2009-01-18 Richard Guenther <rguenther@suse.de>
[official-gcc.git] / gcc / tree-ssa-pre.c
blob0717a366bf380b6ab8d6fa78615374ed55dae69f
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_expression:
1092 if (nary->opcode == TRUTH_NOT_EXPR)
1093 goto do_unary;
1094 if (nary->opcode != TRUTH_AND_EXPR
1095 && nary->opcode != TRUTH_OR_EXPR
1096 && nary->opcode != TRUTH_XOR_EXPR)
1097 return e;
1098 /* Fallthrough. */
1099 case tcc_binary:
1100 case tcc_comparison:
1102 /* We have to go from trees to pre exprs to value ids to
1103 constants. */
1104 tree naryop0 = nary->op[0];
1105 tree naryop1 = nary->op[1];
1106 tree result;
1107 if (!is_gimple_min_invariant (naryop0))
1109 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1110 unsigned int vrep0 = get_expr_value_id (rep0);
1111 tree const0 = get_constant_for_value_id (vrep0);
1112 if (const0)
1113 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1115 if (!is_gimple_min_invariant (naryop1))
1117 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1118 unsigned int vrep1 = get_expr_value_id (rep1);
1119 tree const1 = get_constant_for_value_id (vrep1);
1120 if (const1)
1121 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1123 result = fold_binary (nary->opcode, nary->type,
1124 naryop0, naryop1);
1125 if (result && is_gimple_min_invariant (result))
1126 return get_or_alloc_expr_for_constant (result);
1127 /* We might have simplified the expression to a
1128 SSA_NAME for example from x_1 * 1. But we cannot
1129 insert a PHI for x_1 unconditionally as x_1 might
1130 not be available readily. */
1131 return e;
1133 case tcc_reference:
1134 if (nary->opcode != REALPART_EXPR
1135 && nary->opcode != IMAGPART_EXPR
1136 && nary->opcode != VIEW_CONVERT_EXPR)
1137 return e;
1138 /* Fallthrough. */
1139 case tcc_unary:
1140 do_unary:
1142 /* We have to go from trees to pre exprs to value ids to
1143 constants. */
1144 tree naryop0 = nary->op[0];
1145 tree const0, result;
1146 if (is_gimple_min_invariant (naryop0))
1147 const0 = naryop0;
1148 else
1150 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1151 unsigned int vrep0 = get_expr_value_id (rep0);
1152 const0 = get_constant_for_value_id (vrep0);
1154 result = NULL;
1155 if (const0)
1157 tree type1 = TREE_TYPE (nary->op[0]);
1158 const0 = fold_convert (type1, const0);
1159 result = fold_unary (nary->opcode, nary->type, const0);
1161 if (result && is_gimple_min_invariant (result))
1162 return get_or_alloc_expr_for_constant (result);
1163 return e;
1165 default:
1166 return e;
1169 case REFERENCE:
1171 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1172 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1173 vn_reference_op_t op;
1175 /* Try to simplify the translated expression if it is
1176 a call to a builtin function with at most two arguments. */
1177 op = VEC_index (vn_reference_op_s, operands, 0);
1178 if (op->opcode == CALL_EXPR
1179 && TREE_CODE (op->op0) == ADDR_EXPR
1180 && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL
1181 && DECL_BUILT_IN (TREE_OPERAND (op->op0, 0))
1182 && VEC_length (vn_reference_op_s, operands) >= 2
1183 && VEC_length (vn_reference_op_s, operands) <= 3)
1185 vn_reference_op_t arg0, arg1 = NULL;
1186 bool anyconst = false;
1187 arg0 = VEC_index (vn_reference_op_s, operands, 1);
1188 if (VEC_length (vn_reference_op_s, operands) > 2)
1189 arg1 = VEC_index (vn_reference_op_s, operands, 2);
1190 if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant
1191 || (arg0->opcode == ADDR_EXPR
1192 && is_gimple_min_invariant (arg0->op0)))
1193 anyconst = true;
1194 if (arg1
1195 && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant
1196 || (arg1->opcode == ADDR_EXPR
1197 && is_gimple_min_invariant (arg1->op0))))
1198 anyconst = true;
1199 if (anyconst)
1201 tree folded = build_call_expr (TREE_OPERAND (op->op0, 0),
1202 arg1 ? 2 : 1,
1203 arg0->op0,
1204 arg1 ? arg1->op0 : NULL);
1205 if (folded
1206 && TREE_CODE (folded) == NOP_EXPR)
1207 folded = TREE_OPERAND (folded, 0);
1208 if (folded
1209 && is_gimple_min_invariant (folded))
1210 return get_or_alloc_expr_for_constant (folded);
1213 return e;
1215 default:
1216 return e;
1218 return e;
1221 /* Translate the vuses in the VUSES vector backwards through phi nodes
1222 in PHIBLOCK, so that they have the value they would have in
1223 BLOCK. */
1225 static VEC(tree, gc) *
1226 translate_vuses_through_block (VEC (tree, gc) *vuses,
1227 basic_block phiblock,
1228 basic_block block)
1230 tree oldvuse;
1231 VEC(tree, gc) *result = NULL;
1232 int i;
1234 for (i = 0; VEC_iterate (tree, vuses, i, oldvuse); i++)
1236 gimple phi = SSA_NAME_DEF_STMT (oldvuse);
1237 if (gimple_code (phi) == GIMPLE_PHI
1238 && gimple_bb (phi) == phiblock)
1240 edge e = find_edge (block, gimple_bb (phi));
1241 if (e)
1243 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1244 if (def != oldvuse)
1246 if (!result)
1247 result = VEC_copy (tree, gc, vuses);
1248 VEC_replace (tree, result, i, def);
1254 /* We avoid creating a new copy of the vuses unless something
1255 actually changed, so result can be NULL. */
1256 if (result)
1258 sort_vuses (result);
1259 return result;
1261 return vuses;
1265 /* Like find_leader, but checks for the value existing in SET1 *or*
1266 SET2. This is used to avoid making a set consisting of the union
1267 of PA_IN and ANTIC_IN during insert. */
1269 static inline pre_expr
1270 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1272 pre_expr result;
1274 result = bitmap_find_leader (set1, val, NULL);
1275 if (!result && set2)
1276 result = bitmap_find_leader (set2, val, NULL);
1277 return result;
1280 /* Get the tree type for our PRE expression e. */
1282 static tree
1283 get_expr_type (const pre_expr e)
1285 switch (e->kind)
1287 case NAME:
1288 return TREE_TYPE (PRE_EXPR_NAME (e));
1289 case CONSTANT:
1290 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1291 case REFERENCE:
1293 vn_reference_op_t vro;
1295 gcc_assert (PRE_EXPR_REFERENCE (e)->operands);
1296 vro = VEC_index (vn_reference_op_s,
1297 PRE_EXPR_REFERENCE (e)->operands,
1299 /* We don't store type along with COMPONENT_REF because it is
1300 always the same as FIELD_DECL's type. */
1301 if (!vro->type)
1303 gcc_assert (vro->opcode == COMPONENT_REF);
1304 return TREE_TYPE (vro->op0);
1306 return vro->type;
1309 case NARY:
1310 return PRE_EXPR_NARY (e)->type;
1312 gcc_unreachable();
1315 /* Get a representative SSA_NAME for a given expression.
1316 Since all of our sub-expressions are treated as values, we require
1317 them to be SSA_NAME's for simplicity.
1318 Prior versions of GVNPRE used to use "value handles" here, so that
1319 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1320 either case, the operands are really values (IE we do not expect
1321 them to be usable without finding leaders). */
1323 static tree
1324 get_representative_for (const pre_expr e)
1326 tree exprtype;
1327 tree name;
1328 unsigned int value_id = get_expr_value_id (e);
1330 switch (e->kind)
1332 case NAME:
1333 return PRE_EXPR_NAME (e);
1334 case CONSTANT:
1335 return PRE_EXPR_CONSTANT (e);
1336 case NARY:
1337 case REFERENCE:
1339 /* Go through all of the expressions representing this value
1340 and pick out an SSA_NAME. */
1341 unsigned int i;
1342 bitmap_iterator bi;
1343 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1344 value_id);
1345 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1347 pre_expr rep = expression_for_id (i);
1348 if (rep->kind == NAME)
1349 return PRE_EXPR_NAME (rep);
1352 break;
1354 /* If we reached here we couldn't find an SSA_NAME. This can
1355 happen when we've discovered a value that has never appeared in
1356 the program as set to an SSA_NAME, most likely as the result of
1357 phi translation. */
1358 if (dump_file)
1360 fprintf (dump_file,
1361 "Could not find SSA_NAME representative for expression:");
1362 print_pre_expr (dump_file, e);
1363 fprintf (dump_file, "\n");
1366 exprtype = get_expr_type (e);
1368 /* Build and insert the assignment of the end result to the temporary
1369 that we will return. */
1370 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1372 pretemp = create_tmp_var (exprtype, "pretmp");
1373 get_var_ann (pretemp);
1376 name = make_ssa_name (pretemp, gimple_build_nop ());
1377 VN_INFO_GET (name)->value_id = value_id;
1378 if (e->kind == CONSTANT)
1379 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1380 else
1381 VN_INFO (name)->valnum = name;
1383 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1384 if (dump_file)
1386 fprintf (dump_file, "Created SSA_NAME representative ");
1387 print_generic_expr (dump_file, name, 0);
1388 fprintf (dump_file, " for expression:");
1389 print_pre_expr (dump_file, e);
1390 fprintf (dump_file, "\n");
1393 return name;
1399 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1400 the phis in PRED. SEEN is a bitmap saying which expression we have
1401 translated since we started translation of the toplevel expression.
1402 Return NULL if we can't find a leader for each part of the
1403 translated expression. */
1405 static pre_expr
1406 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1407 basic_block pred, basic_block phiblock, bitmap seen)
1409 pre_expr oldexpr = expr;
1410 pre_expr phitrans;
1412 if (!expr)
1413 return NULL;
1415 if (value_id_constant_p (get_expr_value_id (expr)))
1416 return expr;
1418 phitrans = phi_trans_lookup (expr, pred);
1419 if (phitrans)
1420 return phitrans;
1422 /* Prevent cycles when we have recursively dependent leaders. This
1423 can only happen when phi translating the maximal set. */
1424 if (seen)
1426 unsigned int expr_id = get_expression_id (expr);
1427 if (bitmap_bit_p (seen, expr_id))
1428 return NULL;
1429 bitmap_set_bit (seen, expr_id);
1432 switch (expr->kind)
1434 /* Constants contain no values that need translation. */
1435 case CONSTANT:
1436 return expr;
1438 case NARY:
1440 unsigned int i;
1441 bool changed = false;
1442 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1443 struct vn_nary_op_s newnary;
1444 /* The NARY structure is only guaranteed to have been
1445 allocated to the nary->length operands. */
1446 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1447 - sizeof (tree) * (4 - nary->length)));
1449 for (i = 0; i < newnary.length; i++)
1451 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1452 continue;
1453 else
1455 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1456 pre_expr leader = find_leader_in_sets (op_val_id, set1, set2);
1457 pre_expr result = phi_translate_1 (leader, set1, set2,
1458 pred, phiblock, seen);
1459 if (result && result != leader)
1461 tree name = get_representative_for (result);
1462 if (!name)
1463 return NULL;
1464 newnary.op[i] = name;
1466 else if (!result)
1467 return NULL;
1469 changed |= newnary.op[i] != nary->op[i];
1472 if (changed)
1474 pre_expr constant;
1476 tree result = vn_nary_op_lookup_pieces (newnary.length,
1477 newnary.opcode,
1478 newnary.type,
1479 newnary.op[0],
1480 newnary.op[1],
1481 newnary.op[2],
1482 newnary.op[3],
1483 &nary);
1484 unsigned int new_val_id;
1486 expr = (pre_expr) pool_alloc (pre_expr_pool);
1487 expr->kind = NARY;
1488 expr->id = 0;
1489 if (result && is_gimple_min_invariant (result))
1490 return get_or_alloc_expr_for_constant (result);
1493 if (nary)
1495 PRE_EXPR_NARY (expr) = nary;
1496 constant = fully_constant_expression (expr);
1497 if (constant != expr)
1498 return constant;
1500 new_val_id = nary->value_id;
1501 get_or_alloc_expression_id (expr);
1503 else
1505 new_val_id = get_next_value_id ();
1506 VEC_safe_grow_cleared (bitmap_set_t, heap,
1507 value_expressions,
1508 get_max_value_id() + 1);
1509 nary = vn_nary_op_insert_pieces (newnary.length,
1510 newnary.opcode,
1511 newnary.type,
1512 newnary.op[0],
1513 newnary.op[1],
1514 newnary.op[2],
1515 newnary.op[3],
1516 result, new_val_id);
1517 PRE_EXPR_NARY (expr) = nary;
1518 constant = fully_constant_expression (expr);
1519 if (constant != expr)
1520 return constant;
1521 get_or_alloc_expression_id (expr);
1523 add_to_value (new_val_id, expr);
1525 phi_trans_add (oldexpr, expr, pred);
1526 return expr;
1528 break;
1530 case REFERENCE:
1532 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1533 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1534 VEC (tree, gc) *vuses = ref->vuses;
1535 VEC (tree, gc) *newvuses = vuses;
1536 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1537 bool changed = false;
1538 unsigned int i;
1539 vn_reference_op_t operand;
1540 vn_reference_t newref;
1542 for (i = 0; VEC_iterate (vn_reference_op_s, operands, i, operand); i++)
1544 pre_expr opresult;
1545 pre_expr leader;
1546 tree oldop0 = operand->op0;
1547 tree oldop1 = operand->op1;
1548 tree oldop2 = operand->op2;
1549 tree op0 = oldop0;
1550 tree op1 = oldop1;
1551 tree op2 = oldop2;
1552 tree type = operand->type;
1553 vn_reference_op_s newop = *operand;
1555 if (op0 && TREE_CODE (op0) == SSA_NAME)
1557 unsigned int op_val_id = VN_INFO (op0)->value_id;
1558 leader = find_leader_in_sets (op_val_id, set1, set2);
1559 opresult = phi_translate_1 (leader, set1, set2,
1560 pred, phiblock, seen);
1561 if (opresult && opresult != leader)
1563 tree name = get_representative_for (opresult);
1564 if (!name)
1565 break;
1566 op0 = name;
1568 else if (!opresult)
1569 break;
1571 changed |= op0 != oldop0;
1573 if (op1 && TREE_CODE (op1) == SSA_NAME)
1575 unsigned int op_val_id = VN_INFO (op1)->value_id;
1576 leader = find_leader_in_sets (op_val_id, set1, set2);
1577 opresult = phi_translate_1 (leader, set1, set2,
1578 pred, phiblock, seen);
1579 if (opresult && opresult != leader)
1581 tree name = get_representative_for (opresult);
1582 if (!name)
1583 break;
1584 op1 = name;
1586 else if (!opresult)
1587 break;
1589 changed |= op1 != oldop1;
1590 if (op2 && TREE_CODE (op2) == SSA_NAME)
1592 unsigned int op_val_id = VN_INFO (op2)->value_id;
1593 leader = find_leader_in_sets (op_val_id, set1, set2);
1594 opresult = phi_translate_1 (leader, set1, set2,
1595 pred, phiblock, seen);
1596 if (opresult && opresult != leader)
1598 tree name = get_representative_for (opresult);
1599 if (!name)
1600 break;
1601 op2 = name;
1603 else if (!opresult)
1604 break;
1606 changed |= op2 != oldop2;
1608 if (!newoperands)
1609 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1610 /* We may have changed from an SSA_NAME to a constant */
1611 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1612 newop.opcode = TREE_CODE (op0);
1613 newop.type = type;
1614 newop.op0 = op0;
1615 newop.op1 = op1;
1616 newop.op2 = op2;
1617 VEC_replace (vn_reference_op_s, newoperands, i, &newop);
1619 if (i != VEC_length (vn_reference_op_s, operands))
1621 if (newoperands)
1622 VEC_free (vn_reference_op_s, heap, newoperands);
1623 return NULL;
1626 newvuses = translate_vuses_through_block (vuses, phiblock, pred);
1627 changed |= newvuses != vuses;
1629 if (changed)
1631 unsigned int new_val_id;
1632 pre_expr constant;
1634 tree result = vn_reference_lookup_pieces (newvuses,
1635 newoperands,
1636 &newref, true);
1637 if (newref)
1638 VEC_free (vn_reference_op_s, heap, newoperands);
1640 if (result && is_gimple_min_invariant (result))
1642 gcc_assert (!newoperands);
1643 return get_or_alloc_expr_for_constant (result);
1646 expr = (pre_expr) pool_alloc (pre_expr_pool);
1647 expr->kind = REFERENCE;
1648 expr->id = 0;
1650 if (newref)
1652 PRE_EXPR_REFERENCE (expr) = newref;
1653 constant = fully_constant_expression (expr);
1654 if (constant != expr)
1655 return constant;
1657 new_val_id = newref->value_id;
1658 get_or_alloc_expression_id (expr);
1660 else
1662 new_val_id = get_next_value_id ();
1663 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
1664 get_max_value_id() + 1);
1665 newref = vn_reference_insert_pieces (newvuses,
1666 newoperands,
1667 result, new_val_id);
1668 newoperands = NULL;
1669 PRE_EXPR_REFERENCE (expr) = newref;
1670 constant = fully_constant_expression (expr);
1671 if (constant != expr)
1672 return constant;
1673 get_or_alloc_expression_id (expr);
1675 add_to_value (new_val_id, expr);
1677 VEC_free (vn_reference_op_s, heap, newoperands);
1678 phi_trans_add (oldexpr, expr, pred);
1679 return expr;
1681 break;
1683 case NAME:
1685 gimple phi = NULL;
1686 edge e;
1687 gimple def_stmt;
1688 tree name = PRE_EXPR_NAME (expr);
1690 def_stmt = SSA_NAME_DEF_STMT (name);
1691 if (gimple_code (def_stmt) == GIMPLE_PHI
1692 && gimple_bb (def_stmt) == phiblock)
1693 phi = def_stmt;
1694 else
1695 return expr;
1697 e = find_edge (pred, gimple_bb (phi));
1698 if (e)
1700 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1701 pre_expr newexpr;
1703 /* Handle constant. */
1704 if (is_gimple_min_invariant (def))
1705 return get_or_alloc_expr_for_constant (def);
1707 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1708 return NULL;
1710 newexpr = get_or_alloc_expr_for_name (def);
1711 return newexpr;
1714 return expr;
1716 default:
1717 gcc_unreachable ();
1721 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1722 the phis in PRED.
1723 Return NULL if we can't find a leader for each part of the
1724 translated expression. */
1726 static pre_expr
1727 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1728 basic_block pred, basic_block phiblock)
1730 bitmap_clear (seen_during_translate);
1731 return phi_translate_1 (expr, set1, set2, pred, phiblock,
1732 seen_during_translate);
1735 /* For each expression in SET, translate the values through phi nodes
1736 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1737 expressions in DEST. */
1739 static void
1740 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1741 basic_block phiblock)
1743 VEC (pre_expr, heap) *exprs;
1744 pre_expr expr;
1745 int i;
1747 if (!phi_nodes (phiblock))
1749 bitmap_set_copy (dest, set);
1750 return;
1753 exprs = sorted_array_from_bitmap_set (set);
1754 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1756 pre_expr translated;
1757 translated = phi_translate (expr, set, NULL, pred, phiblock);
1759 /* Don't add empty translations to the cache */
1760 if (translated)
1761 phi_trans_add (expr, translated, pred);
1763 if (translated != NULL)
1764 bitmap_value_insert_into_set (dest, translated);
1766 VEC_free (pre_expr, heap, exprs);
1769 /* Find the leader for a value (i.e., the name representing that
1770 value) in a given set, and return it. If STMT is non-NULL it
1771 makes sure the defining statement for the leader dominates it.
1772 Return NULL if no leader is found. */
1774 static pre_expr
1775 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1777 if (value_id_constant_p (val))
1779 unsigned int i;
1780 bitmap_iterator bi;
1781 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1783 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1785 pre_expr expr = expression_for_id (i);
1786 if (expr->kind == CONSTANT)
1787 return expr;
1790 if (bitmap_set_contains_value (set, val))
1792 /* Rather than walk the entire bitmap of expressions, and see
1793 whether any of them has the value we are looking for, we look
1794 at the reverse mapping, which tells us the set of expressions
1795 that have a given value (IE value->expressions with that
1796 value) and see if any of those expressions are in our set.
1797 The number of expressions per value is usually significantly
1798 less than the number of expressions in the set. In fact, for
1799 large testcases, doing it this way is roughly 5-10x faster
1800 than walking the bitmap.
1801 If this is somehow a significant lose for some cases, we can
1802 choose which set to walk based on which set is smaller. */
1803 unsigned int i;
1804 bitmap_iterator bi;
1805 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1807 EXECUTE_IF_AND_IN_BITMAP (exprset->expressions,
1808 set->expressions, 0, i, bi)
1810 pre_expr val = expression_for_id (i);
1811 /* At the point where stmt is not null, there should always
1812 be an SSA_NAME first in the list of expressions. */
1813 if (stmt)
1815 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1816 if (gimple_code (def_stmt) != GIMPLE_PHI
1817 && gimple_bb (def_stmt) == gimple_bb (stmt)
1818 && gimple_uid (def_stmt) >= gimple_uid (stmt))
1819 continue;
1821 return val;
1824 return NULL;
1827 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1828 BLOCK by seeing if it is not killed in the block. Note that we are
1829 only determining whether there is a store that kills it. Because
1830 of the order in which clean iterates over values, we are guaranteed
1831 that altered operands will have caused us to be eliminated from the
1832 ANTIC_IN set already. */
1834 static bool
1835 value_dies_in_block_x (pre_expr expr, basic_block block)
1837 int i;
1838 tree vuse;
1839 VEC (tree, gc) *vuses = PRE_EXPR_REFERENCE (expr)->vuses;
1841 /* Conservatively, a value dies if it's vuses are defined in this
1842 block, unless they come from phi nodes (which are merge operations,
1843 rather than stores. */
1844 for (i = 0; VEC_iterate (tree, vuses, i, vuse); i++)
1846 gimple def = SSA_NAME_DEF_STMT (vuse);
1848 if (gimple_bb (def) != block)
1849 continue;
1850 if (gimple_code (def) == GIMPLE_PHI)
1851 continue;
1852 return true;
1854 return false;
1858 #define union_contains_value(SET1, SET2, VAL) \
1859 (bitmap_set_contains_value ((SET1), (VAL)) \
1860 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
1862 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
1864 static bool
1865 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
1866 vn_reference_op_t vro)
1868 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
1870 struct pre_expr_d temp;
1871 temp.kind = NAME;
1872 temp.id = 0;
1873 PRE_EXPR_NAME (&temp) = vro->op0;
1874 temp.id = lookup_expression_id (&temp);
1875 if (temp.id == 0)
1876 return false;
1877 if (!union_contains_value (set1, set2,
1878 get_expr_value_id (&temp)))
1879 return false;
1881 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
1883 struct pre_expr_d temp;
1884 temp.kind = NAME;
1885 temp.id = 0;
1886 PRE_EXPR_NAME (&temp) = vro->op1;
1887 temp.id = lookup_expression_id (&temp);
1888 if (temp.id == 0)
1889 return false;
1890 if (!union_contains_value (set1, set2,
1891 get_expr_value_id (&temp)))
1892 return false;
1895 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
1897 struct pre_expr_d temp;
1898 temp.kind = NAME;
1899 temp.id = 0;
1900 PRE_EXPR_NAME (&temp) = vro->op2;
1901 temp.id = lookup_expression_id (&temp);
1902 if (temp.id == 0)
1903 return false;
1904 if (!union_contains_value (set1, set2,
1905 get_expr_value_id (&temp)))
1906 return false;
1909 return true;
1912 /* Determine if the expression EXPR is valid in SET1 U SET2.
1913 ONLY SET2 CAN BE NULL.
1914 This means that we have a leader for each part of the expression
1915 (if it consists of values), or the expression is an SSA_NAME.
1916 For loads/calls, we also see if the vuses are killed in this block.
1919 static bool
1920 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
1921 basic_block block)
1923 switch (expr->kind)
1925 case NAME:
1926 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
1927 case NARY:
1929 unsigned int i;
1930 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1931 for (i = 0; i < nary->length; i++)
1933 if (TREE_CODE (nary->op[i]) == SSA_NAME)
1935 struct pre_expr_d temp;
1936 temp.kind = NAME;
1937 temp.id = 0;
1938 PRE_EXPR_NAME (&temp) = nary->op[i];
1939 temp.id = lookup_expression_id (&temp);
1940 if (temp.id == 0)
1941 return false;
1942 if (!union_contains_value (set1, set2,
1943 get_expr_value_id (&temp)))
1944 return false;
1947 return true;
1949 break;
1950 case REFERENCE:
1952 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1953 vn_reference_op_t vro;
1954 unsigned int i;
1956 for (i = 0; VEC_iterate (vn_reference_op_s, ref->operands, i, vro); i++)
1958 if (!vro_valid_in_sets (set1, set2, vro))
1959 return false;
1961 return !value_dies_in_block_x (expr, block);
1963 default:
1964 gcc_unreachable ();
1968 /* Clean the set of expressions that are no longer valid in SET1 or
1969 SET2. This means expressions that are made up of values we have no
1970 leaders for in SET1 or SET2. This version is used for partial
1971 anticipation, which means it is not valid in either ANTIC_IN or
1972 PA_IN. */
1974 static void
1975 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
1977 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
1978 pre_expr expr;
1979 int i;
1981 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1983 if (!valid_in_sets (set1, set2, expr, block))
1984 bitmap_remove_from_set (set1, expr);
1986 VEC_free (pre_expr, heap, exprs);
1989 /* Clean the set of expressions that are no longer valid in SET. This
1990 means expressions that are made up of values we have no leaders for
1991 in SET. */
1993 static void
1994 clean (bitmap_set_t set, basic_block block)
1996 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
1997 pre_expr expr;
1998 int i;
2000 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2002 if (!valid_in_sets (set, NULL, expr, block))
2003 bitmap_remove_from_set (set, expr);
2005 VEC_free (pre_expr, heap, exprs);
2008 static sbitmap has_abnormal_preds;
2010 /* List of blocks that may have changed during ANTIC computation and
2011 thus need to be iterated over. */
2013 static sbitmap changed_blocks;
2015 /* Decide whether to defer a block for a later iteration, or PHI
2016 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2017 should defer the block, and true if we processed it. */
2019 static bool
2020 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2021 basic_block block, basic_block phiblock)
2023 if (!BB_VISITED (phiblock))
2025 SET_BIT (changed_blocks, block->index);
2026 BB_VISITED (block) = 0;
2027 BB_DEFERRED (block) = 1;
2028 return false;
2030 else
2031 phi_translate_set (dest, source, block, phiblock);
2032 return true;
2035 /* Compute the ANTIC set for BLOCK.
2037 If succs(BLOCK) > 1 then
2038 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2039 else if succs(BLOCK) == 1 then
2040 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2042 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2045 static bool
2046 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2048 bool changed = false;
2049 bitmap_set_t S, old, ANTIC_OUT;
2050 bitmap_iterator bi;
2051 unsigned int bii;
2052 edge e;
2053 edge_iterator ei;
2055 old = ANTIC_OUT = S = NULL;
2056 BB_VISITED (block) = 1;
2058 /* If any edges from predecessors are abnormal, antic_in is empty,
2059 so do nothing. */
2060 if (block_has_abnormal_pred_edge)
2061 goto maybe_dump_sets;
2063 old = ANTIC_IN (block);
2064 ANTIC_OUT = bitmap_set_new ();
2066 /* If the block has no successors, ANTIC_OUT is empty. */
2067 if (EDGE_COUNT (block->succs) == 0)
2069 /* If we have one successor, we could have some phi nodes to
2070 translate through. */
2071 else if (single_succ_p (block))
2073 basic_block succ_bb = single_succ (block);
2075 /* We trade iterations of the dataflow equations for having to
2076 phi translate the maximal set, which is incredibly slow
2077 (since the maximal set often has 300+ members, even when you
2078 have a small number of blocks).
2079 Basically, we defer the computation of ANTIC for this block
2080 until we have processed it's successor, which will inevitably
2081 have a *much* smaller set of values to phi translate once
2082 clean has been run on it.
2083 The cost of doing this is that we technically perform more
2084 iterations, however, they are lower cost iterations.
2086 Timings for PRE on tramp3d-v4:
2087 without maximal set fix: 11 seconds
2088 with maximal set fix/without deferring: 26 seconds
2089 with maximal set fix/with deferring: 11 seconds
2092 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2093 block, succ_bb))
2095 changed = true;
2096 goto maybe_dump_sets;
2099 /* If we have multiple successors, we take the intersection of all of
2100 them. Note that in the case of loop exit phi nodes, we may have
2101 phis to translate through. */
2102 else
2104 VEC(basic_block, heap) * worklist;
2105 size_t i;
2106 basic_block bprime, first;
2108 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2109 FOR_EACH_EDGE (e, ei, block->succs)
2110 VEC_quick_push (basic_block, worklist, e->dest);
2111 first = VEC_index (basic_block, worklist, 0);
2113 if (phi_nodes (first))
2115 bitmap_set_t from = ANTIC_IN (first);
2117 if (!BB_VISITED (first))
2118 from = maximal_set;
2119 phi_translate_set (ANTIC_OUT, from, block, first);
2121 else
2123 if (!BB_VISITED (first))
2124 bitmap_set_copy (ANTIC_OUT, maximal_set);
2125 else
2126 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2129 for (i = 1; VEC_iterate (basic_block, worklist, i, bprime); i++)
2131 if (phi_nodes (bprime))
2133 bitmap_set_t tmp = bitmap_set_new ();
2134 bitmap_set_t from = ANTIC_IN (bprime);
2136 if (!BB_VISITED (bprime))
2137 from = maximal_set;
2138 phi_translate_set (tmp, from, block, bprime);
2139 bitmap_set_and (ANTIC_OUT, tmp);
2140 bitmap_set_free (tmp);
2142 else
2144 if (!BB_VISITED (bprime))
2145 bitmap_set_and (ANTIC_OUT, maximal_set);
2146 else
2147 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2150 VEC_free (basic_block, heap, worklist);
2153 /* Generate ANTIC_OUT - TMP_GEN. */
2154 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2156 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2157 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2158 TMP_GEN (block));
2160 /* Then union in the ANTIC_OUT - TMP_GEN values,
2161 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2162 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2163 bitmap_value_insert_into_set (ANTIC_IN (block),
2164 expression_for_id (bii));
2166 clean (ANTIC_IN (block), block);
2168 /* !old->expressions can happen when we deferred a block. */
2169 if (!old->expressions || !bitmap_set_equal (old, ANTIC_IN (block)))
2171 changed = true;
2172 SET_BIT (changed_blocks, block->index);
2173 FOR_EACH_EDGE (e, ei, block->preds)
2174 SET_BIT (changed_blocks, e->src->index);
2176 else
2177 RESET_BIT (changed_blocks, block->index);
2179 maybe_dump_sets:
2180 if (dump_file && (dump_flags & TDF_DETAILS))
2182 if (!BB_DEFERRED (block) || BB_VISITED (block))
2184 if (ANTIC_OUT)
2185 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2187 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2188 block->index);
2190 if (S)
2191 print_bitmap_set (dump_file, S, "S", block->index);
2193 else
2195 fprintf (dump_file,
2196 "Block %d was deferred for a future iteration.\n",
2197 block->index);
2200 if (old)
2201 bitmap_set_free (old);
2202 if (S)
2203 bitmap_set_free (S);
2204 if (ANTIC_OUT)
2205 bitmap_set_free (ANTIC_OUT);
2206 return changed;
2209 /* Compute PARTIAL_ANTIC for BLOCK.
2211 If succs(BLOCK) > 1 then
2212 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2213 in ANTIC_OUT for all succ(BLOCK)
2214 else if succs(BLOCK) == 1 then
2215 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2217 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2218 - ANTIC_IN[BLOCK])
2221 static bool
2222 compute_partial_antic_aux (basic_block block,
2223 bool block_has_abnormal_pred_edge)
2225 bool changed = false;
2226 bitmap_set_t old_PA_IN;
2227 bitmap_set_t PA_OUT;
2228 edge e;
2229 edge_iterator ei;
2230 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2232 old_PA_IN = PA_OUT = NULL;
2234 /* If any edges from predecessors are abnormal, antic_in is empty,
2235 so do nothing. */
2236 if (block_has_abnormal_pred_edge)
2237 goto maybe_dump_sets;
2239 /* If there are too many partially anticipatable values in the
2240 block, phi_translate_set can take an exponential time: stop
2241 before the translation starts. */
2242 if (max_pa
2243 && single_succ_p (block)
2244 && bitmap_count_bits (PA_IN (single_succ (block))->values) > max_pa)
2245 goto maybe_dump_sets;
2247 old_PA_IN = PA_IN (block);
2248 PA_OUT = bitmap_set_new ();
2250 /* If the block has no successors, ANTIC_OUT is empty. */
2251 if (EDGE_COUNT (block->succs) == 0)
2253 /* If we have one successor, we could have some phi nodes to
2254 translate through. Note that we can't phi translate across DFS
2255 back edges in partial antic, because it uses a union operation on
2256 the successors. For recurrences like IV's, we will end up
2257 generating a new value in the set on each go around (i + 3 (VH.1)
2258 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2259 else if (single_succ_p (block))
2261 basic_block succ = single_succ (block);
2262 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2263 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2265 /* If we have multiple successors, we take the union of all of
2266 them. */
2267 else
2269 VEC(basic_block, heap) * worklist;
2270 size_t i;
2271 basic_block bprime;
2273 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2274 FOR_EACH_EDGE (e, ei, block->succs)
2276 if (e->flags & EDGE_DFS_BACK)
2277 continue;
2278 VEC_quick_push (basic_block, worklist, e->dest);
2280 if (VEC_length (basic_block, worklist) > 0)
2282 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2284 unsigned int i;
2285 bitmap_iterator bi;
2287 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2288 bitmap_value_insert_into_set (PA_OUT,
2289 expression_for_id (i));
2290 if (phi_nodes (bprime))
2292 bitmap_set_t pa_in = bitmap_set_new ();
2293 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2294 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2295 bitmap_value_insert_into_set (PA_OUT,
2296 expression_for_id (i));
2297 bitmap_set_free (pa_in);
2299 else
2300 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2301 bitmap_value_insert_into_set (PA_OUT,
2302 expression_for_id (i));
2305 VEC_free (basic_block, heap, worklist);
2308 /* PA_IN starts with PA_OUT - TMP_GEN.
2309 Then we subtract things from ANTIC_IN. */
2310 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2312 /* For partial antic, we want to put back in the phi results, since
2313 we will properly avoid making them partially antic over backedges. */
2314 bitmap_ior_into (PA_IN (block)->values, PHI_GEN (block)->values);
2315 bitmap_ior_into (PA_IN (block)->expressions, PHI_GEN (block)->expressions);
2317 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2318 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2320 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2322 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2324 changed = true;
2325 SET_BIT (changed_blocks, block->index);
2326 FOR_EACH_EDGE (e, ei, block->preds)
2327 SET_BIT (changed_blocks, e->src->index);
2329 else
2330 RESET_BIT (changed_blocks, block->index);
2332 maybe_dump_sets:
2333 if (dump_file && (dump_flags & TDF_DETAILS))
2335 if (PA_OUT)
2336 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2338 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2340 if (old_PA_IN)
2341 bitmap_set_free (old_PA_IN);
2342 if (PA_OUT)
2343 bitmap_set_free (PA_OUT);
2344 return changed;
2347 /* Compute ANTIC and partial ANTIC sets. */
2349 static void
2350 compute_antic (void)
2352 bool changed = true;
2353 int num_iterations = 0;
2354 basic_block block;
2355 int i;
2357 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2358 We pre-build the map of blocks with incoming abnormal edges here. */
2359 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2360 sbitmap_zero (has_abnormal_preds);
2362 FOR_EACH_BB (block)
2364 edge_iterator ei;
2365 edge e;
2367 FOR_EACH_EDGE (e, ei, block->preds)
2369 e->flags &= ~EDGE_DFS_BACK;
2370 if (e->flags & EDGE_ABNORMAL)
2372 SET_BIT (has_abnormal_preds, block->index);
2373 break;
2377 BB_VISITED (block) = 0;
2378 BB_DEFERRED (block) = 0;
2379 /* While we are here, give empty ANTIC_IN sets to each block. */
2380 ANTIC_IN (block) = bitmap_set_new ();
2381 PA_IN (block) = bitmap_set_new ();
2384 /* At the exit block we anticipate nothing. */
2385 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2386 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2387 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2389 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2390 sbitmap_ones (changed_blocks);
2391 while (changed)
2393 if (dump_file && (dump_flags & TDF_DETAILS))
2394 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2395 num_iterations++;
2396 changed = false;
2397 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2399 if (TEST_BIT (changed_blocks, postorder[i]))
2401 basic_block block = BASIC_BLOCK (postorder[i]);
2402 changed |= compute_antic_aux (block,
2403 TEST_BIT (has_abnormal_preds,
2404 block->index));
2407 #ifdef ENABLE_CHECKING
2408 /* Theoretically possible, but *highly* unlikely. */
2409 gcc_assert (num_iterations < 500);
2410 #endif
2413 statistics_histogram_event (cfun, "compute_antic iterations",
2414 num_iterations);
2416 if (do_partial_partial)
2418 sbitmap_ones (changed_blocks);
2419 mark_dfs_back_edges ();
2420 num_iterations = 0;
2421 changed = true;
2422 while (changed)
2424 if (dump_file && (dump_flags & TDF_DETAILS))
2425 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2426 num_iterations++;
2427 changed = false;
2428 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2430 if (TEST_BIT (changed_blocks, postorder[i]))
2432 basic_block block = BASIC_BLOCK (postorder[i]);
2433 changed
2434 |= compute_partial_antic_aux (block,
2435 TEST_BIT (has_abnormal_preds,
2436 block->index));
2439 #ifdef ENABLE_CHECKING
2440 /* Theoretically possible, but *highly* unlikely. */
2441 gcc_assert (num_iterations < 500);
2442 #endif
2444 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2445 num_iterations);
2447 sbitmap_free (has_abnormal_preds);
2448 sbitmap_free (changed_blocks);
2451 /* Return true if we can value number the call in STMT. This is true
2452 if we have a pure or constant call. */
2454 static bool
2455 can_value_number_call (gimple stmt)
2457 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2458 return true;
2459 return false;
2462 /* Return true if OP is an exception handler related operation, such as
2463 FILTER_EXPR or EXC_PTR_EXPR. */
2465 static bool
2466 is_exception_related (gimple stmt)
2468 return (is_gimple_assign (stmt)
2469 && (gimple_assign_rhs_code (stmt) == FILTER_EXPR
2470 || gimple_assign_rhs_code (stmt) == EXC_PTR_EXPR));
2473 /* Return true if OP is a tree which we can perform PRE on
2474 on. This may not match the operations we can value number, but in
2475 a perfect world would. */
2477 static bool
2478 can_PRE_operation (tree op)
2480 return UNARY_CLASS_P (op)
2481 || BINARY_CLASS_P (op)
2482 || COMPARISON_CLASS_P (op)
2483 || TREE_CODE (op) == INDIRECT_REF
2484 || TREE_CODE (op) == COMPONENT_REF
2485 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2486 || TREE_CODE (op) == CALL_EXPR
2487 || TREE_CODE (op) == ARRAY_REF;
2491 /* Inserted expressions are placed onto this worklist, which is used
2492 for performing quick dead code elimination of insertions we made
2493 that didn't turn out to be necessary. */
2494 static VEC(gimple,heap) *inserted_exprs;
2496 /* Pool allocated fake store expressions are placed onto this
2497 worklist, which, after performing dead code elimination, is walked
2498 to see which expressions need to be put into GC'able memory */
2499 static VEC(gimple, heap) *need_creation;
2501 /* The actual worker for create_component_ref_by_pieces. */
2503 static tree
2504 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2505 unsigned int *operand, gimple_seq *stmts,
2506 gimple domstmt)
2508 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2509 *operand);
2510 tree genop;
2511 ++*operand;
2512 switch (currop->opcode)
2514 case CALL_EXPR:
2516 tree folded, sc = currop->op1;
2517 unsigned int nargs = 0;
2518 tree *args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2519 ref->operands) - 1);
2520 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2522 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2523 operand, stmts,
2524 domstmt);
2525 nargs++;
2527 folded = build_call_array (currop->type,
2528 TREE_CODE (currop->op0) == FUNCTION_DECL
2529 ? build_fold_addr_expr (currop->op0)
2530 : currop->op0,
2531 nargs, args);
2532 free (args);
2533 if (sc)
2535 pre_expr scexpr = get_or_alloc_expr_for (sc);
2536 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2537 if (!sc)
2538 return NULL_TREE;
2539 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2541 return folded;
2543 break;
2544 case ADDR_EXPR:
2545 if (currop->op0)
2547 gcc_assert (is_gimple_min_invariant (currop->op0));
2548 return currop->op0;
2550 /* Fallthrough. */
2551 case REALPART_EXPR:
2552 case IMAGPART_EXPR:
2553 case VIEW_CONVERT_EXPR:
2555 tree folded;
2556 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2557 operand,
2558 stmts, domstmt);
2559 if (!genop0)
2560 return NULL_TREE;
2561 folded = fold_build1 (currop->opcode, currop->type,
2562 genop0);
2563 return folded;
2565 break;
2566 case ALIGN_INDIRECT_REF:
2567 case MISALIGNED_INDIRECT_REF:
2568 case INDIRECT_REF:
2570 tree folded;
2571 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2572 operand,
2573 stmts, domstmt);
2574 if (!genop1)
2575 return NULL_TREE;
2576 genop1 = fold_convert (build_pointer_type (currop->type),
2577 genop1);
2579 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2580 folded = fold_build2 (currop->opcode, currop->type,
2581 genop1, currop->op1);
2582 else
2583 folded = fold_build1 (currop->opcode, currop->type,
2584 genop1);
2585 return folded;
2587 break;
2588 case BIT_FIELD_REF:
2590 tree folded;
2591 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2592 stmts, domstmt);
2593 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2594 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2595 tree genop1;
2596 tree genop2;
2598 if (!genop0)
2599 return NULL_TREE;
2600 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2601 if (!genop1)
2602 return NULL_TREE;
2603 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2604 if (!genop2)
2605 return NULL_TREE;
2606 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2607 genop2);
2608 return folded;
2611 /* For array ref vn_reference_op's, operand 1 of the array ref
2612 is op0 of the reference op and operand 3 of the array ref is
2613 op1. */
2614 case ARRAY_RANGE_REF:
2615 case ARRAY_REF:
2617 tree genop0;
2618 tree genop1 = currop->op0;
2619 pre_expr op1expr;
2620 tree genop2 = currop->op1;
2621 pre_expr op2expr;
2622 tree genop3;
2623 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2624 stmts, domstmt);
2625 if (!genop0)
2626 return NULL_TREE;
2627 op1expr = get_or_alloc_expr_for (genop1);
2628 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2629 if (!genop1)
2630 return NULL_TREE;
2631 if (genop2)
2633 op2expr = get_or_alloc_expr_for (genop2);
2634 genop2 = find_or_generate_expression (block, op2expr, stmts,
2635 domstmt);
2636 if (!genop2)
2637 return NULL_TREE;
2640 genop3 = currop->op2;
2641 return build4 (currop->opcode, currop->type, genop0, genop1,
2642 genop2, genop3);
2644 case COMPONENT_REF:
2646 tree op0;
2647 tree op1;
2648 tree genop2 = currop->op1;
2649 pre_expr op2expr;
2650 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2651 stmts, domstmt);
2652 if (!op0)
2653 return NULL_TREE;
2654 /* op1 should be a FIELD_DECL, which are represented by
2655 themselves. */
2656 op1 = currop->op0;
2657 if (genop2)
2659 op2expr = get_or_alloc_expr_for (genop2);
2660 genop2 = find_or_generate_expression (block, op2expr, stmts,
2661 domstmt);
2662 if (!genop2)
2663 return NULL_TREE;
2666 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2667 genop2);
2669 break;
2670 case SSA_NAME:
2672 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2673 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2674 return genop;
2676 case STRING_CST:
2677 case INTEGER_CST:
2678 case COMPLEX_CST:
2679 case VECTOR_CST:
2680 case REAL_CST:
2681 case CONSTRUCTOR:
2682 case VAR_DECL:
2683 case PARM_DECL:
2684 case CONST_DECL:
2685 case RESULT_DECL:
2686 case FUNCTION_DECL:
2687 return currop->op0;
2689 default:
2690 gcc_unreachable ();
2694 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2695 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2696 trying to rename aggregates into ssa form directly, which is a no no.
2698 Thus, this routine doesn't create temporaries, it just builds a
2699 single access expression for the array, calling
2700 find_or_generate_expression to build the innermost pieces.
2702 This function is a subroutine of create_expression_by_pieces, and
2703 should not be called on it's own unless you really know what you
2704 are doing. */
2706 static tree
2707 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2708 gimple_seq *stmts, gimple domstmt)
2710 unsigned int op = 0;
2711 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2714 /* Find a leader for an expression, or generate one using
2715 create_expression_by_pieces if it's ANTIC but
2716 complex.
2717 BLOCK is the basic_block we are looking for leaders in.
2718 EXPR is the expression to find a leader or generate for.
2719 STMTS is the statement list to put the inserted expressions on.
2720 Returns the SSA_NAME of the LHS of the generated expression or the
2721 leader.
2722 DOMSTMT if non-NULL is a statement that should be dominated by
2723 all uses in the generated expression. If DOMSTMT is non-NULL this
2724 routine can fail and return NULL_TREE. Otherwise it will assert
2725 on failure. */
2727 static tree
2728 find_or_generate_expression (basic_block block, pre_expr expr,
2729 gimple_seq *stmts, gimple domstmt)
2731 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2732 get_expr_value_id (expr), domstmt);
2733 tree genop = NULL;
2734 if (leader)
2736 if (leader->kind == NAME)
2737 genop = PRE_EXPR_NAME (leader);
2738 else if (leader->kind == CONSTANT)
2739 genop = PRE_EXPR_CONSTANT (leader);
2742 /* If it's still NULL, it must be a complex expression, so generate
2743 it recursively. Not so for FRE though. */
2744 if (genop == NULL
2745 && !in_fre)
2747 bitmap_set_t exprset;
2748 unsigned int lookfor = get_expr_value_id (expr);
2749 bool handled = false;
2750 bitmap_iterator bi;
2751 unsigned int i;
2753 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2754 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2756 pre_expr temp = expression_for_id (i);
2757 if (temp->kind != NAME)
2759 handled = true;
2760 genop = create_expression_by_pieces (block, temp, stmts,
2761 domstmt,
2762 get_expr_type (expr));
2763 break;
2766 if (!handled && domstmt)
2767 return NULL_TREE;
2769 gcc_assert (handled);
2771 return genop;
2774 #define NECESSARY GF_PLF_1
2776 /* Create an expression in pieces, so that we can handle very complex
2777 expressions that may be ANTIC, but not necessary GIMPLE.
2778 BLOCK is the basic block the expression will be inserted into,
2779 EXPR is the expression to insert (in value form)
2780 STMTS is a statement list to append the necessary insertions into.
2782 This function will die if we hit some value that shouldn't be
2783 ANTIC but is (IE there is no leader for it, or its components).
2784 This function may also generate expressions that are themselves
2785 partially or fully redundant. Those that are will be either made
2786 fully redundant during the next iteration of insert (for partially
2787 redundant ones), or eliminated by eliminate (for fully redundant
2788 ones).
2790 If DOMSTMT is non-NULL then we make sure that all uses in the
2791 expressions dominate that statement. In this case the function
2792 can return NULL_TREE to signal failure. */
2794 static tree
2795 create_expression_by_pieces (basic_block block, pre_expr expr,
2796 gimple_seq *stmts, gimple domstmt, tree type)
2798 tree temp, name;
2799 tree folded, newexpr;
2800 gimple_seq forced_stmts;
2801 unsigned int value_id;
2802 gimple_stmt_iterator gsi;
2803 tree exprtype = type ? type : get_expr_type (expr);
2804 pre_expr nameexpr;
2805 gimple newstmt;
2807 switch (expr->kind)
2809 /* We may hit the NAME/CONSTANT case if we have to convert types
2810 that value numbering saw through. */
2811 case NAME:
2812 folded = PRE_EXPR_NAME (expr);
2813 break;
2814 case CONSTANT:
2815 folded = PRE_EXPR_CONSTANT (expr);
2816 break;
2817 case REFERENCE:
2819 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2820 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
2822 break;
2823 case NARY:
2825 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2826 switch (nary->length)
2828 case 2:
2830 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2831 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
2832 tree genop1 = find_or_generate_expression (block, op1,
2833 stmts, domstmt);
2834 tree genop2 = find_or_generate_expression (block, op2,
2835 stmts, domstmt);
2836 if (!genop1 || !genop2)
2837 return NULL_TREE;
2838 genop1 = fold_convert (TREE_TYPE (nary->op[0]),
2839 genop1);
2840 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
2841 may be a constant with the wrong type. */
2842 if (nary->opcode == POINTER_PLUS_EXPR)
2843 genop2 = fold_convert (sizetype, genop2);
2844 else
2845 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
2847 folded = fold_build2 (nary->opcode, nary->type,
2848 genop1, genop2);
2850 break;
2851 case 1:
2853 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2854 tree genop1 = find_or_generate_expression (block, op1,
2855 stmts, domstmt);
2856 if (!genop1)
2857 return NULL_TREE;
2858 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
2860 folded = fold_build1 (nary->opcode, nary->type,
2861 genop1);
2863 break;
2864 default:
2865 return NULL_TREE;
2868 break;
2869 default:
2870 return NULL_TREE;
2872 folded = fold_convert (exprtype, folded);
2873 /* Force the generated expression to be a sequence of GIMPLE
2874 statements.
2875 We have to call unshare_expr because force_gimple_operand may
2876 modify the tree we pass to it. */
2877 newexpr = force_gimple_operand (unshare_expr (folded), &forced_stmts,
2878 false, NULL);
2880 /* If we have any intermediate expressions to the value sets, add them
2881 to the value sets and chain them in the instruction stream. */
2882 if (forced_stmts)
2884 gsi = gsi_start (forced_stmts);
2885 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2887 gimple stmt = gsi_stmt (gsi);
2888 tree forcedname = gimple_get_lhs (stmt);
2889 pre_expr nameexpr;
2891 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
2892 if (TREE_CODE (forcedname) == SSA_NAME)
2894 VN_INFO_GET (forcedname)->valnum = forcedname;
2895 VN_INFO (forcedname)->value_id = get_next_value_id ();
2896 nameexpr = get_or_alloc_expr_for_name (forcedname);
2897 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2898 if (!in_fre)
2899 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2900 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2902 mark_symbols_for_renaming (stmt);
2904 gimple_seq_add_seq (stmts, forced_stmts);
2907 /* Build and insert the assignment of the end result to the temporary
2908 that we will return. */
2909 if (!pretemp || exprtype != TREE_TYPE (pretemp))
2911 pretemp = create_tmp_var (exprtype, "pretmp");
2912 get_var_ann (pretemp);
2915 temp = pretemp;
2916 add_referenced_var (temp);
2918 if (TREE_CODE (exprtype) == COMPLEX_TYPE
2919 || TREE_CODE (exprtype) == VECTOR_TYPE)
2920 DECL_GIMPLE_REG_P (temp) = 1;
2922 newstmt = gimple_build_assign (temp, newexpr);
2923 name = make_ssa_name (temp, newstmt);
2924 gimple_assign_set_lhs (newstmt, name);
2925 gimple_set_plf (newstmt, NECESSARY, false);
2927 gimple_seq_add_stmt (stmts, newstmt);
2928 VEC_safe_push (gimple, heap, inserted_exprs, newstmt);
2930 /* All the symbols in NEWEXPR should be put into SSA form. */
2931 mark_symbols_for_renaming (newstmt);
2933 /* Add a value number to the temporary.
2934 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2935 we are creating the expression by pieces, and this particular piece of
2936 the expression may have been represented. There is no harm in replacing
2937 here. */
2938 VN_INFO_GET (name)->valnum = name;
2939 value_id = get_expr_value_id (expr);
2940 VN_INFO (name)->value_id = value_id;
2941 nameexpr = get_or_alloc_expr_for_name (name);
2942 add_to_value (value_id, nameexpr);
2943 if (!in_fre)
2944 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2945 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2947 pre_stats.insertions++;
2948 if (dump_file && (dump_flags & TDF_DETAILS))
2950 fprintf (dump_file, "Inserted ");
2951 print_gimple_stmt (dump_file, newstmt, 0, 0);
2952 fprintf (dump_file, " in predecessor %d\n", block->index);
2955 return name;
2959 /* Insert the to-be-made-available values of expression EXPRNUM for each
2960 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
2961 merge the result with a phi node, given the same value number as
2962 NODE. Return true if we have inserted new stuff. */
2964 static bool
2965 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
2966 pre_expr *avail)
2968 pre_expr expr = expression_for_id (exprnum);
2969 pre_expr newphi;
2970 unsigned int val = get_expr_value_id (expr);
2971 edge pred;
2972 bool insertions = false;
2973 bool nophi = false;
2974 basic_block bprime;
2975 pre_expr eprime;
2976 edge_iterator ei;
2977 tree type = get_expr_type (expr);
2978 tree temp, res;
2979 gimple phi;
2981 if (dump_file && (dump_flags & TDF_DETAILS))
2983 fprintf (dump_file, "Found partial redundancy for expression ");
2984 print_pre_expr (dump_file, expr);
2985 fprintf (dump_file, " (%04d)\n", val);
2988 /* Make sure we aren't creating an induction variable. */
2989 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2
2990 && expr->kind != REFERENCE)
2992 bool firstinsideloop = false;
2993 bool secondinsideloop = false;
2994 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
2995 EDGE_PRED (block, 0)->src);
2996 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
2997 EDGE_PRED (block, 1)->src);
2998 /* Induction variables only have one edge inside the loop. */
2999 if (firstinsideloop ^ secondinsideloop)
3001 if (dump_file && (dump_flags & TDF_DETAILS))
3002 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3003 nophi = true;
3007 /* Make sure we are not inserting trapping expressions. */
3008 FOR_EACH_EDGE (pred, ei, block->preds)
3010 bprime = pred->src;
3011 eprime = avail[bprime->index];
3012 if (eprime->kind == NARY
3013 && vn_nary_may_trap (PRE_EXPR_NARY (eprime)))
3014 return false;
3017 /* Make the necessary insertions. */
3018 FOR_EACH_EDGE (pred, ei, block->preds)
3020 gimple_seq stmts = NULL;
3021 tree builtexpr;
3022 bprime = pred->src;
3023 eprime = avail[bprime->index];
3025 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3027 builtexpr = create_expression_by_pieces (bprime,
3028 eprime,
3029 &stmts, NULL,
3030 type);
3031 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3032 gsi_insert_seq_on_edge (pred, stmts);
3033 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3034 insertions = true;
3036 else if (eprime->kind == CONSTANT)
3038 /* Constants may not have the right type, fold_convert
3039 should give us back a constant with the right type.
3041 tree constant = PRE_EXPR_CONSTANT (eprime);
3042 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3044 tree builtexpr = fold_convert (type, constant);
3045 if (!is_gimple_min_invariant (builtexpr))
3047 tree forcedexpr = force_gimple_operand (builtexpr,
3048 &stmts, true,
3049 NULL);
3050 if (!is_gimple_min_invariant (forcedexpr))
3052 if (forcedexpr != builtexpr)
3054 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3055 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3057 if (stmts)
3059 gimple_stmt_iterator gsi;
3060 gsi = gsi_start (stmts);
3061 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3063 gimple stmt = gsi_stmt (gsi);
3064 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3065 gimple_set_plf (stmt, NECESSARY, false);
3067 gsi_insert_seq_on_edge (pred, stmts);
3069 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3074 else if (eprime->kind == NAME)
3076 /* We may have to do a conversion because our value
3077 numbering can look through types in certain cases, but
3078 our IL requires all operands of a phi node have the same
3079 type. */
3080 tree name = PRE_EXPR_NAME (eprime);
3081 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3083 tree builtexpr;
3084 tree forcedexpr;
3085 builtexpr = fold_convert (type, name);
3086 forcedexpr = force_gimple_operand (builtexpr,
3087 &stmts, true,
3088 NULL);
3090 if (forcedexpr != name)
3092 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3093 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3096 if (stmts)
3098 gimple_stmt_iterator gsi;
3099 gsi = gsi_start (stmts);
3100 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3102 gimple stmt = gsi_stmt (gsi);
3103 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3104 gimple_set_plf (stmt, NECESSARY, false);
3106 gsi_insert_seq_on_edge (pred, stmts);
3108 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3112 /* If we didn't want a phi node, and we made insertions, we still have
3113 inserted new stuff, and thus return true. If we didn't want a phi node,
3114 and didn't make insertions, we haven't added anything new, so return
3115 false. */
3116 if (nophi && insertions)
3117 return true;
3118 else if (nophi && !insertions)
3119 return false;
3121 /* Now build a phi for the new variable. */
3122 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3124 prephitemp = create_tmp_var (type, "prephitmp");
3125 get_var_ann (prephitemp);
3128 temp = prephitemp;
3129 add_referenced_var (temp);
3131 if (TREE_CODE (type) == COMPLEX_TYPE
3132 || TREE_CODE (type) == VECTOR_TYPE)
3133 DECL_GIMPLE_REG_P (temp) = 1;
3135 phi = create_phi_node (temp, block);
3136 FOR_EACH_EDGE (pred, ei, block->preds)
3138 pre_expr ae = avail[pred->src->index];
3139 gcc_assert (get_expr_type (ae) == type
3140 || useless_type_conversion_p (type, get_expr_type (ae)));
3141 if (ae->kind == CONSTANT)
3142 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred);
3143 else
3144 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred);
3146 /* If the PHI node is already available, use it. */
3147 if ((res = vn_phi_lookup (phi)) != NULL_TREE)
3149 gimple_stmt_iterator gsi = gsi_for_stmt (phi);
3150 remove_phi_node (&gsi, true);
3151 release_defs (phi);
3152 add_to_value (val, get_or_alloc_expr_for_name (res));
3153 return false;
3156 gimple_set_plf (phi, NECESSARY, false);
3157 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3158 VN_INFO (gimple_phi_result (phi))->value_id = val;
3159 VEC_safe_push (gimple, heap, inserted_exprs, phi);
3161 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3162 add_to_value (val, newphi);
3164 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3165 this insertion, since we test for the existence of this value in PHI_GEN
3166 before proceeding with the partial redundancy checks in insert_aux.
3168 The value may exist in AVAIL_OUT, in particular, it could be represented
3169 by the expression we are trying to eliminate, in which case we want the
3170 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3171 inserted there.
3173 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3174 this block, because if it did, it would have existed in our dominator's
3175 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3178 bitmap_insert_into_set (PHI_GEN (block), newphi);
3179 bitmap_value_replace_in_set (AVAIL_OUT (block),
3180 newphi);
3181 bitmap_insert_into_set (NEW_SETS (block),
3182 newphi);
3184 if (dump_file && (dump_flags & TDF_DETAILS))
3186 fprintf (dump_file, "Created phi ");
3187 print_gimple_stmt (dump_file, phi, 0, 0);
3188 fprintf (dump_file, " in block %d\n", block->index);
3190 pre_stats.phis++;
3191 return true;
3196 /* Perform insertion of partially redundant values.
3197 For BLOCK, do the following:
3198 1. Propagate the NEW_SETS of the dominator into the current block.
3199 If the block has multiple predecessors,
3200 2a. Iterate over the ANTIC expressions for the block to see if
3201 any of them are partially redundant.
3202 2b. If so, insert them into the necessary predecessors to make
3203 the expression fully redundant.
3204 2c. Insert a new PHI merging the values of the predecessors.
3205 2d. Insert the new PHI, and the new expressions, into the
3206 NEW_SETS set.
3207 3. Recursively call ourselves on the dominator children of BLOCK.
3209 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3210 do_regular_insertion and do_partial_insertion.
3214 static bool
3215 do_regular_insertion (basic_block block, basic_block dom)
3217 bool new_stuff = false;
3218 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3219 pre_expr expr;
3220 int i;
3222 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3224 if (expr->kind != NAME)
3226 pre_expr *avail;
3227 unsigned int val;
3228 bool by_some = false;
3229 bool cant_insert = false;
3230 bool all_same = true;
3231 pre_expr first_s = NULL;
3232 edge pred;
3233 basic_block bprime;
3234 pre_expr eprime = NULL;
3235 edge_iterator ei;
3236 pre_expr edoubleprime = NULL;
3238 val = get_expr_value_id (expr);
3239 if (bitmap_set_contains_value (PHI_GEN (block), val))
3240 continue;
3241 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3243 if (dump_file && (dump_flags & TDF_DETAILS))
3244 fprintf (dump_file, "Found fully redundant value\n");
3245 continue;
3248 avail = XCNEWVEC (pre_expr, last_basic_block);
3249 FOR_EACH_EDGE (pred, ei, block->preds)
3251 unsigned int vprime;
3253 /* We should never run insertion for the exit block
3254 and so not come across fake pred edges. */
3255 gcc_assert (!(pred->flags & EDGE_FAKE));
3256 bprime = pred->src;
3257 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3258 bprime, block);
3260 /* eprime will generally only be NULL if the
3261 value of the expression, translated
3262 through the PHI for this predecessor, is
3263 undefined. If that is the case, we can't
3264 make the expression fully redundant,
3265 because its value is undefined along a
3266 predecessor path. We can thus break out
3267 early because it doesn't matter what the
3268 rest of the results are. */
3269 if (eprime == NULL)
3271 cant_insert = true;
3272 break;
3275 eprime = fully_constant_expression (eprime);
3276 vprime = get_expr_value_id (eprime);
3277 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3278 vprime, NULL);
3279 if (edoubleprime == NULL)
3281 avail[bprime->index] = eprime;
3282 all_same = false;
3284 else
3286 avail[bprime->index] = edoubleprime;
3287 by_some = true;
3288 if (first_s == NULL)
3289 first_s = edoubleprime;
3290 else if (!pre_expr_eq (first_s, edoubleprime))
3291 all_same = false;
3294 /* If we can insert it, it's not the same value
3295 already existing along every predecessor, and
3296 it's defined by some predecessor, it is
3297 partially redundant. */
3298 if (!cant_insert && !all_same && by_some && dbg_cnt (treepre_insert))
3300 if (insert_into_preds_of_block (block, get_expression_id (expr),
3301 avail))
3302 new_stuff = true;
3304 /* If all edges produce the same value and that value is
3305 an invariant, then the PHI has the same value on all
3306 edges. Note this. */
3307 else if (!cant_insert && all_same && eprime
3308 && (edoubleprime->kind == CONSTANT
3309 || edoubleprime->kind == NAME)
3310 && !value_id_constant_p (val))
3312 unsigned int j;
3313 bitmap_iterator bi;
3314 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3315 value_expressions, val);
3317 unsigned int new_val = get_expr_value_id (edoubleprime);
3318 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3320 pre_expr expr = expression_for_id (j);
3322 if (expr->kind == NAME)
3324 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3325 /* Just reset the value id and valnum so it is
3326 the same as the constant we have discovered. */
3327 if (edoubleprime->kind == CONSTANT)
3329 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3330 pre_stats.constified++;
3332 else
3333 info->valnum = PRE_EXPR_NAME (edoubleprime);
3334 info->value_id = new_val;
3338 free (avail);
3342 VEC_free (pre_expr, heap, exprs);
3343 return new_stuff;
3347 /* Perform insertion for partially anticipatable expressions. There
3348 is only one case we will perform insertion for these. This case is
3349 if the expression is partially anticipatable, and fully available.
3350 In this case, we know that putting it earlier will enable us to
3351 remove the later computation. */
3354 static bool
3355 do_partial_partial_insertion (basic_block block, basic_block dom)
3357 bool new_stuff = false;
3358 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3359 pre_expr expr;
3360 int i;
3362 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3364 if (expr->kind != NAME)
3366 pre_expr *avail;
3367 unsigned int val;
3368 bool by_all = true;
3369 bool cant_insert = false;
3370 edge pred;
3371 basic_block bprime;
3372 pre_expr eprime = NULL;
3373 edge_iterator ei;
3375 val = get_expr_value_id (expr);
3376 if (bitmap_set_contains_value (PHI_GEN (block), val))
3377 continue;
3378 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3379 continue;
3381 avail = XCNEWVEC (pre_expr, last_basic_block);
3382 FOR_EACH_EDGE (pred, ei, block->preds)
3384 unsigned int vprime;
3385 pre_expr edoubleprime;
3387 /* We should never run insertion for the exit block
3388 and so not come across fake pred edges. */
3389 gcc_assert (!(pred->flags & EDGE_FAKE));
3390 bprime = pred->src;
3391 eprime = phi_translate (expr, ANTIC_IN (block),
3392 PA_IN (block),
3393 bprime, block);
3395 /* eprime will generally only be NULL if the
3396 value of the expression, translated
3397 through the PHI for this predecessor, is
3398 undefined. If that is the case, we can't
3399 make the expression fully redundant,
3400 because its value is undefined along a
3401 predecessor path. We can thus break out
3402 early because it doesn't matter what the
3403 rest of the results are. */
3404 if (eprime == NULL)
3406 cant_insert = true;
3407 break;
3410 eprime = fully_constant_expression (eprime);
3411 vprime = get_expr_value_id (eprime);
3412 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3413 vprime, NULL);
3414 if (edoubleprime == NULL)
3416 by_all = false;
3417 break;
3419 else
3420 avail[bprime->index] = edoubleprime;
3424 /* If we can insert it, it's not the same value
3425 already existing along every predecessor, and
3426 it's defined by some predecessor, it is
3427 partially redundant. */
3428 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3430 pre_stats.pa_insert++;
3431 if (insert_into_preds_of_block (block, get_expression_id (expr),
3432 avail))
3433 new_stuff = true;
3435 free (avail);
3439 VEC_free (pre_expr, heap, exprs);
3440 return new_stuff;
3443 static bool
3444 insert_aux (basic_block block)
3446 basic_block son;
3447 bool new_stuff = false;
3449 if (block)
3451 basic_block dom;
3452 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3453 if (dom)
3455 unsigned i;
3456 bitmap_iterator bi;
3457 bitmap_set_t newset = NEW_SETS (dom);
3458 if (newset)
3460 /* Note that we need to value_replace both NEW_SETS, and
3461 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3462 represented by some non-simple expression here that we want
3463 to replace it with. */
3464 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3466 pre_expr expr = expression_for_id (i);
3467 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3468 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3471 if (!single_pred_p (block))
3473 new_stuff |= do_regular_insertion (block, dom);
3474 if (do_partial_partial)
3475 new_stuff |= do_partial_partial_insertion (block, dom);
3479 for (son = first_dom_son (CDI_DOMINATORS, block);
3480 son;
3481 son = next_dom_son (CDI_DOMINATORS, son))
3483 new_stuff |= insert_aux (son);
3486 return new_stuff;
3489 /* Perform insertion of partially redundant values. */
3491 static void
3492 insert (void)
3494 bool new_stuff = true;
3495 basic_block bb;
3496 int num_iterations = 0;
3498 FOR_ALL_BB (bb)
3499 NEW_SETS (bb) = bitmap_set_new ();
3501 while (new_stuff)
3503 num_iterations++;
3504 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3506 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3510 /* Add OP to EXP_GEN (block), and possibly to the maximal set if it is
3511 not defined by a phi node.
3512 PHI nodes can't go in the maximal sets because they are not in
3513 TMP_GEN, so it is possible to get into non-monotonic situations
3514 during ANTIC calculation, because it will *add* bits. */
3516 static void
3517 add_to_exp_gen (basic_block block, tree op)
3519 if (!in_fre)
3521 pre_expr result;
3522 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3523 return;
3524 result = get_or_alloc_expr_for_name (op);
3525 bitmap_value_insert_into_set (EXP_GEN (block), result);
3526 if (TREE_CODE (op) != SSA_NAME
3527 || gimple_code (SSA_NAME_DEF_STMT (op)) != GIMPLE_PHI)
3528 bitmap_value_insert_into_set (maximal_set, result);
3532 /* Create value ids for PHI in BLOCK. */
3534 static void
3535 make_values_for_phi (gimple phi, basic_block block)
3537 tree result = gimple_phi_result (phi);
3539 /* We have no need for virtual phis, as they don't represent
3540 actual computations. */
3541 if (is_gimple_reg (result))
3543 pre_expr e = get_or_alloc_expr_for_name (result);
3544 add_to_value (get_expr_value_id (e), e);
3545 bitmap_insert_into_set (PHI_GEN (block), e);
3546 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3550 /* Compute the AVAIL set for all basic blocks.
3552 This function performs value numbering of the statements in each basic
3553 block. The AVAIL sets are built from information we glean while doing
3554 this value numbering, since the AVAIL sets contain only one entry per
3555 value.
3557 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3558 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3560 static void
3561 compute_avail (void)
3564 basic_block block, son;
3565 basic_block *worklist;
3566 size_t sp = 0;
3567 tree param;
3569 /* For arguments with default definitions, we pretend they are
3570 defined in the entry block. */
3571 for (param = DECL_ARGUMENTS (current_function_decl);
3572 param;
3573 param = TREE_CHAIN (param))
3575 if (gimple_default_def (cfun, param) != NULL)
3577 tree def = gimple_default_def (cfun, param);
3578 pre_expr e = get_or_alloc_expr_for_name (def);
3580 add_to_value (get_expr_value_id (e), e);
3581 if (!in_fre)
3583 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3584 bitmap_value_insert_into_set (maximal_set, e);
3586 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3590 /* Likewise for the static chain decl. */
3591 if (cfun->static_chain_decl)
3593 param = cfun->static_chain_decl;
3594 if (gimple_default_def (cfun, param) != NULL)
3596 tree def = gimple_default_def (cfun, param);
3597 pre_expr e = get_or_alloc_expr_for_name (def);
3599 add_to_value (get_expr_value_id (e), e);
3600 if (!in_fre)
3602 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3603 bitmap_value_insert_into_set (maximal_set, e);
3605 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3609 /* Allocate the worklist. */
3610 worklist = XNEWVEC (basic_block, n_basic_blocks);
3612 /* Seed the algorithm by putting the dominator children of the entry
3613 block on the worklist. */
3614 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3615 son;
3616 son = next_dom_son (CDI_DOMINATORS, son))
3617 worklist[sp++] = son;
3619 /* Loop until the worklist is empty. */
3620 while (sp)
3622 gimple_stmt_iterator gsi;
3623 gimple stmt;
3624 basic_block dom;
3625 unsigned int stmt_uid = 1;
3627 /* Pick a block from the worklist. */
3628 block = worklist[--sp];
3630 /* Initially, the set of available values in BLOCK is that of
3631 its immediate dominator. */
3632 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3633 if (dom)
3634 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3636 /* Generate values for PHI nodes. */
3637 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3638 make_values_for_phi (gsi_stmt (gsi), block);
3640 /* Now compute value numbers and populate value sets with all
3641 the expressions computed in BLOCK. */
3642 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3644 ssa_op_iter iter;
3645 tree op;
3647 stmt = gsi_stmt (gsi);
3648 gimple_set_uid (stmt, stmt_uid++);
3650 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3652 pre_expr e = get_or_alloc_expr_for_name (op);
3654 add_to_value (get_expr_value_id (e), e);
3655 if (!in_fre)
3656 bitmap_insert_into_set (TMP_GEN (block), e);
3657 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3660 if (gimple_has_volatile_ops (stmt)
3661 || stmt_could_throw_p (stmt))
3662 continue;
3664 switch (gimple_code (stmt))
3666 case GIMPLE_RETURN:
3667 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3668 add_to_exp_gen (block, op);
3669 continue;
3671 case GIMPLE_CALL:
3673 vn_reference_t ref;
3674 unsigned int i;
3675 vn_reference_op_t vro;
3676 pre_expr result = NULL;
3677 VEC(vn_reference_op_s, heap) *ops = NULL;
3679 if (!can_value_number_call (stmt))
3680 continue;
3682 copy_reference_ops_from_call (stmt, &ops);
3683 vn_reference_lookup_pieces (shared_vuses_from_stmt (stmt),
3684 ops, &ref, false);
3685 VEC_free (vn_reference_op_s, heap, ops);
3686 if (!ref)
3687 continue;
3689 for (i = 0; VEC_iterate (vn_reference_op_s,
3690 ref->operands, i,
3691 vro); i++)
3693 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3694 add_to_exp_gen (block, vro->op0);
3695 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3696 add_to_exp_gen (block, vro->op1);
3697 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3698 add_to_exp_gen (block, vro->op2);
3700 result = (pre_expr) pool_alloc (pre_expr_pool);
3701 result->kind = REFERENCE;
3702 result->id = 0;
3703 PRE_EXPR_REFERENCE (result) = ref;
3705 get_or_alloc_expression_id (result);
3706 add_to_value (get_expr_value_id (result), result);
3707 if (!in_fre)
3709 bitmap_value_insert_into_set (EXP_GEN (block),
3710 result);
3711 bitmap_value_insert_into_set (maximal_set, result);
3713 continue;
3716 case GIMPLE_ASSIGN:
3718 pre_expr result = NULL;
3719 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
3721 case tcc_unary:
3722 if (is_exception_related (stmt))
3723 continue;
3724 case tcc_binary:
3725 case tcc_comparison:
3727 vn_nary_op_t nary;
3728 unsigned int i;
3730 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
3731 gimple_assign_rhs_code (stmt),
3732 gimple_expr_type (stmt),
3733 gimple_assign_rhs1 (stmt),
3734 gimple_assign_rhs2 (stmt),
3735 NULL_TREE, NULL_TREE, &nary);
3737 if (!nary)
3738 continue;
3740 for (i = 0; i < nary->length; i++)
3741 if (TREE_CODE (nary->op[i]) == SSA_NAME)
3742 add_to_exp_gen (block, nary->op[i]);
3744 result = (pre_expr) pool_alloc (pre_expr_pool);
3745 result->kind = NARY;
3746 result->id = 0;
3747 PRE_EXPR_NARY (result) = nary;
3748 break;
3751 case tcc_declaration:
3752 case tcc_reference:
3754 vn_reference_t ref;
3755 unsigned int i;
3756 vn_reference_op_t vro;
3758 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3759 shared_vuses_from_stmt (stmt),
3760 false, &ref);
3761 if (!ref)
3762 continue;
3764 for (i = 0; VEC_iterate (vn_reference_op_s,
3765 ref->operands, i,
3766 vro); i++)
3768 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3769 add_to_exp_gen (block, vro->op0);
3770 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3771 add_to_exp_gen (block, vro->op1);
3772 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3773 add_to_exp_gen (block, vro->op2);
3775 result = (pre_expr) pool_alloc (pre_expr_pool);
3776 result->kind = REFERENCE;
3777 result->id = 0;
3778 PRE_EXPR_REFERENCE (result) = ref;
3779 break;
3782 default:
3783 /* For any other statement that we don't
3784 recognize, simply add all referenced
3785 SSA_NAMEs to EXP_GEN. */
3786 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3787 add_to_exp_gen (block, op);
3788 continue;
3791 get_or_alloc_expression_id (result);
3792 add_to_value (get_expr_value_id (result), result);
3793 if (!in_fre)
3795 bitmap_value_insert_into_set (EXP_GEN (block), result);
3796 bitmap_value_insert_into_set (maximal_set, result);
3799 continue;
3801 default:
3802 break;
3806 /* Put the dominator children of BLOCK on the worklist of blocks
3807 to compute available sets for. */
3808 for (son = first_dom_son (CDI_DOMINATORS, block);
3809 son;
3810 son = next_dom_son (CDI_DOMINATORS, son))
3811 worklist[sp++] = son;
3814 free (worklist);
3817 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
3818 than the available expressions for it. The insertion point is
3819 right before the first use in STMT. Returns the SSA_NAME that should
3820 be used for replacement. */
3822 static tree
3823 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
3825 basic_block bb = gimple_bb (stmt);
3826 gimple_stmt_iterator gsi;
3827 gimple_seq stmts = NULL;
3828 tree expr;
3829 pre_expr e;
3831 /* First create a value expression from the expression we want
3832 to insert and associate it with the value handle for SSA_VN. */
3833 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
3834 if (e == NULL)
3835 return NULL_TREE;
3837 /* Then use create_expression_by_pieces to generate a valid
3838 expression to insert at this point of the IL stream. */
3839 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
3840 if (expr == NULL_TREE)
3841 return NULL_TREE;
3842 gsi = gsi_for_stmt (stmt);
3843 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
3845 return expr;
3848 /* Eliminate fully redundant computations. */
3850 static unsigned int
3851 eliminate (void)
3853 basic_block b;
3854 unsigned int todo = 0;
3856 FOR_EACH_BB (b)
3858 gimple_stmt_iterator i;
3860 for (i = gsi_start_bb (b); !gsi_end_p (i); gsi_next (&i))
3862 gimple stmt = gsi_stmt (i);
3864 /* Lookup the RHS of the expression, see if we have an
3865 available computation for it. If so, replace the RHS with
3866 the available computation. */
3867 if (gimple_has_lhs (stmt)
3868 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
3869 && !gimple_assign_ssa_name_copy_p (stmt)
3870 && (!gimple_assign_single_p (stmt)
3871 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
3872 && !gimple_has_volatile_ops (stmt)
3873 && !has_zero_uses (gimple_get_lhs (stmt)))
3875 tree lhs = gimple_get_lhs (stmt);
3876 tree rhs = NULL_TREE;
3877 tree sprime = NULL;
3878 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
3879 pre_expr sprimeexpr;
3881 if (gimple_assign_single_p (stmt))
3882 rhs = gimple_assign_rhs1 (stmt);
3884 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
3885 get_expr_value_id (lhsexpr),
3886 NULL);
3888 if (sprimeexpr)
3890 if (sprimeexpr->kind == CONSTANT)
3891 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
3892 else if (sprimeexpr->kind == NAME)
3893 sprime = PRE_EXPR_NAME (sprimeexpr);
3894 else
3895 gcc_unreachable ();
3898 /* If there is no existing leader but SCCVN knows this
3899 value is constant, use that constant. */
3900 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
3902 sprime = fold_convert (TREE_TYPE (lhs),
3903 VN_INFO (lhs)->valnum);
3905 if (dump_file && (dump_flags & TDF_DETAILS))
3907 fprintf (dump_file, "Replaced ");
3908 print_gimple_expr (dump_file, stmt, 0, 0);
3909 fprintf (dump_file, " with ");
3910 print_generic_expr (dump_file, sprime, 0);
3911 fprintf (dump_file, " in ");
3912 print_gimple_stmt (dump_file, stmt, 0, 0);
3914 pre_stats.eliminations++;
3915 propagate_tree_value_into_stmt (&i, sprime);
3916 stmt = gsi_stmt (i);
3917 update_stmt (stmt);
3918 continue;
3921 /* If there is no existing usable leader but SCCVN thinks
3922 it has an expression it wants to use as replacement,
3923 insert that. */
3924 if (!sprime || sprime == lhs)
3926 tree val = VN_INFO (lhs)->valnum;
3927 if (val != VN_TOP
3928 && TREE_CODE (val) == SSA_NAME
3929 && VN_INFO (val)->needs_insertion
3930 && can_PRE_operation (vn_get_expr_for (val)))
3931 sprime = do_SCCVN_insertion (stmt, val);
3933 if (sprime
3934 && sprime != lhs
3935 && (rhs == NULL_TREE
3936 || TREE_CODE (rhs) != SSA_NAME
3937 || may_propagate_copy (rhs, sprime)))
3939 gcc_assert (sprime != rhs);
3941 if (dump_file && (dump_flags & TDF_DETAILS))
3943 fprintf (dump_file, "Replaced ");
3944 print_gimple_expr (dump_file, stmt, 0, 0);
3945 fprintf (dump_file, " with ");
3946 print_generic_expr (dump_file, sprime, 0);
3947 fprintf (dump_file, " in ");
3948 print_gimple_stmt (dump_file, stmt, 0, 0);
3951 if (TREE_CODE (sprime) == SSA_NAME)
3952 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
3953 NECESSARY, true);
3954 /* We need to make sure the new and old types actually match,
3955 which may require adding a simple cast, which fold_convert
3956 will do for us. */
3957 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
3958 && !useless_type_conversion_p (gimple_expr_type (stmt),
3959 TREE_TYPE (sprime)))
3960 sprime = fold_convert (gimple_expr_type (stmt), sprime);
3962 pre_stats.eliminations++;
3963 propagate_tree_value_into_stmt (&i, sprime);
3964 stmt = gsi_stmt (i);
3965 update_stmt (stmt);
3967 /* If we removed EH side effects from the statement, clean
3968 its EH information. */
3969 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
3971 bitmap_set_bit (need_eh_cleanup,
3972 gimple_bb (stmt)->index);
3973 if (dump_file && (dump_flags & TDF_DETAILS))
3974 fprintf (dump_file, " Removed EH side effects.\n");
3978 /* Visit COND_EXPRs and fold the comparison with the
3979 available value-numbers. */
3980 else if (gimple_code (stmt) == GIMPLE_COND)
3982 tree op0 = gimple_cond_lhs (stmt);
3983 tree op1 = gimple_cond_rhs (stmt);
3984 tree result;
3986 if (TREE_CODE (op0) == SSA_NAME)
3987 op0 = VN_INFO (op0)->valnum;
3988 if (TREE_CODE (op1) == SSA_NAME)
3989 op1 = VN_INFO (op1)->valnum;
3990 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
3991 op0, op1);
3992 if (result && TREE_CODE (result) == INTEGER_CST)
3994 if (integer_zerop (result))
3995 gimple_cond_make_false (stmt);
3996 else
3997 gimple_cond_make_true (stmt);
3998 update_stmt (stmt);
3999 todo = TODO_cleanup_cfg;
4005 return todo;
4008 /* Borrow a bit of tree-ssa-dce.c for the moment.
4009 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4010 this may be a bit faster, and we may want critical edges kept split. */
4012 /* If OP's defining statement has not already been determined to be necessary,
4013 mark that statement necessary. Return the stmt, if it is newly
4014 necessary. */
4016 static inline gimple
4017 mark_operand_necessary (tree op)
4019 gimple stmt;
4021 gcc_assert (op);
4023 if (TREE_CODE (op) != SSA_NAME)
4024 return NULL;
4026 stmt = SSA_NAME_DEF_STMT (op);
4027 gcc_assert (stmt);
4029 if (gimple_plf (stmt, NECESSARY)
4030 || gimple_nop_p (stmt))
4031 return NULL;
4033 gimple_set_plf (stmt, NECESSARY, true);
4034 return stmt;
4037 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4038 to insert PHI nodes sometimes, and because value numbering of casts isn't
4039 perfect, we sometimes end up inserting dead code. This simple DCE-like
4040 pass removes any insertions we made that weren't actually used. */
4042 static void
4043 remove_dead_inserted_code (void)
4045 VEC(gimple,heap) *worklist = NULL;
4046 int i;
4047 gimple t;
4049 worklist = VEC_alloc (gimple, heap, VEC_length (gimple, inserted_exprs));
4050 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4052 if (gimple_plf (t, NECESSARY))
4053 VEC_quick_push (gimple, worklist, t);
4055 while (VEC_length (gimple, worklist) > 0)
4057 t = VEC_pop (gimple, worklist);
4059 /* PHI nodes are somewhat special in that each PHI alternative has
4060 data and control dependencies. All the statements feeding the
4061 PHI node's arguments are always necessary. */
4062 if (gimple_code (t) == GIMPLE_PHI)
4064 unsigned k;
4066 VEC_reserve (gimple, heap, worklist, gimple_phi_num_args (t));
4067 for (k = 0; k < gimple_phi_num_args (t); k++)
4069 tree arg = PHI_ARG_DEF (t, k);
4070 if (TREE_CODE (arg) == SSA_NAME)
4072 gimple n = mark_operand_necessary (arg);
4073 if (n)
4074 VEC_quick_push (gimple, worklist, n);
4078 else
4080 /* Propagate through the operands. Examine all the USE, VUSE and
4081 VDEF operands in this statement. Mark all the statements
4082 which feed this statement's uses as necessary. */
4083 ssa_op_iter iter;
4084 tree use;
4086 /* The operands of VDEF expressions are also needed as they
4087 represent potential definitions that may reach this
4088 statement (VDEF operands allow us to follow def-def
4089 links). */
4091 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4093 gimple n = mark_operand_necessary (use);
4094 if (n)
4095 VEC_safe_push (gimple, heap, worklist, n);
4100 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4102 if (!gimple_plf (t, NECESSARY))
4104 gimple_stmt_iterator gsi;
4106 if (dump_file && (dump_flags & TDF_DETAILS))
4108 fprintf (dump_file, "Removing unnecessary insertion:");
4109 print_gimple_stmt (dump_file, t, 0, 0);
4112 gsi = gsi_for_stmt (t);
4113 if (gimple_code (t) == GIMPLE_PHI)
4114 remove_phi_node (&gsi, true);
4115 else
4116 gsi_remove (&gsi, true);
4117 release_defs (t);
4120 VEC_free (gimple, heap, worklist);
4123 /* Initialize data structures used by PRE. */
4125 static void
4126 init_pre (bool do_fre)
4128 basic_block bb;
4130 next_expression_id = 1;
4131 expressions = NULL;
4132 VEC_safe_push (pre_expr, heap, expressions, NULL);
4133 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4134 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4135 get_max_value_id() + 1);
4137 in_fre = do_fre;
4139 inserted_exprs = NULL;
4140 need_creation = NULL;
4141 pretemp = NULL_TREE;
4142 storetemp = NULL_TREE;
4143 prephitemp = NULL_TREE;
4145 connect_infinite_loops_to_exit ();
4146 memset (&pre_stats, 0, sizeof (pre_stats));
4149 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4150 post_order_compute (postorder, false, false);
4152 FOR_ALL_BB (bb)
4153 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4155 calculate_dominance_info (CDI_POST_DOMINATORS);
4156 calculate_dominance_info (CDI_DOMINATORS);
4158 bitmap_obstack_initialize (&grand_bitmap_obstack);
4159 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4160 expr_pred_trans_eq, free);
4161 expression_to_id = htab_create (num_ssa_names * 3,
4162 pre_expr_hash,
4163 pre_expr_eq, NULL);
4164 seen_during_translate = BITMAP_ALLOC (&grand_bitmap_obstack);
4165 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4166 sizeof (struct bitmap_set), 30);
4167 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4168 sizeof (struct pre_expr_d), 30);
4169 FOR_ALL_BB (bb)
4171 EXP_GEN (bb) = bitmap_set_new ();
4172 PHI_GEN (bb) = bitmap_set_new ();
4173 TMP_GEN (bb) = bitmap_set_new ();
4174 AVAIL_OUT (bb) = bitmap_set_new ();
4176 maximal_set = in_fre ? NULL : bitmap_set_new ();
4178 need_eh_cleanup = BITMAP_ALLOC (NULL);
4182 /* Deallocate data structures used by PRE. */
4184 static void
4185 fini_pre (bool do_fre)
4187 basic_block bb;
4189 free (postorder);
4190 VEC_free (bitmap_set_t, heap, value_expressions);
4191 VEC_free (gimple, heap, inserted_exprs);
4192 VEC_free (gimple, heap, need_creation);
4193 bitmap_obstack_release (&grand_bitmap_obstack);
4194 free_alloc_pool (bitmap_set_pool);
4195 free_alloc_pool (pre_expr_pool);
4196 htab_delete (phi_translate_table);
4197 htab_delete (expression_to_id);
4199 FOR_ALL_BB (bb)
4201 free (bb->aux);
4202 bb->aux = NULL;
4205 free_dominance_info (CDI_POST_DOMINATORS);
4207 if (!bitmap_empty_p (need_eh_cleanup))
4209 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4210 cleanup_tree_cfg ();
4213 BITMAP_FREE (need_eh_cleanup);
4215 if (!do_fre)
4216 loop_optimizer_finalize ();
4219 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4220 only wants to do full redundancy elimination. */
4222 static unsigned int
4223 execute_pre (bool do_fre ATTRIBUTE_UNUSED)
4225 unsigned int todo = 0;
4227 do_partial_partial = optimize > 2;
4229 /* This has to happen before SCCVN runs because
4230 loop_optimizer_init may create new phis, etc. */
4231 if (!do_fre)
4232 loop_optimizer_init (LOOPS_NORMAL);
4234 if (!run_scc_vn (do_fre))
4236 if (!do_fre)
4238 remove_dead_inserted_code ();
4239 loop_optimizer_finalize ();
4242 return 0;
4244 init_pre (do_fre);
4247 /* Collect and value number expressions computed in each basic block. */
4248 compute_avail ();
4250 if (dump_file && (dump_flags & TDF_DETAILS))
4252 basic_block bb;
4254 FOR_ALL_BB (bb)
4256 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4257 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen",
4258 bb->index);
4259 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out",
4260 bb->index);
4264 /* Insert can get quite slow on an incredibly large number of basic
4265 blocks due to some quadratic behavior. Until this behavior is
4266 fixed, don't run it when he have an incredibly large number of
4267 bb's. If we aren't going to run insert, there is no point in
4268 computing ANTIC, either, even though it's plenty fast. */
4269 if (!do_fre && n_basic_blocks < 4000)
4271 compute_antic ();
4272 insert ();
4275 /* Remove all the redundant expressions. */
4276 todo |= eliminate ();
4278 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4279 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4280 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4281 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4282 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4284 /* Make sure to remove fake edges before committing our inserts.
4285 This makes sure we don't end up with extra critical edges that
4286 we would need to split. */
4287 remove_fake_exit_edges ();
4288 gsi_commit_edge_inserts ();
4290 clear_expression_ids ();
4291 free_scc_vn ();
4292 if (!do_fre)
4293 remove_dead_inserted_code ();
4295 fini_pre (do_fre);
4297 return todo;
4300 /* Gate and execute functions for PRE. */
4302 static unsigned int
4303 do_pre (void)
4305 return TODO_rebuild_alias | execute_pre (false);
4308 static bool
4309 gate_pre (void)
4311 /* PRE tends to generate bigger code. */
4312 return flag_tree_pre != 0 && optimize_function_for_speed_p (cfun);
4315 struct gimple_opt_pass pass_pre =
4318 GIMPLE_PASS,
4319 "pre", /* name */
4320 gate_pre, /* gate */
4321 do_pre, /* execute */
4322 NULL, /* sub */
4323 NULL, /* next */
4324 0, /* static_pass_number */
4325 TV_TREE_PRE, /* tv_id */
4326 PROP_no_crit_edges | PROP_cfg
4327 | PROP_ssa | PROP_alias, /* properties_required */
4328 0, /* properties_provided */
4329 0, /* properties_destroyed */
4330 0, /* todo_flags_start */
4331 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4332 | TODO_verify_ssa /* todo_flags_finish */
4337 /* Gate and execute functions for FRE. */
4339 static unsigned int
4340 execute_fre (void)
4342 return execute_pre (true);
4345 static bool
4346 gate_fre (void)
4348 return flag_tree_fre != 0;
4351 struct gimple_opt_pass pass_fre =
4354 GIMPLE_PASS,
4355 "fre", /* name */
4356 gate_fre, /* gate */
4357 execute_fre, /* execute */
4358 NULL, /* sub */
4359 NULL, /* next */
4360 0, /* static_pass_number */
4361 TV_TREE_FRE, /* tv_id */
4362 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
4363 0, /* properties_provided */
4364 0, /* properties_destroyed */
4365 0, /* todo_flags_start */
4366 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */