Do so by
4267 replacing the PHI with a single copy if possible.
4268 Do not touch inserted, single-argument or virtual PHIs. */
4269 if (gimple_phi_num_args (phi) == 1
4270 || !is_gimple_reg (res)
4271 || bitmap_bit_p (inserted_phi_names, SSA_NAME_VERSION (res)))
4273 gsi_next (&gsi);
4274 continue;
4277 resexpr = get_or_alloc_expr_for_name (res);
4278 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4279 get_expr_value_id (resexpr), NULL);
4280 if (sprimeexpr)
4282 if (sprimeexpr->kind == CONSTANT)
4283 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4284 else if (sprimeexpr->kind == NAME)
4285 sprime = PRE_EXPR_NAME (sprimeexpr);
4286 else
4287 gcc_unreachable ();
4289 if (!sprimeexpr
4290 || sprime == res)
4292 gsi_next (&gsi);
4293 continue;
4296 if (dump_file && (dump_flags & TDF_DETAILS))
4298 fprintf (dump_file, "Replaced redundant PHI node defining ");
4299 print_generic_expr (dump_file, res, 0);
4300 fprintf (dump_file, " with ");
4301 print_generic_expr (dump_file, sprime, 0);
4302 fprintf (dump_file, "\n");
4305 remove_phi_node (&gsi, false);
4307 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4308 sprime = fold_convert (TREE_TYPE (res), sprime);
4309 stmt = gimple_build_assign (res, sprime);
4310 SSA_NAME_DEF_STMT (res) = stmt;
4311 if (TREE_CODE (sprime) == SSA_NAME)
4312 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4313 NECESSARY, true);
4314 gsi2 = gsi_after_labels (b);
4315 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4316 /* Queue the copy for eventual removal. */
4317 VEC_safe_push (gimple, heap, to_remove, stmt);
4318 pre_stats.eliminations++;
4322 /* We cannot remove stmts during BB walk, especially not release SSA
4323 names there as this confuses the VN machinery. The stmts ending
4324 up in to_remove are either stores or simple copies. */
4325 for (i = 0; VEC_iterate (gimple, to_remove, i, stmt); ++i)
4327 tree lhs = gimple_assign_lhs (stmt);
4328 use_operand_p use_p;
4329 gimple use_stmt;
4331 /* If there is a single use only, propagate the equivalency
4332 instead of keeping the copy. */
4333 if (TREE_CODE (lhs) == SSA_NAME
4334 && single_imm_use (lhs, &use_p, &use_stmt)
4335 && may_propagate_copy (USE_FROM_PTR (use_p),
4336 gimple_assign_rhs1 (stmt)))
4338 SET_USE (use_p, gimple_assign_rhs1 (stmt));
4339 update_stmt (use_stmt);
4342 /* If this is a store or a now unused copy, remove it. */
4343 if (TREE_CODE (lhs) != SSA_NAME
4344 || has_zero_uses (lhs))
4346 gsi = gsi_for_stmt (stmt);
4347 unlink_stmt_vdef (stmt);
4348 gsi_remove (&gsi, true);
4349 release_defs (stmt);
4352 VEC_free (gimple, heap, to_remove);
4354 return todo;
4357 /* Borrow a bit of tree-ssa-dce.c for the moment.
4358 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4359 this may be a bit faster, and we may want critical edges kept split. */
4361 /* If OP's defining statement has not already been determined to be necessary,
4362 mark that statement necessary. Return the stmt, if it is newly
4363 necessary. */
4365 static inline gimple
4366 mark_operand_necessary (tree op)
4368 gimple stmt;
4370 gcc_assert (op);
4372 if (TREE_CODE (op) != SSA_NAME)
4373 return NULL;
4375 stmt = SSA_NAME_DEF_STMT (op);
4376 gcc_assert (stmt);
4378 if (gimple_plf (stmt, NECESSARY)
4379 || gimple_nop_p (stmt))
4380 return NULL;
4382 gimple_set_plf (stmt, NECESSARY, true);
4383 return stmt;
4386 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4387 to insert PHI nodes sometimes, and because value numbering of casts isn't
4388 perfect, we sometimes end up inserting dead code. This simple DCE-like
4389 pass removes any insertions we made that weren't actually used. */
4391 static void
4392 remove_dead_inserted_code (void)
4394 VEC(gimple,heap) *worklist = NULL;
4395 int i;
4396 gimple t;
4398 worklist = VEC_alloc (gimple, heap, VEC_length (gimple, inserted_exprs));
4399 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4401 if (gimple_plf (t, NECESSARY))
4402 VEC_quick_push (gimple, worklist, t);
4404 while (VEC_length (gimple, worklist) > 0)
4406 t = VEC_pop (gimple, worklist);
4408 /* PHI nodes are somewhat special in that each PHI alternative has
4409 data and control dependencies. All the statements feeding the
4410 PHI node's arguments are always necessary. */
4411 if (gimple_code (t) == GIMPLE_PHI)
4413 unsigned k;
4415 VEC_reserve (gimple, heap, worklist, gimple_phi_num_args (t));
4416 for (k = 0; k < gimple_phi_num_args (t); k++)
4418 tree arg = PHI_ARG_DEF (t, k);
4419 if (TREE_CODE (arg) == SSA_NAME)
4421 gimple n = mark_operand_necessary (arg);
4422 if (n)
4423 VEC_quick_push (gimple, worklist, n);
4427 else
4429 /* Propagate through the operands. Examine all the USE, VUSE and
4430 VDEF operands in this statement. Mark all the statements
4431 which feed this statement's uses as necessary. */
4432 ssa_op_iter iter;
4433 tree use;
4435 /* The operands of VDEF expressions are also needed as they
4436 represent potential definitions that may reach this
4437 statement (VDEF operands allow us to follow def-def
4438 links). */
4440 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4442 gimple n = mark_operand_necessary (use);
4443 if (n)
4444 VEC_safe_push (gimple, heap, worklist, n);
4449 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4451 if (!gimple_plf (t, NECESSARY))
4453 gimple_stmt_iterator gsi;
4455 if (dump_file && (dump_flags & TDF_DETAILS))
4457 fprintf (dump_file, "Removing unnecessary insertion:");
4458 print_gimple_stmt (dump_file, t, 0, 0);
4461 gsi = gsi_for_stmt (t);
4462 if (gimple_code (t) == GIMPLE_PHI)
4463 remove_phi_node (&gsi, true);
4464 else
4466 gsi_remove (&gsi, true);
4467 release_defs (t);
4471 VEC_free (gimple, heap, worklist);
4474 /* Initialize data structures used by PRE. */
4476 static void
4477 init_pre (bool do_fre)
4479 basic_block bb;
4481 next_expression_id = 1;
4482 expressions = NULL;
4483 VEC_safe_push (pre_expr, heap, expressions, NULL);
4484 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4485 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4486 get_max_value_id() + 1);
4488 in_fre = do_fre;
4490 inserted_exprs = NULL;
4491 need_creation = NULL;
4492 pretemp = NULL_TREE;
4493 storetemp = NULL_TREE;
4494 prephitemp = NULL_TREE;
4496 connect_infinite_loops_to_exit ();
4497 memset (&pre_stats, 0, sizeof (pre_stats));
4500 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4501 post_order_compute (postorder, false, false);
4503 FOR_ALL_BB (bb)
4504 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4506 calculate_dominance_info (CDI_POST_DOMINATORS);
4507 calculate_dominance_info (CDI_DOMINATORS);
4509 bitmap_obstack_initialize (&grand_bitmap_obstack);
4510 inserted_phi_names = BITMAP_ALLOC (&grand_bitmap_obstack);
4511 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4512 expr_pred_trans_eq, free);
4513 expression_to_id = htab_create (num_ssa_names * 3,
4514 pre_expr_hash,
4515 pre_expr_eq, NULL);
4516 seen_during_translate = BITMAP_ALLOC (&grand_bitmap_obstack);
4517 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4518 sizeof (struct bitmap_set), 30);
4519 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4520 sizeof (struct pre_expr_d), 30);
4521 FOR_ALL_BB (bb)
4523 EXP_GEN (bb) = bitmap_set_new ();
4524 PHI_GEN (bb) = bitmap_set_new ();
4525 TMP_GEN (bb) = bitmap_set_new ();
4526 AVAIL_OUT (bb) = bitmap_set_new ();
4529 need_eh_cleanup = BITMAP_ALLOC (NULL);
4533 /* Deallocate data structures used by PRE. */
4535 static void
4536 fini_pre (bool do_fre)
4538 basic_block bb;
4540 free (postorder);
4541 VEC_free (bitmap_set_t, heap, value_expressions);
4542 VEC_free (gimple, heap, inserted_exprs);
4543 VEC_free (gimple, heap, need_creation);
4544 bitmap_obstack_release (&grand_bitmap_obstack);
4545 free_alloc_pool (bitmap_set_pool);
4546 free_alloc_pool (pre_expr_pool);
4547 htab_delete (phi_translate_table);
4548 htab_delete (expression_to_id);
4550 FOR_ALL_BB (bb)
4552 free (bb->aux);
4553 bb->aux = NULL;
4556 free_dominance_info (CDI_POST_DOMINATORS);
4558 if (!bitmap_empty_p (need_eh_cleanup))
4560 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4561 cleanup_tree_cfg ();
4564 BITMAP_FREE (need_eh_cleanup);
4566 if (!do_fre)
4567 loop_optimizer_finalize ();
4570 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4571 only wants to do full redundancy elimination. */
4573 static unsigned int
4574 execute_pre (bool do_fre)
4576 unsigned int todo = 0;
4578 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4580 /* This has to happen before SCCVN runs because
4581 loop_optimizer_init may create new phis, etc. */
4582 if (!do_fre)
4583 loop_optimizer_init (LOOPS_NORMAL);
4585 if (!run_scc_vn (do_fre))
4587 if (!do_fre)
4589 remove_dead_inserted_code ();
4590 loop_optimizer_finalize ();
4593 return 0;
4595 init_pre (do_fre);
4596 scev_initialize ();
4599 /* Collect and value number expressions computed in each basic block. */
4600 compute_avail ();
4602 if (dump_file && (dump_flags & TDF_DETAILS))
4604 basic_block bb;
4606 FOR_ALL_BB (bb)
4608 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4609 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4610 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4611 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4615 /* Insert can get quite slow on an incredibly large number of basic
4616 blocks due to some quadratic behavior. Until this behavior is
4617 fixed, don't run it when he have an incredibly large number of
4618 bb's. If we aren't going to run insert, there is no point in
4619 computing ANTIC, either, even though it's plenty fast. */
4620 if (!do_fre && n_basic_blocks < 4000)
4622 compute_antic ();
4623 insert ();
4626 /* Remove all the redundant expressions. */
4627 todo |= eliminate ();
4629 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4630 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4631 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4632 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4633 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4635 /* Make sure to remove fake edges before committing our inserts.
4636 This makes sure we don't end up with extra critical edges that
4637 we would need to split. */
4638 remove_fake_exit_edges ();
4639 gsi_commit_edge_inserts ();
4641 clear_expression_ids ();
4642 free_scc_vn ();
4643 if (!do_fre)
4644 remove_dead_inserted_code ();
4646 scev_finalize ();
4647 fini_pre (do_fre);
4649 return todo;
4652 /* Gate and execute functions for PRE. */
4654 static unsigned int
4655 do_pre (void)
4657 return execute_pre (false);
4660 static bool
4661 gate_pre (void)
4663 return flag_tree_pre != 0;
4666 struct gimple_opt_pass pass_pre =
4669 GIMPLE_PASS,
4670 "pre", /* name */
4671 gate_pre, /* gate */
4672 do_pre, /* execute */
4673 NULL, /* sub */
4674 NULL, /* next */
4675 0, /* static_pass_number */
4676 TV_TREE_PRE, /* tv_id */
4677 PROP_no_crit_edges | PROP_cfg
4678 | PROP_ssa, /* properties_required */
4679 0, /* properties_provided */
4680 0, /* properties_destroyed */
4681 TODO_rebuild_alias, /* todo_flags_start */
4682 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4683 | TODO_verify_ssa /* todo_flags_finish */
4688 /* Gate and execute functions for FRE. */
4690 static unsigned int
4691 execute_fre (void)
4693 return execute_pre (true);
4696 static bool
4697 gate_fre (void)
4699 return flag_tree_fre != 0;
4702 struct gimple_opt_pass pass_fre =
4705 GIMPLE_PASS,
4706 "fre", /* name */
4707 gate_fre, /* gate */
4708 execute_fre, /* execute */
4709 NULL, /* sub */
4710 NULL, /* next */
4711 0, /* static_pass_number */
4712 TV_TREE_FRE, /* tv_id */
4713 PROP_cfg | PROP_ssa, /* properties_required */
4714 0, /* properties_provided */
4715 0, /* properties_destroyed */
4716 0, /* todo_flags_start */
4717 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */