2008-09-22 Richard Guenther <rguenther@suse.de>
[official-gcc.git] / gcc / tree-ssa-pre.c
blobdd61349a3d3931874a3d095f6660bb0d21a3c617
1 /* SSA-PRE for trees.
2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007
3 Free Software Foundation, Inc.
4 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
5 <stevenb@suse.de>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
12 any later version.
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 #include "config.h"
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "ggc.h"
28 #include "tree.h"
29 #include "basic-block.h"
30 #include "diagnostic.h"
31 #include "tree-inline.h"
32 #include "tree-flow.h"
33 #include "gimple.h"
34 #include "tree-dump.h"
35 #include "timevar.h"
36 #include "fibheap.h"
37 #include "hashtab.h"
38 #include "tree-iterator.h"
39 #include "real.h"
40 #include "alloc-pool.h"
41 #include "obstack.h"
42 #include "tree-pass.h"
43 #include "flags.h"
44 #include "bitmap.h"
45 #include "langhooks.h"
46 #include "cfgloop.h"
47 #include "tree-ssa-sccvn.h"
48 #include "params.h"
49 #include "dbgcnt.h"
51 /* TODO:
53 1. Avail sets can be shared by making an avail_find_leader that
54 walks up the dominator tree and looks in those avail sets.
55 This might affect code optimality, it's unclear right now.
56 2. Strength reduction can be performed by anticipating expressions
57 we can repair later on.
58 3. We can do back-substitution or smarter value numbering to catch
59 commutative expressions split up over multiple statements.
62 /* For ease of terminology, "expression node" in the below refers to
63 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
64 represent the actual statement containing the expressions we care about,
65 and we cache the value number by putting it in the expression. */
67 /* Basic algorithm
69 First we walk the statements to generate the AVAIL sets, the
70 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
71 generation of values/expressions by a given block. We use them
72 when computing the ANTIC sets. The AVAIL sets consist of
73 SSA_NAME's that represent values, so we know what values are
74 available in what blocks. AVAIL is a forward dataflow problem. In
75 SSA, values are never killed, so we don't need a kill set, or a
76 fixpoint iteration, in order to calculate the AVAIL sets. In
77 traditional parlance, AVAIL sets tell us the downsafety of the
78 expressions/values.
80 Next, we generate the ANTIC sets. These sets represent the
81 anticipatable expressions. ANTIC is a backwards dataflow
82 problem. An expression is anticipatable in a given block if it could
83 be generated in that block. This means that if we had to perform
84 an insertion in that block, of the value of that expression, we
85 could. Calculating the ANTIC sets requires phi translation of
86 expressions, because the flow goes backwards through phis. We must
87 iterate to a fixpoint of the ANTIC sets, because we have a kill
88 set. Even in SSA form, values are not live over the entire
89 function, only from their definition point onwards. So we have to
90 remove values from the ANTIC set once we go past the definition
91 point of the leaders that make them up.
92 compute_antic/compute_antic_aux performs this computation.
94 Third, we perform insertions to make partially redundant
95 expressions fully redundant.
97 An expression is partially redundant (excluding partial
98 anticipation) if:
100 1. It is AVAIL in some, but not all, of the predecessors of a
101 given block.
102 2. It is ANTIC in all the predecessors.
104 In order to make it fully redundant, we insert the expression into
105 the predecessors where it is not available, but is ANTIC.
107 For the partial anticipation case, we only perform insertion if it
108 is partially anticipated in some block, and fully available in all
109 of the predecessors.
111 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
112 performs these steps.
114 Fourth, we eliminate fully redundant expressions.
115 This is a simple statement walk that replaces redundant
116 calculations with the now available values. */
118 /* Representations of value numbers:
120 Value numbers are represented by a representative SSA_NAME. We
121 will create fake SSA_NAME's in situations where we need a
122 representative but do not have one (because it is a complex
123 expression). In order to facilitate storing the value numbers in
124 bitmaps, and keep the number of wasted SSA_NAME's down, we also
125 associate a value_id with each value number, and create full blown
126 ssa_name's only where we actually need them (IE in operands of
127 existing expressions).
129 Theoretically you could replace all the value_id's with
130 SSA_NAME_VERSION, but this would allocate a large number of
131 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
132 It would also require an additional indirection at each point we
133 use the value id. */
135 /* Representation of expressions on value numbers:
137 Expressions consisting of value numbers are represented the same
138 way as our VN internally represents them, with an additional
139 "pre_expr" wrapping around them in order to facilitate storing all
140 of the expressions in the same sets. */
142 /* Representation of sets:
144 The dataflow sets do not need to be sorted in any particular order
145 for the majority of their lifetime, are simply represented as two
146 bitmaps, one that keeps track of values present in the set, and one
147 that keeps track of expressions present in the set.
149 When we need them in topological order, we produce it on demand by
150 transforming the bitmap into an array and sorting it into topo
151 order. */
153 /* Type of expression, used to know which member of the PRE_EXPR union
154 is valid. */
156 enum pre_expr_kind
158 NAME,
159 NARY,
160 REFERENCE,
161 CONSTANT
164 typedef union pre_expr_union_d
166 tree name;
167 tree constant;
168 vn_nary_op_t nary;
169 vn_reference_t reference;
170 } pre_expr_union;
172 typedef struct pre_expr_d
174 enum pre_expr_kind kind;
175 unsigned int id;
176 pre_expr_union u;
177 } *pre_expr;
179 #define PRE_EXPR_NAME(e) (e)->u.name
180 #define PRE_EXPR_NARY(e) (e)->u.nary
181 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
182 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
184 static int
185 pre_expr_eq (const void *p1, const void *p2)
187 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
188 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
190 if (e1->kind != e2->kind)
191 return false;
193 switch (e1->kind)
195 case CONSTANT:
196 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
197 PRE_EXPR_CONSTANT (e2));
198 case NAME:
199 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
200 case NARY:
201 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
202 case REFERENCE:
203 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
204 PRE_EXPR_REFERENCE (e2));
205 default:
206 abort();
210 static hashval_t
211 pre_expr_hash (const void *p1)
213 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
214 switch (e->kind)
216 case CONSTANT:
217 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
218 case NAME:
219 return iterative_hash_expr (PRE_EXPR_NAME (e), 0);
220 case NARY:
221 return vn_nary_op_compute_hash (PRE_EXPR_NARY (e));
222 case REFERENCE:
223 return vn_reference_compute_hash (PRE_EXPR_REFERENCE (e));
224 default:
225 abort ();
230 /* Next global expression id number. */
231 static unsigned int next_expression_id;
233 /* Mapping from expression to id number we can use in bitmap sets. */
234 DEF_VEC_P (pre_expr);
235 DEF_VEC_ALLOC_P (pre_expr, heap);
236 static VEC(pre_expr, heap) *expressions;
237 static htab_t expression_to_id;
239 /* Allocate an expression id for EXPR. */
241 static inline unsigned int
242 alloc_expression_id (pre_expr expr)
244 void **slot;
245 /* Make sure we won't overflow. */
246 gcc_assert (next_expression_id + 1 > next_expression_id);
247 expr->id = next_expression_id++;
248 VEC_safe_push (pre_expr, heap, expressions, expr);
249 slot = htab_find_slot (expression_to_id, expr, INSERT);
250 gcc_assert (!*slot);
251 *slot = expr;
252 return next_expression_id - 1;
255 /* Return the expression id for tree EXPR. */
257 static inline unsigned int
258 get_expression_id (const pre_expr expr)
260 return expr->id;
263 static inline unsigned int
264 lookup_expression_id (const pre_expr expr)
266 void **slot;
268 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
269 if (!slot)
270 return 0;
271 return ((pre_expr)*slot)->id;
274 /* Return the existing expression id for EXPR, or create one if one
275 does not exist yet. */
277 static inline unsigned int
278 get_or_alloc_expression_id (pre_expr expr)
280 unsigned int id = lookup_expression_id (expr);
281 if (id == 0)
282 return alloc_expression_id (expr);
283 return expr->id = id;
286 /* Return the expression that has expression id ID */
288 static inline pre_expr
289 expression_for_id (unsigned int id)
291 return VEC_index (pre_expr, expressions, id);
294 /* Free the expression id field in all of our expressions,
295 and then destroy the expressions array. */
297 static void
298 clear_expression_ids (void)
300 VEC_free (pre_expr, heap, expressions);
303 static alloc_pool pre_expr_pool;
305 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
307 static pre_expr
308 get_or_alloc_expr_for_name (tree name)
310 pre_expr result = (pre_expr) pool_alloc (pre_expr_pool);
311 unsigned int result_id;
313 result->kind = NAME;
314 result->id = 0;
315 PRE_EXPR_NAME (result) = name;
316 result_id = lookup_expression_id (result);
317 if (result_id != 0)
319 pool_free (pre_expr_pool, result);
320 result = expression_for_id (result_id);
321 return result;
323 get_or_alloc_expression_id (result);
324 return result;
327 static bool in_fre = false;
329 /* An unordered bitmap set. One bitmap tracks values, the other,
330 expressions. */
331 typedef struct bitmap_set
333 bitmap expressions;
334 bitmap values;
335 } *bitmap_set_t;
337 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
338 EXECUTE_IF_SET_IN_BITMAP((set)->expressions, 0, (id), (bi))
340 /* Mapping from value id to expressions with that value_id. */
341 DEF_VEC_P (bitmap_set_t);
342 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
343 static VEC(bitmap_set_t, heap) *value_expressions;
345 /* Sets that we need to keep track of. */
346 typedef struct bb_bitmap_sets
348 /* The EXP_GEN set, which represents expressions/values generated in
349 a basic block. */
350 bitmap_set_t exp_gen;
352 /* The PHI_GEN set, which represents PHI results generated in a
353 basic block. */
354 bitmap_set_t phi_gen;
356 /* The TMP_GEN set, which represents results/temporaries generated
357 in a basic block. IE the LHS of an expression. */
358 bitmap_set_t tmp_gen;
360 /* The AVAIL_OUT set, which represents which values are available in
361 a given basic block. */
362 bitmap_set_t avail_out;
364 /* The ANTIC_IN set, which represents which values are anticipatable
365 in a given basic block. */
366 bitmap_set_t antic_in;
368 /* The PA_IN set, which represents which values are
369 partially anticipatable in a given basic block. */
370 bitmap_set_t pa_in;
372 /* The NEW_SETS set, which is used during insertion to augment the
373 AVAIL_OUT set of blocks with the new insertions performed during
374 the current iteration. */
375 bitmap_set_t new_sets;
377 /* True if we have visited this block during ANTIC calculation. */
378 unsigned int visited:1;
380 /* True we have deferred processing this block during ANTIC
381 calculation until its successor is processed. */
382 unsigned int deferred : 1;
383 } *bb_value_sets_t;
385 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
386 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
387 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
388 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
389 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
390 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
391 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
392 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
393 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
396 /* Maximal set of values, used to initialize the ANTIC problem, which
397 is an intersection problem. */
398 static bitmap_set_t maximal_set;
400 /* Basic block list in postorder. */
401 static int *postorder;
403 /* This structure is used to keep track of statistics on what
404 optimization PRE was able to perform. */
405 static struct
407 /* The number of RHS computations eliminated by PRE. */
408 int eliminations;
410 /* The number of new expressions/temporaries generated by PRE. */
411 int insertions;
413 /* The number of inserts found due to partial anticipation */
414 int pa_insert;
416 /* The number of new PHI nodes added by PRE. */
417 int phis;
419 /* The number of values found constant. */
420 int constified;
422 } pre_stats;
424 static bool do_partial_partial;
425 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
426 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
427 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
428 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
429 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
430 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
431 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr, bool);
432 static bitmap_set_t bitmap_set_new (void);
433 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
434 gimple, tree);
435 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
436 gimple);
438 /* We can add and remove elements and entries to and from sets
439 and hash tables, so we use alloc pools for them. */
441 static alloc_pool bitmap_set_pool;
442 static bitmap_obstack grand_bitmap_obstack;
444 /* To avoid adding 300 temporary variables when we only need one, we
445 only create one temporary variable, on demand, and build ssa names
446 off that. We do have to change the variable if the types don't
447 match the current variable's type. */
448 static tree pretemp;
449 static tree storetemp;
450 static tree prephitemp;
452 /* Set of blocks with statements that have had its EH information
453 cleaned up. */
454 static bitmap need_eh_cleanup;
456 /* Which expressions have been seen during a given phi translation. */
457 static bitmap seen_during_translate;
459 /* The phi_translate_table caches phi translations for a given
460 expression and predecessor. */
462 static htab_t phi_translate_table;
464 /* A three tuple {e, pred, v} used to cache phi translations in the
465 phi_translate_table. */
467 typedef struct expr_pred_trans_d
469 /* The expression. */
470 pre_expr e;
472 /* The predecessor block along which we translated the expression. */
473 basic_block pred;
475 /* The value that resulted from the translation. */
476 pre_expr v;
478 /* The hashcode for the expression, pred pair. This is cached for
479 speed reasons. */
480 hashval_t hashcode;
481 } *expr_pred_trans_t;
482 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
484 /* Return the hash value for a phi translation table entry. */
486 static hashval_t
487 expr_pred_trans_hash (const void *p)
489 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
490 return ve->hashcode;
493 /* Return true if two phi translation table entries are the same.
494 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
496 static int
497 expr_pred_trans_eq (const void *p1, const void *p2)
499 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
500 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
501 basic_block b1 = ve1->pred;
502 basic_block b2 = ve2->pred;
504 /* If they are not translations for the same basic block, they can't
505 be equal. */
506 if (b1 != b2)
507 return false;
508 return pre_expr_eq (ve1->e, ve2->e);
511 /* Search in the phi translation table for the translation of
512 expression E in basic block PRED.
