Daily bump.
[official-gcc.git] / gcc / tree-ssa-pre.c
blob4d97b13c999a4bf444a7f9a4c7cb6d7caf595eee
1 /* SSA-PRE for trees.
2 Copyright (C) 2001-2013 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
4 <stevenb@suse.de>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
11 any later version.
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "tree.h"
27 #include "basic-block.h"
28 #include "gimple-pretty-print.h"
29 #include "tree-inline.h"
30 #include "gimple.h"
31 #include "gimple-ssa.h"
32 #include "tree-cfg.h"
33 #include "tree-phinodes.h"
34 #include "ssa-iterators.h"
35 #include "tree-ssanames.h"
36 #include "tree-ssa-loop.h"
37 #include "tree-into-ssa.h"
38 #include "tree-dfa.h"
39 #include "tree-ssa.h"
40 #include "hash-table.h"
41 #include "tree-iterator.h"
42 #include "alloc-pool.h"
43 #include "obstack.h"
44 #include "tree-pass.h"
45 #include "flags.h"
46 #include "langhooks.h"
47 #include "cfgloop.h"
48 #include "tree-ssa-sccvn.h"
49 #include "tree-scalar-evolution.h"
50 #include "params.h"
51 #include "dbgcnt.h"
52 #include "domwalk.h"
53 #include "ipa-prop.h"
54 #include "tree-ssa-propagate.h"
56 /* TODO:
58 1. Avail sets can be shared by making an avail_find_leader that
59 walks up the dominator tree and looks in those avail sets.
60 This might affect code optimality, it's unclear right now.
61 2. Strength reduction can be performed by anticipating expressions
62 we can repair later on.
63 3. We can do back-substitution or smarter value numbering to catch
64 commutative expressions split up over multiple statements.
67 /* For ease of terminology, "expression node" in the below refers to
68 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
69 represent the actual statement containing the expressions we care about,
70 and we cache the value number by putting it in the expression. */
72 /* Basic algorithm
74 First we walk the statements to generate the AVAIL sets, the
75 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
76 generation of values/expressions by a given block. We use them
77 when computing the ANTIC sets. The AVAIL sets consist of
78 SSA_NAME's that represent values, so we know what values are
79 available in what blocks. AVAIL is a forward dataflow problem. In
80 SSA, values are never killed, so we don't need a kill set, or a
81 fixpoint iteration, in order to calculate the AVAIL sets. In
82 traditional parlance, AVAIL sets tell us the downsafety of the
83 expressions/values.
85 Next, we generate the ANTIC sets. These sets represent the
86 anticipatable expressions. ANTIC is a backwards dataflow
87 problem. An expression is anticipatable in a given block if it could
88 be generated in that block. This means that if we had to perform
89 an insertion in that block, of the value of that expression, we
90 could. Calculating the ANTIC sets requires phi translation of
91 expressions, because the flow goes backwards through phis. We must
92 iterate to a fixpoint of the ANTIC sets, because we have a kill
93 set. Even in SSA form, values are not live over the entire
94 function, only from their definition point onwards. So we have to
95 remove values from the ANTIC set once we go past the definition
96 point of the leaders that make them up.
97 compute_antic/compute_antic_aux performs this computation.
99 Third, we perform insertions to make partially redundant
100 expressions fully redundant.
102 An expression is partially redundant (excluding partial
103 anticipation) if:
105 1. It is AVAIL in some, but not all, of the predecessors of a
106 given block.
107 2. It is ANTIC in all the predecessors.
109 In order to make it fully redundant, we insert the expression into
110 the predecessors where it is not available, but is ANTIC.
112 For the partial anticipation case, we only perform insertion if it
113 is partially anticipated in some block, and fully available in all
114 of the predecessors.
116 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
117 performs these steps.
119 Fourth, we eliminate fully redundant expressions.
120 This is a simple statement walk that replaces redundant
121 calculations with the now available values. */
123 /* Representations of value numbers:
125 Value numbers are represented by a representative SSA_NAME. We
126 will create fake SSA_NAME's in situations where we need a
127 representative but do not have one (because it is a complex
128 expression). In order to facilitate storing the value numbers in
129 bitmaps, and keep the number of wasted SSA_NAME's down, we also
130 associate a value_id with each value number, and create full blown
131 ssa_name's only where we actually need them (IE in operands of
132 existing expressions).
134 Theoretically you could replace all the value_id's with
135 SSA_NAME_VERSION, but this would allocate a large number of
136 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
137 It would also require an additional indirection at each point we
138 use the value id. */
140 /* Representation of expressions on value numbers:
142 Expressions consisting of value numbers are represented the same
143 way as our VN internally represents them, with an additional
144 "pre_expr" wrapping around them in order to facilitate storing all
145 of the expressions in the same sets. */
147 /* Representation of sets:
149 The dataflow sets do not need to be sorted in any particular order
150 for the majority of their lifetime, are simply represented as two
151 bitmaps, one that keeps track of values present in the set, and one
152 that keeps track of expressions present in the set.
154 When we need them in topological order, we produce it on demand by
155 transforming the bitmap into an array and sorting it into topo
156 order. */
158 /* Type of expression, used to know which member of the PRE_EXPR union
159 is valid. */
161 enum pre_expr_kind
163 NAME,
164 NARY,
165 REFERENCE,
166 CONSTANT
169 typedef union pre_expr_union_d
171 tree name;
172 tree constant;
173 vn_nary_op_t nary;
174 vn_reference_t reference;
175 } pre_expr_union;
177 typedef struct pre_expr_d : typed_noop_remove <pre_expr_d>
179 enum pre_expr_kind kind;
180 unsigned int id;
181 pre_expr_union u;
183 /* hash_table support. */
184 typedef pre_expr_d value_type;
185 typedef pre_expr_d compare_type;
186 static inline hashval_t hash (const pre_expr_d *);
187 static inline int equal (const pre_expr_d *, const pre_expr_d *);
188 } *pre_expr;
190 #define PRE_EXPR_NAME(e) (e)->u.name
191 #define PRE_EXPR_NARY(e) (e)->u.nary
192 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
193 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
195 /* Compare E1 and E1 for equality. */
197 inline int
198 pre_expr_d::equal (const value_type *e1, const compare_type *e2)
200 if (e1->kind != e2->kind)
201 return false;
203 switch (e1->kind)
205 case CONSTANT:
206 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
207 PRE_EXPR_CONSTANT (e2));
208 case NAME:
209 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
210 case NARY:
211 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
212 case REFERENCE:
213 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
214 PRE_EXPR_REFERENCE (e2));
215 default:
216 gcc_unreachable ();
220 /* Hash E. */
222 inline hashval_t
223 pre_expr_d::hash (const value_type *e)
225 switch (e->kind)
227 case CONSTANT:
228 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
229 case NAME:
230 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
231 case NARY:
232 return PRE_EXPR_NARY (e)->hashcode;
233 case REFERENCE:
234 return PRE_EXPR_REFERENCE (e)->hashcode;
235 default:
236 gcc_unreachable ();
240 /* Next global expression id number. */
241 static unsigned int next_expression_id;
243 /* Mapping from expression to id number we can use in bitmap sets. */
244 static vec<pre_expr> expressions;
245 static hash_table <pre_expr_d> expression_to_id;
246 static vec<unsigned> name_to_id;
248 /* Allocate an expression id for EXPR. */
250 static inline unsigned int
251 alloc_expression_id (pre_expr expr)
253 struct pre_expr_d **slot;
254 /* Make sure we won't overflow. */
255 gcc_assert (next_expression_id + 1 > next_expression_id);
256 expr->id = next_expression_id++;
257 expressions.safe_push (expr);
258 if (expr->kind == NAME)
260 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
261 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
262 re-allocations by using vec::reserve upfront. There is no
263 vec::quick_grow_cleared unfortunately. */
264 unsigned old_len = name_to_id.length ();
265 name_to_id.reserve (num_ssa_names - old_len);
266 name_to_id.safe_grow_cleared (num_ssa_names);
267 gcc_assert (name_to_id[version] == 0);
268 name_to_id[version] = expr->id;
270 else
272 slot = expression_to_id.find_slot (expr, INSERT);
273 gcc_assert (!*slot);
274 *slot = expr;
276 return next_expression_id - 1;
279 /* Return the expression id for tree EXPR. */
281 static inline unsigned int
282 get_expression_id (const pre_expr expr)
284 return expr->id;
287 static inline unsigned int
288 lookup_expression_id (const pre_expr expr)
290 struct pre_expr_d **slot;
292 if (expr->kind == NAME)
294 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
295 if (name_to_id.length () <= version)
296 return 0;
297 return name_to_id[version];
299 else
301 slot = expression_to_id.find_slot (expr, NO_INSERT);
302 if (!slot)
303 return 0;
304 return ((pre_expr)*slot)->id;
308 /* Return the existing expression id for EXPR, or create one if one
309 does not exist yet. */
311 static inline unsigned int
312 get_or_alloc_expression_id (pre_expr expr)
314 unsigned int id = lookup_expression_id (expr);
315 if (id == 0)
316 return alloc_expression_id (expr);
317 return expr->id = id;
320 /* Return the expression that has expression id ID */
322 static inline pre_expr
323 expression_for_id (unsigned int id)
325 return expressions[id];
328 /* Free the expression id field in all of our expressions,
329 and then destroy the expressions array. */
331 static void
332 clear_expression_ids (void)
334 expressions.release ();
337 static alloc_pool pre_expr_pool;
339 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
341 static pre_expr
342 get_or_alloc_expr_for_name (tree name)
344 struct pre_expr_d expr;
345 pre_expr result;
346 unsigned int result_id;
348 expr.kind = NAME;
349 expr.id = 0;
350 PRE_EXPR_NAME (&expr) = name;
351 result_id = lookup_expression_id (&expr);
352 if (result_id != 0)
353 return expression_for_id (result_id);
355 result = (pre_expr) pool_alloc (pre_expr_pool);
356 result->kind = NAME;
357 PRE_EXPR_NAME (result) = name;
358 alloc_expression_id (result);
359 return result;
362 /* An unordered bitmap set. One bitmap tracks values, the other,
363 expressions. */
364 typedef struct bitmap_set
366 bitmap_head expressions;
367 bitmap_head values;
368 } *bitmap_set_t;
370 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
371 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
373 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
374 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
376 /* Mapping from value id to expressions with that value_id. */
377 static vec<bitmap> value_expressions;
379 /* Sets that we need to keep track of. */
380 typedef struct bb_bitmap_sets
382 /* The EXP_GEN set, which represents expressions/values generated in
383 a basic block. */
384 bitmap_set_t exp_gen;
386 /* The PHI_GEN set, which represents PHI results generated in a
387 basic block. */
388 bitmap_set_t phi_gen;
390 /* The TMP_GEN set, which represents results/temporaries generated
391 in a basic block. IE the LHS of an expression. */
392 bitmap_set_t tmp_gen;
394 /* The AVAIL_OUT set, which represents which values are available in
395 a given basic block. */
396 bitmap_set_t avail_out;
398 /* The ANTIC_IN set, which represents which values are anticipatable
399 in a given basic block. */
400 bitmap_set_t antic_in;
402 /* The PA_IN set, which represents which values are
403 partially anticipatable in a given basic block. */
404 bitmap_set_t pa_in;
406 /* The NEW_SETS set, which is used during insertion to augment the
407 AVAIL_OUT set of blocks with the new insertions performed during
408 the current iteration. */
409 bitmap_set_t new_sets;
411 /* A cache for value_dies_in_block_x. */
412 bitmap expr_dies;
414 /* True if we have visited this block during ANTIC calculation. */
415 unsigned int visited : 1;
417 /* True we have deferred processing this block during ANTIC
418 calculation until its successor is processed. */
419 unsigned int deferred : 1;
421 /* True when the block contains a call that might not return. */
422 unsigned int contains_may_not_return_call : 1;
423 } *bb_value_sets_t;
425 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
426 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
427 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
428 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
429 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
430 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
431 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
432 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
433 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
434 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
435 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
438 /* Basic block list in postorder. */
439 static int *postorder;
440 static int postorder_num;
442 /* This structure is used to keep track of statistics on what
443 optimization PRE was able to perform. */
444 static struct
446 /* The number of RHS computations eliminated by PRE. */
447 int eliminations;
449 /* The number of new expressions/temporaries generated by PRE. */
450 int insertions;
452 /* The number of inserts found due to partial anticipation */
453 int pa_insert;
455 /* The number of new PHI nodes added by PRE. */
456 int phis;
457 } pre_stats;
459 static bool do_partial_partial;
460 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int);
461 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
462 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
463 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
464 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
465 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
466 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
467 unsigned int, bool);
468 static bitmap_set_t bitmap_set_new (void);
469 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
470 tree);
471 static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
472 static unsigned int get_expr_value_id (pre_expr);
474 /* We can add and remove elements and entries to and from sets
475 and hash tables, so we use alloc pools for them. */
477 static alloc_pool bitmap_set_pool;
478 static bitmap_obstack grand_bitmap_obstack;
480 /* Set of blocks with statements that have had their EH properties changed. */
481 static bitmap need_eh_cleanup;
483 /* Set of blocks with statements that have had their AB properties changed. */
484 static bitmap need_ab_cleanup;
486 /* A three tuple {e, pred, v} used to cache phi translations in the
487 phi_translate_table. */
489 typedef struct expr_pred_trans_d : typed_free_remove<expr_pred_trans_d>
491 /* The expression. */
492 pre_expr e;
494 /* The predecessor block along which we translated the expression. */
495 basic_block pred;
497 /* The value that resulted from the translation. */
498 pre_expr v;
500 /* The hashcode for the expression, pred pair. This is cached for
501 speed reasons. */
502 hashval_t hashcode;
504 /* hash_table support. */
505 typedef expr_pred_trans_d value_type;
506 typedef expr_pred_trans_d compare_type;
507 static inline hashval_t hash (const value_type *);
508 static inline int equal (const value_type *, const compare_type *);
509 } *expr_pred_trans_t;
510 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
512 inline hashval_t
513 expr_pred_trans_d::hash (const expr_pred_trans_d *e)
515 return e->hashcode;
518 inline int
519 expr_pred_trans_d::equal (const value_type *ve1,
520 const compare_type *ve2)
522 basic_block b1 = ve1->pred;
523 basic_block b2 = ve2->pred;
525 /* If they are not translations for the same basic block, they can't
526 be equal. */
527 if (b1 != b2)
528 return false;
529 return pre_expr_d::equal (ve1->e, ve2->e);
532 /* The phi_translate_table caches phi translations for a given
533 expression and predecessor. */
534 static hash_table <expr_pred_trans_d> phi_translate_table;
536 /* Add the tuple mapping from {expression E, basic block PRED} to
537 the phi translation table and return whether it pre-existed. */
539 static inline bool
540 phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred)
542 expr_pred_trans_t *slot;
543 expr_pred_trans_d tem;
544 hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e),
545 pred->index);
546 tem.e = e;
547 tem.pred = pred;
548 tem.hashcode = hash;
549 slot = phi_translate_table.find_slot_with_hash (&tem, hash, INSERT);
550 if (*slot)
552 *entry = *slot;
553 return true;
556 *entry = *slot = XNEW (struct expr_pred_trans_d);
557 (*entry)->e = e;
558 (*entry)->pred = pred;
559 (*entry)->hashcode = hash;
560 return false;
564 /* Add expression E to the expression set of value id V. */
566 static void
567 add_to_value (unsigned int v, pre_expr e)
569 bitmap set;
571 gcc_checking_assert (get_expr_value_id (e) == v);
573 if (v >= value_expressions.length ())
575 value_expressions.safe_grow_cleared (v + 1);
578 set = value_expressions[v];
579 if (!set)
581 set = BITMAP_ALLOC (&grand_bitmap_obstack);
582 value_expressions[v] = set;
585 bitmap_set_bit (set, get_or_alloc_expression_id (e));
588 /* Create a new bitmap set and return it. */
590 static bitmap_set_t
591 bitmap_set_new (void)
593 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
594 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
595 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
596 return ret;
599 /* Return the value id for a PRE expression EXPR. */
601 static unsigned int
602 get_expr_value_id (pre_expr expr)
604 unsigned int id;
605 switch (expr->kind)
607 case CONSTANT:
608 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
609 break;
610 case NAME:
611 id = VN_INFO (PRE_EXPR_NAME (expr))->value_id;
612 break;
613 case NARY:
614 id = PRE_EXPR_NARY (expr)->value_id;
615 break;
616 case REFERENCE:
617 id = PRE_EXPR_REFERENCE (expr)->value_id;
618 break;
619 default:
620 gcc_unreachable ();
622 /* ??? We cannot assert that expr has a value-id (it can be 0), because
623 we assign value-ids only to expressions that have a result
624 in set_hashtable_value_ids. */
625 return id;
628 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
630 static tree
631 sccvn_valnum_from_value_id (unsigned int val)
633 bitmap_iterator bi;
634 unsigned int i;
635 bitmap exprset = value_expressions[val];
636 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
638 pre_expr vexpr = expression_for_id (i);
639 if (vexpr->kind == NAME)
640 return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
641 else if (vexpr->kind == CONSTANT)
642 return PRE_EXPR_CONSTANT (vexpr);
644 return NULL_TREE;
647 /* Remove an expression EXPR from a bitmapped set. */
649 static void
650 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
652 unsigned int val = get_expr_value_id (expr);
653 if (!value_id_constant_p (val))
655 bitmap_clear_bit (&set->values, val);
656 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
660 static void
661 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
662 unsigned int val, bool allow_constants)
664 if (allow_constants || !value_id_constant_p (val))
666 /* We specifically expect this and only this function to be able to
667 insert constants into a set. */
668 bitmap_set_bit (&set->values, val);
669 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
673 /* Insert an expression EXPR into a bitmapped set. */
675 static void
676 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
678 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
681 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
683 static void
684 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
686 bitmap_copy (&dest->expressions, &orig->expressions);
687 bitmap_copy (&dest->values, &orig->values);
691 /* Free memory used up by SET. */
692 static void
693 bitmap_set_free (bitmap_set_t set)
695 bitmap_clear (&set->expressions);
696 bitmap_clear (&set->values);
700 /* Generate an topological-ordered array of bitmap set SET. */
702 static vec<pre_expr>
703 sorted_array_from_bitmap_set (bitmap_set_t set)
705 unsigned int i, j;
706 bitmap_iterator bi, bj;
707 vec<pre_expr> result;
709 /* Pre-allocate roughly enough space for the array. */
710 result.create (bitmap_count_bits (&set->values));
712 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
714 /* The number of expressions having a given value is usually
715 relatively small. Thus, rather than making a vector of all
716 the expressions and sorting it by value-id, we walk the values
717 and check in the reverse mapping that tells us what expressions
718 have a given value, to filter those in our set. As a result,
719 the expressions are inserted in value-id order, which means
720 topological order.
722 If this is somehow a significant lose for some cases, we can
723 choose which set to walk based on the set size. */
724 bitmap exprset = value_expressions[i];
725 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj)
727 if (bitmap_bit_p (&set->expressions, j))
728 result.safe_push (expression_for_id (j));
732 return result;
735 /* Perform bitmapped set operation DEST &= ORIG. */
737 static void
738 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
740 bitmap_iterator bi;
741 unsigned int i;
743 if (dest != orig)
745 bitmap_head temp;
746 bitmap_initialize (&temp, &grand_bitmap_obstack);
748 bitmap_and_into (&dest->values, &orig->values);
749 bitmap_copy (&temp, &dest->expressions);
750 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
752 pre_expr expr = expression_for_id (i);
753 unsigned int value_id = get_expr_value_id (expr);
754 if (!bitmap_bit_p (&dest->values, value_id))
755 bitmap_clear_bit (&dest->expressions, i);
757 bitmap_clear (&temp);
761 /* Subtract all values and expressions contained in ORIG from DEST. */
763 static bitmap_set_t
764 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
766 bitmap_set_t result = bitmap_set_new ();
767 bitmap_iterator bi;
768 unsigned int i;
770 bitmap_and_compl (&result->expressions, &dest->expressions,
771 &orig->expressions);
773 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
775 pre_expr expr = expression_for_id (i);
776 unsigned int value_id = get_expr_value_id (expr);
777 bitmap_set_bit (&result->values, value_id);
780 return result;
783 /* Subtract all the values in bitmap set B from bitmap set A. */
785 static void
786 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
788 unsigned int i;
789 bitmap_iterator bi;
790 bitmap_head temp;
792 bitmap_initialize (&temp, &grand_bitmap_obstack);
794 bitmap_copy (&temp, &a->expressions);
795 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
797 pre_expr expr = expression_for_id (i);
798 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
799 bitmap_remove_from_set (a, expr);
801 bitmap_clear (&temp);
805 /* Return true if bitmapped set SET contains the value VALUE_ID. */
807 static bool
808 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
810 if (value_id_constant_p (value_id))
811 return true;
813 if (!set || bitmap_empty_p (&set->expressions))
814 return false;
816 return bitmap_bit_p (&set->values, value_id);
819 static inline bool
820 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
822 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
825 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
827 static void
828 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
829 const pre_expr expr)
831 bitmap exprset;
832 unsigned int i;
833 bitmap_iterator bi;
835 if (value_id_constant_p (lookfor))
836 return;
838 if (!bitmap_set_contains_value (set, lookfor))
839 return;
841 /* The number of expressions having a given value is usually
842 significantly less than the total number of expressions in SET.
843 Thus, rather than check, for each expression in SET, whether it
844 has the value LOOKFOR, we walk the reverse mapping that tells us
845 what expressions have a given value, and see if any of those
846 expressions are in our set. For large testcases, this is about
847 5-10x faster than walking the bitmap. If this is somehow a
848 significant lose for some cases, we can choose which set to walk
849 based on the set size. */
850 exprset = value_expressions[lookfor];
851 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
853 if (bitmap_clear_bit (&set->expressions, i))
855 bitmap_set_bit (&set->expressions, get_expression_id (expr));
856 return;
860 gcc_unreachable ();
863 /* Return true if two bitmap sets are equal. */
865 static bool
866 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
868 return bitmap_equal_p (&a->values, &b->values);
871 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
872 and add it otherwise. */
874 static void
875 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
877 unsigned int val = get_expr_value_id (expr);
879 if (bitmap_set_contains_value (set, val))
880 bitmap_set_replace_value (set, val, expr);
881 else
882 bitmap_insert_into_set (set, expr);
885 /* Insert EXPR into SET if EXPR's value is not already present in
886 SET. */
888 static void
889 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
891 unsigned int val = get_expr_value_id (expr);
893 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
895 /* Constant values are always considered to be part of the set. */
896 if (value_id_constant_p (val))
897 return;
899 /* If the value membership changed, add the expression. */
900 if (bitmap_set_bit (&set->values, val))
901 bitmap_set_bit (&set->expressions, expr->id);
904 /* Print out EXPR to outfile. */
906 static void
907 print_pre_expr (FILE *outfile, const pre_expr expr)
909 switch (expr->kind)
911 case CONSTANT:
912 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
913 break;
914 case NAME:
915 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
916 break;
917 case NARY:
919 unsigned int i;
920 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
921 fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode));
922 for (i = 0; i < nary->length; i++)
924 print_generic_expr (outfile, nary->op[i], 0);
925 if (i != (unsigned) nary->length - 1)
926 fprintf (outfile, ",");
928 fprintf (outfile, "}");
930 break;
932 case REFERENCE:
934 vn_reference_op_t vro;
935 unsigned int i;
936 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
937 fprintf (outfile, "{");
938 for (i = 0;
939 ref->operands.iterate (i, &vro);
940 i++)
942 bool closebrace = false;
943 if (vro->opcode != SSA_NAME
944 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
946 fprintf (outfile, "%s", get_tree_code_name (vro->opcode));
947 if (vro->op0)
949 fprintf (outfile, "<");
950 closebrace = true;
953 if (vro->op0)
955 print_generic_expr (outfile, vro->op0, 0);
956 if (vro->op1)
958 fprintf (outfile, ",");
959 print_generic_expr (outfile, vro->op1, 0);
961 if (vro->op2)
963 fprintf (outfile, ",");
964 print_generic_expr (outfile, vro->op2, 0);
967 if (closebrace)
968 fprintf (outfile, ">");
969 if (i != ref->operands.length () - 1)
970 fprintf (outfile, ",");
972 fprintf (outfile, "}");
973 if (ref->vuse)
975 fprintf (outfile, "@");
976 print_generic_expr (outfile, ref->vuse, 0);
979 break;
982 void debug_pre_expr (pre_expr);
984 /* Like print_pre_expr but always prints to stderr. */
985 DEBUG_FUNCTION void
986 debug_pre_expr (pre_expr e)
988 print_pre_expr (stderr, e);
989 fprintf (stderr, "\n");
992 /* Print out SET to OUTFILE. */
994 static void
995 print_bitmap_set (FILE *outfile, bitmap_set_t set,
996 const char *setname, int blockindex)
998 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
999 if (set)
1001 bool first = true;
1002 unsigned i;
1003 bitmap_iterator bi;
1005 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1007 const pre_expr expr = expression_for_id (i);
1009 if (!first)
1010 fprintf (outfile, ", ");
1011 first = false;
1012 print_pre_expr (outfile, expr);
1014 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1017 fprintf (outfile, " }\n");
1020 void debug_bitmap_set (bitmap_set_t);
1022 DEBUG_FUNCTION void
1023 debug_bitmap_set (bitmap_set_t set)
1025 print_bitmap_set (stderr, set, "debug", 0);
1028 void debug_bitmap_sets_for (basic_block);
1030 DEBUG_FUNCTION void
1031 debug_bitmap_sets_for (basic_block bb)
1033 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1034 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1035 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1036 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1037 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1038 if (do_partial_partial)
1039 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1040 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1043 /* Print out the expressions that have VAL to OUTFILE. */
1045 static void
1046 print_value_expressions (FILE *outfile, unsigned int val)
1048 bitmap set = value_expressions[val];
1049 if (set)
1051 bitmap_set x;
1052 char s[10];
1053 sprintf (s, "%04d", val);
1054 x.expressions = *set;
1055 print_bitmap_set (outfile, &x, s, 0);
1060 DEBUG_FUNCTION void
1061 debug_value_expressions (unsigned int val)
1063 print_value_expressions (stderr, val);
1066 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1067 represent it. */
1069 static pre_expr
1070 get_or_alloc_expr_for_constant (tree constant)
1072 unsigned int result_id;
1073 unsigned int value_id;
1074 struct pre_expr_d expr;
1075 pre_expr newexpr;
1077 expr.kind = CONSTANT;
1078 PRE_EXPR_CONSTANT (&expr) = constant;
1079 result_id = lookup_expression_id (&expr);
1080 if (result_id != 0)
1081 return expression_for_id (result_id);
1083 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1084 newexpr->kind = CONSTANT;
1085 PRE_EXPR_CONSTANT (newexpr) = constant;
1086 alloc_expression_id (newexpr);
1087 value_id = get_or_alloc_constant_value_id (constant);
1088 add_to_value (value_id, newexpr);
1089 return newexpr;
1092 /* Given a value id V, find the actual tree representing the constant
1093 value if there is one, and return it. Return NULL if we can't find
1094 a constant. */
1096 static tree
1097 get_constant_for_value_id (unsigned int v)
1099 if (value_id_constant_p (v))
1101 unsigned int i;
1102 bitmap_iterator bi;
1103 bitmap exprset = value_expressions[v];
1105 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1107 pre_expr expr = expression_for_id (i);
1108 if (expr->kind == CONSTANT)
1109 return PRE_EXPR_CONSTANT (expr);
1112 return NULL;
1115 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1116 Currently only supports constants and SSA_NAMES. */
1117 static pre_expr
1118 get_or_alloc_expr_for (tree t)
1120 if (TREE_CODE (t) == SSA_NAME)
1121 return get_or_alloc_expr_for_name (t);
1122 else if (is_gimple_min_invariant (t))
1123 return get_or_alloc_expr_for_constant (t);
1124 else
1126 /* More complex expressions can result from SCCVN expression
1127 simplification that inserts values for them. As they all
1128 do not have VOPs the get handled by the nary ops struct. */
1129 vn_nary_op_t result;
1130 unsigned int result_id;
1131 vn_nary_op_lookup (t, &result);
1132 if (result != NULL)
1134 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1135 e->kind = NARY;
1136 PRE_EXPR_NARY (e) = result;
1137 result_id = lookup_expression_id (e);
1138 if (result_id != 0)
1140 pool_free (pre_expr_pool, e);
1141 e = expression_for_id (result_id);
1142 return e;
1144 alloc_expression_id (e);
1145 return e;
1148 return NULL;
1151 /* Return the folded version of T if T, when folded, is a gimple
1152 min_invariant. Otherwise, return T. */
1154 static pre_expr
1155 fully_constant_expression (pre_expr e)
1157 switch (e->kind)
1159 case CONSTANT:
1160 return e;
1161 case NARY:
1163 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1164 switch (TREE_CODE_CLASS (nary->opcode))
1166 case tcc_binary:
1167 case tcc_comparison:
1169 /* We have to go from trees to pre exprs to value ids to
1170 constants. */
1171 tree naryop0 = nary->op[0];
1172 tree naryop1 = nary->op[1];
1173 tree result;
1174 if (!is_gimple_min_invariant (naryop0))
1176 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1177 unsigned int vrep0 = get_expr_value_id (rep0);
1178 tree const0 = get_constant_for_value_id (vrep0);
1179 if (const0)
1180 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1182 if (!is_gimple_min_invariant (naryop1))
1184 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1185 unsigned int vrep1 = get_expr_value_id (rep1);
1186 tree const1 = get_constant_for_value_id (vrep1);
1187 if (const1)
1188 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1190 result = fold_binary (nary->opcode, nary->type,
1191 naryop0, naryop1);
1192 if (result && is_gimple_min_invariant (result))
1193 return get_or_alloc_expr_for_constant (result);
1194 /* We might have simplified the expression to a
1195 SSA_NAME for example from x_1 * 1. But we cannot
1196 insert a PHI for x_1 unconditionally as x_1 might
1197 not be available readily. */
1198 return e;
1200 case tcc_reference:
1201 if (nary->opcode != REALPART_EXPR
1202 && nary->opcode != IMAGPART_EXPR
1203 && nary->opcode != VIEW_CONVERT_EXPR)
1204 return e;
1205 /* Fallthrough. */
1206 case tcc_unary:
1208 /* We have to go from trees to pre exprs to value ids to
1209 constants. */
1210 tree naryop0 = nary->op[0];
1211 tree const0, result;
1212 if (is_gimple_min_invariant (naryop0))
1213 const0 = naryop0;
1214 else
1216 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1217 unsigned int vrep0 = get_expr_value_id (rep0);
1218 const0 = get_constant_for_value_id (vrep0);
1220 result = NULL;
1221 if (const0)
1223 tree type1 = TREE_TYPE (nary->op[0]);
1224 const0 = fold_convert (type1, const0);
1225 result = fold_unary (nary->opcode, nary->type, const0);
1227 if (result && is_gimple_min_invariant (result))
1228 return get_or_alloc_expr_for_constant (result);
1229 return e;
1231 default:
1232 return e;
1235 case REFERENCE:
1237 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1238 tree folded;
1239 if ((folded = fully_constant_vn_reference_p (ref)))
1240 return get_or_alloc_expr_for_constant (folded);
1241 return e;
1243 default:
1244 return e;
1246 return e;
1249 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1250 it has the value it would have in BLOCK. Set *SAME_VALID to true
1251 in case the new vuse doesn't change the value id of the OPERANDS. */
1253 static tree
1254 translate_vuse_through_block (vec<vn_reference_op_s> operands,
1255 alias_set_type set, tree type, tree vuse,
1256 basic_block phiblock,
1257 basic_block block, bool *same_valid)
1259 gimple phi = SSA_NAME_DEF_STMT (vuse);
1260 ao_ref ref;
1261 edge e = NULL;
1262 bool use_oracle;
1264 *same_valid = true;
1266 if (gimple_bb (phi) != phiblock)
1267 return vuse;
1269 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1271 /* Use the alias-oracle to find either the PHI node in this block,
1272 the first VUSE used in this block that is equivalent to vuse or
1273 the first VUSE which definition in this block kills the value. */
1274 if (gimple_code (phi) == GIMPLE_PHI)
1275 e = find_edge (block, phiblock);
1276 else if (use_oracle)
1277 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1279 vuse = gimple_vuse (phi);
1280 phi = SSA_NAME_DEF_STMT (vuse);
1281 if (gimple_bb (phi) != phiblock)
1282 return vuse;
1283 if (gimple_code (phi) == GIMPLE_PHI)
1285 e = find_edge (block, phiblock);
1286 break;
1289 else
1290 return NULL_TREE;
1292 if (e)
1294 if (use_oracle)
1296 bitmap visited = NULL;
1297 unsigned int cnt;
1298 /* Try to find a vuse that dominates this phi node by skipping
1299 non-clobbering statements. */
1300 vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false);
1301 if (visited)
1302 BITMAP_FREE (visited);
1304 else
1305 vuse = NULL_TREE;
1306 if (!vuse)
1308 /* If we didn't find any, the value ID can't stay the same,
1309 but return the translated vuse. */
1310 *same_valid = false;
1311 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1313 /* ??? We would like to return vuse here as this is the canonical
1314 upmost vdef that this reference is associated with. But during
1315 insertion of the references into the hash tables we only ever
1316 directly insert with their direct gimple_vuse, hence returning
1317 something else would make us not find the other expression. */
1318 return PHI_ARG_DEF (phi, e->dest_idx);
1321 return NULL_TREE;
1324 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1325 SET2. This is used to avoid making a set consisting of the union
1326 of PA_IN and ANTIC_IN during insert. */
1328 static inline pre_expr
1329 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1331 pre_expr result;
1333 result = bitmap_find_leader (set1, val);
1334 if (!result && set2)
1335 result = bitmap_find_leader (set2, val);
1336 return result;
1339 /* Get the tree type for our PRE expression e. */
1341 static tree
1342 get_expr_type (const pre_expr e)
1344 switch (e->kind)
1346 case NAME:
1347 return TREE_TYPE (PRE_EXPR_NAME (e));
1348 case CONSTANT:
1349 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1350 case REFERENCE:
1351 return PRE_EXPR_REFERENCE (e)->type;
1352 case NARY:
1353 return PRE_EXPR_NARY (e)->type;
1355 gcc_unreachable ();
1358 /* Get a representative SSA_NAME for a given expression.
1359 Since all of our sub-expressions are treated as values, we require
1360 them to be SSA_NAME's for simplicity.
1361 Prior versions of GVNPRE used to use "value handles" here, so that
1362 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1363 either case, the operands are really values (IE we do not expect
1364 them to be usable without finding leaders). */
1366 static tree
1367 get_representative_for (const pre_expr e)
1369 tree name;
1370 unsigned int value_id = get_expr_value_id (e);
1372 switch (e->kind)
1374 case NAME:
1375 return PRE_EXPR_NAME (e);
1376 case CONSTANT:
1377 return PRE_EXPR_CONSTANT (e);
1378 case NARY:
1379 case REFERENCE:
1381 /* Go through all of the expressions representing this value
1382 and pick out an SSA_NAME. */
1383 unsigned int i;
1384 bitmap_iterator bi;
1385 bitmap exprs = value_expressions[value_id];
1386 EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1388 pre_expr rep = expression_for_id (i);
1389 if (rep->kind == NAME)
1390 return PRE_EXPR_NAME (rep);
1391 else if (rep->kind == CONSTANT)
1392 return PRE_EXPR_CONSTANT (rep);
1395 break;
1398 /* If we reached here we couldn't find an SSA_NAME. This can
1399 happen when we've discovered a value that has never appeared in
1400 the program as set to an SSA_NAME, as the result of phi translation.
1401 Create one here.
1402 ??? We should be able to re-use this when we insert the statement
1403 to compute it. */
1404 name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1405 VN_INFO_GET (name)->value_id = value_id;
1406 VN_INFO (name)->valnum = name;
1407 /* ??? For now mark this SSA name for release by SCCVN. */
1408 VN_INFO (name)->needs_insertion = true;
1409 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1410 if (dump_file && (dump_flags & TDF_DETAILS))
1412 fprintf (dump_file, "Created SSA_NAME representative ");
1413 print_generic_expr (dump_file, name, 0);
1414 fprintf (dump_file, " for expression:");
1415 print_pre_expr (dump_file, e);
1416 fprintf (dump_file, " (%04d)\n", value_id);
1419 return name;
1424 static pre_expr
1425 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1426 basic_block pred, basic_block phiblock);
1428 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1429 the phis in PRED. Return NULL if we can't find a leader for each part
1430 of the translated expression. */
1432 static pre_expr
1433 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1434 basic_block pred, basic_block phiblock)
1436 switch (expr->kind)
1438 case NARY:
1440 unsigned int i;
1441 bool changed = false;
1442 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1443 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1444 sizeof_vn_nary_op (nary->length));
1445 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1447 for (i = 0; i < newnary->length; i++)
1449 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1450 continue;
1451 else
1453 pre_expr leader, result;
1454 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1455 leader = find_leader_in_sets (op_val_id, set1, set2);
1456 result = phi_translate (leader, set1, set2, pred, phiblock);
1457 if (result && result != leader)
1459 tree name = get_representative_for (result);
1460 if (!name)
1461 return NULL;
1462 newnary->op[i] = name;
1464 else if (!result)
1465 return NULL;
1467 changed |= newnary->op[i] != nary->op[i];
1470 if (changed)
1472 pre_expr constant;
1473 unsigned int new_val_id;
1475 tree result = vn_nary_op_lookup_pieces (newnary->length,
1476 newnary->opcode,
1477 newnary->type,
1478 &newnary->op[0],
1479 &nary);
1480 if (result && is_gimple_min_invariant (result))
1481 return get_or_alloc_expr_for_constant (result);
1483 expr = (pre_expr) pool_alloc (pre_expr_pool);
1484 expr->kind = NARY;
1485 expr->id = 0;
1486 if (nary)
1488 PRE_EXPR_NARY (expr) = nary;
1489 constant = fully_constant_expression (expr);
1490 if (constant != expr)
1491 return constant;
1493 new_val_id = nary->value_id;
1494 get_or_alloc_expression_id (expr);
1496 else
1498 new_val_id = get_next_value_id ();
1499 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
1500 nary = vn_nary_op_insert_pieces (newnary->length,
1501 newnary->opcode,
1502 newnary->type,
1503 &newnary->op[0],
1504 result, new_val_id);
1505 PRE_EXPR_NARY (expr) = nary;
1506 constant = fully_constant_expression (expr);
1507 if (constant != expr)
1508 return constant;
1509 get_or_alloc_expression_id (expr);
1511 add_to_value (new_val_id, expr);
1513 return expr;
1515 break;
1517 case REFERENCE:
1519 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1520 vec<vn_reference_op_s> operands = ref->operands;
1521 tree vuse = ref->vuse;
1522 tree newvuse = vuse;
1523 vec<vn_reference_op_s> newoperands = vNULL;
1524 bool changed = false, same_valid = true;
1525 unsigned int i, j, n;
1526 vn_reference_op_t operand;
1527 vn_reference_t newref;
1529 for (i = 0, j = 0;
1530 operands.iterate (i, &operand); i++, j++)
1532 pre_expr opresult;
1533 pre_expr leader;
1534 tree op[3];
1535 tree type = operand->type;
1536 vn_reference_op_s newop = *operand;
1537 op[0] = operand->op0;
1538 op[1] = operand->op1;
1539 op[2] = operand->op2;
1540 for (n = 0; n < 3; ++n)
1542 unsigned int op_val_id;
1543 if (!op[n])
1544 continue;
1545 if (TREE_CODE (op[n]) != SSA_NAME)
1547 /* We can't possibly insert these. */
1548 if (n != 0
1549 && !is_gimple_min_invariant (op[n]))
1550 break;
1551 continue;
1553 op_val_id = VN_INFO (op[n])->value_id;
1554 leader = find_leader_in_sets (op_val_id, set1, set2);
1555 if (!leader)
1556 break;
1557 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1558 if (!opresult)
1559 break;
1560 if (opresult != leader)
1562 tree name = get_representative_for (opresult);
1563 if (!name)
1564 break;
1565 changed |= name != op[n];
1566 op[n] = name;
1569 if (n != 3)
1571 newoperands.release ();
1572 return NULL;
1574 if (!newoperands.exists ())
1575 newoperands = operands.copy ();
1576 /* We may have changed from an SSA_NAME to a constant */
1577 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1578 newop.opcode = TREE_CODE (op[0]);
1579 newop.type = type;
1580 newop.op0 = op[0];
1581 newop.op1 = op[1];
1582 newop.op2 = op[2];
1583 /* If it transforms a non-constant ARRAY_REF into a constant
1584 one, adjust the constant offset. */
1585 if (newop.opcode == ARRAY_REF
1586 && newop.off == -1
1587 && TREE_CODE (op[0]) == INTEGER_CST
1588 && TREE_CODE (op[1]) == INTEGER_CST
1589 && TREE_CODE (op[2]) == INTEGER_CST)
1591 double_int off = tree_to_double_int (op[0]);
1592 off += -tree_to_double_int (op[1]);
1593 off *= tree_to_double_int (op[2]);
1594 if (off.fits_shwi ())
1595 newop.off = off.low;
1597 newoperands[j] = newop;
1598 /* If it transforms from an SSA_NAME to an address, fold with
1599 a preceding indirect reference. */
1600 if (j > 0 && op[0] && TREE_CODE (op[0]) == ADDR_EXPR
1601 && newoperands[j - 1].opcode == MEM_REF)
1602 vn_reference_fold_indirect (&newoperands, &j);
1604 if (i != operands.length ())
1606 newoperands.release ();
1607 return NULL;
1610 if (vuse)
1612 newvuse = translate_vuse_through_block (newoperands,
1613 ref->set, ref->type,
1614 vuse, phiblock, pred,
1615 &same_valid);
1616 if (newvuse == NULL_TREE)
1618 newoperands.release ();
1619 return NULL;
1623 if (changed || newvuse != vuse)
1625 unsigned int new_val_id;
1626 pre_expr constant;
1628 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1629 ref->type,
1630 newoperands,
1631 &newref, VN_WALK);
1632 if (result)
1633 newoperands.release ();
1635 /* We can always insert constants, so if we have a partial
1636 redundant constant load of another type try to translate it
1637 to a constant of appropriate type. */
1638 if (result && is_gimple_min_invariant (result))
1640 tree tem = result;
1641 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1643 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1644 if (tem && !is_gimple_min_invariant (tem))
1645 tem = NULL_TREE;
1647 if (tem)
1648 return get_or_alloc_expr_for_constant (tem);
1651 /* If we'd have to convert things we would need to validate
1652 if we can insert the translated expression. So fail
1653 here for now - we cannot insert an alias with a different
1654 type in the VN tables either, as that would assert. */
1655 if (result
1656 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1657 return NULL;
1658 else if (!result && newref
1659 && !useless_type_conversion_p (ref->type, newref->type))
1661 newoperands.release ();
1662 return NULL;
1665 expr = (pre_expr) pool_alloc (pre_expr_pool);
1666 expr->kind = REFERENCE;
1667 expr->id = 0;
1669 if (newref)
1671 PRE_EXPR_REFERENCE (expr) = newref;
1672 constant = fully_constant_expression (expr);
1673 if (constant != expr)
1674 return constant;
1676 new_val_id = newref->value_id;
1677 get_or_alloc_expression_id (expr);
1679 else
1681 if (changed || !same_valid)
1683 new_val_id = get_next_value_id ();
1684 value_expressions.safe_grow_cleared
1685 (get_max_value_id () + 1);
1687 else
1688 new_val_id = ref->value_id;
1689 newref = vn_reference_insert_pieces (newvuse, ref->set,
1690 ref->type,
1691 newoperands,
1692 result, new_val_id);
1693 newoperands.create (0);
1694 PRE_EXPR_REFERENCE (expr) = newref;
1695 constant = fully_constant_expression (expr);
1696 if (constant != expr)
1697 return constant;
1698 get_or_alloc_expression_id (expr);
1700 add_to_value (new_val_id, expr);
1702 newoperands.release ();
1703 return expr;
1705 break;
1707 case NAME:
1709 tree name = PRE_EXPR_NAME (expr);
1710 gimple def_stmt = SSA_NAME_DEF_STMT (name);
1711 /* If the SSA name is defined by a PHI node in this block,
1712 translate it. */
1713 if (gimple_code (def_stmt) == GIMPLE_PHI
1714 && gimple_bb (def_stmt) == phiblock)
1716 edge e = find_edge (pred, gimple_bb (def_stmt));
1717 tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1719 /* Handle constant. */
1720 if (is_gimple_min_invariant (def))
1721 return get_or_alloc_expr_for_constant (def);
1723 return get_or_alloc_expr_for_name (def);
1725 /* Otherwise return it unchanged - it will get cleaned if its
1726 value is not available in PREDs AVAIL_OUT set of expressions. */
1727 return expr;
1730 default:
1731 gcc_unreachable ();
1735 /* Wrapper around phi_translate_1 providing caching functionality. */
1737 static pre_expr
1738 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1739 basic_block pred, basic_block phiblock)
1741 expr_pred_trans_t slot = NULL;
1742 pre_expr phitrans;
1744 if (!expr)
1745 return NULL;
1747 /* Constants contain no values that need translation. */
1748 if (expr->kind == CONSTANT)
1749 return expr;
1751 if (value_id_constant_p (get_expr_value_id (expr)))
1752 return expr;
1754 /* Don't add translations of NAMEs as those are cheap to translate. */
1755 if (expr->kind != NAME)
1757 if (phi_trans_add (&slot, expr, pred))
1758 return slot->v;
1759 /* Store NULL for the value we want to return in the case of
1760 recursing. */
1761 slot->v = NULL;
1764 /* Translate. */
1765 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1767 if (slot)
1769 if (phitrans)
1770 slot->v = phitrans;
1771 else
1772 /* Remove failed translations again, they cause insert
1773 iteration to not pick up new opportunities reliably. */
1774 phi_translate_table.remove_elt_with_hash (slot, slot->hashcode);
1777 return phitrans;
1781 /* For each expression in SET, translate the values through phi nodes
1782 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1783 expressions in DEST. */
1785 static void
1786 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1787 basic_block phiblock)
1789 vec<pre_expr> exprs;
1790 pre_expr expr;
1791 int i;
1793 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1795 bitmap_set_copy (dest, set);
1796 return;
1799 exprs = sorted_array_from_bitmap_set (set);
1800 FOR_EACH_VEC_ELT (exprs, i, expr)
1802 pre_expr translated;
1803 translated = phi_translate (expr, set, NULL, pred, phiblock);
1804 if (!translated)
1805 continue;
1807 /* We might end up with multiple expressions from SET being
1808 translated to the same value. In this case we do not want
1809 to retain the NARY or REFERENCE expression but prefer a NAME
1810 which would be the leader. */
1811 if (translated->kind == NAME)
1812 bitmap_value_replace_in_set (dest, translated);
1813 else
1814 bitmap_value_insert_into_set (dest, translated);
1816 exprs.release ();
1819 /* Find the leader for a value (i.e., the name representing that
1820 value) in a given set, and return it. Return NULL if no leader
1821 is found. */
1823 static pre_expr
1824 bitmap_find_leader (bitmap_set_t set, unsigned int val)
1826 if (value_id_constant_p (val))
1828 unsigned int i;
1829 bitmap_iterator bi;
1830 bitmap exprset = value_expressions[val];
1832 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1834 pre_expr expr = expression_for_id (i);
1835 if (expr->kind == CONSTANT)
1836 return expr;
1839 if (bitmap_set_contains_value (set, val))
1841 /* Rather than walk the entire bitmap of expressions, and see
1842 whether any of them has the value we are looking for, we look
1843 at the reverse mapping, which tells us the set of expressions
1844 that have a given value (IE value->expressions with that
1845 value) and see if any of those expressions are in our set.
1846 The number of expressions per value is usually significantly
1847 less than the number of expressions in the set. In fact, for
1848 large testcases, doing it this way is roughly 5-10x faster
1849 than walking the bitmap.
1850 If this is somehow a significant lose for some cases, we can
1851 choose which set to walk based on which set is smaller. */
1852 unsigned int i;
1853 bitmap_iterator bi;
1854 bitmap exprset = value_expressions[val];
1856 EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1857 return expression_for_id (i);
1859 return NULL;
1862 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1863 BLOCK by seeing if it is not killed in the block. Note that we are
1864 only determining whether there is a store that kills it. Because
1865 of the order in which clean iterates over values, we are guaranteed
1866 that altered operands will have caused us to be eliminated from the
1867 ANTIC_IN set already. */
1869 static bool
1870 value_dies_in_block_x (pre_expr expr, basic_block block)
1872 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1873 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1874 gimple def;
1875 gimple_stmt_iterator gsi;
1876 unsigned id = get_expression_id (expr);
1877 bool res = false;
1878 ao_ref ref;
1880 if (!vuse)
1881 return false;
1883 /* Lookup a previously calculated result. */
1884 if (EXPR_DIES (block)
1885 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1886 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1888 /* A memory expression {e, VUSE} dies in the block if there is a
1889 statement that may clobber e. If, starting statement walk from the
1890 top of the basic block, a statement uses VUSE there can be no kill
1891 inbetween that use and the original statement that loaded {e, VUSE},
1892 so we can stop walking. */
1893 ref.base = NULL_TREE;
1894 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1896 tree def_vuse, def_vdef;
1897 def = gsi_stmt (gsi);
1898 def_vuse = gimple_vuse (def);
1899 def_vdef = gimple_vdef (def);
1901 /* Not a memory statement. */
1902 if (!def_vuse)
1903 continue;
1905 /* Not a may-def. */
1906 if (!def_vdef)
1908 /* A load with the same VUSE, we're done. */
1909 if (def_vuse == vuse)
1910 break;
1912 continue;
1915 /* Init ref only if we really need it. */
1916 if (ref.base == NULL_TREE
1917 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1918 refx->operands))
1920 res = true;
1921 break;
1923 /* If the statement may clobber expr, it dies. */
1924 if (stmt_may_clobber_ref_p_1 (def, &ref))
1926 res = true;
1927 break;
1931 /* Remember the result. */
1932 if (!EXPR_DIES (block))
1933 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1934 bitmap_set_bit (EXPR_DIES (block), id * 2);
1935 if (res)
1936 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1938 return res;
1942 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1943 contains its value-id. */
1945 static bool
1946 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1948 if (op && TREE_CODE (op) == SSA_NAME)
1950 unsigned int value_id = VN_INFO (op)->value_id;
1951 if (!(bitmap_set_contains_value (set1, value_id)
1952 || (set2 && bitmap_set_contains_value (set2, value_id))))
1953 return false;
1955 return true;
1958 /* Determine if the expression EXPR is valid in SET1 U SET2.
1959 ONLY SET2 CAN BE NULL.
1960 This means that we have a leader for each part of the expression
1961 (if it consists of values), or the expression is an SSA_NAME.
1962 For loads/calls, we also see if the vuse is killed in this block. */
1964 static bool
1965 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
1966 basic_block block)
1968 switch (expr->kind)
1970 case NAME:
1971 return bitmap_find_leader (AVAIL_OUT (block),
1972 get_expr_value_id (expr)) != NULL;
1973 case NARY:
1975 unsigned int i;
1976 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1977 for (i = 0; i < nary->length; i++)
1978 if (!op_valid_in_sets (set1, set2, nary->op[i]))
1979 return false;
1980 return true;
1982 break;
1983 case REFERENCE:
1985 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1986 vn_reference_op_t vro;
1987 unsigned int i;
1989 FOR_EACH_VEC_ELT (ref->operands, i, vro)
1991 if (!op_valid_in_sets (set1, set2, vro->op0)
1992 || !op_valid_in_sets (set1, set2, vro->op1)
1993 || !op_valid_in_sets (set1, set2, vro->op2))
1994 return false;
1996 return true;
1998 default:
1999 gcc_unreachable ();
2003 /* Clean the set of expressions that are no longer valid in SET1 or
2004 SET2. This means expressions that are made up of values we have no
2005 leaders for in SET1 or SET2. This version is used for partial
2006 anticipation, which means it is not valid in either ANTIC_IN or
2007 PA_IN. */
2009 static void
2010 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2012 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
2013 pre_expr expr;
2014 int i;
2016 FOR_EACH_VEC_ELT (exprs, i, expr)
2018 if (!valid_in_sets (set1, set2, expr, block))
2019 bitmap_remove_from_set (set1, expr);
2021 exprs.release ();
2024 /* Clean the set of expressions that are no longer valid in SET. This
2025 means expressions that are made up of values we have no leaders for
2026 in SET. */
2028 static void
2029 clean (bitmap_set_t set, basic_block block)
2031 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set);
2032 pre_expr expr;
2033 int i;
2035 FOR_EACH_VEC_ELT (exprs, i, expr)
2037 if (!valid_in_sets (set, NULL, expr, block))
2038 bitmap_remove_from_set (set, expr);
2040 exprs.release ();
2043 /* Clean the set of expressions that are no longer valid in SET because
2044 they are clobbered in BLOCK or because they trap and may not be executed. */
2046 static void
2047 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2049 bitmap_iterator bi;
2050 unsigned i;
2052 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2054 pre_expr expr = expression_for_id (i);
2055 if (expr->kind == REFERENCE)
2057 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2058 if (ref->vuse)
2060 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2061 if (!gimple_nop_p (def_stmt)
2062 && ((gimple_bb (def_stmt) != block
2063 && !dominated_by_p (CDI_DOMINATORS,
2064 block, gimple_bb (def_stmt)))
2065 || (gimple_bb (def_stmt) == block
2066 && value_dies_in_block_x (expr, block))))
2067 bitmap_remove_from_set (set, expr);
2070 else if (expr->kind == NARY)
2072 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2073 /* If the NARY may trap make sure the block does not contain
2074 a possible exit point.
2075 ??? This is overly conservative if we translate AVAIL_OUT
2076 as the available expression might be after the exit point. */
2077 if (BB_MAY_NOTRETURN (block)
2078 && vn_nary_may_trap (nary))
2079 bitmap_remove_from_set (set, expr);
2084 static sbitmap has_abnormal_preds;
2086 /* List of blocks that may have changed during ANTIC computation and
2087 thus need to be iterated over. */
2089 static sbitmap changed_blocks;
2091 /* Decide whether to defer a block for a later iteration, or PHI
2092 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2093 should defer the block, and true if we processed it. */
2095 static bool
2096 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2097 basic_block block, basic_block phiblock)
2099 if (!BB_VISITED (phiblock))
2101 bitmap_set_bit (changed_blocks, block->index);
2102 BB_VISITED (block) = 0;
2103 BB_DEFERRED (block) = 1;
2104 return false;
2106 else
2107 phi_translate_set (dest, source, block, phiblock);
2108 return true;
2111 /* Compute the ANTIC set for BLOCK.
2113 If succs(BLOCK) > 1 then
2114 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2115 else if succs(BLOCK) == 1 then
2116 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2118 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2121 static bool
2122 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2124 bool changed = false;
2125 bitmap_set_t S, old, ANTIC_OUT;
2126 bitmap_iterator bi;
2127 unsigned int bii;
2128 edge e;
2129 edge_iterator ei;
2131 old = ANTIC_OUT = S = NULL;
2132 BB_VISITED (block) = 1;
2134 /* If any edges from predecessors are abnormal, antic_in is empty,
2135 so do nothing. */
2136 if (block_has_abnormal_pred_edge)
2137 goto maybe_dump_sets;
2139 old = ANTIC_IN (block);
2140 ANTIC_OUT = bitmap_set_new ();
2142 /* If the block has no successors, ANTIC_OUT is empty. */
2143 if (EDGE_COUNT (block->succs) == 0)
2145 /* If we have one successor, we could have some phi nodes to
2146 translate through. */
2147 else if (single_succ_p (block))
2149 basic_block succ_bb = single_succ (block);
2151 /* We trade iterations of the dataflow equations for having to
2152 phi translate the maximal set, which is incredibly slow
2153 (since the maximal set often has 300+ members, even when you
2154 have a small number of blocks).
2155 Basically, we defer the computation of ANTIC for this block
2156 until we have processed it's successor, which will inevitably
2157 have a *much* smaller set of values to phi translate once
2158 clean has been run on it.
2159 The cost of doing this is that we technically perform more
2160 iterations, however, they are lower cost iterations.
2162 Timings for PRE on tramp3d-v4:
2163 without maximal set fix: 11 seconds
2164 with maximal set fix/without deferring: 26 seconds
2165 with maximal set fix/with deferring: 11 seconds
2168 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2169 block, succ_bb))
2171 changed = true;
2172 goto maybe_dump_sets;
2175 /* If we have multiple successors, we take the intersection of all of
2176 them. Note that in the case of loop exit phi nodes, we may have
2177 phis to translate through. */
2178 else
2180 vec<basic_block> worklist;
2181 size_t i;
2182 basic_block bprime, first = NULL;
2184 worklist.create (EDGE_COUNT (block->succs));
2185 FOR_EACH_EDGE (e, ei, block->succs)
2187 if (!first
2188 && BB_VISITED (e->dest))
2189 first = e->dest;
2190 else if (BB_VISITED (e->dest))
2191 worklist.quick_push (e->dest);
2194 /* Of multiple successors we have to have visited one already. */
2195 if (!first)
2197 bitmap_set_bit (changed_blocks, block->index);
2198 BB_VISITED (block) = 0;
2199 BB_DEFERRED (block) = 1;
2200 changed = true;
2201 worklist.release ();
2202 goto maybe_dump_sets;
2205 if (!gimple_seq_empty_p (phi_nodes (first)))
2206 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2207 else
2208 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2210 FOR_EACH_VEC_ELT (worklist, i, bprime)
2212 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2214 bitmap_set_t tmp = bitmap_set_new ();
2215 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2216 bitmap_set_and (ANTIC_OUT, tmp);
2217 bitmap_set_free (tmp);
2219 else
2220 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2222 worklist.release ();
2225 /* Prune expressions that are clobbered in block and thus become
2226 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2227 prune_clobbered_mems (ANTIC_OUT, block);
2229 /* Generate ANTIC_OUT - TMP_GEN. */
2230 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2232 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2233 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2234 TMP_GEN (block));
2236 /* Then union in the ANTIC_OUT - TMP_GEN values,
2237 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2238 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2239 bitmap_value_insert_into_set (ANTIC_IN (block),
2240 expression_for_id (bii));
2242 clean (ANTIC_IN (block), block);
2244 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2246 changed = true;
2247 bitmap_set_bit (changed_blocks, block->index);
2248 FOR_EACH_EDGE (e, ei, block->preds)
2249 bitmap_set_bit (changed_blocks, e->src->index);
2251 else
2252 bitmap_clear_bit (changed_blocks, block->index);
2254 maybe_dump_sets:
2255 if (dump_file && (dump_flags & TDF_DETAILS))
2257 if (!BB_DEFERRED (block) || BB_VISITED (block))
2259 if (ANTIC_OUT)
2260 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2262 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2263 block->index);
2265 if (S)
2266 print_bitmap_set (dump_file, S, "S", block->index);
2268 else
2270 fprintf (dump_file,
2271 "Block %d was deferred for a future iteration.\n",
2272 block->index);
2275 if (old)
2276 bitmap_set_free (old);
2277 if (S)
2278 bitmap_set_free (S);
2279 if (ANTIC_OUT)
2280 bitmap_set_free (ANTIC_OUT);
2281 return changed;
2284 /* Compute PARTIAL_ANTIC for BLOCK.
2286 If succs(BLOCK) > 1 then
2287 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2288 in ANTIC_OUT for all succ(BLOCK)
2289 else if succs(BLOCK) == 1 then
2290 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2292 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2293 - ANTIC_IN[BLOCK])
2296 static bool
2297 compute_partial_antic_aux (basic_block block,
2298 bool block_has_abnormal_pred_edge)
2300 bool changed = false;
2301 bitmap_set_t old_PA_IN;
2302 bitmap_set_t PA_OUT;
2303 edge e;
2304 edge_iterator ei;
2305 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2307 old_PA_IN = PA_OUT = NULL;
2309 /* If any edges from predecessors are abnormal, antic_in is empty,
2310 so do nothing. */
2311 if (block_has_abnormal_pred_edge)
2312 goto maybe_dump_sets;
2314 /* If there are too many partially anticipatable values in the
2315 block, phi_translate_set can take an exponential time: stop
2316 before the translation starts. */
2317 if (max_pa
2318 && single_succ_p (block)
2319 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2320 goto maybe_dump_sets;
2322 old_PA_IN = PA_IN (block);
2323 PA_OUT = bitmap_set_new ();
2325 /* If the block has no successors, ANTIC_OUT is empty. */
2326 if (EDGE_COUNT (block->succs) == 0)
2328 /* If we have one successor, we could have some phi nodes to
2329 translate through. Note that we can't phi translate across DFS
2330 back edges in partial antic, because it uses a union operation on
2331 the successors. For recurrences like IV's, we will end up
2332 generating a new value in the set on each go around (i + 3 (VH.1)
2333 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2334 else if (single_succ_p (block))
2336 basic_block succ = single_succ (block);
2337 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2338 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2340 /* If we have multiple successors, we take the union of all of
2341 them. */
2342 else
2344 vec<basic_block> worklist;
2345 size_t i;
2346 basic_block bprime;
2348 worklist.create (EDGE_COUNT (block->succs));
2349 FOR_EACH_EDGE (e, ei, block->succs)
2351 if (e->flags & EDGE_DFS_BACK)
2352 continue;
2353 worklist.quick_push (e->dest);
2355 if (worklist.length () > 0)
2357 FOR_EACH_VEC_ELT (worklist, i, bprime)
2359 unsigned int i;
2360 bitmap_iterator bi;
2362 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2363 bitmap_value_insert_into_set (PA_OUT,
2364 expression_for_id (i));
2365 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2367 bitmap_set_t pa_in = bitmap_set_new ();
2368 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2369 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2370 bitmap_value_insert_into_set (PA_OUT,
2371 expression_for_id (i));
2372 bitmap_set_free (pa_in);
2374 else
2375 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2376 bitmap_value_insert_into_set (PA_OUT,
2377 expression_for_id (i));
2380 worklist.release ();
2383 /* Prune expressions that are clobbered in block and thus become
2384 invalid if translated from PA_OUT to PA_IN. */
2385 prune_clobbered_mems (PA_OUT, block);
2387 /* PA_IN starts with PA_OUT - TMP_GEN.
2388 Then we subtract things from ANTIC_IN. */
2389 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2391 /* For partial antic, we want to put back in the phi results, since
2392 we will properly avoid making them partially antic over backedges. */
2393 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2394 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2396 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2397 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2399 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2401 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2403 changed = true;
2404 bitmap_set_bit (changed_blocks, block->index);
2405 FOR_EACH_EDGE (e, ei, block->preds)
2406 bitmap_set_bit (changed_blocks, e->src->index);
2408 else
2409 bitmap_clear_bit (changed_blocks, block->index);
2411 maybe_dump_sets:
2412 if (dump_file && (dump_flags & TDF_DETAILS))
2414 if (PA_OUT)
2415 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2417 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2419 if (old_PA_IN)
2420 bitmap_set_free (old_PA_IN);
2421 if (PA_OUT)
2422 bitmap_set_free (PA_OUT);
2423 return changed;
2426 /* Compute ANTIC and partial ANTIC sets. */
2428 static void
2429 compute_antic (void)
2431 bool changed = true;
2432 int num_iterations = 0;
2433 basic_block block;
2434 int i;
2436 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2437 We pre-build the map of blocks with incoming abnormal edges here. */
2438 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2439 bitmap_clear (has_abnormal_preds);
2441 FOR_ALL_BB (block)
2443 edge_iterator ei;
2444 edge e;
2446 FOR_EACH_EDGE (e, ei, block->preds)
2448 e->flags &= ~EDGE_DFS_BACK;
2449 if (e->flags & EDGE_ABNORMAL)
2451 bitmap_set_bit (has_abnormal_preds, block->index);
2452 break;
2456 BB_VISITED (block) = 0;
2457 BB_DEFERRED (block) = 0;
2459 /* While we are here, give empty ANTIC_IN sets to each block. */
2460 ANTIC_IN (block) = bitmap_set_new ();
2461 PA_IN (block) = bitmap_set_new ();
2464 /* At the exit block we anticipate nothing. */
2465 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2467 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2468 bitmap_ones (changed_blocks);
2469 while (changed)
2471 if (dump_file && (dump_flags & TDF_DETAILS))
2472 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2473 /* ??? We need to clear our PHI translation cache here as the
2474 ANTIC sets shrink and we restrict valid translations to
2475 those having operands with leaders in ANTIC. Same below
2476 for PA ANTIC computation. */
2477 num_iterations++;
2478 changed = false;
2479 for (i = postorder_num - 1; i >= 0; i--)
2481 if (bitmap_bit_p (changed_blocks, postorder[i]))
2483 basic_block block = BASIC_BLOCK (postorder[i]);
2484 changed |= compute_antic_aux (block,
2485 bitmap_bit_p (has_abnormal_preds,
2486 block->index));
2489 /* Theoretically possible, but *highly* unlikely. */
2490 gcc_checking_assert (num_iterations < 500);
2493 statistics_histogram_event (cfun, "compute_antic iterations",
2494 num_iterations);
2496 if (do_partial_partial)
2498 bitmap_ones (changed_blocks);
2499 mark_dfs_back_edges ();
2500 num_iterations = 0;
2501 changed = true;
2502 while (changed)
2504 if (dump_file && (dump_flags & TDF_DETAILS))
2505 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2506 num_iterations++;
2507 changed = false;
2508 for (i = postorder_num - 1 ; i >= 0; i--)
2510 if (bitmap_bit_p (changed_blocks, postorder[i]))
2512 basic_block block = BASIC_BLOCK (postorder[i]);
2513 changed
2514 |= compute_partial_antic_aux (block,
2515 bitmap_bit_p (has_abnormal_preds,
2516 block->index));
2519 /* Theoretically possible, but *highly* unlikely. */
2520 gcc_checking_assert (num_iterations < 500);
2522 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2523 num_iterations);
2525 sbitmap_free (has_abnormal_preds);
2526 sbitmap_free (changed_blocks);
2530 /* Inserted expressions are placed onto this worklist, which is used
2531 for performing quick dead code elimination of insertions we made
2532 that didn't turn out to be necessary. */
2533 static bitmap inserted_exprs;
2535 /* The actual worker for create_component_ref_by_pieces. */
2537 static tree
2538 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2539 unsigned int *operand, gimple_seq *stmts)
2541 vn_reference_op_t currop = &ref->operands[*operand];
2542 tree genop;
2543 ++*operand;
2544 switch (currop->opcode)
2546 case CALL_EXPR:
2548 tree folded, sc = NULL_TREE;
2549 unsigned int nargs = 0;
2550 tree fn, *args;
2551 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2552 fn = currop->op0;
2553 else
2554 fn = find_or_generate_expression (block, currop->op0, stmts);
2555 if (!fn)
2556 return NULL_TREE;
2557 if (currop->op1)
2559 sc = find_or_generate_expression (block, currop->op1, stmts);
2560 if (!sc)
2561 return NULL_TREE;
2563 args = XNEWVEC (tree, ref->operands.length () - 1);
2564 while (*operand < ref->operands.length ())
2566 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2567 operand, stmts);
2568 if (!args[nargs])
2569 return NULL_TREE;
2570 nargs++;
2572 folded = build_call_array (currop->type,
2573 (TREE_CODE (fn) == FUNCTION_DECL
2574 ? build_fold_addr_expr (fn) : fn),
2575 nargs, args);
2576 free (args);
2577 if (sc)
2578 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2579 return folded;
2582 case MEM_REF:
2584 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2585 stmts);
2586 if (!baseop)
2587 return NULL_TREE;
2588 tree offset = currop->op0;
2589 if (TREE_CODE (baseop) == ADDR_EXPR
2590 && handled_component_p (TREE_OPERAND (baseop, 0)))
2592 HOST_WIDE_INT off;
2593 tree base;
2594 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2595 &off);
2596 gcc_assert (base);
2597 offset = int_const_binop (PLUS_EXPR, offset,
2598 build_int_cst (TREE_TYPE (offset),
2599 off));
2600 baseop = build_fold_addr_expr (base);
2602 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2605 case TARGET_MEM_REF:
2607 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2608 vn_reference_op_t nextop = &ref->operands[++*operand];
2609 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2610 stmts);
2611 if (!baseop)
2612 return NULL_TREE;
2613 if (currop->op0)
2615 genop0 = find_or_generate_expression (block, currop->op0, stmts);
2616 if (!genop0)
2617 return NULL_TREE;
2619 if (nextop->op0)
2621 genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2622 if (!genop1)
2623 return NULL_TREE;
2625 return build5 (TARGET_MEM_REF, currop->type,
2626 baseop, currop->op2, genop0, currop->op1, genop1);
2629 case ADDR_EXPR:
2630 if (currop->op0)
2632 gcc_assert (is_gimple_min_invariant (currop->op0));
2633 return currop->op0;
2635 /* Fallthrough. */
2636 case REALPART_EXPR:
2637 case IMAGPART_EXPR:
2638 case VIEW_CONVERT_EXPR:
2640 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2641 stmts);
2642 if (!genop0)
2643 return NULL_TREE;
2644 return fold_build1 (currop->opcode, currop->type, genop0);
2647 case WITH_SIZE_EXPR:
2649 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2650 stmts);
2651 if (!genop0)
2652 return NULL_TREE;
2653 tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2654 if (!genop1)
2655 return NULL_TREE;
2656 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2659 case BIT_FIELD_REF:
2661 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2662 stmts);
2663 if (!genop0)
2664 return NULL_TREE;
2665 tree op1 = currop->op0;
2666 tree op2 = currop->op1;
2667 return fold_build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2670 /* For array ref vn_reference_op's, operand 1 of the array ref
2671 is op0 of the reference op and operand 3 of the array ref is
2672 op1. */
2673 case ARRAY_RANGE_REF:
2674 case ARRAY_REF:
2676 tree genop0;
2677 tree genop1 = currop->op0;
2678 tree genop2 = currop->op1;
2679 tree genop3 = currop->op2;
2680 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2681 stmts);
2682 if (!genop0)
2683 return NULL_TREE;
2684 genop1 = find_or_generate_expression (block, genop1, stmts);
2685 if (!genop1)
2686 return NULL_TREE;
2687 if (genop2)
2689 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2690 /* Drop zero minimum index if redundant. */
2691 if (integer_zerop (genop2)
2692 && (!domain_type
2693 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2694 genop2 = NULL_TREE;
2695 else
2697 genop2 = find_or_generate_expression (block, genop2, stmts);
2698 if (!genop2)
2699 return NULL_TREE;
2702 if (genop3)
2704 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2705 /* We can't always put a size in units of the element alignment
2706 here as the element alignment may be not visible. See
2707 PR43783. Simply drop the element size for constant
2708 sizes. */
2709 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2710 genop3 = NULL_TREE;
2711 else
2713 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2714 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2715 genop3 = find_or_generate_expression (block, genop3, stmts);
2716 if (!genop3)
2717 return NULL_TREE;
2720 return build4 (currop->opcode, currop->type, genop0, genop1,
2721 genop2, genop3);
2723 case COMPONENT_REF:
2725 tree op0;
2726 tree op1;
2727 tree genop2 = currop->op1;
2728 op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2729 if (!op0)
2730 return NULL_TREE;
2731 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2732 op1 = currop->op0;
2733 if (genop2)
2735 genop2 = find_or_generate_expression (block, genop2, stmts);
2736 if (!genop2)
2737 return NULL_TREE;
2739 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2742 case SSA_NAME:
2744 genop = find_or_generate_expression (block, currop->op0, stmts);
2745 return genop;
2747 case STRING_CST:
2748 case INTEGER_CST:
2749 case COMPLEX_CST:
2750 case VECTOR_CST:
2751 case REAL_CST:
2752 case CONSTRUCTOR:
2753 case VAR_DECL:
2754 case PARM_DECL:
2755 case CONST_DECL:
2756 case RESULT_DECL:
2757 case FUNCTION_DECL:
2758 return currop->op0;
2760 default:
2761 gcc_unreachable ();
2765 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2766 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2767 trying to rename aggregates into ssa form directly, which is a no no.
2769 Thus, this routine doesn't create temporaries, it just builds a
2770 single access expression for the array, calling
2771 find_or_generate_expression to build the innermost pieces.
2773 This function is a subroutine of create_expression_by_pieces, and
2774 should not be called on it's own unless you really know what you
2775 are doing. */
2777 static tree
2778 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2779 gimple_seq *stmts)
2781 unsigned int op = 0;
2782 return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2785 /* Find a simple leader for an expression, or generate one using
2786 create_expression_by_pieces from a NARY expression for the value.
2787 BLOCK is the basic_block we are looking for leaders in.
2788 OP is the tree expression to find a leader for or generate.
2789 Returns the leader or NULL_TREE on failure. */
2791 static tree
2792 find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2794 pre_expr expr = get_or_alloc_expr_for (op);
2795 unsigned int lookfor = get_expr_value_id (expr);
2796 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2797 if (leader)
2799 if (leader->kind == NAME)
2800 return PRE_EXPR_NAME (leader);
2801 else if (leader->kind == CONSTANT)
2802 return PRE_EXPR_CONSTANT (leader);
2804 /* Defer. */
2805 return NULL_TREE;
2808 /* It must be a complex expression, so generate it recursively. Note
2809 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2810 where the insert algorithm fails to insert a required expression. */
2811 bitmap exprset = value_expressions[lookfor];
2812 bitmap_iterator bi;
2813 unsigned int i;
2814 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2816 pre_expr temp = expression_for_id (i);
2817 /* We cannot insert random REFERENCE expressions at arbitrary
2818 places. We can insert NARYs which eventually re-materializes
2819 its operand values. */
2820 if (temp->kind == NARY)
2821 return create_expression_by_pieces (block, temp, stmts,
2822 get_expr_type (expr));
2825 /* Defer. */
2826 return NULL_TREE;
2829 #define NECESSARY GF_PLF_1
2831 /* Create an expression in pieces, so that we can handle very complex
2832 expressions that may be ANTIC, but not necessary GIMPLE.
2833 BLOCK is the basic block the expression will be inserted into,
2834 EXPR is the expression to insert (in value form)
2835 STMTS is a statement list to append the necessary insertions into.
2837 This function will die if we hit some value that shouldn't be
2838 ANTIC but is (IE there is no leader for it, or its components).
2839 The function returns NULL_TREE in case a different antic expression
2840 has to be inserted first.
2841 This function may also generate expressions that are themselves
2842 partially or fully redundant. Those that are will be either made
2843 fully redundant during the next iteration of insert (for partially
2844 redundant ones), or eliminated by eliminate (for fully redundant
2845 ones). */
2847 static tree
2848 create_expression_by_pieces (basic_block block, pre_expr expr,
2849 gimple_seq *stmts, tree type)
2851 tree name;
2852 tree folded;
2853 gimple_seq forced_stmts = NULL;
2854 unsigned int value_id;
2855 gimple_stmt_iterator gsi;
2856 tree exprtype = type ? type : get_expr_type (expr);
2857 pre_expr nameexpr;
2858 gimple newstmt;
2860 switch (expr->kind)
2862 /* We may hit the NAME/CONSTANT case if we have to convert types
2863 that value numbering saw through. */
2864 case NAME:
2865 folded = PRE_EXPR_NAME (expr);
2866 break;
2867 case CONSTANT:
2868 folded = PRE_EXPR_CONSTANT (expr);
2869 break;
2870 case REFERENCE:
2872 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2873 folded = create_component_ref_by_pieces (block, ref, stmts);
2874 if (!folded)
2875 return NULL_TREE;
2877 break;
2878 case NARY:
2880 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2881 tree *genop = XALLOCAVEC (tree, nary->length);
2882 unsigned i;
2883 for (i = 0; i < nary->length; ++i)
2885 genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2886 if (!genop[i])
2887 return NULL_TREE;
2888 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2889 may have conversions stripped. */
2890 if (nary->opcode == POINTER_PLUS_EXPR)
2892 if (i == 0)
2893 genop[i] = fold_convert (nary->type, genop[i]);
2894 else if (i == 1)
2895 genop[i] = convert_to_ptrofftype (genop[i]);
2897 else
2898 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]);
2900 if (nary->opcode == CONSTRUCTOR)
2902 vec<constructor_elt, va_gc> *elts = NULL;
2903 for (i = 0; i < nary->length; ++i)
2904 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2905 folded = build_constructor (nary->type, elts);
2907 else
2909 switch (nary->length)
2911 case 1:
2912 folded = fold_build1 (nary->opcode, nary->type,
2913 genop[0]);
2914 break;
2915 case 2:
2916 folded = fold_build2 (nary->opcode, nary->type,
2917 genop[0], genop[1]);
2918 break;
2919 case 3:
2920 folded = fold_build3 (nary->opcode, nary->type,
2921 genop[0], genop[1], genop[2]);
2922 break;
2923 default:
2924 gcc_unreachable ();
2928 break;
2929 default:
2930 gcc_unreachable ();
2933 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2934 folded = fold_convert (exprtype, folded);
2936 /* Force the generated expression to be a sequence of GIMPLE
2937 statements.
2938 We have to call unshare_expr because force_gimple_operand may
2939 modify the tree we pass to it. */
2940 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
2941 false, NULL);
2943 /* If we have any intermediate expressions to the value sets, add them
2944 to the value sets and chain them in the instruction stream. */
2945 if (forced_stmts)
2947 gsi = gsi_start (forced_stmts);
2948 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2950 gimple stmt = gsi_stmt (gsi);
2951 tree forcedname = gimple_get_lhs (stmt);
2952 pre_expr nameexpr;
2954 if (TREE_CODE (forcedname) == SSA_NAME)
2956 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
2957 VN_INFO_GET (forcedname)->valnum = forcedname;
2958 VN_INFO (forcedname)->value_id = get_next_value_id ();
2959 nameexpr = get_or_alloc_expr_for_name (forcedname);
2960 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2961 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2962 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2965 gimple_seq_add_seq (stmts, forced_stmts);
2968 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2969 newstmt = gimple_build_assign (name, folded);
2970 gimple_set_plf (newstmt, NECESSARY, false);
2972 gimple_seq_add_stmt (stmts, newstmt);
2973 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
2975 /* Fold the last statement. */
2976 gsi = gsi_last (*stmts);
2977 if (fold_stmt_inplace (&gsi))
2978 update_stmt (gsi_stmt (gsi));
2980 /* Add a value number to the temporary.
2981 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2982 we are creating the expression by pieces, and this particular piece of
2983 the expression may have been represented. There is no harm in replacing
2984 here. */
2985 value_id = get_expr_value_id (expr);
2986 VN_INFO_GET (name)->value_id = value_id;
2987 VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id);
2988 if (VN_INFO (name)->valnum == NULL_TREE)
2989 VN_INFO (name)->valnum = name;
2990 gcc_assert (VN_INFO (name)->valnum != NULL_TREE);
2991 nameexpr = get_or_alloc_expr_for_name (name);
2992 add_to_value (value_id, nameexpr);
2993 if (NEW_SETS (block))
2994 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2995 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2997 pre_stats.insertions++;
2998 if (dump_file && (dump_flags & TDF_DETAILS))
3000 fprintf (dump_file, "Inserted ");
3001 print_gimple_stmt (dump_file, newstmt, 0, 0);
3002 fprintf (dump_file, " in predecessor %d (%04d)\n",
3003 block->index, value_id);
3006 return name;
3010 /* Returns true if we want to inhibit the insertions of PHI nodes
3011 for the given EXPR for basic block BB (a member of a loop).
3012 We want to do this, when we fear that the induction variable we
3013 create might inhibit vectorization. */
3015 static bool
3016 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3018 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3019 vec<vn_reference_op_s> ops = vr->operands;
3020 vn_reference_op_t op;
3021 unsigned i;
3023 /* If we aren't going to vectorize we don't inhibit anything. */
3024 if (!flag_tree_loop_vectorize)
3025 return false;
3027 /* Otherwise we inhibit the insertion when the address of the
3028 memory reference is a simple induction variable. In other
3029 cases the vectorizer won't do anything anyway (either it's
3030 loop invariant or a complicated expression). */
3031 FOR_EACH_VEC_ELT (ops, i, op)
3033 switch (op->opcode)
3035 case CALL_EXPR:
3036 /* Calls are not a problem. */
3037 return false;
3039 case ARRAY_REF:
3040 case ARRAY_RANGE_REF:
3041 if (TREE_CODE (op->op0) != SSA_NAME)
3042 break;
3043 /* Fallthru. */
3044 case SSA_NAME:
3046 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3047 affine_iv iv;
3048 /* Default defs are loop invariant. */
3049 if (!defbb)
3050 break;
3051 /* Defined outside this loop, also loop invariant. */
3052 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3053 break;
3054 /* If it's a simple induction variable inhibit insertion,
3055 the vectorizer might be interested in this one. */
3056 if (simple_iv (bb->loop_father, bb->loop_father,
3057 op->op0, &iv, true))
3058 return true;
3059 /* No simple IV, vectorizer can't do anything, hence no
3060 reason to inhibit the transformation for this operand. */
3061 break;
3063 default:
3064 break;
3067 return false;
3070 /* Insert the to-be-made-available values of expression EXPRNUM for each
3071 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3072 merge the result with a phi node, given the same value number as
3073 NODE. Return true if we have inserted new stuff. */
3075 static bool
3076 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3077 vec<pre_expr> avail)
3079 pre_expr expr = expression_for_id (exprnum);
3080 pre_expr newphi;
3081 unsigned int val = get_expr_value_id (expr);
3082 edge pred;
3083 bool insertions = false;
3084 bool nophi = false;
3085 basic_block bprime;
3086 pre_expr eprime;
3087 edge_iterator ei;
3088 tree type = get_expr_type (expr);
3089 tree temp;
3090 gimple phi;
3092 /* Make sure we aren't creating an induction variable. */
3093 if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
3095 bool firstinsideloop = false;
3096 bool secondinsideloop = false;
3097 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3098 EDGE_PRED (block, 0)->src);
3099 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3100 EDGE_PRED (block, 1)->src);
3101 /* Induction variables only have one edge inside the loop. */
3102 if ((firstinsideloop ^ secondinsideloop)
3103 && (expr->kind != REFERENCE
3104 || inhibit_phi_insertion (block, expr)))
3106 if (dump_file && (dump_flags & TDF_DETAILS))
3107 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3108 nophi = true;
3112 /* Make the necessary insertions. */
3113 FOR_EACH_EDGE (pred, ei, block->preds)
3115 gimple_seq stmts = NULL;
3116 tree builtexpr;
3117 bprime = pred->src;
3118 eprime = avail[pred->dest_idx];
3120 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3122 builtexpr = create_expression_by_pieces (bprime, eprime,
3123 &stmts, type);
3124 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3125 gsi_insert_seq_on_edge (pred, stmts);
3126 if (!builtexpr)
3128 /* We cannot insert a PHI node if we failed to insert
3129 on one edge. */
3130 nophi = true;
3131 continue;
3133 avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3134 insertions = true;
3136 else if (eprime->kind == CONSTANT)
3138 /* Constants may not have the right type, fold_convert
3139 should give us back a constant with the right type. */
3140 tree constant = PRE_EXPR_CONSTANT (eprime);
3141 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3143 tree builtexpr = fold_convert (type, constant);
3144 if (!is_gimple_min_invariant (builtexpr))
3146 tree forcedexpr = force_gimple_operand (builtexpr,
3147 &stmts, true,
3148 NULL);
3149 if (!is_gimple_min_invariant (forcedexpr))
3151 if (forcedexpr != builtexpr)
3153 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3154 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3156 if (stmts)
3158 gimple_stmt_iterator gsi;
3159 gsi = gsi_start (stmts);
3160 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3162 gimple stmt = gsi_stmt (gsi);
3163 tree lhs = gimple_get_lhs (stmt);
3164 if (TREE_CODE (lhs) == SSA_NAME)
3165 bitmap_set_bit (inserted_exprs,
3166 SSA_NAME_VERSION (lhs));
3167 gimple_set_plf (stmt, NECESSARY, false);
3169 gsi_insert_seq_on_edge (pred, stmts);
3171 avail[pred->dest_idx]
3172 = get_or_alloc_expr_for_name (forcedexpr);
3175 else
3176 avail[pred->dest_idx]
3177 = get_or_alloc_expr_for_constant (builtexpr);
3180 else if (eprime->kind == NAME)
3182 /* We may have to do a conversion because our value
3183 numbering can look through types in certain cases, but
3184 our IL requires all operands of a phi node have the same
3185 type. */
3186 tree name = PRE_EXPR_NAME (eprime);
3187 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3189 tree builtexpr;
3190 tree forcedexpr;
3191 builtexpr = fold_convert (type, name);
3192 forcedexpr = force_gimple_operand (builtexpr,
3193 &stmts, true,
3194 NULL);
3196 if (forcedexpr != name)
3198 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3199 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3202 if (stmts)
3204 gimple_stmt_iterator gsi;
3205 gsi = gsi_start (stmts);
3206 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3208 gimple stmt = gsi_stmt (gsi);
3209 tree lhs = gimple_get_lhs (stmt);
3210 if (TREE_CODE (lhs) == SSA_NAME)
3211 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3212 gimple_set_plf (stmt, NECESSARY, false);
3214 gsi_insert_seq_on_edge (pred, stmts);
3216 avail[pred->dest_idx] = get_or_alloc_expr_for_name (forcedexpr);
3220 /* If we didn't want a phi node, and we made insertions, we still have
3221 inserted new stuff, and thus return true. If we didn't want a phi node,
3222 and didn't make insertions, we haven't added anything new, so return
3223 false. */
3224 if (nophi && insertions)
3225 return true;
3226 else if (nophi && !insertions)
3227 return false;
3229 /* Now build a phi for the new variable. */
3230 temp = make_temp_ssa_name (type, NULL, "prephitmp");
3231 phi = create_phi_node (temp, block);
3233 gimple_set_plf (phi, NECESSARY, false);
3234 VN_INFO_GET (temp)->value_id = val;
3235 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3236 if (VN_INFO (temp)->valnum == NULL_TREE)
3237 VN_INFO (temp)->valnum = temp;
3238 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3239 FOR_EACH_EDGE (pred, ei, block->preds)
3241 pre_expr ae = avail[pred->dest_idx];
3242 gcc_assert (get_expr_type (ae) == type
3243 || useless_type_conversion_p (type, get_expr_type (ae)));
3244 if (ae->kind == CONSTANT)
3245 add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3246 pred, UNKNOWN_LOCATION);
3247 else
3248 add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3251 newphi = get_or_alloc_expr_for_name (temp);
3252 add_to_value (val, newphi);
3254 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3255 this insertion, since we test for the existence of this value in PHI_GEN
3256 before proceeding with the partial redundancy checks in insert_aux.
3258 The value may exist in AVAIL_OUT, in particular, it could be represented
3259 by the expression we are trying to eliminate, in which case we want the
3260 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3261 inserted there.
3263 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3264 this block, because if it did, it would have existed in our dominator's
3265 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3268 bitmap_insert_into_set (PHI_GEN (block), newphi);
3269 bitmap_value_replace_in_set (AVAIL_OUT (block),
3270 newphi);
3271 bitmap_insert_into_set (NEW_SETS (block),
3272 newphi);
3274 if (dump_file && (dump_flags & TDF_DETAILS))
3276 fprintf (dump_file, "Created phi ");
3277 print_gimple_stmt (dump_file, phi, 0, 0);
3278 fprintf (dump_file, " in block %d (%04d)\n", block->index, val);
3280 pre_stats.phis++;
3281 return true;
3286 /* Perform insertion of partially redundant values.
3287 For BLOCK, do the following:
3288 1. Propagate the NEW_SETS of the dominator into the current block.
3289 If the block has multiple predecessors,
3290 2a. Iterate over the ANTIC expressions for the block to see if
3291 any of them are partially redundant.
3292 2b. If so, insert them into the necessary predecessors to make
3293 the expression fully redundant.
3294 2c. Insert a new PHI merging the values of the predecessors.
3295 2d. Insert the new PHI, and the new expressions, into the
3296 NEW_SETS set.
3297 3. Recursively call ourselves on the dominator children of BLOCK.
3299 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3300 do_regular_insertion and do_partial_insertion.
3304 static bool
3305 do_regular_insertion (basic_block block, basic_block dom)
3307 bool new_stuff = false;
3308 vec<pre_expr> exprs;
3309 pre_expr expr;
3310 vec<pre_expr> avail = vNULL;
3311 int i;
3313 exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3314 avail.safe_grow (EDGE_COUNT (block->preds));
3316 FOR_EACH_VEC_ELT (exprs, i, expr)
3318 if (expr->kind == NARY
3319 || expr->kind == REFERENCE)
3321 unsigned int val;
3322 bool by_some = false;
3323 bool cant_insert = false;
3324 bool all_same = true;
3325 pre_expr first_s = NULL;
3326 edge pred;
3327 basic_block bprime;
3328 pre_expr eprime = NULL;
3329 edge_iterator ei;
3330 pre_expr edoubleprime = NULL;
3331 bool do_insertion = false;
3333 val = get_expr_value_id (expr);
3334 if (bitmap_set_contains_value (PHI_GEN (block), val))
3335 continue;
3336 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3338 if (dump_file && (dump_flags & TDF_DETAILS))
3340 fprintf (dump_file, "Found fully redundant value: ");
3341 print_pre_expr (dump_file, expr);
3342 fprintf (dump_file, "\n");
3344 continue;
3347 FOR_EACH_EDGE (pred, ei, block->preds)
3349 unsigned int vprime;
3351 /* We should never run insertion for the exit block
3352 and so not come across fake pred edges. */
3353 gcc_assert (!(pred->flags & EDGE_FAKE));
3354 bprime = pred->src;
3355 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3356 bprime, block);
3358 /* eprime will generally only be NULL if the
3359 value of the expression, translated
3360 through the PHI for this predecessor, is
3361 undefined. If that is the case, we can't
3362 make the expression fully redundant,
3363 because its value is undefined along a
3364 predecessor path. We can thus break out
3365 early because it doesn't matter what the
3366 rest of the results are. */
3367 if (eprime == NULL)
3369 avail[pred->dest_idx] = NULL;
3370 cant_insert = true;
3371 break;
3374 eprime = fully_constant_expression (eprime);
3375 vprime = get_expr_value_id (eprime);
3376 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3377 vprime);
3378 if (edoubleprime == NULL)
3380 avail[pred->dest_idx] = eprime;
3381 all_same = false;
3383 else
3385 avail[pred->dest_idx] = edoubleprime;
3386 by_some = true;
3387 /* We want to perform insertions to remove a redundancy on
3388 a path in the CFG we want to optimize for speed. */
3389 if (optimize_edge_for_speed_p (pred))
3390 do_insertion = true;
3391 if (first_s == NULL)
3392 first_s = edoubleprime;
3393 else if (!pre_expr_d::equal (first_s, edoubleprime))
3394 all_same = false;
3397 /* If we can insert it, it's not the same value
3398 already existing along every predecessor, and
3399 it's defined by some predecessor, it is
3400 partially redundant. */
3401 if (!cant_insert && !all_same && by_some)
3403 if (!do_insertion)
3405 if (dump_file && (dump_flags & TDF_DETAILS))
3407 fprintf (dump_file, "Skipping partial redundancy for "
3408 "expression ");
3409 print_pre_expr (dump_file, expr);
3410 fprintf (dump_file, " (%04d), no redundancy on to be "
3411 "optimized for speed edge\n", val);
3414 else if (dbg_cnt (treepre_insert))
3416 if (dump_file && (dump_flags & TDF_DETAILS))
3418 fprintf (dump_file, "Found partial redundancy for "
3419 "expression ");
3420 print_pre_expr (dump_file, expr);
3421 fprintf (dump_file, " (%04d)\n",
3422 get_expr_value_id (expr));
3424 if (insert_into_preds_of_block (block,
3425 get_expression_id (expr),
3426 avail))
3427 new_stuff = true;
3430 /* If all edges produce the same value and that value is
3431 an invariant, then the PHI has the same value on all
3432 edges. Note this. */
3433 else if (!cant_insert && all_same)
3435 gcc_assert (edoubleprime->kind == CONSTANT
3436 || edoubleprime->kind == NAME);
3438 tree temp = make_temp_ssa_name (get_expr_type (expr),
3439 NULL, "pretmp");
3440 gimple assign = gimple_build_assign (temp,
3441 edoubleprime->kind == CONSTANT ? PRE_EXPR_CONSTANT (edoubleprime) : PRE_EXPR_NAME (edoubleprime));
3442 gimple_stmt_iterator gsi = gsi_after_labels (block);
3443 gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3445 gimple_set_plf (assign, NECESSARY, false);
3446 VN_INFO_GET (temp)->value_id = val;
3447 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3448 if (VN_INFO (temp)->valnum == NULL_TREE)
3449 VN_INFO (temp)->valnum = temp;
3450 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3451 pre_expr newe = get_or_alloc_expr_for_name (temp);
3452 add_to_value (val, newe);
3453 bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3454 bitmap_insert_into_set (NEW_SETS (block), newe);
3459 exprs.release ();
3460 avail.release ();
3461 return new_stuff;
3465 /* Perform insertion for partially anticipatable expressions. There
3466 is only one case we will perform insertion for these. This case is
3467 if the expression is partially anticipatable, and fully available.
3468 In this case, we know that putting it earlier will enable us to
3469 remove the later computation. */
3472 static bool
3473 do_partial_partial_insertion (basic_block block, basic_block dom)
3475 bool new_stuff = false;
3476 vec<pre_expr> exprs;
3477 pre_expr expr;
3478 vec<pre_expr> avail = vNULL;
3479 int i;
3481 exprs = sorted_array_from_bitmap_set (PA_IN (block));
3482 avail.safe_grow (EDGE_COUNT (block->preds));
3484 FOR_EACH_VEC_ELT (exprs, i, expr)
3486 if (expr->kind == NARY
3487 || expr->kind == REFERENCE)
3489 unsigned int val;
3490 bool by_all = true;
3491 bool cant_insert = false;
3492 edge pred;
3493 basic_block bprime;
3494 pre_expr eprime = NULL;
3495 edge_iterator ei;
3497 val = get_expr_value_id (expr);
3498 if (bitmap_set_contains_value (PHI_GEN (block), val))
3499 continue;
3500 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3501 continue;
3503 FOR_EACH_EDGE (pred, ei, block->preds)
3505 unsigned int vprime;
3506 pre_expr edoubleprime;
3508 /* We should never run insertion for the exit block
3509 and so not come across fake pred edges. */
3510 gcc_assert (!(pred->flags & EDGE_FAKE));
3511 bprime = pred->src;
3512 eprime = phi_translate (expr, ANTIC_IN (block),
3513 PA_IN (block),
3514 bprime, block);
3516 /* eprime will generally only be NULL if the
3517 value of the expression, translated
3518 through the PHI for this predecessor, is
3519 undefined. If that is the case, we can't
3520 make the expression fully redundant,
3521 because its value is undefined along a
3522 predecessor path. We can thus break out
3523 early because it doesn't matter what the
3524 rest of the results are. */
3525 if (eprime == NULL)
3527 avail[pred->dest_idx] = NULL;
3528 cant_insert = true;
3529 break;
3532 eprime = fully_constant_expression (eprime);
3533 vprime = get_expr_value_id (eprime);
3534 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3535 avail[pred->dest_idx] = edoubleprime;
3536 if (edoubleprime == NULL)
3538 by_all = false;
3539 break;
3543 /* If we can insert it, it's not the same value
3544 already existing along every predecessor, and
3545 it's defined by some predecessor, it is
3546 partially redundant. */
3547 if (!cant_insert && by_all)
3549 edge succ;
3550 bool do_insertion = false;
3552 /* Insert only if we can remove a later expression on a path
3553 that we want to optimize for speed.
3554 The phi node that we will be inserting in BLOCK is not free,
3555 and inserting it for the sake of !optimize_for_speed successor
3556 may cause regressions on the speed path. */
3557 FOR_EACH_EDGE (succ, ei, block->succs)
3559 if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3560 || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3562 if (optimize_edge_for_speed_p (succ))
3563 do_insertion = true;
3567 if (!do_insertion)
3569 if (dump_file && (dump_flags & TDF_DETAILS))
3571 fprintf (dump_file, "Skipping partial partial redundancy "
3572 "for expression ");
3573 print_pre_expr (dump_file, expr);
3574 fprintf (dump_file, " (%04d), not (partially) anticipated "
3575 "on any to be optimized for speed edges\n", val);
3578 else if (dbg_cnt (treepre_insert))
3580 pre_stats.pa_insert++;
3581 if (dump_file && (dump_flags & TDF_DETAILS))
3583 fprintf (dump_file, "Found partial partial redundancy "
3584 "for expression ");
3585 print_pre_expr (dump_file, expr);
3586 fprintf (dump_file, " (%04d)\n",
3587 get_expr_value_id (expr));
3589 if (insert_into_preds_of_block (block,
3590 get_expression_id (expr),
3591 avail))
3592 new_stuff = true;
3598 exprs.release ();
3599 avail.release ();
3600 return new_stuff;
3603 static bool
3604 insert_aux (basic_block block)
3606 basic_block son;
3607 bool new_stuff = false;
3609 if (block)
3611 basic_block dom;
3612 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3613 if (dom)
3615 unsigned i;
3616 bitmap_iterator bi;
3617 bitmap_set_t newset = NEW_SETS (dom);
3618 if (newset)
3620 /* Note that we need to value_replace both NEW_SETS, and
3621 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3622 represented by some non-simple expression here that we want
3623 to replace it with. */
3624 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3626 pre_expr expr = expression_for_id (i);
3627 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3628 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3631 if (!single_pred_p (block))
3633 new_stuff |= do_regular_insertion (block, dom);
3634 if (do_partial_partial)
3635 new_stuff |= do_partial_partial_insertion (block, dom);
3639 for (son = first_dom_son (CDI_DOMINATORS, block);
3640 son;
3641 son = next_dom_son (CDI_DOMINATORS, son))
3643 new_stuff |= insert_aux (son);
3646 return new_stuff;
3649 /* Perform insertion of partially redundant values. */
3651 static void
3652 insert (void)
3654 bool new_stuff = true;
3655 basic_block bb;
3656 int num_iterations = 0;
3658 FOR_ALL_BB (bb)
3659 NEW_SETS (bb) = bitmap_set_new ();
3661 while (new_stuff)
3663 num_iterations++;
3664 if (dump_file && dump_flags & TDF_DETAILS)
3665 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3666 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3668 /* Clear the NEW sets before the next iteration. We have already
3669 fully propagated its contents. */
3670 if (new_stuff)
3671 FOR_ALL_BB (bb)
3672 bitmap_set_free (NEW_SETS (bb));
3674 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3678 /* Compute the AVAIL set for all basic blocks.
3680 This function performs value numbering of the statements in each basic
3681 block. The AVAIL sets are built from information we glean while doing
3682 this value numbering, since the AVAIL sets contain only one entry per
3683 value.
3685 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3686 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3688 static void
3689 compute_avail (void)
3692 basic_block block, son;
3693 basic_block *worklist;
3694 size_t sp = 0;
3695 unsigned i;
3697 /* We pretend that default definitions are defined in the entry block.
3698 This includes function arguments and the static chain decl. */
3699 for (i = 1; i < num_ssa_names; ++i)
3701 tree name = ssa_name (i);
3702 pre_expr e;
3703 if (!name
3704 || !SSA_NAME_IS_DEFAULT_DEF (name)
3705 || has_zero_uses (name)
3706 || virtual_operand_p (name))
3707 continue;
3709 e = get_or_alloc_expr_for_name (name);
3710 add_to_value (get_expr_value_id (e), e);
3711 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3712 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3715 if (dump_file && (dump_flags & TDF_DETAILS))
3717 print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR),
3718 "tmp_gen", ENTRY_BLOCK);
3719 print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR),
3720 "avail_out", ENTRY_BLOCK);
3723 /* Allocate the worklist. */
3724 worklist = XNEWVEC (basic_block, n_basic_blocks);
3726 /* Seed the algorithm by putting the dominator children of the entry
3727 block on the worklist. */
3728 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3729 son;
3730 son = next_dom_son (CDI_DOMINATORS, son))
3731 worklist[sp++] = son;
3733 /* Loop until the worklist is empty. */
3734 while (sp)
3736 gimple_stmt_iterator gsi;
3737 gimple stmt;
3738 basic_block dom;
3740 /* Pick a block from the worklist. */
3741 block = worklist[--sp];
3743 /* Initially, the set of available values in BLOCK is that of
3744 its immediate dominator. */
3745 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3746 if (dom)
3747 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3749 /* Generate values for PHI nodes. */
3750 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3752 tree result = gimple_phi_result (gsi_stmt (gsi));
3754 /* We have no need for virtual phis, as they don't represent
3755 actual computations. */
3756 if (virtual_operand_p (result))
3757 continue;
3759 pre_expr e = get_or_alloc_expr_for_name (result);
3760 add_to_value (get_expr_value_id (e), e);
3761 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3762 bitmap_insert_into_set (PHI_GEN (block), e);
3765 BB_MAY_NOTRETURN (block) = 0;
3767 /* Now compute value numbers and populate value sets with all
3768 the expressions computed in BLOCK. */
3769 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3771 ssa_op_iter iter;
3772 tree op;
3774 stmt = gsi_stmt (gsi);
3776 /* Cache whether the basic-block has any non-visible side-effect
3777 or control flow.
3778 If this isn't a call or it is the last stmt in the
3779 basic-block then the CFG represents things correctly. */
3780 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3782 /* Non-looping const functions always return normally.
3783 Otherwise the call might not return or have side-effects
3784 that forbids hoisting possibly trapping expressions
3785 before it. */
3786 int flags = gimple_call_flags (stmt);
3787 if (!(flags & ECF_CONST)
3788 || (flags & ECF_LOOPING_CONST_OR_PURE))
3789 BB_MAY_NOTRETURN (block) = 1;
3792 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3794 pre_expr e = get_or_alloc_expr_for_name (op);
3796 add_to_value (get_expr_value_id (e), e);
3797 bitmap_insert_into_set (TMP_GEN (block), e);
3798 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3801 if (gimple_has_side_effects (stmt)
3802 || stmt_could_throw_p (stmt)
3803 || is_gimple_debug (stmt))
3804 continue;
3806 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3808 if (ssa_undefined_value_p (op))
3809 continue;
3810 pre_expr e = get_or_alloc_expr_for_name (op);
3811 bitmap_value_insert_into_set (EXP_GEN (block), e);
3814 switch (gimple_code (stmt))
3816 case GIMPLE_RETURN:
3817 continue;
3819 case GIMPLE_CALL:
3821 vn_reference_t ref;
3822 pre_expr result = NULL;
3823 vec<vn_reference_op_s> ops = vNULL;
3825 /* We can value number only calls to real functions. */
3826 if (gimple_call_internal_p (stmt))
3827 continue;
3829 copy_reference_ops_from_call (stmt, &ops);
3830 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3831 gimple_expr_type (stmt),
3832 ops, &ref, VN_NOWALK);
3833 ops.release ();
3834 if (!ref)
3835 continue;
3837 /* If the value of the call is not invalidated in
3838 this block until it is computed, add the expression
3839 to EXP_GEN. */
3840 if (!gimple_vuse (stmt)
3841 || gimple_code
3842 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
3843 || gimple_bb (SSA_NAME_DEF_STMT
3844 (gimple_vuse (stmt))) != block)
3846 result = (pre_expr) pool_alloc (pre_expr_pool);
3847 result->kind = REFERENCE;
3848 result->id = 0;
3849 PRE_EXPR_REFERENCE (result) = ref;
3851 get_or_alloc_expression_id (result);
3852 add_to_value (get_expr_value_id (result), result);
3853 bitmap_value_insert_into_set (EXP_GEN (block), result);
3855 continue;
3858 case GIMPLE_ASSIGN:
3860 pre_expr result = NULL;
3861 switch (vn_get_stmt_kind (stmt))
3863 case VN_NARY:
3865 enum tree_code code = gimple_assign_rhs_code (stmt);
3866 vn_nary_op_t nary;
3868 /* COND_EXPR and VEC_COND_EXPR are awkward in
3869 that they contain an embedded complex expression.
3870 Don't even try to shove those through PRE. */
3871 if (code == COND_EXPR
3872 || code == VEC_COND_EXPR)
3873 continue;
3875 vn_nary_op_lookup_stmt (stmt, &nary);
3876 if (!nary)
3877 continue;
3879 /* If the NARY traps and there was a preceding
3880 point in the block that might not return avoid
3881 adding the nary to EXP_GEN. */
3882 if (BB_MAY_NOTRETURN (block)
3883 && vn_nary_may_trap (nary))
3884 continue;
3886 result = (pre_expr) pool_alloc (pre_expr_pool);
3887 result->kind = NARY;
3888 result->id = 0;
3889 PRE_EXPR_NARY (result) = nary;
3890 break;
3893 case VN_REFERENCE:
3895 vn_reference_t ref;
3896 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3897 gimple_vuse (stmt),
3898 VN_WALK, &ref);
3899 if (!ref)
3900 continue;
3902 /* If the value of the reference is not invalidated in
3903 this block until it is computed, add the expression
3904 to EXP_GEN. */
3905 if (gimple_vuse (stmt))
3907 gimple def_stmt;
3908 bool ok = true;
3909 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
3910 while (!gimple_nop_p (def_stmt)
3911 && gimple_code (def_stmt) != GIMPLE_PHI
3912 && gimple_bb (def_stmt) == block)
3914 if (stmt_may_clobber_ref_p
3915 (def_stmt, gimple_assign_rhs1 (stmt)))
3917 ok = false;
3918 break;
3920 def_stmt
3921 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
3923 if (!ok)
3924 continue;
3927 result = (pre_expr) pool_alloc (pre_expr_pool);
3928 result->kind = REFERENCE;
3929 result->id = 0;
3930 PRE_EXPR_REFERENCE (result) = ref;
3931 break;
3934 default:
3935 continue;
3938 get_or_alloc_expression_id (result);
3939 add_to_value (get_expr_value_id (result), result);
3940 bitmap_value_insert_into_set (EXP_GEN (block), result);
3941 continue;
3943 default:
3944 break;
3948 if (dump_file && (dump_flags & TDF_DETAILS))
3950 print_bitmap_set (dump_file, EXP_GEN (block),
3951 "exp_gen", block->index);
3952 print_bitmap_set (dump_file, PHI_GEN (block),
3953 "phi_gen", block->index);
3954 print_bitmap_set (dump_file, TMP_GEN (block),
3955 "tmp_gen", block->index);
3956 print_bitmap_set (dump_file, AVAIL_OUT (block),
3957 "avail_out", block->index);
3960 /* Put the dominator children of BLOCK on the worklist of blocks
3961 to compute available sets for. */
3962 for (son = first_dom_son (CDI_DOMINATORS, block);
3963 son;
3964 son = next_dom_son (CDI_DOMINATORS, son))
3965 worklist[sp++] = son;
3968 free (worklist);
3972 /* Local state for the eliminate domwalk. */
3973 static vec<gimple> el_to_remove;
3974 static vec<gimple> el_to_update;
3975 static unsigned int el_todo;
3976 static vec<tree> el_avail;
3977 static vec<tree> el_avail_stack;
3979 /* Return a leader for OP that is available at the current point of the
3980 eliminate domwalk. */
3982 static tree
3983 eliminate_avail (tree op)
3985 tree valnum = VN_INFO (op)->valnum;
3986 if (TREE_CODE (valnum) == SSA_NAME)
3988 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
3989 return valnum;
3990 if (el_avail.length () > SSA_NAME_VERSION (valnum))
3991 return el_avail[SSA_NAME_VERSION (valnum)];
3993 else if (is_gimple_min_invariant (valnum))
3994 return valnum;
3995 return NULL_TREE;
3998 /* At the current point of the eliminate domwalk make OP available. */
4000 static void
4001 eliminate_push_avail (tree op)
4003 tree valnum = VN_INFO (op)->valnum;
4004 if (TREE_CODE (valnum) == SSA_NAME)
4006 if (el_avail.length () <= SSA_NAME_VERSION (valnum))
4007 el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1);
4008 el_avail[SSA_NAME_VERSION (valnum)] = op;
4009 el_avail_stack.safe_push (op);
4013 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
4014 the leader for the expression if insertion was successful. */
4016 static tree
4017 eliminate_insert (gimple_stmt_iterator *gsi, tree val)
4019 tree expr = vn_get_expr_for (val);
4020 if (!CONVERT_EXPR_P (expr)
4021 && TREE_CODE (expr) != VIEW_CONVERT_EXPR)
4022 return NULL_TREE;
4024 tree op = TREE_OPERAND (expr, 0);
4025 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op;
4026 if (!leader)
4027 return NULL_TREE;
4029 tree res = make_temp_ssa_name (TREE_TYPE (val), NULL, "pretmp");
4030 gimple tem = gimple_build_assign (res,
4031 fold_build1 (TREE_CODE (expr),
4032 TREE_TYPE (expr), leader));
4033 gsi_insert_before (gsi, tem, GSI_SAME_STMT);
4034 VN_INFO_GET (res)->valnum = val;
4036 if (TREE_CODE (leader) == SSA_NAME)
4037 gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true);
4039 pre_stats.insertions++;
4040 if (dump_file && (dump_flags & TDF_DETAILS))
4042 fprintf (dump_file, "Inserted ");
4043 print_gimple_stmt (dump_file, tem, 0, 0);
4046 return res;
4049 class eliminate_dom_walker : public dom_walker
4051 public:
4052 eliminate_dom_walker (cdi_direction direction) : dom_walker (direction) {}
4054 virtual void before_dom_children (basic_block);
4055 virtual void after_dom_children (basic_block);
4058 /* Perform elimination for the basic-block B during the domwalk. */
4060 void
4061 eliminate_dom_walker::before_dom_children (basic_block b)
4063 gimple_stmt_iterator gsi;
4064 gimple stmt;
4066 /* Mark new bb. */
4067 el_avail_stack.safe_push (NULL_TREE);
4069 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4071 gimple stmt, phi = gsi_stmt (gsi);
4072 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4073 gimple_stmt_iterator gsi2;
4075 /* We want to perform redundant PHI elimination. Do so by
4076 replacing the PHI with a single copy if possible.
4077 Do not touch inserted, single-argument or virtual PHIs. */
4078 if (gimple_phi_num_args (phi) == 1
4079 || virtual_operand_p (res))
4081 gsi_next (&gsi);
4082 continue;
4085 sprime = eliminate_avail (res);
4086 if (!sprime
4087 || sprime == res)
4089 eliminate_push_avail (res);
4090 gsi_next (&gsi);
4091 continue;
4093 else if (is_gimple_min_invariant (sprime))
4095 if (!useless_type_conversion_p (TREE_TYPE (res),
4096 TREE_TYPE (sprime)))
4097 sprime = fold_convert (TREE_TYPE (res), sprime);
4100 if (dump_file && (dump_flags & TDF_DETAILS))
4102 fprintf (dump_file, "Replaced redundant PHI node defining ");
4103 print_generic_expr (dump_file, res, 0);
4104 fprintf (dump_file, " with ");
4105 print_generic_expr (dump_file, sprime, 0);
4106 fprintf (dump_file, "\n");
4109 remove_phi_node (&gsi, false);
4111 if (inserted_exprs
4112 && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4113 && TREE_CODE (sprime) == SSA_NAME)
4114 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4116 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4117 sprime = fold_convert (TREE_TYPE (res), sprime);
4118 stmt = gimple_build_assign (res, sprime);
4119 SSA_NAME_DEF_STMT (res) = stmt;
4120 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4122 gsi2 = gsi_after_labels (b);
4123 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4124 /* Queue the copy for eventual removal. */
4125 el_to_remove.safe_push (stmt);
4126 /* If we inserted this PHI node ourself, it's not an elimination. */
4127 if (inserted_exprs
4128 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4129 pre_stats.phis--;
4130 else
4131 pre_stats.eliminations++;
4134 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4136 tree lhs = NULL_TREE;
4137 tree rhs = NULL_TREE;
4139 stmt = gsi_stmt (gsi);
4141 if (gimple_has_lhs (stmt))
4142 lhs = gimple_get_lhs (stmt);
4144 if (gimple_assign_single_p (stmt))
4145 rhs = gimple_assign_rhs1 (stmt);
4147 /* Lookup the RHS of the expression, see if we have an
4148 available computation for it. If so, replace the RHS with
4149 the available computation. */
4150 if (gimple_has_lhs (stmt)
4151 && TREE_CODE (lhs) == SSA_NAME
4152 && !gimple_has_volatile_ops (stmt))
4154 tree sprime;
4155 gimple orig_stmt = stmt;
4157 sprime = eliminate_avail (lhs);
4158 /* If there is no usable leader mark lhs as leader for its value. */
4159 if (!sprime)
4160 eliminate_push_avail (lhs);
4162 /* See PR43491. Do not replace a global register variable when
4163 it is a the RHS of an assignment. Do replace local register
4164 variables since gcc does not guarantee a local variable will
4165 be allocated in register.
4166 Do not perform copy propagation or undo constant propagation. */
4167 if (gimple_assign_single_p (stmt)
4168 && (TREE_CODE (rhs) == SSA_NAME
4169 || is_gimple_min_invariant (rhs)
4170 || (TREE_CODE (rhs) == VAR_DECL
4171 && is_global_var (rhs)
4172 && DECL_HARD_REGISTER (rhs))))
4173 continue;
4175 if (!sprime)
4177 /* If there is no existing usable leader but SCCVN thinks
4178 it has an expression it wants to use as replacement,
4179 insert that. */
4180 tree val = VN_INFO (lhs)->valnum;
4181 if (val != VN_TOP
4182 && TREE_CODE (val) == SSA_NAME
4183 && VN_INFO (val)->needs_insertion
4184 && VN_INFO (val)->expr != NULL_TREE
4185 && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE)
4186 eliminate_push_avail (sprime);
4188 else if (is_gimple_min_invariant (sprime))
4190 /* If there is no existing leader but SCCVN knows this
4191 value is constant, use that constant. */
4192 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4193 TREE_TYPE (sprime)))
4194 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4196 if (dump_file && (dump_flags & TDF_DETAILS))
4198 fprintf (dump_file, "Replaced ");
4199 print_gimple_expr (dump_file, stmt, 0, 0);
4200 fprintf (dump_file, " with ");
4201 print_generic_expr (dump_file, sprime, 0);
4202 fprintf (dump_file, " in ");
4203 print_gimple_stmt (dump_file, stmt, 0, 0);
4205 pre_stats.eliminations++;
4206 propagate_tree_value_into_stmt (&gsi, sprime);
4207 stmt = gsi_stmt (gsi);
4208 update_stmt (stmt);
4210 /* If we removed EH side-effects from the statement, clean
4211 its EH information. */
4212 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4214 bitmap_set_bit (need_eh_cleanup,
4215 gimple_bb (stmt)->index);
4216 if (dump_file && (dump_flags & TDF_DETAILS))
4217 fprintf (dump_file, " Removed EH side-effects.\n");
4219 continue;
4222 if (sprime
4223 && sprime != lhs
4224 && (rhs == NULL_TREE
4225 || TREE_CODE (rhs) != SSA_NAME
4226 || may_propagate_copy (rhs, sprime)))
4228 bool can_make_abnormal_goto
4229 = is_gimple_call (stmt)
4230 && stmt_can_make_abnormal_goto (stmt);
4232 gcc_assert (sprime != rhs);
4234 if (dump_file && (dump_flags & TDF_DETAILS))
4236 fprintf (dump_file, "Replaced ");
4237 print_gimple_expr (dump_file, stmt, 0, 0);
4238 fprintf (dump_file, " with ");
4239 print_generic_expr (dump_file, sprime, 0);
4240 fprintf (dump_file, " in ");
4241 print_gimple_stmt (dump_file, stmt, 0, 0);
4244 if (TREE_CODE (sprime) == SSA_NAME)
4245 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4246 NECESSARY, true);
4247 /* We need to make sure the new and old types actually match,
4248 which may require adding a simple cast, which fold_convert
4249 will do for us. */
4250 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4251 && !useless_type_conversion_p (gimple_expr_type (stmt),
4252 TREE_TYPE (sprime)))
4253 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4255 pre_stats.eliminations++;
4256 propagate_tree_value_into_stmt (&gsi, sprime);
4257 stmt = gsi_stmt (gsi);
4258 update_stmt (stmt);
4260 /* If we removed EH side-effects from the statement, clean
4261 its EH information. */
4262 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4264 bitmap_set_bit (need_eh_cleanup,
4265 gimple_bb (stmt)->index);
4266 if (dump_file && (dump_flags & TDF_DETAILS))
4267 fprintf (dump_file, " Removed EH side-effects.\n");
4270 /* Likewise for AB side-effects. */
4271 if (can_make_abnormal_goto
4272 && !stmt_can_make_abnormal_goto (stmt))
4274 bitmap_set_bit (need_ab_cleanup,
4275 gimple_bb (stmt)->index);
4276 if (dump_file && (dump_flags & TDF_DETAILS))
4277 fprintf (dump_file, " Removed AB side-effects.\n");
4281 /* If the statement is a scalar store, see if the expression
4282 has the same value number as its rhs. If so, the store is
4283 dead. */
4284 else if (gimple_assign_single_p (stmt)
4285 && !gimple_has_volatile_ops (stmt)
4286 && !is_gimple_reg (gimple_assign_lhs (stmt))
4287 && (TREE_CODE (rhs) == SSA_NAME
4288 || is_gimple_min_invariant (rhs)))
4290 tree val;
4291 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4292 gimple_vuse (stmt), VN_WALK, NULL);
4293 if (TREE_CODE (rhs) == SSA_NAME)
4294 rhs = VN_INFO (rhs)->valnum;
4295 if (val
4296 && operand_equal_p (val, rhs, 0))
4298 if (dump_file && (dump_flags & TDF_DETAILS))
4300 fprintf (dump_file, "Deleted redundant store ");
4301 print_gimple_stmt (dump_file, stmt, 0, 0);
4304 /* Queue stmt for removal. */
4305 el_to_remove.safe_push (stmt);
4308 /* Visit COND_EXPRs and fold the comparison with the
4309 available value-numbers. */
4310 else if (gimple_code (stmt) == GIMPLE_COND)
4312 tree op0 = gimple_cond_lhs (stmt);
4313 tree op1 = gimple_cond_rhs (stmt);
4314 tree result;
4316 if (TREE_CODE (op0) == SSA_NAME)
4317 op0 = VN_INFO (op0)->valnum;
4318 if (TREE_CODE (op1) == SSA_NAME)
4319 op1 = VN_INFO (op1)->valnum;
4320 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4321 op0, op1);
4322 if (result && TREE_CODE (result) == INTEGER_CST)
4324 if (integer_zerop (result))
4325 gimple_cond_make_false (stmt);
4326 else
4327 gimple_cond_make_true (stmt);
4328 update_stmt (stmt);
4329 el_todo = TODO_cleanup_cfg;
4332 /* Visit indirect calls and turn them into direct calls if
4333 possible. */
4334 if (is_gimple_call (stmt))
4336 tree orig_fn = gimple_call_fn (stmt);
4337 tree fn;
4338 if (!orig_fn)
4339 continue;
4340 if (TREE_CODE (orig_fn) == SSA_NAME)
4341 fn = VN_INFO (orig_fn)->valnum;
4342 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4343 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4345 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4346 if (!gimple_call_addr_fndecl (fn))
4348 fn = ipa_intraprocedural_devirtualization (stmt);
4349 if (fn)
4350 fn = build_fold_addr_expr (fn);
4353 else
4354 continue;
4355 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4356 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4357 TREE_TYPE (fn)))
4359 bool can_make_abnormal_goto
4360 = stmt_can_make_abnormal_goto (stmt);
4361 bool was_noreturn = gimple_call_noreturn_p (stmt);
4363 if (dump_file && (dump_flags & TDF_DETAILS))
4365 fprintf (dump_file, "Replacing call target with ");
4366 print_generic_expr (dump_file, fn, 0);
4367 fprintf (dump_file, " in ");
4368 print_gimple_stmt (dump_file, stmt, 0, 0);
4371 gimple_call_set_fn (stmt, fn);
4372 el_to_update.safe_push (stmt);
4374 /* When changing a call into a noreturn call, cfg cleanup
4375 is needed to fix up the noreturn call. */
4376 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4377 el_todo |= TODO_cleanup_cfg;
4379 /* If we removed EH side-effects from the statement, clean
4380 its EH information. */
4381 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4383 bitmap_set_bit (need_eh_cleanup,
4384 gimple_bb (stmt)->index);
4385 if (dump_file && (dump_flags & TDF_DETAILS))
4386 fprintf (dump_file, " Removed EH side-effects.\n");
4389 /* Likewise for AB side-effects. */
4390 if (can_make_abnormal_goto
4391 && !stmt_can_make_abnormal_goto (stmt))
4393 bitmap_set_bit (need_ab_cleanup,
4394 gimple_bb (stmt)->index);
4395 if (dump_file && (dump_flags & TDF_DETAILS))
4396 fprintf (dump_file, " Removed AB side-effects.\n");
4399 /* Changing an indirect call to a direct call may
4400 have exposed different semantics. This may
4401 require an SSA update. */
4402 el_todo |= TODO_update_ssa_only_virtuals;
4408 /* Make no longer available leaders no longer available. */
4410 void
4411 eliminate_dom_walker::after_dom_children (basic_block)
4413 tree entry;
4414 while ((entry = el_avail_stack.pop ()) != NULL_TREE)
4415 el_avail[SSA_NAME_VERSION (VN_INFO (entry)->valnum)] = NULL_TREE;
4418 /* Eliminate fully redundant computations. */
4420 static unsigned int
4421 eliminate (void)
4423 gimple_stmt_iterator gsi;
4424 gimple stmt;
4425 unsigned i;
4427 need_eh_cleanup = BITMAP_ALLOC (NULL);
4428 need_ab_cleanup = BITMAP_ALLOC (NULL);
4430 el_to_remove.create (0);
4431 el_to_update.create (0);
4432 el_todo = 0;
4433 el_avail.create (0);
4434 el_avail_stack.create (0);
4436 eliminate_dom_walker (CDI_DOMINATORS).walk (cfun->cfg->x_entry_block_ptr);
4438 el_avail.release ();
4439 el_avail_stack.release ();
4441 /* We cannot remove stmts during BB walk, especially not release SSA
4442 names there as this confuses the VN machinery. The stmts ending
4443 up in el_to_remove are either stores or simple copies. */
4444 FOR_EACH_VEC_ELT (el_to_remove, i, stmt)
4446 tree lhs = gimple_assign_lhs (stmt);
4447 tree rhs = gimple_assign_rhs1 (stmt);
4448 use_operand_p use_p;
4449 gimple use_stmt;
4451 /* If there is a single use only, propagate the equivalency
4452 instead of keeping the copy. */
4453 if (TREE_CODE (lhs) == SSA_NAME
4454 && TREE_CODE (rhs) == SSA_NAME
4455 && single_imm_use (lhs, &use_p, &use_stmt)
4456 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4458 SET_USE (use_p, rhs);
4459 update_stmt (use_stmt);
4460 if (inserted_exprs
4461 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4462 && TREE_CODE (rhs) == SSA_NAME)
4463 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4466 /* If this is a store or a now unused copy, remove it. */
4467 if (TREE_CODE (lhs) != SSA_NAME
4468 || has_zero_uses (lhs))
4470 basic_block bb = gimple_bb (stmt);
4471 gsi = gsi_for_stmt (stmt);
4472 unlink_stmt_vdef (stmt);
4473 if (gsi_remove (&gsi, true))
4474 bitmap_set_bit (need_eh_cleanup, bb->index);
4475 if (inserted_exprs
4476 && TREE_CODE (lhs) == SSA_NAME)
4477 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4478 release_defs (stmt);
4481 el_to_remove.release ();
4483 /* We cannot update call statements with virtual operands during
4484 SSA walk. This might remove them which in turn makes our
4485 VN lattice invalid. */
4486 FOR_EACH_VEC_ELT (el_to_update, i, stmt)
4487 update_stmt (stmt);
4488 el_to_update.release ();
4490 return el_todo;
4493 /* Perform CFG cleanups made necessary by elimination. */
4495 static unsigned
4496 fini_eliminate (void)
4498 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4499 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4501 if (do_eh_cleanup)
4502 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4504 if (do_ab_cleanup)
4505 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4507 BITMAP_FREE (need_eh_cleanup);
4508 BITMAP_FREE (need_ab_cleanup);
4510 if (do_eh_cleanup || do_ab_cleanup)
4511 return TODO_cleanup_cfg;
4512 return 0;
4515 /* Borrow a bit of tree-ssa-dce.c for the moment.
4516 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4517 this may be a bit faster, and we may want critical edges kept split. */
4519 /* If OP's defining statement has not already been determined to be necessary,
4520 mark that statement necessary. Return the stmt, if it is newly
4521 necessary. */
4523 static inline gimple
4524 mark_operand_necessary (tree op)
4526 gimple stmt;
4528 gcc_assert (op);
4530 if (TREE_CODE (op) != SSA_NAME)
4531 return NULL;
4533 stmt = SSA_NAME_DEF_STMT (op);
4534 gcc_assert (stmt);
4536 if (gimple_plf (stmt, NECESSARY)
4537 || gimple_nop_p (stmt))
4538 return NULL;
4540 gimple_set_plf (stmt, NECESSARY, true);
4541 return stmt;
4544 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4545 to insert PHI nodes sometimes, and because value numbering of casts isn't
4546 perfect, we sometimes end up inserting dead code. This simple DCE-like
4547 pass removes any insertions we made that weren't actually used. */
4549 static void
4550 remove_dead_inserted_code (void)
4552 bitmap worklist;
4553 unsigned i;
4554 bitmap_iterator bi;
4555 gimple t;
4557 worklist = BITMAP_ALLOC (NULL);
4558 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4560 t = SSA_NAME_DEF_STMT (ssa_name (i));
4561 if (gimple_plf (t, NECESSARY))
4562 bitmap_set_bit (worklist, i);
4564 while (!bitmap_empty_p (worklist))
4566 i = bitmap_first_set_bit (worklist);
4567 bitmap_clear_bit (worklist, i);
4568 t = SSA_NAME_DEF_STMT (ssa_name (i));
4570 /* PHI nodes are somewhat special in that each PHI alternative has
4571 data and control dependencies. All the statements feeding the
4572 PHI node's arguments are always necessary. */
4573 if (gimple_code (t) == GIMPLE_PHI)
4575 unsigned k;
4577 for (k = 0; k < gimple_phi_num_args (t); k++)
4579 tree arg = PHI_ARG_DEF (t, k);
4580 if (TREE_CODE (arg) == SSA_NAME)
4582 gimple n = mark_operand_necessary (arg);
4583 if (n)
4584 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4588 else
4590 /* Propagate through the operands. Examine all the USE, VUSE and
4591 VDEF operands in this statement. Mark all the statements
4592 which feed this statement's uses as necessary. */
4593 ssa_op_iter iter;
4594 tree use;
4596 /* The operands of VDEF expressions are also needed as they
4597 represent potential definitions that may reach this
4598 statement (VDEF operands allow us to follow def-def
4599 links). */
4601 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4603 gimple n = mark_operand_necessary (use);
4604 if (n)
4605 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4610 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4612 t = SSA_NAME_DEF_STMT (ssa_name (i));
4613 if (!gimple_plf (t, NECESSARY))
4615 gimple_stmt_iterator gsi;
4617 if (dump_file && (dump_flags & TDF_DETAILS))
4619 fprintf (dump_file, "Removing unnecessary insertion:");
4620 print_gimple_stmt (dump_file, t, 0, 0);
4623 gsi = gsi_for_stmt (t);
4624 if (gimple_code (t) == GIMPLE_PHI)
4625 remove_phi_node (&gsi, true);
4626 else
4628 gsi_remove (&gsi, true);
4629 release_defs (t);
4633 BITMAP_FREE (worklist);
4637 /* Initialize data structures used by PRE. */
4639 static void
4640 init_pre (void)
4642 basic_block bb;
4644 next_expression_id = 1;
4645 expressions.create (0);
4646 expressions.safe_push (NULL);
4647 value_expressions.create (get_max_value_id () + 1);
4648 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
4649 name_to_id.create (0);
4651 inserted_exprs = BITMAP_ALLOC (NULL);
4653 connect_infinite_loops_to_exit ();
4654 memset (&pre_stats, 0, sizeof (pre_stats));
4656 postorder = XNEWVEC (int, n_basic_blocks);
4657 postorder_num = inverted_post_order_compute (postorder);
4659 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4661 calculate_dominance_info (CDI_POST_DOMINATORS);
4662 calculate_dominance_info (CDI_DOMINATORS);
4664 bitmap_obstack_initialize (&grand_bitmap_obstack);
4665 phi_translate_table.create (5110);
4666 expression_to_id.create (num_ssa_names * 3);
4667 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4668 sizeof (struct bitmap_set), 30);
4669 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4670 sizeof (struct pre_expr_d), 30);
4671 FOR_ALL_BB (bb)
4673 EXP_GEN (bb) = bitmap_set_new ();
4674 PHI_GEN (bb) = bitmap_set_new ();
4675 TMP_GEN (bb) = bitmap_set_new ();
4676 AVAIL_OUT (bb) = bitmap_set_new ();
4681 /* Deallocate data structures used by PRE. */
4683 static void
4684 fini_pre ()
4686 free (postorder);
4687 value_expressions.release ();
4688 BITMAP_FREE (inserted_exprs);
4689 bitmap_obstack_release (&grand_bitmap_obstack);
4690 free_alloc_pool (bitmap_set_pool);
4691 free_alloc_pool (pre_expr_pool);
4692 phi_translate_table.dispose ();
4693 expression_to_id.dispose ();
4694 name_to_id.release ();
4696 free_aux_for_blocks ();
4698 free_dominance_info (CDI_POST_DOMINATORS);
4701 /* Gate and execute functions for PRE. */
4703 static unsigned int
4704 do_pre (void)
4706 unsigned int todo = 0;
4708 do_partial_partial =
4709 flag_tree_partial_pre && optimize_function_for_speed_p (cfun);
4711 /* This has to happen before SCCVN runs because
4712 loop_optimizer_init may create new phis, etc. */
4713 loop_optimizer_init (LOOPS_NORMAL);
4715 if (!run_scc_vn (VN_WALK))
4717 loop_optimizer_finalize ();
4718 return 0;
4721 init_pre ();
4722 scev_initialize ();
4724 /* Collect and value number expressions computed in each basic block. */
4725 compute_avail ();
4727 /* Insert can get quite slow on an incredibly large number of basic
4728 blocks due to some quadratic behavior. Until this behavior is
4729 fixed, don't run it when he have an incredibly large number of
4730 bb's. If we aren't going to run insert, there is no point in
4731 computing ANTIC, either, even though it's plenty fast. */
4732 if (n_basic_blocks < 4000)
4734 compute_antic ();
4735 insert ();
4738 /* Make sure to remove fake edges before committing our inserts.
4739 This makes sure we don't end up with extra critical edges that
4740 we would need to split. */
4741 remove_fake_exit_edges ();
4742 gsi_commit_edge_inserts ();
4744 /* Remove all the redundant expressions. */
4745 todo |= eliminate ();
4747 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4748 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4749 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4750 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4752 clear_expression_ids ();
4753 remove_dead_inserted_code ();
4754 todo |= TODO_verify_flow;
4756 scev_finalize ();
4757 fini_pre ();
4758 todo |= fini_eliminate ();
4759 loop_optimizer_finalize ();
4761 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4762 case we can merge the block with the remaining predecessor of the block.
4763 It should either:
4764 - call merge_blocks after each tail merge iteration
4765 - call merge_blocks after all tail merge iterations
4766 - mark TODO_cleanup_cfg when necessary
4767 - share the cfg cleanup with fini_pre. */
4768 todo |= tail_merge_optimize (todo);
4770 free_scc_vn ();
4772 /* Tail merging invalidates the virtual SSA web, together with
4773 cfg-cleanup opportunities exposed by PRE this will wreck the
4774 SSA updating machinery. So make sure to run update-ssa
4775 manually, before eventually scheduling cfg-cleanup as part of
4776 the todo. */
4777 update_ssa (TODO_update_ssa_only_virtuals);
4779 return todo;
4782 static bool
4783 gate_pre (void)
4785 return flag_tree_pre != 0;
4788 namespace {
4790 const pass_data pass_data_pre =
4792 GIMPLE_PASS, /* type */
4793 "pre", /* name */
4794 OPTGROUP_NONE, /* optinfo_flags */
4795 true, /* has_gate */
4796 true, /* has_execute */
4797 TV_TREE_PRE, /* tv_id */
4798 ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */
4799 0, /* properties_provided */
4800 0, /* properties_destroyed */
4801 TODO_rebuild_alias, /* todo_flags_start */
4802 TODO_verify_ssa, /* todo_flags_finish */
4805 class pass_pre : public gimple_opt_pass
4807 public:
4808 pass_pre (gcc::context *ctxt)
4809 : gimple_opt_pass (pass_data_pre, ctxt)
4812 /* opt_pass methods: */
4813 bool gate () { return gate_pre (); }
4814 unsigned int execute () { return do_pre (); }
4816 }; // class pass_pre
4818 } // anon namespace
4820 gimple_opt_pass *
4821 make_pass_pre (gcc::context *ctxt)
4823 return new pass_pre (ctxt);
4827 /* Gate and execute functions for FRE. */
4829 static unsigned int
4830 execute_fre (void)
4832 unsigned int todo = 0;
4834 if (!run_scc_vn (VN_WALKREWRITE))
4835 return 0;
4837 memset (&pre_stats, 0, sizeof (pre_stats));
4839 /* Remove all the redundant expressions. */
4840 todo |= eliminate ();
4842 todo |= fini_eliminate ();
4844 free_scc_vn ();
4846 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4847 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4849 return todo;
4852 static bool
4853 gate_fre (void)
4855 return flag_tree_fre != 0;
4858 namespace {
4860 const pass_data pass_data_fre =
4862 GIMPLE_PASS, /* type */
4863 "fre", /* name */
4864 OPTGROUP_NONE, /* optinfo_flags */
4865 true, /* has_gate */
4866 true, /* has_execute */
4867 TV_TREE_FRE, /* tv_id */
4868 ( PROP_cfg | PROP_ssa ), /* properties_required */
4869 0, /* properties_provided */
4870 0, /* properties_destroyed */
4871 0, /* todo_flags_start */
4872 TODO_verify_ssa, /* todo_flags_finish */
4875 class pass_fre : public gimple_opt_pass
4877 public:
4878 pass_fre (gcc::context *ctxt)
4879 : gimple_opt_pass (pass_data_fre, ctxt)
4882 /* opt_pass methods: */
4883 opt_pass * clone () { return new pass_fre (m_ctxt); }
4884 bool gate () { return gate_fre (); }
4885 unsigned int execute () { return execute_fre (); }
4887 }; // class pass_fre
4889 } // anon namespace
4891 gimple_opt_pass *
4892 make_pass_fre (gcc::context *ctxt)
4894 return new pass_fre (ctxt);