513 Return the translated value, if found, NULL otherwise. */
515 static inline pre_expr
516 phi_trans_lookup (pre_expr e, basic_block pred)
518 void **slot;
519 struct expr_pred_trans_d ept;
521 ept.e = e;
522 ept.pred = pred;
523 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
524 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
525 NO_INSERT);
526 if (!slot)
527 return NULL;
528 else
529 return ((expr_pred_trans_t) *slot)->v;
533 /* Add the tuple mapping from {expression E, basic block PRED} to
534 value V, to the phi translation table. */
536 static inline void
537 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
539 void **slot;
540 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
541 new_pair->e = e;
542 new_pair->pred = pred;
543 new_pair->v = v;
544 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
545 pred->index);
547 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
548 new_pair->hashcode, INSERT);
549 if (*slot)
550 free (*slot);
551 *slot = (void *) new_pair;
555 /* Add expression E to the expression set of value id V. */
557 void
558 add_to_value (unsigned int v, pre_expr e)
560 bitmap_set_t set;
562 if (v >= VEC_length (bitmap_set_t, value_expressions))
564 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
565 v + 1);
568 set = VEC_index (bitmap_set_t, value_expressions, v);
569 if (!set)
571 set = bitmap_set_new ();
572 VEC_replace (bitmap_set_t, value_expressions, v, set);
575 bitmap_insert_into_set_1 (set, e, true);
578 /* Create a new bitmap set and return it. */
580 static bitmap_set_t
581 bitmap_set_new (void)
583 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
584 ret->expressions = BITMAP_ALLOC (&grand_bitmap_obstack);
585 ret->values = BITMAP_ALLOC (&grand_bitmap_obstack);
586 return ret;
589 /* Return the value id for a PRE expression EXPR. */
591 static unsigned int
592 get_expr_value_id (pre_expr expr)
594 switch (expr->kind)
596 case CONSTANT:
598 unsigned int id;
599 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
600 if (id == 0)
602 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
603 add_to_value (id, expr);
605 return id;
607 case NAME:
608 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
609 case NARY:
610 return PRE_EXPR_NARY (expr)->value_id;
611 case REFERENCE:
612 return PRE_EXPR_REFERENCE (expr)->value_id;
613 default:
614 gcc_unreachable ();
618 /* Remove an expression EXPR from a bitmapped set. */
620 static void
621 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
623 unsigned int val = get_expr_value_id (expr);
624 if (!value_id_constant_p (val))
626 bitmap_clear_bit (set->values, val);
627 bitmap_clear_bit (set->expressions, get_expression_id (expr));
631 static void
632 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
633 bool allow_constants)
635 unsigned int val = get_expr_value_id (expr);
636 if (allow_constants || !value_id_constant_p (val))
638 /* We specifically expect this and only this function to be able to
639 insert constants into a set. */
640 bitmap_set_bit (set->values, val);
641 bitmap_set_bit (set->expressions, get_or_alloc_expression_id (expr));
645 /* Insert an expression EXPR into a bitmapped set. */
647 static void
648 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
650 bitmap_insert_into_set_1 (set, expr, false);
653 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
655 static void
656 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
658 bitmap_copy (dest->expressions, orig->expressions);
659 bitmap_copy (dest->values, orig->values);
663 /* Free memory used up by SET. */
664 static void
665 bitmap_set_free (bitmap_set_t set)
667 BITMAP_FREE (set->expressions);
668 BITMAP_FREE (set->values);
672 /* A comparison function for use in qsort to top sort a bitmap set. Simply
673 subtracts value ids, since they are created with leaves before
674 their parent users (IE topological order). */
676 static int
677 value_id_compare (const void *pa, const void *pb)
679 const unsigned int vha = get_expr_value_id (*((const pre_expr *)pa));
680 const unsigned int vhb = get_expr_value_id (*((const pre_expr *)pb));
682 return vha - vhb;
685 /* Generate an topological-ordered array of bitmap set SET. */
687 static VEC(pre_expr, heap) *
688 sorted_array_from_bitmap_set (bitmap_set_t set)
690 unsigned int i;
691 bitmap_iterator bi;
692 VEC(pre_expr, heap) *result = NULL;
694 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
695 VEC_safe_push (pre_expr, heap, result, expression_for_id (i));
697 qsort (VEC_address (pre_expr, result), VEC_length (pre_expr, result),
698 sizeof (pre_expr), value_id_compare);
700 return result;
703 /* Perform bitmapped set operation DEST &= ORIG. */
705 static void
706 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
708 bitmap_iterator bi;
709 unsigned int i;
711 if (dest != orig)
713 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
715 bitmap_and_into (dest->values, orig->values);
716 bitmap_copy (temp, dest->expressions);
717 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
719 pre_expr expr = expression_for_id (i);
720 unsigned int value_id = get_expr_value_id (expr);
721 if (!bitmap_bit_p (dest->values, value_id))
722 bitmap_clear_bit (dest->expressions, i);
724 BITMAP_FREE (temp);
728 /* Subtract all values and expressions contained in ORIG from DEST. */
730 static bitmap_set_t
731 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
733 bitmap_set_t result = bitmap_set_new ();
734 bitmap_iterator bi;
735 unsigned int i;
737 bitmap_and_compl (result->expressions, dest->expressions,
738 orig->expressions);
740 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
742 pre_expr expr = expression_for_id (i);
743 unsigned int value_id = get_expr_value_id (expr);
744 bitmap_set_bit (result->values, value_id);
747 return result;
750 /* Subtract all the values in bitmap set B from bitmap set A. */
752 static void
753 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
755 unsigned int i;
756 bitmap_iterator bi;
757 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
759 bitmap_copy (temp, a->expressions);
760 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
762 pre_expr expr = expression_for_id (i);
763 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
764 bitmap_remove_from_set (a, expr);
766 BITMAP_FREE (temp);
770 /* Return true if bitmapped set SET contains the value VALUE_ID. */
772 static bool
773 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
775 if (value_id_constant_p (value_id))
776 return true;
778 if (!set || bitmap_empty_p (set->expressions))
779 return false;
781 return bitmap_bit_p (set->values, value_id);
784 static inline bool
785 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
787 return bitmap_bit_p (set->expressions, get_expression_id (expr));
790 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
792 static void
793 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
794 const pre_expr expr)
796 bitmap_set_t exprset;
797 unsigned int i;
798 bitmap_iterator bi;
800 if (value_id_constant_p (lookfor))
801 return;
803 if (!bitmap_set_contains_value (set, lookfor))
804 return;
806 /* The number of expressions having a given value is usually
807 significantly less than the total number of expressions in SET.
808 Thus, rather than check, for each expression in SET, whether it
809 has the value LOOKFOR, we walk the reverse mapping that tells us
810 what expressions have a given value, and see if any of those
811 expressions are in our set. For large testcases, this is about
812 5-10x faster than walking the bitmap. If this is somehow a
813 significant lose for some cases, we can choose which set to walk
814 based on the set size. */
815 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
816 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
818 if (bitmap_bit_p (set->expressions, i))
820 bitmap_clear_bit (set->expressions, i);
821 bitmap_set_bit (set->expressions, get_expression_id (expr));
822 return;
827 /* Return true if two bitmap sets are equal. */
829 static bool
830 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
832 return bitmap_equal_p (a->values, b->values);
835 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
836 and add it otherwise. */
838 static void
839 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
841 unsigned int val = get_expr_value_id (expr);
843 if (bitmap_set_contains_value (set, val))
844 bitmap_set_replace_value (set, val, expr);
845 else
846 bitmap_insert_into_set (set, expr);
849 /* Insert EXPR into SET if EXPR's value is not already present in
850 SET. */
852 static void
853 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
855 unsigned int val = get_expr_value_id (expr);
857 if (value_id_constant_p (val))
858 return;
860 if (!bitmap_set_contains_value (set, val))
861 bitmap_insert_into_set (set, expr);
864 /* Print out EXPR to outfile. */
866 static void
867 print_pre_expr (FILE *outfile, const pre_expr expr)
869 switch (expr->kind)
871 case CONSTANT:
872 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
873 break;
874 case NAME:
875 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
876 break;
877 case NARY:
879 unsigned int i;
880 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
881 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
882 for (i = 0; i < nary->length; i++)
884 print_generic_expr (outfile, nary->op[i], 0);
885 if (i != (unsigned) nary->length - 1)
886 fprintf (outfile, ",");
888 fprintf (outfile, "}");
890 break;
892 case REFERENCE:
894 vn_reference_op_t vro;
895 unsigned int i;
896 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
897 fprintf (outfile, "{");
898 for (i = 0;
899 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
900 i++)
902 if (vro->opcode != SSA_NAME
903 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
904 fprintf (outfile, "%s ", tree_code_name [vro->opcode]);
905 if (vro->op0)
907 if (vro->op1)
908 fprintf (outfile, "<");
909 print_generic_expr (outfile, vro->op0, 0);
910 if (vro->op1)
912 fprintf (outfile, ",");
913 print_generic_expr (outfile, vro->op1, 0);
915 if (vro->op1)
916 fprintf (outfile, ">");
918 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
919 fprintf (outfile, ",");
921 fprintf (outfile, "}");
923 break;
926 void debug_pre_expr (pre_expr);
928 /* Like print_pre_expr but always prints to stderr. */
929 void
930 debug_pre_expr (pre_expr e)
932 print_pre_expr (stderr, e);
933 fprintf (stderr, "\n");
936 /* Print out SET to OUTFILE. */
938 static void
939 print_bitmap_set (FILE *outfile, bitmap_set_t set,
940 const char *setname, int blockindex)
942 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
943 if (set)
945 bool first = true;
946 unsigned i;
947 bitmap_iterator bi;
949 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
951 const pre_expr expr = expression_for_id (i);
953 if (!first)
954 fprintf (outfile, ", ");
955 first = false;
956 print_pre_expr (outfile, expr);
958 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
961 fprintf (outfile, " }\n");
964 void debug_bitmap_set (bitmap_set_t);
966 void
967 debug_bitmap_set (bitmap_set_t set)
969 print_bitmap_set (stderr, set, "debug", 0);
972 /* Print out the expressions that have VAL to OUTFILE. */
974 void
975 print_value_expressions (FILE *outfile, unsigned int val)
977 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
978 if (set)
980 char s[10];
981 sprintf (s, "%04d", val);
982 print_bitmap_set (outfile, set, s, 0);
987 void
988 debug_value_expressions (unsigned int val)
990 print_value_expressions (stderr, val);
993 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
994 represent it. */
996 static pre_expr
997 get_or_alloc_expr_for_constant (tree constant)
999 unsigned int result_id;
1000 unsigned int value_id;
1001 pre_expr newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1002 newexpr->kind = CONSTANT;
1003 PRE_EXPR_CONSTANT (newexpr) = constant;
1004 result_id = lookup_expression_id (newexpr);
1005 if (result_id != 0)
1007 pool_free (pre_expr_pool, newexpr);
1008 newexpr = expression_for_id (result_id);
1009 return newexpr;
1011 value_id = get_or_alloc_constant_value_id (constant);
1012 get_or_alloc_expression_id (newexpr);
1013 add_to_value (value_id, newexpr);
1014 return newexpr;
1017 /* Given a value id V, find the actual tree representing the constant
1018 value if there is one, and return it. Return NULL if we can't find
1019 a constant. */
1021 static tree
1022 get_constant_for_value_id (unsigned int v)
1024 if (value_id_constant_p (v))
1026 unsigned int i;
1027 bitmap_iterator bi;
1028 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1030 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1032 pre_expr expr = expression_for_id (i);
1033 if (expr->kind == CONSTANT)
1034 return PRE_EXPR_CONSTANT (expr);
1037 return NULL;
1040 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1041 Currently only supports constants and SSA_NAMES. */
1042 static pre_expr
1043 get_or_alloc_expr_for (tree t)
1045 if (TREE_CODE (t) == SSA_NAME)
1046 return get_or_alloc_expr_for_name (t);
1047 else if (is_gimple_min_invariant (t))
1048 return get_or_alloc_expr_for_constant (t);
1049 else
1051 /* More complex expressions can result from SCCVN expression
1052 simplification that inserts values for them. As they all
1053 do not have VOPs the get handled by the nary ops struct. */
1054 vn_nary_op_t result;
1055 unsigned int result_id;
1056 vn_nary_op_lookup (t, &result);
1057 if (result != NULL)
1059 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1060 e->kind = NARY;
1061 PRE_EXPR_NARY (e) = result;
1062 result_id = lookup_expression_id (e);
1063 if (result_id != 0)
1065 pool_free (pre_expr_pool, e);
1066 e = expression_for_id (result_id);
1067 return e;
1069 alloc_expression_id (e);
1070 return e;
1073 return NULL;
1076 /* Return the folded version of T if T, when folded, is a gimple
1077 min_invariant. Otherwise, return T. */
1079 static pre_expr
1080 fully_constant_expression (pre_expr e)
1082 switch (e->kind)
1084 case CONSTANT:
1085 return e;
1086 case NARY:
1088 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1089 switch (TREE_CODE_CLASS (nary->opcode))
1091 case tcc_binary:
1093 /* We have to go from trees to pre exprs to value ids to
1094 constants. */
1095 tree naryop0 = nary->op[0];
1096 tree naryop1 = nary->op[1];
1097 tree const0, const1, result;
1098 if (is_gimple_min_invariant (naryop0))
1099 const0 = naryop0;
1100 else
1102 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1103 unsigned int vrep0 = get_expr_value_id (rep0);
1104 const0 = get_constant_for_value_id (vrep0);
1106 if (is_gimple_min_invariant (naryop1))
1107 const1 = naryop1;
1108 else
1110 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1111 unsigned int vrep1 = get_expr_value_id (rep1);
1112 const1 = get_constant_for_value_id (vrep1);
1114 result = NULL;
1115 if (const0 && const1)
1117 tree type1 = TREE_TYPE (nary->op[0]);
1118 tree type2 = TREE_TYPE (nary->op[1]);
1119 const0 = fold_convert (type1, const0);
1120 const1 = fold_convert (type2, const1);
1121 result = fold_binary (nary->opcode, nary->type, const0,
1122 const1);
1124 if (result && is_gimple_min_invariant (result))
1125 return get_or_alloc_expr_for_constant (result);
1126 return e;
1128 case tcc_unary:
1130 /* We have to go from trees to pre exprs to value ids to
1131 constants. */
1132 tree naryop0 = nary->op[0];
1133 tree const0, result;
1134 if (is_gimple_min_invariant (naryop0))
1135 const0 = naryop0;
1136 else
1138 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1139 unsigned int vrep0 = get_expr_value_id (rep0);
1140 const0 = get_constant_for_value_id (vrep0);
1142 result = NULL;
1143 if (const0)
1145 tree type1 = TREE_TYPE (nary->op[0]);
1146 const0 = fold_convert (type1, const0);
1147 result = fold_unary (nary->opcode, nary->type, const0);
1150 if (result && is_gimple_min_invariant (result))
1151 return get_or_alloc_expr_for_constant (result);
1152 return e;
1154 default:
1155 return e;
1158 case REFERENCE:
1160 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1161 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1162 vn_reference_op_t op;
1164 /* Try to simplify the translated expression if it is
1165 a call to a builtin function with at most two arguments. */
1166 op = VEC_index (vn_reference_op_s, operands, 0);
1167 if (op->opcode == CALL_EXPR
1168 && TREE_CODE (op->op0) == ADDR_EXPR
1169 && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL
1170 && DECL_BUILT_IN (TREE_OPERAND (op->op0, 0))
1171 && VEC_length (vn_reference_op_s, operands) >= 2
1172 && VEC_length (vn_reference_op_s, operands) <= 3)
1174 vn_reference_op_t arg0, arg1 = NULL;
1175 bool anyconst = false;
1176 arg0 = VEC_index (vn_reference_op_s, operands, 1);
1177 if (VEC_length (vn_reference_op_s, operands) > 2)
1178 arg1 = VEC_index (vn_reference_op_s, operands, 2);
1179 if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant
1180 || (arg0->opcode == ADDR_EXPR
1181 && is_gimple_min_invariant (arg0->op0)))
1182 anyconst = true;
1183 if (arg1
1184 && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant
1185 || (arg1->opcode == ADDR_EXPR
1186 && is_gimple_min_invariant (arg1->op0))))
1187 anyconst = true;
1188 if (anyconst)
1190 tree folded = build_call_expr (TREE_OPERAND (op->op0, 0),
1191 arg1 ? 2 : 1,
1192 arg0->op0,
1193 arg1 ? arg1->op0 : NULL);
1194 if (folded
1195 && TREE_CODE (folded) == NOP_EXPR)
1196 folded = TREE_OPERAND (folded, 0);
1197 if (folded
1198 && is_gimple_min_invariant (folded))
1199 return get_or_alloc_expr_for_constant (folded);
1202 return e;
1204 default:
1205 return e;
1207 return e;
1210 /* Translate the vuses in the VUSES vector backwards through phi nodes
1211 in PHIBLOCK, so that they have the value they would have in
1212 BLOCK. */
1214 static VEC(tree, gc) *
1215 translate_vuses_through_block (VEC (tree, gc) *vuses,
1216 basic_block phiblock,
1217 basic_block block)
1219 tree oldvuse;
1220 VEC(tree, gc) *result = NULL;
1221 int i;
1223 for (i = 0; VEC_iterate (tree, vuses, i, oldvuse); i++)
1225 gimple phi = SSA_NAME_DEF_STMT (oldvuse);
1226 if (gimple_code (phi) == GIMPLE_PHI
1227 && gimple_bb (phi) == phiblock)
1229 edge e = find_edge (block, gimple_bb (phi));
1230 if (e)
1232 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1233 if (def != oldvuse)
1235 if (!result)
1236 result = VEC_copy (tree, gc, vuses);
1237 VEC_replace (tree, result, i, def);
1243 /* We avoid creating a new copy of the vuses unless something
1244 actually changed, so result can be NULL. */
1245 if (result)
1247 sort_vuses (result);
1248 return result;
1250 return vuses;
1254 /* Like find_leader, but checks for the value existing in SET1 *or*
1255 SET2. This is used to avoid making a set consisting of the union
1256 of PA_IN and ANTIC_IN during insert. */
1258 static inline pre_expr
1259 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1261 pre_expr result;
1263 result = bitmap_find_leader (set1, val, NULL);
1264 if (!result && set2)
1265 result = bitmap_find_leader (set2, val, NULL);
1266 return result;
1269 /* Get the tree type for our PRE expression e. */
1271 static tree
1272 get_expr_type (const pre_expr e)
1274 switch (e->kind)
1276 case NAME:
1277 return TREE_TYPE (PRE_EXPR_NAME (e));
1278 case CONSTANT:
1279 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1280 case REFERENCE:
1282 vn_reference_op_t vro;
1284 gcc_assert (PRE_EXPR_REFERENCE (e)->operands);
1285 vro = VEC_index (vn_reference_op_s,
1286 PRE_EXPR_REFERENCE (e)->operands,
1288 /* We don't store type along with COMPONENT_REF because it is
1289 always the same as FIELD_DECL's type. */
1290 if (!vro->type)
1292 gcc_assert (vro->opcode == COMPONENT_REF);
1293 return TREE_TYPE (vro->op0);
1295 return vro->type;
1298 case NARY:
1299 return PRE_EXPR_NARY (e)->type;
1301 gcc_unreachable();
1304 /* Get a representative SSA_NAME for a given expression.
1305 Since all of our sub-expressions are treated as values, we require
1306 them to be SSA_NAME's for simplicity.
1307 Prior versions of GVNPRE used to use "value handles" here, so that
1308 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1309 either case, the operands are really values (IE we do not expect
1310 them to be usable without finding leaders). */
1312 static tree
1313 get_representative_for (const pre_expr e)
1315 tree exprtype;
1316 tree name;
1317 unsigned int value_id = get_expr_value_id (e);
1319 switch (e->kind)
1321 case NAME:
1322 return PRE_EXPR_NAME (e);
1323 case CONSTANT:
1324 return PRE_EXPR_CONSTANT (e);
1325 case NARY:
1326 case REFERENCE:
1328 /* Go through all of the expressions representing this value
1329 and pick out an SSA_NAME. */
1330 unsigned int i;
1331 bitmap_iterator bi;
1332 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1333 value_id);
1334 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1336 pre_expr rep = expression_for_id (i);
1337 if (rep->kind == NAME)
1338 return PRE_EXPR_NAME (rep);
1341 break;
1343 /* If we reached here we couldn't find an SSA_NAME. This can
1344 happen when we've discovered a value that has never appeared in
1345 the program as set to an SSA_NAME, most likely as the result of
1346 phi translation. */
1347 if (dump_file)
1349 fprintf (dump_file,
1350 "Could not find SSA_NAME representative for expression:");
1351 print_pre_expr (dump_file, e);
1352 fprintf (dump_file, "\n");
1355 exprtype = get_expr_type (e);
1357 /* Build and insert the assignment of the end result to the temporary
1358 that we will return. */
1359 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1361 pretemp = create_tmp_var (exprtype, "pretmp");
1362 get_var_ann (pretemp);
1365 name = make_ssa_name (pretemp, gimple_build_nop ());
1366 VN_INFO_GET (name)->value_id = value_id;
1367 if (e->kind == CONSTANT)
1368 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1369 else
1370 VN_INFO (name)->valnum = name;
1372 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1373 if (dump_file)
1375 fprintf (dump_file, "Created SSA_NAME representative ");
1376 print_generic_expr (dump_file, name, 0);
1377 fprintf (dump_file, " for expression:");
1378 print_pre_expr (dump_file, e);
1379 fprintf (dump_file, "\n");
1382 return name;
1388 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1389 the phis in PRED. SEEN is a bitmap saying which expression we have
1390 translated since we started translation of the toplevel expression.
1391 Return NULL if we can't find a leader for each part of the
1392 translated expression. */
1394 static pre_expr
1395 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1396 basic_block pred, basic_block phiblock, bitmap seen)
1398 pre_expr oldexpr = expr;
1399 pre_expr phitrans;
1401 if (!expr)
1402 return NULL;
1404 if (value_id_constant_p (get_expr_value_id (expr)))
1405 return expr;
1407 phitrans = phi_trans_lookup (expr, pred);
1408 if (phitrans)
1409 return phitrans;
1411 /* Prevent cycles when we have recursively dependent leaders. This
1412 can only happen when phi translating the maximal set. */
1413 if (seen)
1415 unsigned int expr_id = get_expression_id (expr);
1416 if (bitmap_bit_p (seen, expr_id))
1417 return NULL;
1418 bitmap_set_bit (seen, expr_id);
1421 switch (expr->kind)
1423 /* Constants contain no values that need translation. */
1424 case CONSTANT:
1425 return expr;
1427 case NARY:
1429 unsigned int i;
1430 bool changed = false;
1431 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1432 struct vn_nary_op_s newnary;
1433 /* The NARY structure is only guaranteed to have been
1434 allocated to the nary->length operands. */
1435 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1436 - sizeof (tree) * (4 - nary->length)));
1438 for (i = 0; i < newnary.length; i++)
1440 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1441 continue;
1442 else
1444 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1445 pre_expr leader = find_leader_in_sets (op_val_id, set1, set2);
1446 pre_expr result = phi_translate_1 (leader, set1, set2,
1447 pred, phiblock, seen);
1448 if (result && result != leader)
1450 tree name = get_representative_for (result);
1451 if (!name)
1452 return NULL;
1453 newnary.op[i] = name;
1455 else if (!result)
1456 return NULL;
1458 changed |= newnary.op[i] != nary->op[i];
1461 if (changed)
1463 pre_expr constant;
1465 tree result = vn_nary_op_lookup_pieces (newnary.length,
1466 newnary.opcode,
1467 newnary.type,
1468 newnary.op[0],
1469 newnary.op[1],
1470 newnary.op[2],
1471 newnary.op[3],
1472 &nary);
1473 unsigned int new_val_id;
1475 expr = (pre_expr) pool_alloc (pre_expr_pool);
1476 expr->kind = NARY;
1477 expr->id = 0;
1478 if (result && is_gimple_min_invariant (result))
1479 return get_or_alloc_expr_for_constant (result);
1482 if (nary)
1484 PRE_EXPR_NARY (expr) = nary;
1485 constant = fully_constant_expression (expr);
1486 if (constant != expr)
1487 return constant;
1489 new_val_id = nary->value_id;
1490 get_or_alloc_expression_id (expr);
1492 else
1494 new_val_id = get_next_value_id ();
1495 VEC_safe_grow_cleared (bitmap_set_t, heap,
1496 value_expressions,
1497 get_max_value_id() + 1);
1498 nary = vn_nary_op_insert_pieces (newnary.length,
1499 newnary.opcode,
1500 newnary.type,
1501 newnary.op[0],
1502 newnary.op[1],
1503 newnary.op[2],
1504 newnary.op[3],
1505 result, new_val_id);
1506 PRE_EXPR_NARY (expr) = nary;
1507 constant = fully_constant_expression (expr);
1508 if (constant != expr)
1509 return constant;
1510 get_or_alloc_expression_id (expr);
1512 add_to_value (new_val_id, expr);
1514 phi_trans_add (oldexpr, expr, pred);
1515 return expr;
1517 break;
1519 case REFERENCE:
1521 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1522 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1523 VEC (tree, gc) *vuses = ref->vuses;
1524 VEC (tree, gc) *newvuses = vuses;
1525 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1526 bool changed = false;
1527 unsigned int i;
1528 vn_reference_op_t operand;
1529 vn_reference_t newref;
1531 for (i = 0; VEC_iterate (vn_reference_op_s, operands, i, operand); i++)
1533 pre_expr opresult;
1534 pre_expr leader;
1535 tree oldop0 = operand->op0;
1536 tree oldop1 = operand->op1;
1537 tree oldop2 = operand->op2;
1538 tree op0 = oldop0;
1539 tree op1 = oldop1;
1540 tree op2 = oldop2;
1541 tree type = operand->type;
1542 vn_reference_op_s newop = *operand;
1544 if (op0 && TREE_CODE (op0) == SSA_NAME)
1546 unsigned int op_val_id = VN_INFO (op0)->value_id;
1547 leader = find_leader_in_sets (op_val_id, set1, set2);
1548 opresult = phi_translate_1 (leader, set1, set2,
1549 pred, phiblock, seen);
1550 if (opresult && opresult != leader)
1552 tree name = get_representative_for (opresult);
1553 if (!name)
1554 break;
1555 op0 = name;
1557 else if (!opresult)
1558 break;
1560 changed |= op0 != oldop0;
1562 if (op1 && TREE_CODE (op1) == SSA_NAME)
1564 unsigned int op_val_id = VN_INFO (op1)->value_id;
1565 leader = find_leader_in_sets (op_val_id, set1, set2);
1566 opresult = phi_translate_1 (leader, set1, set2,
1567 pred, phiblock, seen);
1568 if (opresult && opresult != leader)
1570 tree name = get_representative_for (opresult);
1571 if (!name)
1572 break;
1573 op1 = name;
1575 else if (!opresult)
1576 break;
1578 changed |= op1 != oldop1;
1579 if (op2 && TREE_CODE (op2) == SSA_NAME)
1581 unsigned int op_val_id = VN_INFO (op2)->value_id;
1582 leader = find_leader_in_sets (op_val_id, set1, set2);
1583 opresult = phi_translate_1 (leader, set1, set2,
1584 pred, phiblock, seen);
1585 if (opresult && opresult != leader)
1587 tree name = get_representative_for (opresult);
1588 if (!name)
1589 break;
1590 op2 = name;
1592 else if (!opresult)
1593 break;
1595 changed |= op2 != oldop2;
1597 if (!newoperands)
1598 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1599 /* We may have changed from an SSA_NAME to a constant */
1600 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1601 newop.opcode = TREE_CODE (op0);
1602 newop.type = type;
1603 newop.op0 = op0;
1604 newop.op1 = op1;
1605 newop.op2 = op2;
1606 VEC_replace (vn_reference_op_s, newoperands, i, &newop);
1608 if (i != VEC_length (vn_reference_op_s, operands))
1610 if (newoperands)
1611 VEC_free (vn_reference_op_s, heap, newoperands);
1612 return NULL;
1615 newvuses = translate_vuses_through_block (vuses, phiblock, pred);
1616 changed |= newvuses != vuses;
1618 if (changed)
1620 unsigned int new_val_id;
1621 pre_expr constant;
1623 tree result = vn_reference_lookup_pieces (newvuses,
1624 newoperands,
1625 &newref, true);
1626 if (newref)
1627 VEC_free (vn_reference_op_s, heap, newoperands);
1629 if (result && is_gimple_min_invariant (result))
1631 gcc_assert (!newoperands);
1632 return get_or_alloc_expr_for_constant (result);
1635 expr = (pre_expr) pool_alloc (pre_expr_pool);
1636 expr->kind = REFERENCE;
1637 expr->id = 0;
1639 if (newref)
1641 PRE_EXPR_REFERENCE (expr) = newref;
1642 constant = fully_constant_expression (expr);
1643 if (constant != expr)
1644 return constant;
1646 new_val_id = newref->value_id;
1647 get_or_alloc_expression_id (expr);
1649 else
1651 new_val_id = get_next_value_id ();
1652 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
1653 get_max_value_id() + 1);
1654 newref = vn_reference_insert_pieces (newvuses,
1655 newoperands,
1656 result, new_val_id);
1657 newoperands = NULL;
1658 PRE_EXPR_REFERENCE (expr) = newref;
1659 constant = fully_constant_expression (expr);
1660 if (constant != expr)
1661 return constant;
1662 get_or_alloc_expression_id (expr);
1664 add_to_value (new_val_id, expr);
1666 VEC_free (vn_reference_op_s, heap, newoperands);
1667 phi_trans_add (oldexpr, expr, pred);
1668 return expr;
1670 break;
1672 case NAME:
1674 gimple phi = NULL;
1675 edge e;
1676 gimple def_stmt;
1677 tree name = PRE_EXPR_NAME (expr);
1679 def_stmt = SSA_NAME_DEF_STMT (name);
1680 if (gimple_code (def_stmt) == GIMPLE_PHI
1681 && gimple_bb (def_stmt) == phiblock)
1682 phi = def_stmt;
1683 else
1684 return expr;
1686 e = find_edge (pred, gimple_bb (phi));
1687 if (e)
1689 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1690 pre_expr newexpr;
1692 /* Handle constant. */
1693 if (is_gimple_min_invariant (def))
1694 return get_or_alloc_expr_for_constant (def);
1696 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1697 return NULL;
1699 newexpr = get_or_alloc_expr_for_name (def);
1700 return newexpr;
1703 return expr;
1705 default:
1706 gcc_unreachable ();
1710 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1711 the phis in PRED.
1712 Return NULL if we can't find a leader for each part of the
1713 translated expression. */
1715 static pre_expr
1716 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1717 basic_block pred, basic_block phiblock)
1719 bitmap_clear (seen_during_translate);
1720 return phi_translate_1 (expr, set1, set2, pred, phiblock,
1721 seen_during_translate);
1724 /* For each expression in SET, translate the values through phi nodes
1725 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1726 expressions in DEST. */
1728 static void
1729 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1730 basic_block phiblock)
1732 VEC (pre_expr, heap) *exprs;
1733 pre_expr expr;
1734 int i;
1736 if (!phi_nodes (phiblock))
1738 bitmap_set_copy (dest, set);
1739 return;
1742 exprs = sorted_array_from_bitmap_set (set);
1743 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1745 pre_expr translated;
1746 translated = phi_translate (expr, set, NULL, pred, phiblock);
1748 /* Don't add constants or empty translations to the cache, since
1749 we won't look them up that way, or use the result, anyway. */
1750 if (translated && !value_id_constant_p (get_expr_value_id (translated)))
1751 phi_trans_add (expr, translated, pred);
1753 if (translated != NULL)
1754 bitmap_value_insert_into_set (dest, translated);
1756 VEC_free (pre_expr, heap, exprs);
1759 /* Find the leader for a value (i.e., the name representing that
1760 value) in a given set, and return it. If STMT is non-NULL it
1761 makes sure the defining statement for the leader dominates it.
1762 Return NULL if no leader is found. */
1764 static pre_expr
1765 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1767 if (value_id_constant_p (val))
1769 unsigned int i;
1770 bitmap_iterator bi;
1771 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1773 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1775 pre_expr expr = expression_for_id (i);
1776 if (expr->kind == CONSTANT)
1777 return expr;
1780 if (bitmap_set_contains_value (set, val))
1782 /* Rather than walk the entire bitmap of expressions, and see
1783 whether any of them has the value we are looking for, we look
1784 at the reverse mapping, which tells us the set of expressions
1785 that have a given value (IE value->expressions with that
1786 value) and see if any of those expressions are in our set.
1787 The number of expressions per value is usually significantly
1788 less than the number of expressions in the set. In fact, for
1789 large testcases, doing it this way is roughly 5-10x faster
1790 than walking the bitmap.
1791 If this is somehow a significant lose for some cases, we can
1792 choose which set to walk based on which set is smaller. */
1793 unsigned int i;
1794 bitmap_iterator bi;
1795 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1797 EXECUTE_IF_AND_IN_BITMAP (exprset->expressions,
1798 set->expressions, 0, i, bi)
1800 pre_expr val = expression_for_id (i);
1801 /* At the point where stmt is not null, there should always
1802 be an SSA_NAME first in the list of expressions. */
1803 if (stmt)
1805 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1806 if (gimple_code (def_stmt) != GIMPLE_PHI
1807 && gimple_bb (def_stmt) == gimple_bb (stmt)
1808 && gimple_uid (def_stmt) >= gimple_uid (stmt))
1809 continue;
1811 return val;
1814 return NULL;
1817 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1818 BLOCK by seeing if it is not killed in the block. Note that we are
1819 only determining whether there is a store that kills it. Because
1820 of the order in which clean iterates over values, we are guaranteed
1821 that altered operands will have caused us to be eliminated from the
1822 ANTIC_IN set already. */
1824 static bool
1825 value_dies_in_block_x (pre_expr expr, basic_block block)
1827 int i;
1828 tree vuse;
1829 VEC (tree, gc) *vuses = PRE_EXPR_REFERENCE (expr)->vuses;
1831 /* Conservatively, a value dies if it's vuses are defined in this
1832 block, unless they come from phi nodes (which are merge operations,
1833 rather than stores. */
1834 for (i = 0; VEC_iterate (tree, vuses, i, vuse); i++)
1836 gimple def = SSA_NAME_DEF_STMT (vuse);
1838 if (gimple_bb (def) != block)
1839 continue;
1840 if (gimple_code (def) == GIMPLE_PHI)
1841 continue;
1842 return true;
1844 return false;
1848 #define union_contains_value(SET1, SET2, VAL) \
1849 (bitmap_set_contains_value ((SET1), (VAL)) \
1850 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
1852 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
1854 static bool
1855 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
1856 vn_reference_op_t vro)
1858 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
1860 struct pre_expr_d temp;
1861 temp.kind = NAME;
1862 temp.id = 0;
1863 PRE_EXPR_NAME (&temp) = vro->op0;
1864 temp.id = lookup_expression_id (&temp);
1865 if (temp.id == 0)
1866 return false;
1867 if (!union_contains_value (set1, set2,
1868 get_expr_value_id (&temp)))
1869 return false;
1871 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
1873 struct pre_expr_d temp;
1874 temp.kind = NAME;
1875 temp.id = 0;
1876 PRE_EXPR_NAME (&temp) = vro->op1;
1877 temp.id = lookup_expression_id (&temp);
1878 if (temp.id == 0)
1879 return false;
1880 if (!union_contains_value (set1, set2,
1881 get_expr_value_id (&temp)))
1882 return false;
1885 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
1887 struct pre_expr_d temp;
1888 temp.kind = NAME;
1889 temp.id = 0;
1890 PRE_EXPR_NAME (&temp) = vro->op2;
1891 temp.id = lookup_expression_id (&temp);
1892 if (temp.id == 0)
1893 return false;
1894 if (!union_contains_value (set1, set2,
1895 get_expr_value_id (&temp)))
1896 return false;
1899 return true;
1902 /* Determine if the expression EXPR is valid in SET1 U SET2.
1903 ONLY SET2 CAN BE NULL.
1904 This means that we have a leader for each part of the expression
1905 (if it consists of values), or the expression is an SSA_NAME.
1906 For loads/calls, we also see if the vuses are killed in this block.
1909 static bool
1910 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
1911 basic_block block)
1913 switch (expr->kind)
1915 case NAME:
1916 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
1917 case NARY:
1919 unsigned int i;
1920 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1921 for (i = 0; i < nary->length; i++)
1923 if (TREE_CODE (nary->op[i]) == SSA_NAME)
1925 struct pre_expr_d temp;
1926 temp.kind = NAME;
1927 temp.id = 0;
1928 PRE_EXPR_NAME (&temp) = nary->op[i];
1929 temp.id = lookup_expression_id (&temp);
1930 if (temp.id == 0)
1931 return false;
1932 if (!union_contains_value (set1, set2,
1933 get_expr_value_id (&temp)))
1934 return false;
1937 return true;
1939 break;
1940 case REFERENCE:
1942 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1943 vn_reference_op_t vro;
1944 unsigned int i;
1946 for (i = 0; VEC_iterate (vn_reference_op_s, ref->operands, i, vro); i++)
1948 if (!vro_valid_in_sets (set1, set2, vro))
1949 return false;
1951 return !value_dies_in_block_x (expr, block);
1953 default:
1954 gcc_unreachable ();
1958 /* Clean the set of expressions that are no longer valid in SET1 or
1959 SET2. This means expressions that are made up of values we have no
1960 leaders for in SET1 or SET2. This version is used for partial
1961 anticipation, which means it is not valid in either ANTIC_IN or
1962 PA_IN. */
1964 static void
1965 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
1967 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
1968 pre_expr expr;
1969 int i;
1971 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1973 if (!valid_in_sets (set1, set2, expr, block))
1974 bitmap_remove_from_set (set1, expr);
1976 VEC_free (pre_expr, heap, exprs);
1979 /* Clean the set of expressions that are no longer valid in SET. This
1980 means expressions that are made up of values we have no leaders for
1981 in SET. */
1983 static void
1984 clean (bitmap_set_t set, basic_block block)
1986 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
1987 pre_expr expr;
1988 int i;
1990 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1992 if (!valid_in_sets (set, NULL, expr, block))
1993 bitmap_remove_from_set (set, expr);
1995 VEC_free (pre_expr, heap, exprs);
1998 static sbitmap has_abnormal_preds;
2000 /* List of blocks that may have changed during ANTIC computation and
2001 thus need to be iterated over. */
2003 static sbitmap changed_blocks;
2005 /* Decide whether to defer a block for a later iteration, or PHI
2006 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2007 should defer the block, and true if we processed it. */
2009 static bool
2010 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2011 basic_block block, basic_block phiblock)
2013 if (!BB_VISITED (phiblock))
2015 SET_BIT (changed_blocks, block->index);
2016 BB_VISITED (block) = 0;
2017 BB_DEFERRED (block) = 1;
2018 return false;
2020 else
2021 phi_translate_set (dest, source, block, phiblock);
2022 return true;
2025 /* Compute the ANTIC set for BLOCK.
2027 If succs(BLOCK) > 1 then
2028 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2029 else if succs(BLOCK) == 1 then
2030 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2032 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2035 static bool
2036 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2038 bool changed = false;
2039 bitmap_set_t S, old, ANTIC_OUT;
2040 bitmap_iterator bi;
2041 unsigned int bii;
2042 edge e;
2043 edge_iterator ei;
2045 old = ANTIC_OUT = S = NULL;
2046 BB_VISITED (block) = 1;
2048 /* If any edges from predecessors are abnormal, antic_in is empty,
2049 so do nothing. */
2050 if (block_has_abnormal_pred_edge)
2051 goto maybe_dump_sets;
2053 old = ANTIC_IN (block);
2054 ANTIC_OUT = bitmap_set_new ();
2056 /* If the block has no successors, ANTIC_OUT is empty. */
2057 if (EDGE_COUNT (block->succs) == 0)
2059 /* If we have one successor, we could have some phi nodes to
2060 translate through. */
2061 else if (single_succ_p (block))
2063 basic_block succ_bb = single_succ (block);
2065 /* We trade iterations of the dataflow equations for having to
2066 phi translate the maximal set, which is incredibly slow
2067 (since the maximal set often has 300+ members, even when you
2068 have a small number of blocks).
2069 Basically, we defer the computation of ANTIC for this block
2070 until we have processed it's successor, which will inevitably
2071 have a *much* smaller set of values to phi translate once
2072 clean has been run on it.
2073 The cost of doing this is that we technically perform more
2074 iterations, however, they are lower cost iterations.
2076 Timings for PRE on tramp3d-v4:
2077 without maximal set fix: 11 seconds
2078 with maximal set fix/without deferring: 26 seconds
2079 with maximal set fix/with deferring: 11 seconds
2082 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2083 block, succ_bb))
2085 changed = true;
2086 goto maybe_dump_sets;
2089 /* If we have multiple successors, we take the intersection of all of
2090 them. Note that in the case of loop exit phi nodes, we may have
2091 phis to translate through. */
2092 else
2094 VEC(basic_block, heap) * worklist;
2095 size_t i;
2096 basic_block bprime, first;
2098 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2099 FOR_EACH_EDGE (e, ei, block->succs)
2100 VEC_quick_push (basic_block, worklist, e->dest);
2101 first = VEC_index (basic_block, worklist, 0);
2103 if (phi_nodes (first))
2105 bitmap_set_t from = ANTIC_IN (first);
2107 if (!BB_VISITED (first))
2108 from = maximal_set;
2109 phi_translate_set (ANTIC_OUT, from, block, first);
2111 else
2113 if (!BB_VISITED (first))
2114 bitmap_set_copy (ANTIC_OUT, maximal_set);
2115 else
2116 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2119 for (i = 1; VEC_iterate (basic_block, worklist, i, bprime); i++)
2121 if (phi_nodes (bprime))
2123 bitmap_set_t tmp = bitmap_set_new ();
2124 bitmap_set_t from = ANTIC_IN (bprime);
2126 if (!BB_VISITED (bprime))
2127 from = maximal_set;
2128 phi_translate_set (tmp, from, block, bprime);
2129 bitmap_set_and (ANTIC_OUT, tmp);
2130 bitmap_set_free (tmp);
2132 else
2134 if (!BB_VISITED (bprime))
2135 bitmap_set_and (ANTIC_OUT, maximal_set);
2136 else
2137 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2140 VEC_free (basic_block, heap, worklist);
2143 /* Generate ANTIC_OUT - TMP_GEN. */
2144 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2146 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2147 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2148 TMP_GEN (block));
2150 /* Then union in the ANTIC_OUT - TMP_GEN values,
2151 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2152 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2153 bitmap_value_insert_into_set (ANTIC_IN (block),
2154 expression_for_id (bii));
2156 clean (ANTIC_IN (block), block);
2158 /* !old->expressions can happen when we deferred a block. */
2159 if (!old->expressions || !bitmap_set_equal (old, ANTIC_IN (block)))
2161 changed = true;
2162 SET_BIT (changed_blocks, block->index);
2163 FOR_EACH_EDGE (e, ei, block->preds)
2164 SET_BIT (changed_blocks, e->src->index);
2166 else
2167 RESET_BIT (changed_blocks, block->index);
2169 maybe_dump_sets:
2170 if (dump_file && (dump_flags & TDF_DETAILS))
2172 if (!BB_DEFERRED (block) || BB_VISITED (block))
2174 if (ANTIC_OUT)
2175 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2177 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2178 block->index);
2180 if (S)
2181 print_bitmap_set (dump_file, S, "S", block->index);
2183 else
2185 fprintf (dump_file,
2186 "Block %d was deferred for a future iteration.\n",
2187 block->index);
2190 if (old)
2191 bitmap_set_free (old);
2192 if (S)
2193 bitmap_set_free (S);
2194 if (ANTIC_OUT)
2195 bitmap_set_free (ANTIC_OUT);
2196 return changed;
2199 /* Compute PARTIAL_ANTIC for BLOCK.
2201 If succs(BLOCK) > 1 then
2202 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2203 in ANTIC_OUT for all succ(BLOCK)
2204 else if succs(BLOCK) == 1 then
2205 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2207 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2208 - ANTIC_IN[BLOCK])
2211 static bool
2212 compute_partial_antic_aux (basic_block block,
2213 bool block_has_abnormal_pred_edge)
2215 bool changed = false;
2216 bitmap_set_t old_PA_IN;
2217 bitmap_set_t PA_OUT;
2218 edge e;
2219 edge_iterator ei;
2220 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2222 old_PA_IN = PA_OUT = NULL;
2224 /* If any edges from predecessors are abnormal, antic_in is empty,
2225 so do nothing. */
2226 if (block_has_abnormal_pred_edge)
2227 goto maybe_dump_sets;
2229 /* If there are too many partially anticipatable values in the
2230 block, phi_translate_set can take an exponential time: stop
2231 before the translation starts. */
2232 if (max_pa
2233 && single_succ_p (block)
2234 && bitmap_count_bits (PA_IN (single_succ (block))->values) > max_pa)
2235 goto maybe_dump_sets;
2237 old_PA_IN = PA_IN (block);
2238 PA_OUT = bitmap_set_new ();
2240 /* If the block has no successors, ANTIC_OUT is empty. */
2241 if (EDGE_COUNT (block->succs) == 0)
2243 /* If we have one successor, we could have some phi nodes to
2244 translate through. Note that we can't phi translate across DFS
2245 back edges in partial antic, because it uses a union operation on
2246 the successors. For recurrences like IV's, we will end up
2247 generating a new value in the set on each go around (i + 3 (VH.1)
2248 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2249 else if (single_succ_p (block))
2251 basic_block succ = single_succ (block);
2252 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2253 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2255 /* If we have multiple successors, we take the union of all of
2256 them. */
2257 else
2259 VEC(basic_block, heap) * worklist;
2260 size_t i;
2261 basic_block bprime;
2263 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2264 FOR_EACH_EDGE (e, ei, block->succs)
2266 if (e->flags & EDGE_DFS_BACK)
2267 continue;
2268 VEC_quick_push (basic_block, worklist, e->dest);
2270 if (VEC_length (basic_block, worklist) > 0)
2272 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2274 unsigned int i;
2275 bitmap_iterator bi;
2277 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2278 bitmap_value_insert_into_set (PA_OUT,
2279 expression_for_id (i));
2280 if (phi_nodes (bprime))
2282 bitmap_set_t pa_in = bitmap_set_new ();
2283 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2284 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2285 bitmap_value_insert_into_set (PA_OUT,
2286 expression_for_id (i));
2287 bitmap_set_free (pa_in);
2289 else
2290 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2291 bitmap_value_insert_into_set (PA_OUT,
2292 expression_for_id (i));
2295 VEC_free (basic_block, heap, worklist);
2298 /* PA_IN starts with PA_OUT - TMP_GEN.
2299 Then we subtract things from ANTIC_IN. */
2300 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2302 /* For partial antic, we want to put back in the phi results, since
2303 we will properly avoid making them partially antic over backedges. */
2304 bitmap_ior_into (PA_IN (block)->values, PHI_GEN (block)->values);
2305 bitmap_ior_into (PA_IN (block)->expressions, PHI_GEN (block)->expressions);
2307 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2308 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2310 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2312 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2314 changed = true;
2315 SET_BIT (changed_blocks, block->index);
2316 FOR_EACH_EDGE (e, ei, block->preds)
2317 SET_BIT (changed_blocks, e->src->index);
2319 else
2320 RESET_BIT (changed_blocks, block->index);
2322 maybe_dump_sets:
2323 if (dump_file && (dump_flags & TDF_DETAILS))
2325 if (PA_OUT)
2326 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2328 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2330 if (old_PA_IN)
2331 bitmap_set_free (old_PA_IN);
2332 if (PA_OUT)
2333 bitmap_set_free (PA_OUT);
2334 return changed;
2337 /* Compute ANTIC and partial ANTIC sets. */
2339 static void
2340 compute_antic (void)
2342 bool changed = true;
2343 int num_iterations = 0;
2344 basic_block block;
2345 int i;
2347 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2348 We pre-build the map of blocks with incoming abnormal edges here. */
2349 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2350 sbitmap_zero (has_abnormal_preds);
2352 FOR_EACH_BB (block)
2354 edge_iterator ei;
2355 edge e;
2357 FOR_EACH_EDGE (e, ei, block->preds)
2359 e->flags &= ~EDGE_DFS_BACK;
2360 if (e->flags & EDGE_ABNORMAL)
2362 SET_BIT (has_abnormal_preds, block->index);
2363 break;
2367 BB_VISITED (block) = 0;
2368 BB_DEFERRED (block) = 0;
2369 /* While we are here, give empty ANTIC_IN sets to each block. */
2370 ANTIC_IN (block) = bitmap_set_new ();
2371 PA_IN (block) = bitmap_set_new ();
2374 /* At the exit block we anticipate nothing. */
2375 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2376 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2377 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2379 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2380 sbitmap_ones (changed_blocks);
2381 while (changed)
2383 if (dump_file && (dump_flags & TDF_DETAILS))
2384 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2385 num_iterations++;
2386 changed = false;
2387 for (i = 0; i < last_basic_block - NUM_FIXED_BLOCKS; i++)
2389 if (TEST_BIT (changed_blocks, postorder[i]))
2391 basic_block block = BASIC_BLOCK (postorder[i]);
2392 changed |= compute_antic_aux (block,
2393 TEST_BIT (has_abnormal_preds,
2394 block->index));
2397 #ifdef ENABLE_CHECKING
2398 /* Theoretically possible, but *highly* unlikely. */
2399 gcc_assert (num_iterations < 500);
2400 #endif
2403 statistics_histogram_event (cfun, "compute_antic iterations",
2404 num_iterations);
2406 if (do_partial_partial)
2408 sbitmap_ones (changed_blocks);
2409 mark_dfs_back_edges ();
2410 num_iterations = 0;
2411 changed = true;
2412 while (changed)
2414 if (dump_file && (dump_flags & TDF_DETAILS))
2415 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2416 num_iterations++;
2417 changed = false;
2418 for (i = 0; i < last_basic_block - NUM_FIXED_BLOCKS; i++)
2420 if (TEST_BIT (changed_blocks, postorder[i]))
2422 basic_block block = BASIC_BLOCK (postorder[i]);
2423 changed
2424 |= compute_partial_antic_aux (block,
2425 TEST_BIT (has_abnormal_preds,
2426 block->index));
2429 #ifdef ENABLE_CHECKING
2430 /* Theoretically possible, but *highly* unlikely. */
2431 gcc_assert (num_iterations < 500);
2432 #endif
2434 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2435 num_iterations);
2437 sbitmap_free (has_abnormal_preds);
2438 sbitmap_free (changed_blocks);
2441 /* Return true if we can value number the call in STMT. This is true
2442 if we have a pure or constant call. */
2444 static bool
2445 can_value_number_call (gimple stmt)
2447 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2448 return true;
2449 return false;
2452 /* Return true if OP is an exception handler related operation, such as
2453 FILTER_EXPR or EXC_PTR_EXPR. */
2455 static bool
2456 is_exception_related (gimple stmt)
2458 return (is_gimple_assign (stmt)
2459 && (gimple_assign_rhs_code (stmt) == FILTER_EXPR
2460 || gimple_assign_rhs_code (stmt) == EXC_PTR_EXPR));
2463 /* Return true if OP is a tree which we can perform PRE on
2464 on. This may not match the operations we can value number, but in
2465 a perfect world would. */
2467 static bool
2468 can_PRE_operation (tree op)
2470 return UNARY_CLASS_P (op)
2471 || BINARY_CLASS_P (op)
2472 || COMPARISON_CLASS_P (op)
2473 || TREE_CODE (op) == INDIRECT_REF
2474 || TREE_CODE (op) == COMPONENT_REF
2475 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2476 || TREE_CODE (op) == CALL_EXPR
2477 || TREE_CODE (op) == ARRAY_REF;
2481 /* Inserted expressions are placed onto this worklist, which is used
2482 for performing quick dead code elimination of insertions we made
2483 that didn't turn out to be necessary. */
2484 static VEC(gimple,heap) *inserted_exprs;
2486 /* Pool allocated fake store expressions are placed onto this
2487 worklist, which, after performing dead code elimination, is walked
2488 to see which expressions need to be put into GC'able memory */
2489 static VEC(gimple, heap) *need_creation;
2491 /* The actual worker for create_component_ref_by_pieces. */
2493 static tree
2494 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2495 unsigned int *operand, gimple_seq *stmts,
2496 gimple domstmt)
2498 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2499 *operand);
2500 tree genop;
2501 ++*operand;
2502 switch (currop->opcode)
2504 case CALL_EXPR:
2506 tree folded, sc = currop->op1;
2507 unsigned int nargs = 0;
2508 tree *args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2509 ref->operands) - 1);
2510 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2512 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2513 operand, stmts,
2514 domstmt);
2515 nargs++;
2517 folded = build_call_array (currop->type,
2518 TREE_CODE (currop->op0) == FUNCTION_DECL
2519 ? build_fold_addr_expr (currop->op0)
2520 : currop->op0,
2521 nargs, args);
2522 free (args);
2523 if (sc)
2525 pre_expr scexpr = get_or_alloc_expr_for (sc);
2526 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2527 if (!sc)
2528 return NULL_TREE;
2529 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2531 return folded;
2533 break;
2534 case ADDR_EXPR:
2535 if (currop->op0)
2537 gcc_assert (is_gimple_min_invariant (currop->op0));
2538 return currop->op0;
2540 /* Fallthrough. */
2541 case REALPART_EXPR:
2542 case IMAGPART_EXPR:
2543 case VIEW_CONVERT_EXPR:
2545 tree folded;
2546 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2547 operand,
2548 stmts, domstmt);
2549 if (!genop0)
2550 return NULL_TREE;
2551 folded = fold_build1 (currop->opcode, currop->type,
2552 genop0);
2553 return folded;
2555 break;
2556 case ALIGN_INDIRECT_REF:
2557 case MISALIGNED_INDIRECT_REF:
2558 case INDIRECT_REF:
2560 tree folded;
2561 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2562 operand,
2563 stmts, domstmt);
2564 if (!genop1)
2565 return NULL_TREE;
2566 genop1 = fold_convert (build_pointer_type (currop->type),
2567 genop1);
2569 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2570 folded = fold_build2 (currop->opcode, currop->type,
2571 genop1, currop->op1);
2572 else
2573 folded = fold_build1 (currop->opcode, currop->type,
2574 genop1);
2575 return folded;
2577 break;
2578 case BIT_FIELD_REF:
2580 tree folded;
2581 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2582 stmts, domstmt);
2583 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2584 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2585 tree genop1;
2586 tree genop2;
2588 if (!genop0)
2589 return NULL_TREE;
2590 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2591 if (!genop1)
2592 return NULL_TREE;
2593 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2594 if (!genop2)
2595 return NULL_TREE;
2596 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2597 genop2);
2598 return folded;
2601 /* For array ref vn_reference_op's, operand 1 of the array ref
2602 is op0 of the reference op and operand 3 of the array ref is
2603 op1. */
2604 case ARRAY_RANGE_REF:
2605 case ARRAY_REF:
2607 tree genop0;
2608 tree genop1 = currop->op0;
2609 pre_expr op1expr;
2610 tree genop2 = currop->op1;
2611 pre_expr op2expr;
2612 tree genop3;
2613 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2614 stmts, domstmt);
2615 if (!genop0)
2616 return NULL_TREE;
2617 op1expr = get_or_alloc_expr_for (genop1);
2618 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2619 if (!genop1)
2620 return NULL_TREE;
2621 if (genop2)
2623 op2expr = get_or_alloc_expr_for (genop2);
2624 genop2 = find_or_generate_expression (block, op2expr, stmts,
2625 domstmt);
2626 if (!genop2)
2627 return NULL_TREE;
2630 genop3 = currop->op2;
2631 return build4 (currop->opcode, currop->type, genop0, genop1,
2632 genop2, genop3);
2634 case COMPONENT_REF:
2636 tree op0;
2637 tree op1;
2638 tree genop2 = currop->op1;
2639 pre_expr op2expr;
2640 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2641 stmts, domstmt);
2642 if (!op0)
2643 return NULL_TREE;
2644 /* op1 should be a FIELD_DECL, which are represented by
2645 themselves. */
2646 op1 = currop->op0;
2647 if (genop2)
2649 op2expr = get_or_alloc_expr_for (genop2);
2650 genop2 = find_or_generate_expression (block, op2expr, stmts,
2651 domstmt);
2652 if (!genop2)
2653 return NULL_TREE;
2656 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2657 genop2);
2659 break;
2660 case SSA_NAME:
2662 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2663 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2664 return genop;
2666 case STRING_CST:
2667 case INTEGER_CST:
2668 case COMPLEX_CST:
2669 case VECTOR_CST:
2670 case REAL_CST:
2671 case CONSTRUCTOR:
2672 case VAR_DECL:
2673 case PARM_DECL:
2674 case CONST_DECL:
2675 case RESULT_DECL:
2676 case FUNCTION_DECL:
2677 return currop->op0;
2679 default:
2680 gcc_unreachable ();
2684 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2685 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2686 trying to rename aggregates into ssa form directly, which is a no no.
2688 Thus, this routine doesn't create temporaries, it just builds a
2689 single access expression for the array, calling
2690 find_or_generate_expression to build the innermost pieces.
2692 This function is a subroutine of create_expression_by_pieces, and
2693 should not be called on it's own unless you really know what you
2694 are doing. */
2696 static tree
2697 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2698 gimple_seq *stmts, gimple domstmt)
2700 unsigned int op = 0;
2701 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2704 /* Find a leader for an expression, or generate one using
2705 create_expression_by_pieces if it's ANTIC but
2706 complex.
2707 BLOCK is the basic_block we are looking for leaders in.
2708 EXPR is the expression to find a leader or generate for.
2709 STMTS is the statement list to put the inserted expressions on.
2710 Returns the SSA_NAME of the LHS of the generated expression or the
2711 leader.
2712 DOMSTMT if non-NULL is a statement that should be dominated by
2713 all uses in the generated expression. If DOMSTMT is non-NULL this
2714 routine can fail and return NULL_TREE. Otherwise it will assert
2715 on failure. */
2717 static tree
2718 find_or_generate_expression (basic_block block, pre_expr expr,
2719 gimple_seq *stmts, gimple domstmt)
2721 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2722 get_expr_value_id (expr), domstmt);
2723 tree genop = NULL;
2724 if (leader)
2726 if (leader->kind == NAME)
2727 genop = PRE_EXPR_NAME (leader);
2728 else if (leader->kind == CONSTANT)
2729 genop = PRE_EXPR_CONSTANT (leader);
2732 /* If it's still NULL, it must be a complex expression, so generate
2733 it recursively. */
2734 if (genop == NULL)
2736 bitmap_set_t exprset;
2737 unsigned int lookfor = get_expr_value_id (expr);
2738 bool handled = false;
2739 bitmap_iterator bi;
2740 unsigned int i;
2742 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2743 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2745 pre_expr temp = expression_for_id (i);
2746 if (temp->kind != NAME)
2748 handled = true;
2749 genop = create_expression_by_pieces (block, temp, stmts,
2750 domstmt,
2751 get_expr_type (expr));
2752 break;
2755 if (!handled && domstmt)
2756 return NULL_TREE;
2758 gcc_assert (handled);
2760 return genop;
2763 #define NECESSARY GF_PLF_1
2765 /* Create an expression in pieces, so that we can handle very complex
2766 expressions that may be ANTIC, but not necessary GIMPLE.
2767 BLOCK is the basic block the expression will be inserted into,
2768 EXPR is the expression to insert (in value form)
2769 STMTS is a statement list to append the necessary insertions into.
2771 This function will die if we hit some value that shouldn't be
2772 ANTIC but is (IE there is no leader for it, or its components).
2773 This function may also generate expressions that are themselves
2774 partially or fully redundant. Those that are will be either made
2775 fully redundant during the next iteration of insert (for partially
2776 redundant ones), or eliminated by eliminate (for fully redundant
2777 ones).
2779 If DOMSTMT is non-NULL then we make sure that all uses in the
2780 expressions dominate that statement. In this case the function
2781 can return NULL_TREE to signal failure. */
2783 static tree
2784 create_expression_by_pieces (basic_block block, pre_expr expr,
2785 gimple_seq *stmts, gimple domstmt, tree type)
2787 tree temp, name;
2788 tree folded, newexpr;
2789 gimple_seq forced_stmts;
2790 unsigned int value_id;
2791 gimple_stmt_iterator gsi;
2792 tree exprtype = type ? type : get_expr_type (expr);
2793 pre_expr nameexpr;
2794 gimple newstmt;
2796 switch (expr->kind)
2798 /* We may hit the NAME/CONSTANT case if we have to convert types
2799 that value numbering saw through. */
2800 case NAME:
2801 folded = PRE_EXPR_NAME (expr);
2802 break;
2803 case CONSTANT:
2804 folded = PRE_EXPR_CONSTANT (expr);
2805 break;
2806 case REFERENCE:
2808 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2809 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
2811 break;
2812 case NARY:
2814 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2815 switch (nary->length)
2817 case 2:
2819 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2820 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
2821 tree genop1 = find_or_generate_expression (block, op1,
2822 stmts, domstmt);
2823 tree genop2 = find_or_generate_expression (block, op2,
2824 stmts, domstmt);
2825 if (!genop1 || !genop2)
2826 return NULL_TREE;
2827 genop1 = fold_convert (TREE_TYPE (nary->op[0]),
2828 genop1);
2829 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
2830 may be a constant with the wrong type. */
2831 if (nary->opcode == POINTER_PLUS_EXPR)
2832 genop2 = fold_convert (sizetype, genop2);
2833 else
2834 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
2836 folded = fold_build2 (nary->opcode, nary->type,
2837 genop1, genop2);
2839 break;
2840 case 1:
2842 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2843 tree genop1 = find_or_generate_expression (block, op1,
2844 stmts, domstmt);
2845 if (!genop1)
2846 return NULL_TREE;
2847 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
2849 folded = fold_build1 (nary->opcode, nary->type,
2850 genop1);
2852 break;
2853 default:
2854 return NULL_TREE;
2857 break;
2858 default:
2859 return NULL_TREE;
2861 folded = fold_convert (exprtype, folded);
2862 /* Force the generated expression to be a sequence of GIMPLE
2863 statements.
2864 We have to call unshare_expr because force_gimple_operand may
2865 modify the tree we pass to it. */
2866 newexpr = force_gimple_operand (unshare_expr (folded), &forced_stmts,
2867 false, NULL);
2869 /* If we have any intermediate expressions to the value sets, add them
2870 to the value sets and chain them in the instruction stream. */
2871 if (forced_stmts)
2873 gsi = gsi_start (forced_stmts);
2874 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2876 gimple stmt = gsi_stmt (gsi);
2877 tree forcedname = gimple_get_lhs (stmt);
2878 pre_expr nameexpr;
2880 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
2881 if (TREE_CODE (forcedname) == SSA_NAME)
2883 VN_INFO_GET (forcedname)->valnum = forcedname;
2884 VN_INFO (forcedname)->value_id = get_next_value_id ();
2885 nameexpr = get_or_alloc_expr_for_name (forcedname);
2886 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2887 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2888 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2890 mark_symbols_for_renaming (stmt);
2892 gimple_seq_add_seq (stmts, forced_stmts);
2895 /* Build and insert the assignment of the end result to the temporary
2896 that we will return. */
2897 if (!pretemp || exprtype != TREE_TYPE (pretemp))
2899 pretemp = create_tmp_var (exprtype, "pretmp");
2900 get_var_ann (pretemp);
2903 temp = pretemp;
2904 add_referenced_var (temp);
2906 if (TREE_CODE (exprtype) == COMPLEX_TYPE
2907 || TREE_CODE (exprtype) == VECTOR_TYPE)
2908 DECL_GIMPLE_REG_P (temp) = 1;
2910 newstmt = gimple_build_assign (temp, newexpr);
2911 name = make_ssa_name (temp, newstmt);
2912 gimple_assign_set_lhs (newstmt, name);
2913 gimple_set_plf (newstmt, NECESSARY, false);
2915 gimple_seq_add_stmt (stmts, newstmt);
2916 VEC_safe_push (gimple, heap, inserted_exprs, newstmt);
2918 /* All the symbols in NEWEXPR should be put into SSA form. */
2919 mark_symbols_for_renaming (newstmt);
2921 /* Add a value number to the temporary.
2922 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2923 we are creating the expression by pieces, and this particular piece of
2924 the expression may have been represented. There is no harm in replacing
2925 here. */
2926 VN_INFO_GET (name)->valnum = name;
2927 value_id = get_expr_value_id (expr);
2928 VN_INFO (name)->value_id = value_id;
2929 nameexpr = get_or_alloc_expr_for_name (name);
2930 add_to_value (value_id, nameexpr);
2931 if (!in_fre)
2932 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2933 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2935 pre_stats.insertions++;
2936 if (dump_file && (dump_flags & TDF_DETAILS))
2938 fprintf (dump_file, "Inserted ");
2939 print_gimple_stmt (dump_file, newstmt, 0, 0);
2940 fprintf (dump_file, " in predecessor %d\n", block->index);
2943 return name;
2947 /* Insert the to-be-made-available values of expression EXPRNUM for each
2948 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
2949 merge the result with a phi node, given the same value number as
2950 NODE. Return true if we have inserted new stuff. */
2952 static bool
2953 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
2954 pre_expr *avail)
2956 pre_expr expr = expression_for_id (exprnum);
2957 pre_expr newphi;
2958 unsigned int val = get_expr_value_id (expr);
2959 edge pred;
2960 bool insertions = false;
2961 bool nophi = false;
2962 basic_block bprime;
2963 pre_expr eprime;
2964 edge_iterator ei;
2965 tree type = get_expr_type (expr);
2966 tree temp, res;
2967 gimple phi;
2969 if (dump_file && (dump_flags & TDF_DETAILS))
2971 fprintf (dump_file, "Found partial redundancy for expression ");
2972 print_pre_expr (dump_file, expr);
2973 fprintf (dump_file, " (%04d)\n", val);
2976 /* Make sure we aren't creating an induction variable. */
2977 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2
2978 && expr->kind != REFERENCE)
2980 bool firstinsideloop = false;
2981 bool secondinsideloop = false;
2982 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
2983 EDGE_PRED (block, 0)->src);
2984 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
2985 EDGE_PRED (block, 1)->src);
2986 /* Induction variables only have one edge inside the loop. */
2987 if (firstinsideloop ^ secondinsideloop)
2989 if (dump_file && (dump_flags & TDF_DETAILS))
2990 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
2991 nophi = true;
2996 /* Make the necessary insertions. */
2997 FOR_EACH_EDGE (pred, ei, block->preds)
2999 gimple_seq stmts = NULL;
3000 tree builtexpr;
3001 bprime = pred->src;
3002 eprime = avail[bprime->index];
3004 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3006 builtexpr = create_expression_by_pieces (bprime,
3007 eprime,
3008 &stmts, NULL,
3009 type);
3010 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3011 gsi_insert_seq_on_edge (pred, stmts);
3012 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3013 insertions = true;
3015 else if (eprime->kind == CONSTANT)
3017 /* Constants may not have the right type, fold_convert
3018 should give us back a constant with the right type.
3020 tree constant = PRE_EXPR_CONSTANT (eprime);
3021 if (TREE_TYPE (constant) != type)
3023 tree builtexpr = fold_convert (type, constant);
3024 if (is_gimple_min_invariant (builtexpr))
3026 PRE_EXPR_CONSTANT (eprime) = builtexpr;
3028 else
3030 tree forcedexpr = force_gimple_operand (builtexpr,
3031 &stmts, true,
3032 NULL);
3033 if (is_gimple_min_invariant (forcedexpr))
3035 PRE_EXPR_CONSTANT (eprime) = forcedexpr;
3037 else
3039 if (forcedexpr != builtexpr)
3041 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3042 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3044 if (stmts)
3046 gimple_stmt_iterator gsi;
3047 gsi = gsi_start (stmts);
3048 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3050 gimple stmt = gsi_stmt (gsi);
3051 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3052 gimple_set_plf (stmt, NECESSARY, false);
3054 gsi_insert_seq_on_edge (pred, stmts);
3056 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3061 else if (eprime->kind == NAME)
3063 /* We may have to do a conversion because our value
3064 numbering can look through types in certain cases, but
3065 our IL requires all operands of a phi node have the same
3066 type. */
3067 tree name = PRE_EXPR_NAME (eprime);
3068 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3070 tree builtexpr;
3071 tree forcedexpr;
3072 builtexpr = fold_convert (type, name);
3073 forcedexpr = force_gimple_operand (builtexpr,
3074 &stmts, true,
3075 NULL);
3077 if (forcedexpr != name)
3079 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3080 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3083 if (stmts)
3085 gimple_stmt_iterator gsi;
3086 gsi = gsi_start (stmts);
3087 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3089 gimple stmt = gsi_stmt (gsi);
3090 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3091 gimple_set_plf (stmt, NECESSARY, false);
3093 gsi_insert_seq_on_edge (pred, stmts);
3095 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3099 /* If we didn't want a phi node, and we made insertions, we still have
3100 inserted new stuff, and thus return true. If we didn't want a phi node,
3101 and didn't make insertions, we haven't added anything new, so return
3102 false. */
3103 if (nophi && insertions)
3104 return true;
3105 else if (nophi && !insertions)
3106 return false;
3108 /* Now build a phi for the new variable. */
3109 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3111 prephitemp = create_tmp_var (type, "prephitmp");
3112 get_var_ann (prephitemp);
3115 temp = prephitemp;
3116 add_referenced_var (temp);
3118 if (TREE_CODE (type) == COMPLEX_TYPE
3119 || TREE_CODE (type) == VECTOR_TYPE)
3120 DECL_GIMPLE_REG_P (temp) = 1;
3122 phi = create_phi_node (temp, block);
3123 FOR_EACH_EDGE (pred, ei, block->preds)
3125 pre_expr ae = avail[pred->src->index];
3126 gcc_assert (get_expr_type (ae) == type
3127 || useless_type_conversion_p (type, get_expr_type (ae)));
3128 if (ae->kind == CONSTANT)
3129 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred);
3130 else
3131 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred);
3133 /* If the PHI node is already available, use it. */
3134 if ((res = vn_phi_lookup (phi)) != NULL_TREE)
3136 gimple_stmt_iterator gsi = gsi_for_stmt (phi);
3137 remove_phi_node (&gsi, true);
3138 release_defs (phi);
3139 add_to_value (val, get_or_alloc_expr_for_name (res));
3140 return false;
3143 gimple_set_plf (phi, NECESSARY, false);
3144 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3145 VN_INFO (gimple_phi_result (phi))->value_id = val;
3146 VEC_safe_push (gimple, heap, inserted_exprs, phi);
3148 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3149 add_to_value (val, newphi);
3151 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3152 this insertion, since we test for the existence of this value in PHI_GEN
3153 before proceeding with the partial redundancy checks in insert_aux.
3155 The value may exist in AVAIL_OUT, in particular, it could be represented
3156 by the expression we are trying to eliminate, in which case we want the
3157 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3158 inserted there.
3160 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3161 this block, because if it did, it would have existed in our dominator's
3162 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3165 bitmap_insert_into_set (PHI_GEN (block), newphi);
3166 bitmap_value_replace_in_set (AVAIL_OUT (block),
3167 newphi);
3168 bitmap_insert_into_set (NEW_SETS (block),
3169 newphi);
3171 if (dump_file && (dump_flags & TDF_DETAILS))
3173 fprintf (dump_file, "Created phi ");
3174 print_gimple_stmt (dump_file, phi, 0, 0);
3175 fprintf (dump_file, " in block %d\n", block->index);
3177 pre_stats.phis++;
3178 return true;
3183 /* Perform insertion of partially redundant values.
3184 For BLOCK, do the following:
3185 1. Propagate the NEW_SETS of the dominator into the current block.
3186 If the block has multiple predecessors,
3187 2a. Iterate over the ANTIC expressions for the block to see if
3188 any of them are partially redundant.
3189 2b. If so, insert them into the necessary predecessors to make
3190 the expression fully redundant.
3191 2c. Insert a new PHI merging the values of the predecessors.
3192 2d. Insert the new PHI, and the new expressions, into the
3193 NEW_SETS set.
3194 3. Recursively call ourselves on the dominator children of BLOCK.
3196 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3197 do_regular_insertion and do_partial_insertion.
3201 static bool
3202 do_regular_insertion (basic_block block, basic_block dom)
3204 bool new_stuff = false;
3205 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3206 pre_expr expr;
3207 int i;
3209 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3211 if (expr->kind != NAME)
3213 pre_expr *avail;
3214 unsigned int val;
3215 bool by_some = false;
3216 bool cant_insert = false;
3217 bool all_same = true;
3218 pre_expr first_s = NULL;
3219 edge pred;
3220 basic_block bprime;
3221 pre_expr eprime = NULL;
3222 edge_iterator ei;
3223 pre_expr edoubleprime;
3225 val = get_expr_value_id (expr);
3226 if (bitmap_set_contains_value (PHI_GEN (block), val))
3227 continue;
3228 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3230 if (dump_file && (dump_flags & TDF_DETAILS))
3231 fprintf (dump_file, "Found fully redundant value\n");
3232 continue;
3235 avail = XCNEWVEC (pre_expr, last_basic_block);
3236 FOR_EACH_EDGE (pred, ei, block->preds)
3238 unsigned int vprime;
3240 /* We should never run insertion for the exit block
3241 and so not come across fake pred edges. */
3242 gcc_assert (!(pred->flags & EDGE_FAKE));
3243 bprime = pred->src;
3244 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3245 bprime, block);
3247 /* eprime will generally only be NULL if the
3248 value of the expression, translated
3249 through the PHI for this predecessor, is
3250 undefined. If that is the case, we can't
3251 make the expression fully redundant,
3252 because its value is undefined along a
3253 predecessor path. We can thus break out
3254 early because it doesn't matter what the
3255 rest of the results are. */
3256 if (eprime == NULL)
3258 cant_insert = true;
3259 break;
3262 eprime = fully_constant_expression (eprime);
3263 vprime = get_expr_value_id (eprime);
3264 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3265 vprime, NULL);
3266 if (edoubleprime == NULL)
3268 avail[bprime->index] = eprime;
3269 all_same = false;
3271 else
3273 avail[bprime->index] = edoubleprime;
3274 by_some = true;
3275 if (first_s == NULL)
3276 first_s = edoubleprime;
3277 else if (!pre_expr_eq (first_s, edoubleprime))
3278 all_same = false;
3281 /* If we can insert it, it's not the same value
3282 already existing along every predecessor, and
3283 it's defined by some predecessor, it is
3284 partially redundant. */
3285 if (!cant_insert && !all_same && by_some && dbg_cnt (treepre_insert))
3287 if (insert_into_preds_of_block (block, get_expression_id (expr),
3288 avail))
3289 new_stuff = true;
3291 /* If all edges produce the same value and that value is
3292 an invariant, then the PHI has the same value on all
3293 edges. Note this. */
3294 else if (!cant_insert && all_same && eprime
3295 && (edoubleprime->kind == CONSTANT
3296 || edoubleprime->kind == NAME)
3297 && !value_id_constant_p (val))
3299 unsigned int j;
3300 bitmap_iterator bi;
3301 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3302 value_expressions, val);
3304 unsigned int new_val = get_expr_value_id (edoubleprime);
3305 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3307 pre_expr expr = expression_for_id (j);
3309 if (expr->kind == NAME)
3311 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3312 /* Just reset the value id and valnum so it is
3313 the same as the constant we have discovered. */
3314 if (edoubleprime->kind == CONSTANT)
3316 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3317 pre_stats.constified++;
3319 else
3320 info->valnum = PRE_EXPR_NAME (edoubleprime);
3321 info->value_id = new_val;
3325 free (avail);
3329 VEC_free (pre_expr, heap, exprs);
3330 return new_stuff;
3334 /* Perform insertion for partially anticipatable expressions. There
3335 is only one case we will perform insertion for these. This case is
3336 if the expression is partially anticipatable, and fully available.
3337 In this case, we know that putting it earlier will enable us to
3338 remove the later computation. */
3341 static bool
3342 do_partial_partial_insertion (basic_block block, basic_block dom)
3344 bool new_stuff = false;
3345 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3346 pre_expr expr;
3347 int i;
3349 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3351 if (expr->kind != NAME)
3353 pre_expr *avail;
3354 unsigned int val;
3355 bool by_all = true;
3356 bool cant_insert = false;
3357 edge pred;
3358 basic_block bprime;
3359 pre_expr eprime = NULL;
3360 edge_iterator ei;
3362 val = get_expr_value_id (expr);
3363 if (bitmap_set_contains_value (PHI_GEN (block), val))
3364 continue;
3365 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3366 continue;
3368 avail = XCNEWVEC (pre_expr, last_basic_block);
3369 FOR_EACH_EDGE (pred, ei, block->preds)
3371 unsigned int vprime;
3372 pre_expr edoubleprime;
3374 /* We should never run insertion for the exit block
3375 and so not come across fake pred edges. */
3376 gcc_assert (!(pred->flags & EDGE_FAKE));
3377 bprime = pred->src;
3378 eprime = phi_translate (expr, ANTIC_IN (block),
3379 PA_IN (block),
3380 bprime, block);
3382 /* eprime will generally only be NULL if the
3383 value of the expression, translated
3384 through the PHI for this predecessor, is
3385 undefined. If that is the case, we can't
3386 make the expression fully redundant,
3387 because its value is undefined along a
3388 predecessor path. We can thus break out
3389 early because it doesn't matter what the
3390 rest of the results are. */
3391 if (eprime == NULL)
3393 cant_insert = true;
3394 break;
3397 eprime = fully_constant_expression (eprime);
3398 vprime = get_expr_value_id (eprime);
3399 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3400 vprime, NULL);
3401 if (edoubleprime == NULL)
3403 by_all = false;
3404 break;
3406 else
3407 avail[bprime->index] = edoubleprime;
3411 /* If we can insert it, it's not the same value
3412 already existing along every predecessor, and
3413 it's defined by some predecessor, it is
3414 partially redundant. */
3415 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3417 pre_stats.pa_insert++;
3418 if (insert_into_preds_of_block (block, get_expression_id (expr),
3419 avail))
3420 new_stuff = true;
3422 free (avail);
3426 VEC_free (pre_expr, heap, exprs);
3427 return new_stuff;
3430 static bool
3431 insert_aux (basic_block block)
3433 basic_block son;
3434 bool new_stuff = false;
3436 if (block)
3438 basic_block dom;
3439 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3440 if (dom)
3442 unsigned i;
3443 bitmap_iterator bi;
3444 bitmap_set_t newset = NEW_SETS (dom);
3445 if (newset)
3447 /* Note that we need to value_replace both NEW_SETS, and
3448 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3449 represented by some non-simple expression here that we want
3450 to replace it with. */
3451 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3453 pre_expr expr = expression_for_id (i);
3454 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3455 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3458 if (!single_pred_p (block))
3460 new_stuff |= do_regular_insertion (block, dom);
3461 if (do_partial_partial)
3462 new_stuff |= do_partial_partial_insertion (block, dom);
3466 for (son = first_dom_son (CDI_DOMINATORS, block);
3467 son;
3468 son = next_dom_son (CDI_DOMINATORS, son))
3470 new_stuff |= insert_aux (son);
3473 return new_stuff;
3476 /* Perform insertion of partially redundant values. */
3478 static void
3479 insert (void)
3481 bool new_stuff = true;
3482 basic_block bb;
3483 int num_iterations = 0;
3485 FOR_ALL_BB (bb)
3486 NEW_SETS (bb) = bitmap_set_new ();
3488 while (new_stuff)
3490 num_iterations++;
3491 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3493 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3497 /* Add OP to EXP_GEN (block), and possibly to the maximal set if it is
3498 not defined by a phi node.
3499 PHI nodes can't go in the maximal sets because they are not in
3500 TMP_GEN, so it is possible to get into non-monotonic situations
3501 during ANTIC calculation, because it will *add* bits. */
3503 static void
3504 add_to_exp_gen (basic_block block, tree op)
3506 if (!in_fre)
3508 pre_expr result;
3509 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3510 return;
3511 result = get_or_alloc_expr_for_name (op);
3512 bitmap_value_insert_into_set (EXP_GEN (block), result);
3513 if (TREE_CODE (op) != SSA_NAME
3514 || gimple_code (SSA_NAME_DEF_STMT (op)) != GIMPLE_PHI)
3515 bitmap_value_insert_into_set (maximal_set, result);
3519 /* Create value ids for PHI in BLOCK. */
3521 static void
3522 make_values_for_phi (gimple phi, basic_block block)
3524 tree result = gimple_phi_result (phi);
3526 /* We have no need for virtual phis, as they don't represent
3527 actual computations. */
3528 if (is_gimple_reg (result))
3530 pre_expr e = get_or_alloc_expr_for_name (result);
3531 add_to_value (get_expr_value_id (e), e);
3532 bitmap_insert_into_set (PHI_GEN (block), e);
3533 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3537 /* Compute the AVAIL set for all basic blocks.
3539 This function performs value numbering of the statements in each basic
3540 block. The AVAIL sets are built from information we glean while doing
3541 this value numbering, since the AVAIL sets contain only one entry per
3542 value.
3544 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3545 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3547 static void
3548 compute_avail (void)
3551 basic_block block, son;
3552 basic_block *worklist;
3553 size_t sp = 0;
3554 tree param;
3556 /* For arguments with default definitions, we pretend they are
3557 defined in the entry block. */
3558 for (param = DECL_ARGUMENTS (current_function_decl);
3559 param;
3560 param = TREE_CHAIN (param))
3562 if (gimple_default_def (cfun, param) != NULL)
3564 tree def = gimple_default_def (cfun, param);
3565 pre_expr e = get_or_alloc_expr_for_name (def);
3567 add_to_value (get_expr_value_id (e), e);
3568 if (!in_fre)
3570 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3571 bitmap_value_insert_into_set (maximal_set, e);
3573 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3577 /* Likewise for the static chain decl. */
3578 if (cfun->static_chain_decl)
3580 param = cfun->static_chain_decl;
3581 if (gimple_default_def (cfun, param) != NULL)
3583 tree def = gimple_default_def (cfun, param);
3584 pre_expr e = get_or_alloc_expr_for_name (def);
3586 add_to_value (get_expr_value_id (e), e);
3587 if (!in_fre)
3589 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3590 bitmap_value_insert_into_set (maximal_set, e);
3592 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3596 /* Allocate the worklist. */
3597 worklist = XNEWVEC (basic_block, n_basic_blocks);
3599 /* Seed the algorithm by putting the dominator children of the entry
3600 block on the worklist. */
3601 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3602 son;
3603 son = next_dom_son (CDI_DOMINATORS, son))
3604 worklist[sp++] = son;
3606 /* Loop until the worklist is empty. */
3607 while (sp)
3609 gimple_stmt_iterator gsi;
3610 gimple stmt;
3611 basic_block dom;
3612 unsigned int stmt_uid = 1;
3614 /* Pick a block from the worklist. */
3615 block = worklist[--sp];
3617 /* Initially, the set of available values in BLOCK is that of
3618 its immediate dominator. */
3619 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3620 if (dom)
3621 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3623 /* Generate values for PHI nodes. */
3624 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3625 make_values_for_phi (gsi_stmt (gsi), block);
3627 /* Now compute value numbers and populate value sets with all
3628 the expressions computed in BLOCK. */
3629 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3631 ssa_op_iter iter;
3632 tree op;
3634 stmt = gsi_stmt (gsi);
3635 gimple_set_uid (stmt, stmt_uid++);
3637 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3639 pre_expr e = get_or_alloc_expr_for_name (op);
3641 add_to_value (get_expr_value_id (e), e);
3642 if (!in_fre)
3644 bitmap_insert_into_set (TMP_GEN (block), e);
3645 bitmap_value_insert_into_set (maximal_set, e);
3647 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3650 if (gimple_has_volatile_ops (stmt)
3651 || stmt_could_throw_p (stmt))
3652 continue;
3654 switch (gimple_code (stmt))
3656 case GIMPLE_RETURN:
3657 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3658 add_to_exp_gen (block, op);
3659 continue;
3661 case GIMPLE_CALL:
3663 vn_reference_t ref;
3664 unsigned int i;
3665 vn_reference_op_t vro;
3666 pre_expr result = NULL;
3667 VEC(vn_reference_op_s, heap) *ops = NULL;
3669 if (!can_value_number_call (stmt))
3670 continue;
3672 copy_reference_ops_from_call (stmt, &ops);
3673 vn_reference_lookup_pieces (shared_vuses_from_stmt (stmt),
3674 ops, &ref, false);
3675 VEC_free (vn_reference_op_s, heap, ops);
3676 if (!ref)
3677 continue;
3679 for (i = 0; VEC_iterate (vn_reference_op_s,
3680 ref->operands, i,
3681 vro); i++)
3683 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3684 add_to_exp_gen (block, vro->op0);
3685 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3686 add_to_exp_gen (block, vro->op1);
3687 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3688 add_to_exp_gen (block, vro->op2);
3690 result = (pre_expr) pool_alloc (pre_expr_pool);
3691 result->kind = REFERENCE;
3692 result->id = 0;
3693 PRE_EXPR_REFERENCE (result) = ref;
3695 get_or_alloc_expression_id (result);
3696 add_to_value (get_expr_value_id (result), result);
3697 if (!in_fre)
3699 bitmap_value_insert_into_set (EXP_GEN (block),
3700 result);
3701 bitmap_value_insert_into_set (maximal_set, result);
3703 continue;
3706 case GIMPLE_ASSIGN:
3708 pre_expr result = NULL;
3709 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
3711 case tcc_unary:
3712 if (is_exception_related (stmt))
3713 continue;
3714 case tcc_binary:
3716 vn_nary_op_t nary;
3717 unsigned int i;
3719 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
3720 gimple_assign_rhs_code (stmt),
3721 gimple_expr_type (stmt),
3722 gimple_assign_rhs1 (stmt),
3723 gimple_assign_rhs2 (stmt),
3724 NULL_TREE, NULL_TREE, &nary);
3726 if (!nary)
3727 continue;
3729 for (i = 0; i < nary->length; i++)
3730 if (TREE_CODE (nary->op[i]) == SSA_NAME)
3731 add_to_exp_gen (block, nary->op[i]);
3733 result = (pre_expr) pool_alloc (pre_expr_pool);
3734 result->kind = NARY;
3735 result->id = 0;
3736 PRE_EXPR_NARY (result) = nary;
3737 break;
3740 case tcc_declaration:
3741 case tcc_reference:
3743 vn_reference_t ref;
3744 unsigned int i;
3745 vn_reference_op_t vro;
3747 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3748 shared_vuses_from_stmt (stmt),
3749 false, &ref);
3750 if (!ref)
3751 continue;
3753 for (i = 0; VEC_iterate (vn_reference_op_s,
3754 ref->operands, i,
3755 vro); i++)
3757 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3758 add_to_exp_gen (block, vro->op0);
3759 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3760 add_to_exp_gen (block, vro->op1);
3761 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3762 add_to_exp_gen (block, vro->op2);
3764 result = (pre_expr) pool_alloc (pre_expr_pool);
3765 result->kind = REFERENCE;
3766 result->id = 0;
3767 PRE_EXPR_REFERENCE (result) = ref;
3768 break;
3771 default:
3772 /* For any other statement that we don't
3773 recognize, simply add all referenced
3774 SSA_NAMEs to EXP_GEN. */
3775 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3776 add_to_exp_gen (block, op);
3777 continue;
3780 get_or_alloc_expression_id (result);
3781 add_to_value (get_expr_value_id (result), result);
3782 if (!in_fre)
3784 bitmap_value_insert_into_set (EXP_GEN (block), result);
3785 bitmap_value_insert_into_set (maximal_set, result);
3788 continue;
3790 default:
3791 break;
3795 /* Put the dominator children of BLOCK on the worklist of blocks
3796 to compute available sets for. */
3797 for (son = first_dom_son (CDI_DOMINATORS, block);
3798 son;
3799 son = next_dom_son (CDI_DOMINATORS, son))
3800 worklist[sp++] = son;
3803 free (worklist);
3806 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
3807 than the available expressions for it. The insertion point is
3808 right before the first use in STMT. Returns the SSA_NAME that should
3809 be used for replacement. */
3811 static tree
3812 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
3814 basic_block bb = gimple_bb (stmt);
3815 gimple_stmt_iterator gsi;
3816 gimple_seq stmts = NULL;
3817 tree expr;
3818 pre_expr e;
3820 /* First create a value expression from the expression we want
3821 to insert and associate it with the value handle for SSA_VN. */
3822 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
3823 if (e == NULL)
3824 return NULL_TREE;
3826 /* Then use create_expression_by_pieces to generate a valid
3827 expression to insert at this point of the IL stream. */
3828 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
3829 if (expr == NULL_TREE)
3830 return NULL_TREE;
3831 gsi = gsi_for_stmt (stmt);
3832 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
3834 return expr;
3837 /* Eliminate fully redundant computations. */
3839 static unsigned int
3840 eliminate (void)
3842 basic_block b;
3843 unsigned int todo = 0;
3845 FOR_EACH_BB (b)
3847 gimple_stmt_iterator i;
3849 for (i = gsi_start_bb (b); !gsi_end_p (i); gsi_next (&i))
3851 gimple stmt = gsi_stmt (i);
3853 /* Lookup the RHS of the expression, see if we have an
3854 available computation for it. If so, replace the RHS with
3855 the available computation. */
3856 if (gimple_has_lhs (stmt)
3857 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
3858 && !gimple_assign_ssa_name_copy_p (stmt)
3859 && (!gimple_assign_single_p (stmt)
3860 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
3861 && !gimple_has_volatile_ops (stmt)
3862 && !has_zero_uses (gimple_get_lhs (stmt)))
3864 tree lhs = gimple_get_lhs (stmt);
3865 tree rhs = NULL_TREE;
3866 tree sprime = NULL;
3867 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
3868 pre_expr sprimeexpr;
3870 if (gimple_assign_single_p (stmt))
3871 rhs = gimple_assign_rhs1 (stmt);
3873 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
3874 get_expr_value_id (lhsexpr),
3875 NULL);
3877 if (sprimeexpr)
3879 if (sprimeexpr->kind == CONSTANT)
3880 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
3881 else if (sprimeexpr->kind == NAME)
3882 sprime = PRE_EXPR_NAME (sprimeexpr);
3883 else
3884 gcc_unreachable ();
3887 /* If there is no existing leader but SCCVN knows this
3888 value is constant, use that constant. */
3889 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
3891 sprime = fold_convert (TREE_TYPE (lhs),
3892 VN_INFO (lhs)->valnum);
3894 if (dump_file && (dump_flags & TDF_DETAILS))
3896 fprintf (dump_file, "Replaced ");
3897 print_gimple_expr (dump_file, stmt, 0, 0);
3898 fprintf (dump_file, " with ");
3899 print_generic_expr (dump_file, sprime, 0);
3900 fprintf (dump_file, " in ");
3901 print_gimple_stmt (dump_file, stmt, 0, 0);
3903 pre_stats.eliminations++;
3904 propagate_tree_value_into_stmt (&i, sprime);
3905 stmt = gsi_stmt (i);
3906 update_stmt (stmt);
3907 continue;
3910 /* If there is no existing usable leader but SCCVN thinks
3911 it has an expression it wants to use as replacement,
3912 insert that. */
3913 if (!sprime || sprime == lhs)
3915 tree val = VN_INFO (lhs)->valnum;
3916 if (val != VN_TOP
3917 && TREE_CODE (val) == SSA_NAME
3918 && VN_INFO (val)->needs_insertion
3919 && can_PRE_operation (vn_get_expr_for (val)))
3920 sprime = do_SCCVN_insertion (stmt, val);
3922 if (sprime
3923 && sprime != lhs
3924 && (rhs == NULL_TREE
3925 || TREE_CODE (rhs) != SSA_NAME
3926 || may_propagate_copy (rhs, sprime)))
3928 gcc_assert (sprime != rhs);
3930 if (dump_file && (dump_flags & TDF_DETAILS))
3932 fprintf (dump_file, "Replaced ");
3933 print_gimple_expr (dump_file, stmt, 0, 0);
3934 fprintf (dump_file, " with ");
3935 print_generic_expr (dump_file, sprime, 0);
3936 fprintf (dump_file, " in ");
3937 print_gimple_stmt (dump_file, stmt, 0, 0);
3940 if (TREE_CODE (sprime) == SSA_NAME)
3941 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
3942 NECESSARY, true);
3943 /* We need to make sure the new and old types actually match,
3944 which may require adding a simple cast, which fold_convert
3945 will do for us. */
3946 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
3947 && !useless_type_conversion_p (gimple_expr_type (stmt),
3948 TREE_TYPE (sprime)))
3949 sprime = fold_convert (gimple_expr_type (stmt), sprime);
3951 pre_stats.eliminations++;
3952 propagate_tree_value_into_stmt (&i, sprime);
3953 stmt = gsi_stmt (i);
3954 update_stmt (stmt);
3956 /* If we removed EH side effects from the statement, clean
3957 its EH information. */
3958 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
3960 bitmap_set_bit (need_eh_cleanup,
3961 gimple_bb (stmt)->index);
3962 if (dump_file && (dump_flags & TDF_DETAILS))
3963 fprintf (dump_file, " Removed EH side effects.\n");
3967 /* Visit COND_EXPRs and fold the comparison with the
3968 available value-numbers. */
3969 else if (gimple_code (stmt) == GIMPLE_COND)
3971 tree op0 = gimple_cond_lhs (stmt);
3972 tree op1 = gimple_cond_rhs (stmt);
3973 tree result;
3975 if (TREE_CODE (op0) == SSA_NAME)
3976 op0 = VN_INFO (op0)->valnum;
3977 if (TREE_CODE (op1) == SSA_NAME)
3978 op1 = VN_INFO (op1)->valnum;
3979 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
3980 op0, op1);
3981 if (result && TREE_CODE (result) == INTEGER_CST)
3983 if (integer_zerop (result))
3984 gimple_cond_make_false (stmt);
3985 else
3986 gimple_cond_make_true (stmt);
3987 update_stmt (stmt);
3988 todo = TODO_cleanup_cfg;
3994 return todo;
3997 /* Borrow a bit of tree-ssa-dce.c for the moment.
3998 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
3999 this may be a bit faster, and we may want critical edges kept split. */
4001 /* If OP's defining statement has not already been determined to be necessary,
4002 mark that statement necessary. Return the stmt, if it is newly
4003 necessary. */
4005 static inline gimple
4006 mark_operand_necessary (tree op)
4008 gimple stmt;
4010 gcc_assert (op);
4012 if (TREE_CODE (op) != SSA_NAME)
4013 return NULL;
4015 stmt = SSA_NAME_DEF_STMT (op);
4016 gcc_assert (stmt);
4018 if (gimple_plf (stmt, NECESSARY)
4019 || gimple_nop_p (stmt))
4020 return NULL;
4022 gimple_set_plf (stmt, NECESSARY, true);
4023 return stmt;
4026 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4027 to insert PHI nodes sometimes, and because value numbering of casts isn't
4028 perfect, we sometimes end up inserting dead code. This simple DCE-like
4029 pass removes any insertions we made that weren't actually used. */
4031 static void
4032 remove_dead_inserted_code (void)
4034 VEC(gimple,heap) *worklist = NULL;
4035 int i;
4036 gimple t;
4038 worklist = VEC_alloc (gimple, heap, VEC_length (gimple, inserted_exprs));
4039 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4041 if (gimple_plf (t, NECESSARY))
4042 VEC_quick_push (gimple, worklist, t);
4044 while (VEC_length (gimple, worklist) > 0)
4046 t = VEC_pop (gimple, worklist);
4048 /* PHI nodes are somewhat special in that each PHI alternative has
4049 data and control dependencies. All the statements feeding the
4050 PHI node's arguments are always necessary. */
4051 if (gimple_code (t) == GIMPLE_PHI)
4053 unsigned k;
4055 VEC_reserve (gimple, heap, worklist, gimple_phi_num_args (t));
4056 for (k = 0; k < gimple_phi_num_args (t); k++)
4058 tree arg = PHI_ARG_DEF (t, k);
4059 if (TREE_CODE (arg) == SSA_NAME)
4061 gimple n = mark_operand_necessary (arg);
4062 if (n)
4063 VEC_quick_push (gimple, worklist, n);
4067 else
4069 /* Propagate through the operands. Examine all the USE, VUSE and
4070 VDEF operands in this statement. Mark all the statements
4071 which feed this statement's uses as necessary. */
4072 ssa_op_iter iter;
4073 tree use;
4075 /* The operands of VDEF expressions are also needed as they
4076 represent potential definitions that may reach this
4077 statement (VDEF operands allow us to follow def-def
4078 links). */
4080 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4082 gimple n = mark_operand_necessary (use);
4083 if (n)
4084 VEC_safe_push (gimple, heap, worklist, n);
4089 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4091 if (!gimple_plf (t, NECESSARY))
4093 gimple_stmt_iterator gsi;
4095 if (dump_file && (dump_flags & TDF_DETAILS))
4097 fprintf (dump_file, "Removing unnecessary insertion:");
4098 print_gimple_stmt (dump_file, t, 0, 0);
4101 gsi = gsi_for_stmt (t);
4102 if (gimple_code (t) == GIMPLE_PHI)
4103 remove_phi_node (&gsi, true);
4104 else
4105 gsi_remove (&gsi, true);
4106 release_defs (t);
4109 VEC_free (gimple, heap, worklist);
4112 /* Initialize data structures used by PRE. */
4114 static void
4115 init_pre (bool do_fre)
4117 basic_block bb;
4119 next_expression_id = 1;
4120 expressions = NULL;
4121 VEC_safe_push (pre_expr, heap, expressions, NULL);
4122 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4123 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4124 get_max_value_id() + 1);
4126 in_fre = do_fre;
4128 inserted_exprs = NULL;
4129 need_creation = NULL;
4130 pretemp = NULL_TREE;
4131 storetemp = NULL_TREE;
4132 prephitemp = NULL_TREE;
4134 connect_infinite_loops_to_exit ();
4135 memset (&pre_stats, 0, sizeof (pre_stats));
4138 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4139 post_order_compute (postorder, false, false);
4141 FOR_ALL_BB (bb)
4142 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4144 calculate_dominance_info (CDI_POST_DOMINATORS);
4145 calculate_dominance_info (CDI_DOMINATORS);
4147 bitmap_obstack_initialize (&grand_bitmap_obstack);
4148 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4149 expr_pred_trans_eq, free);
4150 expression_to_id = htab_create (num_ssa_names * 3,
4151 pre_expr_hash,
4152 pre_expr_eq, NULL);
4153 seen_during_translate = BITMAP_ALLOC (&grand_bitmap_obstack);
4154 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4155 sizeof (struct bitmap_set), 30);
4156 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4157 sizeof (struct pre_expr_d), 30);
4158 FOR_ALL_BB (bb)
4160 EXP_GEN (bb) = bitmap_set_new ();
4161 PHI_GEN (bb) = bitmap_set_new ();
4162 TMP_GEN (bb) = bitmap_set_new ();
4163 AVAIL_OUT (bb) = bitmap_set_new ();
4165 maximal_set = in_fre ? NULL : bitmap_set_new ();
4167 need_eh_cleanup = BITMAP_ALLOC (NULL);
4171 /* Deallocate data structures used by PRE. */
4173 static void
4174 fini_pre (bool do_fre)
4176 basic_block bb;
4178 free (postorder);
4179 VEC_free (bitmap_set_t, heap, value_expressions);
4180 VEC_free (gimple, heap, inserted_exprs);
4181 VEC_free (gimple, heap, need_creation);
4182 bitmap_obstack_release (&grand_bitmap_obstack);
4183 free_alloc_pool (bitmap_set_pool);
4184 free_alloc_pool (pre_expr_pool);
4185 htab_delete (phi_translate_table);
4186 htab_delete (expression_to_id);
4188 FOR_ALL_BB (bb)
4190 free (bb->aux);
4191 bb->aux = NULL;
4194 free_dominance_info (CDI_POST_DOMINATORS);
4196 if (!bitmap_empty_p (need_eh_cleanup))
4198 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4199 cleanup_tree_cfg ();
4202 BITMAP_FREE (need_eh_cleanup);
4204 if (!do_fre)
4205 loop_optimizer_finalize ();
4208 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4209 only wants to do full redundancy elimination. */
4211 static unsigned int
4212 execute_pre (bool do_fre ATTRIBUTE_UNUSED)
4214 unsigned int todo = 0;
4216 do_partial_partial = optimize > 2;
4218 /* This has to happen before SCCVN runs because
4219 loop_optimizer_init may create new phis, etc. */
4220 if (!do_fre)
4221 loop_optimizer_init (LOOPS_NORMAL);
4223 if (!run_scc_vn (do_fre))
4225 if (!do_fre)
4227 remove_dead_inserted_code ();
4228 loop_optimizer_finalize ();
4231 return 0;
4233 init_pre (do_fre);
4236 /* Collect and value number expressions computed in each basic block. */
4237 compute_avail ();
4239 if (dump_file && (dump_flags & TDF_DETAILS))
4241 basic_block bb;
4243 FOR_ALL_BB (bb)
4245 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4246 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen",
4247 bb->index);
4248 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out",
4249 bb->index);
4253 /* Insert can get quite slow on an incredibly large number of basic
4254 blocks due to some quadratic behavior. Until this behavior is
4255 fixed, don't run it when he have an incredibly large number of
4256 bb's. If we aren't going to run insert, there is no point in
4257 computing ANTIC, either, even though it's plenty fast. */
4258 if (!do_fre && n_basic_blocks < 4000)
4260 compute_antic ();
4261 insert ();
4264 /* Remove all the redundant expressions. */
4265 todo |= eliminate ();
4267 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4268 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4269 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4270 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4271 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4273 /* Make sure to remove fake edges before committing our inserts.
4274 This makes sure we don't end up with extra critical edges that
4275 we would need to split. */
4276 remove_fake_exit_edges ();
4277 gsi_commit_edge_inserts ();
4279 clear_expression_ids ();
4280 free_scc_vn ();
4281 if (!do_fre)
4282 remove_dead_inserted_code ();
4284 fini_pre (do_fre);
4286 return todo;
4289 /* Gate and execute functions for PRE. */
4291 static unsigned int
4292 do_pre (void)
4294 return TODO_rebuild_alias | execute_pre (false);
4297 static bool
4298 gate_pre (void)
4300 /* PRE tends to generate bigger code. */
4301 return flag_tree_pre != 0 && optimize_function_for_speed_p (cfun);
4304 struct gimple_opt_pass pass_pre =
4307 GIMPLE_PASS,
4308 "pre", /* name */
4309 gate_pre, /* gate */
4310 do_pre, /* execute */
4311 NULL, /* sub */
4312 NULL, /* next */
4313 0, /* static_pass_number */
4314 TV_TREE_PRE, /* tv_id */
4315 PROP_no_crit_edges | PROP_cfg
4316 | PROP_ssa | PROP_alias, /* properties_required */
4317 0, /* properties_provided */
4318 0, /* properties_destroyed */
4319 0, /* todo_flags_start */
4320 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4321 | TODO_verify_ssa /* todo_flags_finish */
4326 /* Gate and execute functions for FRE. */
4328 static unsigned int
4329 execute_fre (void)
4331 return execute_pre (true);
4334 static bool
4335 gate_fre (void)
4337 return flag_tree_fre != 0;
4340 struct gimple_opt_pass pass_fre =
4343 GIMPLE_PASS,
4344 "fre", /* name */
4345 gate_fre, /* gate */
4346 execute_fre, /* execute */
4347 NULL, /* sub */
4348 NULL, /* next */
4349 0, /* static_pass_number */
4350 TV_TREE_FRE, /* tv_id */
4351 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
4352 0, /* properties_provided */
4353 0, /* properties_destroyed */
4354 0, /* todo_flags_start */
4355 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */