2013-11-29 Richard Biener <rguenther@suse.de>
[official-gcc.git] / gcc / tree-ssa-pre.c
blobf9ac337fef9586ffdb41fbad95dfe92185bd54c3
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 "hash-table.h"
31 #include "tree-ssa-alias.h"
32 #include "internal-fn.h"
33 #include "gimple-fold.h"
34 #include "tree-eh.h"
35 #include "gimple-expr.h"
36 #include "is-a.h"
37 #include "gimple.h"
38 #include "gimplify.h"
39 #include "gimple-iterator.h"
40 #include "gimplify-me.h"
41 #include "gimple-ssa.h"
42 #include "tree-cfg.h"
43 #include "tree-phinodes.h"
44 #include "ssa-iterators.h"
45 #include "stringpool.h"
46 #include "tree-ssanames.h"
47 #include "tree-ssa-loop.h"
48 #include "tree-into-ssa.h"
49 #include "expr.h"
50 #include "tree-dfa.h"
51 #include "tree-ssa.h"
52 #include "tree-iterator.h"
53 #include "alloc-pool.h"
54 #include "obstack.h"
55 #include "tree-pass.h"
56 #include "flags.h"
57 #include "langhooks.h"
58 #include "cfgloop.h"
59 #include "tree-ssa-sccvn.h"
60 #include "tree-scalar-evolution.h"
61 #include "params.h"
62 #include "dbgcnt.h"
63 #include "domwalk.h"
64 #include "ipa-prop.h"
65 #include "tree-ssa-propagate.h"
67 /* TODO:
69 1. Avail sets can be shared by making an avail_find_leader that
70 walks up the dominator tree and looks in those avail sets.
71 This might affect code optimality, it's unclear right now.
72 2. Strength reduction can be performed by anticipating expressions
73 we can repair later on.
74 3. We can do back-substitution or smarter value numbering to catch
75 commutative expressions split up over multiple statements.
78 /* For ease of terminology, "expression node" in the below refers to
79 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
80 represent the actual statement containing the expressions we care about,
81 and we cache the value number by putting it in the expression. */
83 /* Basic algorithm
85 First we walk the statements to generate the AVAIL sets, the
86 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
87 generation of values/expressions by a given block. We use them
88 when computing the ANTIC sets. The AVAIL sets consist of
89 SSA_NAME's that represent values, so we know what values are
90 available in what blocks. AVAIL is a forward dataflow problem. In
91 SSA, values are never killed, so we don't need a kill set, or a
92 fixpoint iteration, in order to calculate the AVAIL sets. In
93 traditional parlance, AVAIL sets tell us the downsafety of the
94 expressions/values.
96 Next, we generate the ANTIC sets. These sets represent the
97 anticipatable expressions. ANTIC is a backwards dataflow
98 problem. An expression is anticipatable in a given block if it could
99 be generated in that block. This means that if we had to perform
100 an insertion in that block, of the value of that expression, we
101 could. Calculating the ANTIC sets requires phi translation of
102 expressions, because the flow goes backwards through phis. We must
103 iterate to a fixpoint of the ANTIC sets, because we have a kill
104 set. Even in SSA form, values are not live over the entire
105 function, only from their definition point onwards. So we have to
106 remove values from the ANTIC set once we go past the definition
107 point of the leaders that make them up.
108 compute_antic/compute_antic_aux performs this computation.
110 Third, we perform insertions to make partially redundant
111 expressions fully redundant.
113 An expression is partially redundant (excluding partial
114 anticipation) if:
116 1. It is AVAIL in some, but not all, of the predecessors of a
117 given block.
118 2. It is ANTIC in all the predecessors.
120 In order to make it fully redundant, we insert the expression into
121 the predecessors where it is not available, but is ANTIC.
123 For the partial anticipation case, we only perform insertion if it
124 is partially anticipated in some block, and fully available in all
125 of the predecessors.
127 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
128 performs these steps.
130 Fourth, we eliminate fully redundant expressions.
131 This is a simple statement walk that replaces redundant
132 calculations with the now available values. */
134 /* Representations of value numbers:
136 Value numbers are represented by a representative SSA_NAME. We
137 will create fake SSA_NAME's in situations where we need a
138 representative but do not have one (because it is a complex
139 expression). In order to facilitate storing the value numbers in
140 bitmaps, and keep the number of wasted SSA_NAME's down, we also
141 associate a value_id with each value number, and create full blown
142 ssa_name's only where we actually need them (IE in operands of
143 existing expressions).
145 Theoretically you could replace all the value_id's with
146 SSA_NAME_VERSION, but this would allocate a large number of
147 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
148 It would also require an additional indirection at each point we
149 use the value id. */
151 /* Representation of expressions on value numbers:
153 Expressions consisting of value numbers are represented the same
154 way as our VN internally represents them, with an additional
155 "pre_expr" wrapping around them in order to facilitate storing all
156 of the expressions in the same sets. */
158 /* Representation of sets:
160 The dataflow sets do not need to be sorted in any particular order
161 for the majority of their lifetime, are simply represented as two
162 bitmaps, one that keeps track of values present in the set, and one
163 that keeps track of expressions present in the set.
165 When we need them in topological order, we produce it on demand by
166 transforming the bitmap into an array and sorting it into topo
167 order. */
169 /* Type of expression, used to know which member of the PRE_EXPR union
170 is valid. */
172 enum pre_expr_kind
174 NAME,
175 NARY,
176 REFERENCE,
177 CONSTANT
180 typedef union pre_expr_union_d
182 tree name;
183 tree constant;
184 vn_nary_op_t nary;
185 vn_reference_t reference;
186 } pre_expr_union;
188 typedef struct pre_expr_d : typed_noop_remove <pre_expr_d>
190 enum pre_expr_kind kind;
191 unsigned int id;
192 pre_expr_union u;
194 /* hash_table support. */
195 typedef pre_expr_d value_type;
196 typedef pre_expr_d compare_type;
197 static inline hashval_t hash (const pre_expr_d *);
198 static inline int equal (const pre_expr_d *, const pre_expr_d *);
199 } *pre_expr;
201 #define PRE_EXPR_NAME(e) (e)->u.name
202 #define PRE_EXPR_NARY(e) (e)->u.nary
203 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
204 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
206 /* Compare E1 and E1 for equality. */
208 inline int
209 pre_expr_d::equal (const value_type *e1, const compare_type *e2)
211 if (e1->kind != e2->kind)
212 return false;
214 switch (e1->kind)
216 case CONSTANT:
217 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
218 PRE_EXPR_CONSTANT (e2));
219 case NAME:
220 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
221 case NARY:
222 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
223 case REFERENCE:
224 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
225 PRE_EXPR_REFERENCE (e2));
226 default:
227 gcc_unreachable ();
231 /* Hash E. */
233 inline hashval_t
234 pre_expr_d::hash (const value_type *e)
236 switch (e->kind)
238 case CONSTANT:
239 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
240 case NAME:
241 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
242 case NARY:
243 return PRE_EXPR_NARY (e)->hashcode;
244 case REFERENCE:
245 return PRE_EXPR_REFERENCE (e)->hashcode;
246 default:
247 gcc_unreachable ();
251 /* Next global expression id number. */
252 static unsigned int next_expression_id;
254 /* Mapping from expression to id number we can use in bitmap sets. */
255 static vec<pre_expr> expressions;
256 static hash_table <pre_expr_d> expression_to_id;
257 static vec<unsigned> name_to_id;
259 /* Allocate an expression id for EXPR. */
261 static inline unsigned int
262 alloc_expression_id (pre_expr expr)
264 struct pre_expr_d **slot;
265 /* Make sure we won't overflow. */
266 gcc_assert (next_expression_id + 1 > next_expression_id);
267 expr->id = next_expression_id++;
268 expressions.safe_push (expr);
269 if (expr->kind == NAME)
271 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
272 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
273 re-allocations by using vec::reserve upfront. There is no
274 vec::quick_grow_cleared unfortunately. */
275 unsigned old_len = name_to_id.length ();
276 name_to_id.reserve (num_ssa_names - old_len);
277 name_to_id.safe_grow_cleared (num_ssa_names);
278 gcc_assert (name_to_id[version] == 0);
279 name_to_id[version] = expr->id;
281 else
283 slot = expression_to_id.find_slot (expr, INSERT);
284 gcc_assert (!*slot);
285 *slot = expr;
287 return next_expression_id - 1;
290 /* Return the expression id for tree EXPR. */
292 static inline unsigned int
293 get_expression_id (const pre_expr expr)
295 return expr->id;
298 static inline unsigned int
299 lookup_expression_id (const pre_expr expr)
301 struct pre_expr_d **slot;
303 if (expr->kind == NAME)
305 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
306 if (name_to_id.length () <= version)
307 return 0;
308 return name_to_id[version];
310 else
312 slot = expression_to_id.find_slot (expr, NO_INSERT);
313 if (!slot)
314 return 0;
315 return ((pre_expr)*slot)->id;
319 /* Return the existing expression id for EXPR, or create one if one
320 does not exist yet. */
322 static inline unsigned int
323 get_or_alloc_expression_id (pre_expr expr)
325 unsigned int id = lookup_expression_id (expr);
326 if (id == 0)
327 return alloc_expression_id (expr);
328 return expr->id = id;
331 /* Return the expression that has expression id ID */
333 static inline pre_expr
334 expression_for_id (unsigned int id)
336 return expressions[id];
339 /* Free the expression id field in all of our expressions,
340 and then destroy the expressions array. */
342 static void
343 clear_expression_ids (void)
345 expressions.release ();
348 static alloc_pool pre_expr_pool;
350 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
352 static pre_expr
353 get_or_alloc_expr_for_name (tree name)
355 struct pre_expr_d expr;
356 pre_expr result;
357 unsigned int result_id;
359 expr.kind = NAME;
360 expr.id = 0;
361 PRE_EXPR_NAME (&expr) = name;
362 result_id = lookup_expression_id (&expr);
363 if (result_id != 0)
364 return expression_for_id (result_id);
366 result = (pre_expr) pool_alloc (pre_expr_pool);
367 result->kind = NAME;
368 PRE_EXPR_NAME (result) = name;
369 alloc_expression_id (result);
370 return result;
373 /* An unordered bitmap set. One bitmap tracks values, the other,
374 expressions. */
375 typedef struct bitmap_set
377 bitmap_head expressions;
378 bitmap_head values;
379 } *bitmap_set_t;
381 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
382 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
384 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
385 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
387 /* Mapping from value id to expressions with that value_id. */
388 static vec<bitmap> value_expressions;
390 /* Sets that we need to keep track of. */
391 typedef struct bb_bitmap_sets
393 /* The EXP_GEN set, which represents expressions/values generated in
394 a basic block. */
395 bitmap_set_t exp_gen;
397 /* The PHI_GEN set, which represents PHI results generated in a
398 basic block. */
399 bitmap_set_t phi_gen;
401 /* The TMP_GEN set, which represents results/temporaries generated
402 in a basic block. IE the LHS of an expression. */
403 bitmap_set_t tmp_gen;
405 /* The AVAIL_OUT set, which represents which values are available in
406 a given basic block. */
407 bitmap_set_t avail_out;
409 /* The ANTIC_IN set, which represents which values are anticipatable
410 in a given basic block. */
411 bitmap_set_t antic_in;
413 /* The PA_IN set, which represents which values are
414 partially anticipatable in a given basic block. */
415 bitmap_set_t pa_in;
417 /* The NEW_SETS set, which is used during insertion to augment the
418 AVAIL_OUT set of blocks with the new insertions performed during
419 the current iteration. */
420 bitmap_set_t new_sets;
422 /* A cache for value_dies_in_block_x. */
423 bitmap expr_dies;
425 /* True if we have visited this block during ANTIC calculation. */
426 unsigned int visited : 1;
428 /* True we have deferred processing this block during ANTIC
429 calculation until its successor is processed. */
430 unsigned int deferred : 1;
432 /* True when the block contains a call that might not return. */
433 unsigned int contains_may_not_return_call : 1;
434 } *bb_value_sets_t;
436 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
437 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
438 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
439 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
440 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
441 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
442 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
443 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
444 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
445 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
446 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
449 /* Basic block list in postorder. */
450 static int *postorder;
451 static int postorder_num;
453 /* This structure is used to keep track of statistics on what
454 optimization PRE was able to perform. */
455 static struct
457 /* The number of RHS computations eliminated by PRE. */
458 int eliminations;
460 /* The number of new expressions/temporaries generated by PRE. */
461 int insertions;
463 /* The number of inserts found due to partial anticipation */
464 int pa_insert;
466 /* The number of new PHI nodes added by PRE. */
467 int phis;
468 } pre_stats;
470 static bool do_partial_partial;
471 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int);
472 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
473 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
474 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
475 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
476 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
477 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
478 unsigned int, bool);
479 static bitmap_set_t bitmap_set_new (void);
480 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
481 tree);
482 static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
483 static unsigned int get_expr_value_id (pre_expr);
485 /* We can add and remove elements and entries to and from sets
486 and hash tables, so we use alloc pools for them. */
488 static alloc_pool bitmap_set_pool;
489 static bitmap_obstack grand_bitmap_obstack;
491 /* Set of blocks with statements that have had their EH properties changed. */
492 static bitmap need_eh_cleanup;
494 /* Set of blocks with statements that have had their AB properties changed. */
495 static bitmap need_ab_cleanup;
497 /* A three tuple {e, pred, v} used to cache phi translations in the
498 phi_translate_table. */
500 typedef struct expr_pred_trans_d : typed_free_remove<expr_pred_trans_d>
502 /* The expression. */
503 pre_expr e;
505 /* The predecessor block along which we translated the expression. */
506 basic_block pred;
508 /* The value that resulted from the translation. */
509 pre_expr v;
511 /* The hashcode for the expression, pred pair. This is cached for
512 speed reasons. */
513 hashval_t hashcode;
515 /* hash_table support. */
516 typedef expr_pred_trans_d value_type;
517 typedef expr_pred_trans_d compare_type;
518 static inline hashval_t hash (const value_type *);
519 static inline int equal (const value_type *, const compare_type *);
520 } *expr_pred_trans_t;
521 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
523 inline hashval_t
524 expr_pred_trans_d::hash (const expr_pred_trans_d *e)
526 return e->hashcode;
529 inline int
530 expr_pred_trans_d::equal (const value_type *ve1,
531 const compare_type *ve2)
533 basic_block b1 = ve1->pred;
534 basic_block b2 = ve2->pred;
536 /* If they are not translations for the same basic block, they can't
537 be equal. */
538 if (b1 != b2)
539 return false;
540 return pre_expr_d::equal (ve1->e, ve2->e);
543 /* The phi_translate_table caches phi translations for a given
544 expression and predecessor. */
545 static hash_table <expr_pred_trans_d> phi_translate_table;
547 /* Add the tuple mapping from {expression E, basic block PRED} to
548 the phi translation table and return whether it pre-existed. */
550 static inline bool
551 phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred)
553 expr_pred_trans_t *slot;
554 expr_pred_trans_d tem;
555 hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e),
556 pred->index);
557 tem.e = e;
558 tem.pred = pred;
559 tem.hashcode = hash;
560 slot = phi_translate_table.find_slot_with_hash (&tem, hash, INSERT);
561 if (*slot)
563 *entry = *slot;
564 return true;
567 *entry = *slot = XNEW (struct expr_pred_trans_d);
568 (*entry)->e = e;
569 (*entry)->pred = pred;
570 (*entry)->hashcode = hash;
571 return false;
575 /* Add expression E to the expression set of value id V. */
577 static void
578 add_to_value (unsigned int v, pre_expr e)
580 bitmap set;
582 gcc_checking_assert (get_expr_value_id (e) == v);
584 if (v >= value_expressions.length ())
586 value_expressions.safe_grow_cleared (v + 1);
589 set = value_expressions[v];
590 if (!set)
592 set = BITMAP_ALLOC (&grand_bitmap_obstack);
593 value_expressions[v] = set;
596 bitmap_set_bit (set, get_or_alloc_expression_id (e));
599 /* Create a new bitmap set and return it. */
601 static bitmap_set_t
602 bitmap_set_new (void)
604 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
605 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
606 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
607 return ret;
610 /* Return the value id for a PRE expression EXPR. */
612 static unsigned int
613 get_expr_value_id (pre_expr expr)
615 unsigned int id;
616 switch (expr->kind)
618 case CONSTANT:
619 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
620 break;
621 case NAME:
622 id = VN_INFO (PRE_EXPR_NAME (expr))->value_id;
623 break;
624 case NARY:
625 id = PRE_EXPR_NARY (expr)->value_id;
626 break;
627 case REFERENCE:
628 id = PRE_EXPR_REFERENCE (expr)->value_id;
629 break;
630 default:
631 gcc_unreachable ();
633 /* ??? We cannot assert that expr has a value-id (it can be 0), because
634 we assign value-ids only to expressions that have a result
635 in set_hashtable_value_ids. */
636 return id;
639 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
641 static tree
642 sccvn_valnum_from_value_id (unsigned int val)
644 bitmap_iterator bi;
645 unsigned int i;
646 bitmap exprset = value_expressions[val];
647 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
649 pre_expr vexpr = expression_for_id (i);
650 if (vexpr->kind == NAME)
651 return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
652 else if (vexpr->kind == CONSTANT)
653 return PRE_EXPR_CONSTANT (vexpr);
655 return NULL_TREE;
658 /* Remove an expression EXPR from a bitmapped set. */
660 static void
661 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
663 unsigned int val = get_expr_value_id (expr);
664 if (!value_id_constant_p (val))
666 bitmap_clear_bit (&set->values, val);
667 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
671 static void
672 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
673 unsigned int val, bool allow_constants)
675 if (allow_constants || !value_id_constant_p (val))
677 /* We specifically expect this and only this function to be able to
678 insert constants into a set. */
679 bitmap_set_bit (&set->values, val);
680 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
684 /* Insert an expression EXPR into a bitmapped set. */
686 static void
687 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
689 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
692 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
694 static void
695 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
697 bitmap_copy (&dest->expressions, &orig->expressions);
698 bitmap_copy (&dest->values, &orig->values);
702 /* Free memory used up by SET. */
703 static void
704 bitmap_set_free (bitmap_set_t set)
706 bitmap_clear (&set->expressions);
707 bitmap_clear (&set->values);
711 /* Generate an topological-ordered array of bitmap set SET. */
713 static vec<pre_expr>
714 sorted_array_from_bitmap_set (bitmap_set_t set)
716 unsigned int i, j;
717 bitmap_iterator bi, bj;
718 vec<pre_expr> result;
720 /* Pre-allocate roughly enough space for the array. */
721 result.create (bitmap_count_bits (&set->values));
723 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
725 /* The number of expressions having a given value is usually
726 relatively small. Thus, rather than making a vector of all
727 the expressions and sorting it by value-id, we walk the values
728 and check in the reverse mapping that tells us what expressions
729 have a given value, to filter those in our set. As a result,
730 the expressions are inserted in value-id order, which means
731 topological order.
733 If this is somehow a significant lose for some cases, we can
734 choose which set to walk based on the set size. */
735 bitmap exprset = value_expressions[i];
736 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj)
738 if (bitmap_bit_p (&set->expressions, j))
739 result.safe_push (expression_for_id (j));
743 return result;
746 /* Perform bitmapped set operation DEST &= ORIG. */
748 static void
749 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
751 bitmap_iterator bi;
752 unsigned int i;
754 if (dest != orig)
756 bitmap_head temp;
757 bitmap_initialize (&temp, &grand_bitmap_obstack);
759 bitmap_and_into (&dest->values, &orig->values);
760 bitmap_copy (&temp, &dest->expressions);
761 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
763 pre_expr expr = expression_for_id (i);
764 unsigned int value_id = get_expr_value_id (expr);
765 if (!bitmap_bit_p (&dest->values, value_id))
766 bitmap_clear_bit (&dest->expressions, i);
768 bitmap_clear (&temp);
772 /* Subtract all values and expressions contained in ORIG from DEST. */
774 static bitmap_set_t
775 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
777 bitmap_set_t result = bitmap_set_new ();
778 bitmap_iterator bi;
779 unsigned int i;
781 bitmap_and_compl (&result->expressions, &dest->expressions,
782 &orig->expressions);
784 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
786 pre_expr expr = expression_for_id (i);
787 unsigned int value_id = get_expr_value_id (expr);
788 bitmap_set_bit (&result->values, value_id);
791 return result;
794 /* Subtract all the values in bitmap set B from bitmap set A. */
796 static void
797 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
799 unsigned int i;
800 bitmap_iterator bi;
801 bitmap_head temp;
803 bitmap_initialize (&temp, &grand_bitmap_obstack);
805 bitmap_copy (&temp, &a->expressions);
806 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
808 pre_expr expr = expression_for_id (i);
809 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
810 bitmap_remove_from_set (a, expr);
812 bitmap_clear (&temp);
816 /* Return true if bitmapped set SET contains the value VALUE_ID. */
818 static bool
819 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
821 if (value_id_constant_p (value_id))
822 return true;
824 if (!set || bitmap_empty_p (&set->expressions))
825 return false;
827 return bitmap_bit_p (&set->values, value_id);
830 static inline bool
831 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
833 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
836 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
838 static void
839 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
840 const pre_expr expr)
842 bitmap exprset;
843 unsigned int i;
844 bitmap_iterator bi;
846 if (value_id_constant_p (lookfor))
847 return;
849 if (!bitmap_set_contains_value (set, lookfor))
850 return;
852 /* The number of expressions having a given value is usually
853 significantly less than the total number of expressions in SET.
854 Thus, rather than check, for each expression in SET, whether it
855 has the value LOOKFOR, we walk the reverse mapping that tells us
856 what expressions have a given value, and see if any of those
857 expressions are in our set. For large testcases, this is about
858 5-10x faster than walking the bitmap. If this is somehow a
859 significant lose for some cases, we can choose which set to walk
860 based on the set size. */
861 exprset = value_expressions[lookfor];
862 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
864 if (bitmap_clear_bit (&set->expressions, i))
866 bitmap_set_bit (&set->expressions, get_expression_id (expr));
867 return;
871 gcc_unreachable ();
874 /* Return true if two bitmap sets are equal. */
876 static bool
877 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
879 return bitmap_equal_p (&a->values, &b->values);
882 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
883 and add it otherwise. */
885 static void
886 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
888 unsigned int val = get_expr_value_id (expr);
890 if (bitmap_set_contains_value (set, val))
891 bitmap_set_replace_value (set, val, expr);
892 else
893 bitmap_insert_into_set (set, expr);
896 /* Insert EXPR into SET if EXPR's value is not already present in
897 SET. */
899 static void
900 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
902 unsigned int val = get_expr_value_id (expr);
904 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
906 /* Constant values are always considered to be part of the set. */
907 if (value_id_constant_p (val))
908 return;
910 /* If the value membership changed, add the expression. */
911 if (bitmap_set_bit (&set->values, val))
912 bitmap_set_bit (&set->expressions, expr->id);
915 /* Print out EXPR to outfile. */
917 static void
918 print_pre_expr (FILE *outfile, const pre_expr expr)
920 switch (expr->kind)
922 case CONSTANT:
923 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
924 break;
925 case NAME:
926 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
927 break;
928 case NARY:
930 unsigned int i;
931 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
932 fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode));
933 for (i = 0; i < nary->length; i++)
935 print_generic_expr (outfile, nary->op[i], 0);
936 if (i != (unsigned) nary->length - 1)
937 fprintf (outfile, ",");
939 fprintf (outfile, "}");
941 break;
943 case REFERENCE:
945 vn_reference_op_t vro;
946 unsigned int i;
947 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
948 fprintf (outfile, "{");
949 for (i = 0;
950 ref->operands.iterate (i, &vro);
951 i++)
953 bool closebrace = false;
954 if (vro->opcode != SSA_NAME
955 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
957 fprintf (outfile, "%s", get_tree_code_name (vro->opcode));
958 if (vro->op0)
960 fprintf (outfile, "<");
961 closebrace = true;
964 if (vro->op0)
966 print_generic_expr (outfile, vro->op0, 0);
967 if (vro->op1)
969 fprintf (outfile, ",");
970 print_generic_expr (outfile, vro->op1, 0);
972 if (vro->op2)
974 fprintf (outfile, ",");
975 print_generic_expr (outfile, vro->op2, 0);
978 if (closebrace)
979 fprintf (outfile, ">");
980 if (i != ref->operands.length () - 1)
981 fprintf (outfile, ",");
983 fprintf (outfile, "}");
984 if (ref->vuse)
986 fprintf (outfile, "@");
987 print_generic_expr (outfile, ref->vuse, 0);
990 break;
993 void debug_pre_expr (pre_expr);
995 /* Like print_pre_expr but always prints to stderr. */
996 DEBUG_FUNCTION void
997 debug_pre_expr (pre_expr e)
999 print_pre_expr (stderr, e);
1000 fprintf (stderr, "\n");
1003 /* Print out SET to OUTFILE. */
1005 static void
1006 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1007 const char *setname, int blockindex)
1009 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1010 if (set)
1012 bool first = true;
1013 unsigned i;
1014 bitmap_iterator bi;
1016 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1018 const pre_expr expr = expression_for_id (i);
1020 if (!first)
1021 fprintf (outfile, ", ");
1022 first = false;
1023 print_pre_expr (outfile, expr);
1025 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1028 fprintf (outfile, " }\n");
1031 void debug_bitmap_set (bitmap_set_t);
1033 DEBUG_FUNCTION void
1034 debug_bitmap_set (bitmap_set_t set)
1036 print_bitmap_set (stderr, set, "debug", 0);
1039 void debug_bitmap_sets_for (basic_block);
1041 DEBUG_FUNCTION void
1042 debug_bitmap_sets_for (basic_block bb)
1044 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1045 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1046 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1047 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1048 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1049 if (do_partial_partial)
1050 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1051 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1054 /* Print out the expressions that have VAL to OUTFILE. */
1056 static void
1057 print_value_expressions (FILE *outfile, unsigned int val)
1059 bitmap set = value_expressions[val];
1060 if (set)
1062 bitmap_set x;
1063 char s[10];
1064 sprintf (s, "%04d", val);
1065 x.expressions = *set;
1066 print_bitmap_set (outfile, &x, s, 0);
1071 DEBUG_FUNCTION void
1072 debug_value_expressions (unsigned int val)
1074 print_value_expressions (stderr, val);
1077 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1078 represent it. */
1080 static pre_expr
1081 get_or_alloc_expr_for_constant (tree constant)
1083 unsigned int result_id;
1084 unsigned int value_id;
1085 struct pre_expr_d expr;
1086 pre_expr newexpr;
1088 expr.kind = CONSTANT;
1089 PRE_EXPR_CONSTANT (&expr) = constant;
1090 result_id = lookup_expression_id (&expr);
1091 if (result_id != 0)
1092 return expression_for_id (result_id);
1094 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1095 newexpr->kind = CONSTANT;
1096 PRE_EXPR_CONSTANT (newexpr) = constant;
1097 alloc_expression_id (newexpr);
1098 value_id = get_or_alloc_constant_value_id (constant);
1099 add_to_value (value_id, newexpr);
1100 return newexpr;
1103 /* Given a value id V, find the actual tree representing the constant
1104 value if there is one, and return it. Return NULL if we can't find
1105 a constant. */
1107 static tree
1108 get_constant_for_value_id (unsigned int v)
1110 if (value_id_constant_p (v))
1112 unsigned int i;
1113 bitmap_iterator bi;
1114 bitmap exprset = value_expressions[v];
1116 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1118 pre_expr expr = expression_for_id (i);
1119 if (expr->kind == CONSTANT)
1120 return PRE_EXPR_CONSTANT (expr);
1123 return NULL;
1126 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1127 Currently only supports constants and SSA_NAMES. */
1128 static pre_expr
1129 get_or_alloc_expr_for (tree t)
1131 if (TREE_CODE (t) == SSA_NAME)
1132 return get_or_alloc_expr_for_name (t);
1133 else if (is_gimple_min_invariant (t))
1134 return get_or_alloc_expr_for_constant (t);
1135 else
1137 /* More complex expressions can result from SCCVN expression
1138 simplification that inserts values for them. As they all
1139 do not have VOPs the get handled by the nary ops struct. */
1140 vn_nary_op_t result;
1141 unsigned int result_id;
1142 vn_nary_op_lookup (t, &result);
1143 if (result != NULL)
1145 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1146 e->kind = NARY;
1147 PRE_EXPR_NARY (e) = result;
1148 result_id = lookup_expression_id (e);
1149 if (result_id != 0)
1151 pool_free (pre_expr_pool, e);
1152 e = expression_for_id (result_id);
1153 return e;
1155 alloc_expression_id (e);
1156 return e;
1159 return NULL;
1162 /* Return the folded version of T if T, when folded, is a gimple
1163 min_invariant. Otherwise, return T. */
1165 static pre_expr
1166 fully_constant_expression (pre_expr e)
1168 switch (e->kind)
1170 case CONSTANT:
1171 return e;
1172 case NARY:
1174 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1175 switch (TREE_CODE_CLASS (nary->opcode))
1177 case tcc_binary:
1178 case tcc_comparison:
1180 /* We have to go from trees to pre exprs to value ids to
1181 constants. */
1182 tree naryop0 = nary->op[0];
1183 tree naryop1 = nary->op[1];
1184 tree result;
1185 if (!is_gimple_min_invariant (naryop0))
1187 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1188 unsigned int vrep0 = get_expr_value_id (rep0);
1189 tree const0 = get_constant_for_value_id (vrep0);
1190 if (const0)
1191 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1193 if (!is_gimple_min_invariant (naryop1))
1195 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1196 unsigned int vrep1 = get_expr_value_id (rep1);
1197 tree const1 = get_constant_for_value_id (vrep1);
1198 if (const1)
1199 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1201 result = fold_binary (nary->opcode, nary->type,
1202 naryop0, naryop1);
1203 if (result && is_gimple_min_invariant (result))
1204 return get_or_alloc_expr_for_constant (result);
1205 /* We might have simplified the expression to a
1206 SSA_NAME for example from x_1 * 1. But we cannot
1207 insert a PHI for x_1 unconditionally as x_1 might
1208 not be available readily. */
1209 return e;
1211 case tcc_reference:
1212 if (nary->opcode != REALPART_EXPR
1213 && nary->opcode != IMAGPART_EXPR
1214 && nary->opcode != VIEW_CONVERT_EXPR)
1215 return e;
1216 /* Fallthrough. */
1217 case tcc_unary:
1219 /* We have to go from trees to pre exprs to value ids to
1220 constants. */
1221 tree naryop0 = nary->op[0];
1222 tree const0, result;
1223 if (is_gimple_min_invariant (naryop0))
1224 const0 = naryop0;
1225 else
1227 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1228 unsigned int vrep0 = get_expr_value_id (rep0);
1229 const0 = get_constant_for_value_id (vrep0);
1231 result = NULL;
1232 if (const0)
1234 tree type1 = TREE_TYPE (nary->op[0]);
1235 const0 = fold_convert (type1, const0);
1236 result = fold_unary (nary->opcode, nary->type, const0);
1238 if (result && is_gimple_min_invariant (result))
1239 return get_or_alloc_expr_for_constant (result);
1240 return e;
1242 default:
1243 return e;
1246 case REFERENCE:
1248 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1249 tree folded;
1250 if ((folded = fully_constant_vn_reference_p (ref)))
1251 return get_or_alloc_expr_for_constant (folded);
1252 return e;
1254 default:
1255 return e;
1257 return e;
1260 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1261 it has the value it would have in BLOCK. Set *SAME_VALID to true
1262 in case the new vuse doesn't change the value id of the OPERANDS. */
1264 static tree
1265 translate_vuse_through_block (vec<vn_reference_op_s> operands,
1266 alias_set_type set, tree type, tree vuse,
1267 basic_block phiblock,
1268 basic_block block, bool *same_valid)
1270 gimple phi = SSA_NAME_DEF_STMT (vuse);
1271 ao_ref ref;
1272 edge e = NULL;
1273 bool use_oracle;
1275 *same_valid = true;
1277 if (gimple_bb (phi) != phiblock)
1278 return vuse;
1280 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1282 /* Use the alias-oracle to find either the PHI node in this block,
1283 the first VUSE used in this block that is equivalent to vuse or
1284 the first VUSE which definition in this block kills the value. */
1285 if (gimple_code (phi) == GIMPLE_PHI)
1286 e = find_edge (block, phiblock);
1287 else if (use_oracle)
1288 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1290 vuse = gimple_vuse (phi);
1291 phi = SSA_NAME_DEF_STMT (vuse);
1292 if (gimple_bb (phi) != phiblock)
1293 return vuse;
1294 if (gimple_code (phi) == GIMPLE_PHI)
1296 e = find_edge (block, phiblock);
1297 break;
1300 else
1301 return NULL_TREE;
1303 if (e)
1305 if (use_oracle)
1307 bitmap visited = NULL;
1308 unsigned int cnt;
1309 /* Try to find a vuse that dominates this phi node by skipping
1310 non-clobbering statements. */
1311 vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false);
1312 if (visited)
1313 BITMAP_FREE (visited);
1315 else
1316 vuse = NULL_TREE;
1317 if (!vuse)
1319 /* If we didn't find any, the value ID can't stay the same,
1320 but return the translated vuse. */
1321 *same_valid = false;
1322 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1324 /* ??? We would like to return vuse here as this is the canonical
1325 upmost vdef that this reference is associated with. But during
1326 insertion of the references into the hash tables we only ever
1327 directly insert with their direct gimple_vuse, hence returning
1328 something else would make us not find the other expression. */
1329 return PHI_ARG_DEF (phi, e->dest_idx);
1332 return NULL_TREE;
1335 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1336 SET2. This is used to avoid making a set consisting of the union
1337 of PA_IN and ANTIC_IN during insert. */
1339 static inline pre_expr
1340 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1342 pre_expr result;
1344 result = bitmap_find_leader (set1, val);
1345 if (!result && set2)
1346 result = bitmap_find_leader (set2, val);
1347 return result;
1350 /* Get the tree type for our PRE expression e. */
1352 static tree
1353 get_expr_type (const pre_expr e)
1355 switch (e->kind)
1357 case NAME:
1358 return TREE_TYPE (PRE_EXPR_NAME (e));
1359 case CONSTANT:
1360 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1361 case REFERENCE:
1362 return PRE_EXPR_REFERENCE (e)->type;
1363 case NARY:
1364 return PRE_EXPR_NARY (e)->type;
1366 gcc_unreachable ();
1369 /* Get a representative SSA_NAME for a given expression.
1370 Since all of our sub-expressions are treated as values, we require
1371 them to be SSA_NAME's for simplicity.
1372 Prior versions of GVNPRE used to use "value handles" here, so that
1373 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1374 either case, the operands are really values (IE we do not expect
1375 them to be usable without finding leaders). */
1377 static tree
1378 get_representative_for (const pre_expr e)
1380 tree name;
1381 unsigned int value_id = get_expr_value_id (e);
1383 switch (e->kind)
1385 case NAME:
1386 return PRE_EXPR_NAME (e);
1387 case CONSTANT:
1388 return PRE_EXPR_CONSTANT (e);
1389 case NARY:
1390 case REFERENCE:
1392 /* Go through all of the expressions representing this value
1393 and pick out an SSA_NAME. */
1394 unsigned int i;
1395 bitmap_iterator bi;
1396 bitmap exprs = value_expressions[value_id];
1397 EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1399 pre_expr rep = expression_for_id (i);
1400 if (rep->kind == NAME)
1401 return PRE_EXPR_NAME (rep);
1402 else if (rep->kind == CONSTANT)
1403 return PRE_EXPR_CONSTANT (rep);
1406 break;
1409 /* If we reached here we couldn't find an SSA_NAME. This can
1410 happen when we've discovered a value that has never appeared in
1411 the program as set to an SSA_NAME, as the result of phi translation.
1412 Create one here.
1413 ??? We should be able to re-use this when we insert the statement
1414 to compute it. */
1415 name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1416 VN_INFO_GET (name)->value_id = value_id;
1417 VN_INFO (name)->valnum = name;
1418 /* ??? For now mark this SSA name for release by SCCVN. */
1419 VN_INFO (name)->needs_insertion = true;
1420 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1421 if (dump_file && (dump_flags & TDF_DETAILS))
1423 fprintf (dump_file, "Created SSA_NAME representative ");
1424 print_generic_expr (dump_file, name, 0);
1425 fprintf (dump_file, " for expression:");
1426 print_pre_expr (dump_file, e);
1427 fprintf (dump_file, " (%04d)\n", value_id);
1430 return name;
1435 static pre_expr
1436 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1437 basic_block pred, basic_block phiblock);
1439 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1440 the phis in PRED. Return NULL if we can't find a leader for each part
1441 of the translated expression. */
1443 static pre_expr
1444 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1445 basic_block pred, basic_block phiblock)
1447 switch (expr->kind)
1449 case NARY:
1451 unsigned int i;
1452 bool changed = false;
1453 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1454 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1455 sizeof_vn_nary_op (nary->length));
1456 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1458 for (i = 0; i < newnary->length; i++)
1460 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1461 continue;
1462 else
1464 pre_expr leader, result;
1465 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1466 leader = find_leader_in_sets (op_val_id, set1, set2);
1467 result = phi_translate (leader, set1, set2, pred, phiblock);
1468 if (result && result != leader)
1470 tree name = get_representative_for (result);
1471 if (!name)
1472 return NULL;
1473 newnary->op[i] = name;
1475 else if (!result)
1476 return NULL;
1478 changed |= newnary->op[i] != nary->op[i];
1481 if (changed)
1483 pre_expr constant;
1484 unsigned int new_val_id;
1486 tree result = vn_nary_op_lookup_pieces (newnary->length,
1487 newnary->opcode,
1488 newnary->type,
1489 &newnary->op[0],
1490 &nary);
1491 if (result && is_gimple_min_invariant (result))
1492 return get_or_alloc_expr_for_constant (result);
1494 expr = (pre_expr) pool_alloc (pre_expr_pool);
1495 expr->kind = NARY;
1496 expr->id = 0;
1497 if (nary)
1499 PRE_EXPR_NARY (expr) = nary;
1500 constant = fully_constant_expression (expr);
1501 if (constant != expr)
1502 return constant;
1504 new_val_id = nary->value_id;
1505 get_or_alloc_expression_id (expr);
1507 else
1509 new_val_id = get_next_value_id ();
1510 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
1511 nary = vn_nary_op_insert_pieces (newnary->length,
1512 newnary->opcode,
1513 newnary->type,
1514 &newnary->op[0],
1515 result, new_val_id);
1516 PRE_EXPR_NARY (expr) = nary;
1517 constant = fully_constant_expression (expr);
1518 if (constant != expr)
1519 return constant;
1520 get_or_alloc_expression_id (expr);
1522 add_to_value (new_val_id, expr);
1524 return expr;
1526 break;
1528 case REFERENCE:
1530 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1531 vec<vn_reference_op_s> operands = ref->operands;
1532 tree vuse = ref->vuse;
1533 tree newvuse = vuse;
1534 vec<vn_reference_op_s> newoperands = vNULL;
1535 bool changed = false, same_valid = true;
1536 unsigned int i, j, n;
1537 vn_reference_op_t operand;
1538 vn_reference_t newref;
1540 for (i = 0, j = 0;
1541 operands.iterate (i, &operand); i++, j++)
1543 pre_expr opresult;
1544 pre_expr leader;
1545 tree op[3];
1546 tree type = operand->type;
1547 vn_reference_op_s newop = *operand;
1548 op[0] = operand->op0;
1549 op[1] = operand->op1;
1550 op[2] = operand->op2;
1551 for (n = 0; n < 3; ++n)
1553 unsigned int op_val_id;
1554 if (!op[n])
1555 continue;
1556 if (TREE_CODE (op[n]) != SSA_NAME)
1558 /* We can't possibly insert these. */
1559 if (n != 0
1560 && !is_gimple_min_invariant (op[n]))
1561 break;
1562 continue;
1564 op_val_id = VN_INFO (op[n])->value_id;
1565 leader = find_leader_in_sets (op_val_id, set1, set2);
1566 if (!leader)
1567 break;
1568 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1569 if (!opresult)
1570 break;
1571 if (opresult != leader)
1573 tree name = get_representative_for (opresult);
1574 if (!name)
1575 break;
1576 changed |= name != op[n];
1577 op[n] = name;
1580 if (n != 3)
1582 newoperands.release ();
1583 return NULL;
1585 if (!newoperands.exists ())
1586 newoperands = operands.copy ();
1587 /* We may have changed from an SSA_NAME to a constant */
1588 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1589 newop.opcode = TREE_CODE (op[0]);
1590 newop.type = type;
1591 newop.op0 = op[0];
1592 newop.op1 = op[1];
1593 newop.op2 = op[2];
1594 /* If it transforms a non-constant ARRAY_REF into a constant
1595 one, adjust the constant offset. */
1596 if (newop.opcode == ARRAY_REF
1597 && newop.off == -1
1598 && TREE_CODE (op[0]) == INTEGER_CST
1599 && TREE_CODE (op[1]) == INTEGER_CST
1600 && TREE_CODE (op[2]) == INTEGER_CST)
1602 double_int off = tree_to_double_int (op[0]);
1603 off += -tree_to_double_int (op[1]);
1604 off *= tree_to_double_int (op[2]);
1605 if (off.fits_shwi ())
1606 newop.off = off.low;
1608 newoperands[j] = newop;
1609 /* If it transforms from an SSA_NAME to an address, fold with
1610 a preceding indirect reference. */
1611 if (j > 0 && op[0] && TREE_CODE (op[0]) == ADDR_EXPR
1612 && newoperands[j - 1].opcode == MEM_REF)
1613 vn_reference_fold_indirect (&newoperands, &j);
1615 if (i != operands.length ())
1617 newoperands.release ();
1618 return NULL;
1621 if (vuse)
1623 newvuse = translate_vuse_through_block (newoperands,
1624 ref->set, ref->type,
1625 vuse, phiblock, pred,
1626 &same_valid);
1627 if (newvuse == NULL_TREE)
1629 newoperands.release ();
1630 return NULL;
1634 if (changed || newvuse != vuse)
1636 unsigned int new_val_id;
1637 pre_expr constant;
1639 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1640 ref->type,
1641 newoperands,
1642 &newref, VN_WALK);
1643 if (result)
1644 newoperands.release ();
1646 /* We can always insert constants, so if we have a partial
1647 redundant constant load of another type try to translate it
1648 to a constant of appropriate type. */
1649 if (result && is_gimple_min_invariant (result))
1651 tree tem = result;
1652 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1654 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1655 if (tem && !is_gimple_min_invariant (tem))
1656 tem = NULL_TREE;
1658 if (tem)
1659 return get_or_alloc_expr_for_constant (tem);
1662 /* If we'd have to convert things we would need to validate
1663 if we can insert the translated expression. So fail
1664 here for now - we cannot insert an alias with a different
1665 type in the VN tables either, as that would assert. */
1666 if (result
1667 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1668 return NULL;
1669 else if (!result && newref
1670 && !useless_type_conversion_p (ref->type, newref->type))
1672 newoperands.release ();
1673 return NULL;
1676 expr = (pre_expr) pool_alloc (pre_expr_pool);
1677 expr->kind = REFERENCE;
1678 expr->id = 0;
1680 if (newref)
1682 PRE_EXPR_REFERENCE (expr) = newref;
1683 constant = fully_constant_expression (expr);
1684 if (constant != expr)
1685 return constant;
1687 new_val_id = newref->value_id;
1688 get_or_alloc_expression_id (expr);
1690 else
1692 if (changed || !same_valid)
1694 new_val_id = get_next_value_id ();
1695 value_expressions.safe_grow_cleared
1696 (get_max_value_id () + 1);
1698 else
1699 new_val_id = ref->value_id;
1700 newref = vn_reference_insert_pieces (newvuse, ref->set,
1701 ref->type,
1702 newoperands,
1703 result, new_val_id);
1704 newoperands.create (0);
1705 PRE_EXPR_REFERENCE (expr) = newref;
1706 constant = fully_constant_expression (expr);
1707 if (constant != expr)
1708 return constant;
1709 get_or_alloc_expression_id (expr);
1711 add_to_value (new_val_id, expr);
1713 newoperands.release ();
1714 return expr;
1716 break;
1718 case NAME:
1720 tree name = PRE_EXPR_NAME (expr);
1721 gimple def_stmt = SSA_NAME_DEF_STMT (name);
1722 /* If the SSA name is defined by a PHI node in this block,
1723 translate it. */
1724 if (gimple_code (def_stmt) == GIMPLE_PHI
1725 && gimple_bb (def_stmt) == phiblock)
1727 edge e = find_edge (pred, gimple_bb (def_stmt));
1728 tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1730 /* Handle constant. */
1731 if (is_gimple_min_invariant (def))
1732 return get_or_alloc_expr_for_constant (def);
1734 return get_or_alloc_expr_for_name (def);
1736 /* Otherwise return it unchanged - it will get cleaned if its
1737 value is not available in PREDs AVAIL_OUT set of expressions. */
1738 return expr;
1741 default:
1742 gcc_unreachable ();
1746 /* Wrapper around phi_translate_1 providing caching functionality. */
1748 static pre_expr
1749 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1750 basic_block pred, basic_block phiblock)
1752 expr_pred_trans_t slot = NULL;
1753 pre_expr phitrans;
1755 if (!expr)
1756 return NULL;
1758 /* Constants contain no values that need translation. */
1759 if (expr->kind == CONSTANT)
1760 return expr;
1762 if (value_id_constant_p (get_expr_value_id (expr)))
1763 return expr;
1765 /* Don't add translations of NAMEs as those are cheap to translate. */
1766 if (expr->kind != NAME)
1768 if (phi_trans_add (&slot, expr, pred))
1769 return slot->v;
1770 /* Store NULL for the value we want to return in the case of
1771 recursing. */
1772 slot->v = NULL;
1775 /* Translate. */
1776 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1778 if (slot)
1780 if (phitrans)
1781 slot->v = phitrans;
1782 else
1783 /* Remove failed translations again, they cause insert
1784 iteration to not pick up new opportunities reliably. */
1785 phi_translate_table.remove_elt_with_hash (slot, slot->hashcode);
1788 return phitrans;
1792 /* For each expression in SET, translate the values through phi nodes
1793 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1794 expressions in DEST. */
1796 static void
1797 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1798 basic_block phiblock)
1800 vec<pre_expr> exprs;
1801 pre_expr expr;
1802 int i;
1804 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1806 bitmap_set_copy (dest, set);
1807 return;
1810 exprs = sorted_array_from_bitmap_set (set);
1811 FOR_EACH_VEC_ELT (exprs, i, expr)
1813 pre_expr translated;
1814 translated = phi_translate (expr, set, NULL, pred, phiblock);
1815 if (!translated)
1816 continue;
1818 /* We might end up with multiple expressions from SET being
1819 translated to the same value. In this case we do not want
1820 to retain the NARY or REFERENCE expression but prefer a NAME
1821 which would be the leader. */
1822 if (translated->kind == NAME)
1823 bitmap_value_replace_in_set (dest, translated);
1824 else
1825 bitmap_value_insert_into_set (dest, translated);
1827 exprs.release ();
1830 /* Find the leader for a value (i.e., the name representing that
1831 value) in a given set, and return it. Return NULL if no leader
1832 is found. */
1834 static pre_expr
1835 bitmap_find_leader (bitmap_set_t set, unsigned int val)
1837 if (value_id_constant_p (val))
1839 unsigned int i;
1840 bitmap_iterator bi;
1841 bitmap exprset = value_expressions[val];
1843 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1845 pre_expr expr = expression_for_id (i);
1846 if (expr->kind == CONSTANT)
1847 return expr;
1850 if (bitmap_set_contains_value (set, val))
1852 /* Rather than walk the entire bitmap of expressions, and see
1853 whether any of them has the value we are looking for, we look
1854 at the reverse mapping, which tells us the set of expressions
1855 that have a given value (IE value->expressions with that
1856 value) and see if any of those expressions are in our set.
1857 The number of expressions per value is usually significantly
1858 less than the number of expressions in the set. In fact, for
1859 large testcases, doing it this way is roughly 5-10x faster
1860 than walking the bitmap.
1861 If this is somehow a significant lose for some cases, we can
1862 choose which set to walk based on which set is smaller. */
1863 unsigned int i;
1864 bitmap_iterator bi;
1865 bitmap exprset = value_expressions[val];
1867 EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1868 return expression_for_id (i);
1870 return NULL;
1873 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1874 BLOCK by seeing if it is not killed in the block. Note that we are
1875 only determining whether there is a store that kills it. Because
1876 of the order in which clean iterates over values, we are guaranteed
1877 that altered operands will have caused us to be eliminated from the
1878 ANTIC_IN set already. */
1880 static bool
1881 value_dies_in_block_x (pre_expr expr, basic_block block)
1883 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1884 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1885 gimple def;
1886 gimple_stmt_iterator gsi;
1887 unsigned id = get_expression_id (expr);
1888 bool res = false;
1889 ao_ref ref;
1891 if (!vuse)
1892 return false;
1894 /* Lookup a previously calculated result. */
1895 if (EXPR_DIES (block)
1896 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1897 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1899 /* A memory expression {e, VUSE} dies in the block if there is a
1900 statement that may clobber e. If, starting statement walk from the
1901 top of the basic block, a statement uses VUSE there can be no kill
1902 inbetween that use and the original statement that loaded {e, VUSE},
1903 so we can stop walking. */
1904 ref.base = NULL_TREE;
1905 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1907 tree def_vuse, def_vdef;
1908 def = gsi_stmt (gsi);
1909 def_vuse = gimple_vuse (def);
1910 def_vdef = gimple_vdef (def);
1912 /* Not a memory statement. */
1913 if (!def_vuse)
1914 continue;
1916 /* Not a may-def. */
1917 if (!def_vdef)
1919 /* A load with the same VUSE, we're done. */
1920 if (def_vuse == vuse)
1921 break;
1923 continue;
1926 /* Init ref only if we really need it. */
1927 if (ref.base == NULL_TREE
1928 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1929 refx->operands))
1931 res = true;
1932 break;
1934 /* If the statement may clobber expr, it dies. */
1935 if (stmt_may_clobber_ref_p_1 (def, &ref))
1937 res = true;
1938 break;
1942 /* Remember the result. */
1943 if (!EXPR_DIES (block))
1944 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1945 bitmap_set_bit (EXPR_DIES (block), id * 2);
1946 if (res)
1947 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1949 return res;
1953 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1954 contains its value-id. */
1956 static bool
1957 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1959 if (op && TREE_CODE (op) == SSA_NAME)
1961 unsigned int value_id = VN_INFO (op)->value_id;
1962 if (!(bitmap_set_contains_value (set1, value_id)
1963 || (set2 && bitmap_set_contains_value (set2, value_id))))
1964 return false;
1966 return true;
1969 /* Determine if the expression EXPR is valid in SET1 U SET2.
1970 ONLY SET2 CAN BE NULL.
1971 This means that we have a leader for each part of the expression
1972 (if it consists of values), or the expression is an SSA_NAME.
1973 For loads/calls, we also see if the vuse is killed in this block. */
1975 static bool
1976 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
1977 basic_block block)
1979 switch (expr->kind)
1981 case NAME:
1982 return bitmap_find_leader (AVAIL_OUT (block),
1983 get_expr_value_id (expr)) != NULL;
1984 case NARY:
1986 unsigned int i;
1987 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1988 for (i = 0; i < nary->length; i++)
1989 if (!op_valid_in_sets (set1, set2, nary->op[i]))
1990 return false;
1991 return true;
1993 break;
1994 case REFERENCE:
1996 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1997 vn_reference_op_t vro;
1998 unsigned int i;
2000 FOR_EACH_VEC_ELT (ref->operands, i, vro)
2002 if (!op_valid_in_sets (set1, set2, vro->op0)
2003 || !op_valid_in_sets (set1, set2, vro->op1)
2004 || !op_valid_in_sets (set1, set2, vro->op2))
2005 return false;
2007 return true;
2009 default:
2010 gcc_unreachable ();
2014 /* Clean the set of expressions that are no longer valid in SET1 or
2015 SET2. This means expressions that are made up of values we have no
2016 leaders for in SET1 or SET2. This version is used for partial
2017 anticipation, which means it is not valid in either ANTIC_IN or
2018 PA_IN. */
2020 static void
2021 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2023 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
2024 pre_expr expr;
2025 int i;
2027 FOR_EACH_VEC_ELT (exprs, i, expr)
2029 if (!valid_in_sets (set1, set2, expr, block))
2030 bitmap_remove_from_set (set1, expr);
2032 exprs.release ();
2035 /* Clean the set of expressions that are no longer valid in SET. This
2036 means expressions that are made up of values we have no leaders for
2037 in SET. */
2039 static void
2040 clean (bitmap_set_t set, basic_block block)
2042 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set);
2043 pre_expr expr;
2044 int i;
2046 FOR_EACH_VEC_ELT (exprs, i, expr)
2048 if (!valid_in_sets (set, NULL, expr, block))
2049 bitmap_remove_from_set (set, expr);
2051 exprs.release ();
2054 /* Clean the set of expressions that are no longer valid in SET because
2055 they are clobbered in BLOCK or because they trap and may not be executed. */
2057 static void
2058 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2060 bitmap_iterator bi;
2061 unsigned i;
2063 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2065 pre_expr expr = expression_for_id (i);
2066 if (expr->kind == REFERENCE)
2068 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2069 if (ref->vuse)
2071 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2072 if (!gimple_nop_p (def_stmt)
2073 && ((gimple_bb (def_stmt) != block
2074 && !dominated_by_p (CDI_DOMINATORS,
2075 block, gimple_bb (def_stmt)))
2076 || (gimple_bb (def_stmt) == block
2077 && value_dies_in_block_x (expr, block))))
2078 bitmap_remove_from_set (set, expr);
2081 else if (expr->kind == NARY)
2083 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2084 /* If the NARY may trap make sure the block does not contain
2085 a possible exit point.
2086 ??? This is overly conservative if we translate AVAIL_OUT
2087 as the available expression might be after the exit point. */
2088 if (BB_MAY_NOTRETURN (block)
2089 && vn_nary_may_trap (nary))
2090 bitmap_remove_from_set (set, expr);
2095 static sbitmap has_abnormal_preds;
2097 /* List of blocks that may have changed during ANTIC computation and
2098 thus need to be iterated over. */
2100 static sbitmap changed_blocks;
2102 /* Decide whether to defer a block for a later iteration, or PHI
2103 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2104 should defer the block, and true if we processed it. */
2106 static bool
2107 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2108 basic_block block, basic_block phiblock)
2110 if (!BB_VISITED (phiblock))
2112 bitmap_set_bit (changed_blocks, block->index);
2113 BB_VISITED (block) = 0;
2114 BB_DEFERRED (block) = 1;
2115 return false;
2117 else
2118 phi_translate_set (dest, source, block, phiblock);
2119 return true;
2122 /* Compute the ANTIC set for BLOCK.
2124 If succs(BLOCK) > 1 then
2125 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2126 else if succs(BLOCK) == 1 then
2127 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2129 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2132 static bool
2133 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2135 bool changed = false;
2136 bitmap_set_t S, old, ANTIC_OUT;
2137 bitmap_iterator bi;
2138 unsigned int bii;
2139 edge e;
2140 edge_iterator ei;
2142 old = ANTIC_OUT = S = NULL;
2143 BB_VISITED (block) = 1;
2145 /* If any edges from predecessors are abnormal, antic_in is empty,
2146 so do nothing. */
2147 if (block_has_abnormal_pred_edge)
2148 goto maybe_dump_sets;
2150 old = ANTIC_IN (block);
2151 ANTIC_OUT = bitmap_set_new ();
2153 /* If the block has no successors, ANTIC_OUT is empty. */
2154 if (EDGE_COUNT (block->succs) == 0)
2156 /* If we have one successor, we could have some phi nodes to
2157 translate through. */
2158 else if (single_succ_p (block))
2160 basic_block succ_bb = single_succ (block);
2162 /* We trade iterations of the dataflow equations for having to
2163 phi translate the maximal set, which is incredibly slow
2164 (since the maximal set often has 300+ members, even when you
2165 have a small number of blocks).
2166 Basically, we defer the computation of ANTIC for this block
2167 until we have processed it's successor, which will inevitably
2168 have a *much* smaller set of values to phi translate once
2169 clean has been run on it.
2170 The cost of doing this is that we technically perform more
2171 iterations, however, they are lower cost iterations.
2173 Timings for PRE on tramp3d-v4:
2174 without maximal set fix: 11 seconds
2175 with maximal set fix/without deferring: 26 seconds
2176 with maximal set fix/with deferring: 11 seconds
2179 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2180 block, succ_bb))
2182 changed = true;
2183 goto maybe_dump_sets;
2186 /* If we have multiple successors, we take the intersection of all of
2187 them. Note that in the case of loop exit phi nodes, we may have
2188 phis to translate through. */
2189 else
2191 size_t i;
2192 basic_block bprime, first = NULL;
2194 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2195 FOR_EACH_EDGE (e, ei, block->succs)
2197 if (!first
2198 && BB_VISITED (e->dest))
2199 first = e->dest;
2200 else if (BB_VISITED (e->dest))
2201 worklist.quick_push (e->dest);
2204 /* Of multiple successors we have to have visited one already. */
2205 if (!first)
2207 bitmap_set_bit (changed_blocks, block->index);
2208 BB_VISITED (block) = 0;
2209 BB_DEFERRED (block) = 1;
2210 changed = true;
2211 worklist.release ();
2212 goto maybe_dump_sets;
2215 if (!gimple_seq_empty_p (phi_nodes (first)))
2216 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2217 else
2218 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2220 FOR_EACH_VEC_ELT (worklist, i, bprime)
2222 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2224 bitmap_set_t tmp = bitmap_set_new ();
2225 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2226 bitmap_set_and (ANTIC_OUT, tmp);
2227 bitmap_set_free (tmp);
2229 else
2230 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2234 /* Prune expressions that are clobbered in block and thus become
2235 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2236 prune_clobbered_mems (ANTIC_OUT, block);
2238 /* Generate ANTIC_OUT - TMP_GEN. */
2239 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2241 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2242 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2243 TMP_GEN (block));
2245 /* Then union in the ANTIC_OUT - TMP_GEN values,
2246 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2247 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2248 bitmap_value_insert_into_set (ANTIC_IN (block),
2249 expression_for_id (bii));
2251 clean (ANTIC_IN (block), block);
2253 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2255 changed = true;
2256 bitmap_set_bit (changed_blocks, block->index);
2257 FOR_EACH_EDGE (e, ei, block->preds)
2258 bitmap_set_bit (changed_blocks, e->src->index);
2260 else
2261 bitmap_clear_bit (changed_blocks, block->index);
2263 maybe_dump_sets:
2264 if (dump_file && (dump_flags & TDF_DETAILS))
2266 if (!BB_DEFERRED (block) || BB_VISITED (block))
2268 if (ANTIC_OUT)
2269 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2271 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2272 block->index);
2274 if (S)
2275 print_bitmap_set (dump_file, S, "S", block->index);
2277 else
2279 fprintf (dump_file,
2280 "Block %d was deferred for a future iteration.\n",
2281 block->index);
2284 if (old)
2285 bitmap_set_free (old);
2286 if (S)
2287 bitmap_set_free (S);
2288 if (ANTIC_OUT)
2289 bitmap_set_free (ANTIC_OUT);
2290 return changed;
2293 /* Compute PARTIAL_ANTIC for BLOCK.
2295 If succs(BLOCK) > 1 then
2296 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2297 in ANTIC_OUT for all succ(BLOCK)
2298 else if succs(BLOCK) == 1 then
2299 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2301 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2302 - ANTIC_IN[BLOCK])
2305 static bool
2306 compute_partial_antic_aux (basic_block block,
2307 bool block_has_abnormal_pred_edge)
2309 bool changed = false;
2310 bitmap_set_t old_PA_IN;
2311 bitmap_set_t PA_OUT;
2312 edge e;
2313 edge_iterator ei;
2314 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2316 old_PA_IN = PA_OUT = NULL;
2318 /* If any edges from predecessors are abnormal, antic_in is empty,
2319 so do nothing. */
2320 if (block_has_abnormal_pred_edge)
2321 goto maybe_dump_sets;
2323 /* If there are too many partially anticipatable values in the
2324 block, phi_translate_set can take an exponential time: stop
2325 before the translation starts. */
2326 if (max_pa
2327 && single_succ_p (block)
2328 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2329 goto maybe_dump_sets;
2331 old_PA_IN = PA_IN (block);
2332 PA_OUT = bitmap_set_new ();
2334 /* If the block has no successors, ANTIC_OUT is empty. */
2335 if (EDGE_COUNT (block->succs) == 0)
2337 /* If we have one successor, we could have some phi nodes to
2338 translate through. Note that we can't phi translate across DFS
2339 back edges in partial antic, because it uses a union operation on
2340 the successors. For recurrences like IV's, we will end up
2341 generating a new value in the set on each go around (i + 3 (VH.1)
2342 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2343 else if (single_succ_p (block))
2345 basic_block succ = single_succ (block);
2346 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2347 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2349 /* If we have multiple successors, we take the union of all of
2350 them. */
2351 else
2353 size_t i;
2354 basic_block bprime;
2356 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2357 FOR_EACH_EDGE (e, ei, block->succs)
2359 if (e->flags & EDGE_DFS_BACK)
2360 continue;
2361 worklist.quick_push (e->dest);
2363 if (worklist.length () > 0)
2365 FOR_EACH_VEC_ELT (worklist, i, bprime)
2367 unsigned int i;
2368 bitmap_iterator bi;
2370 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2371 bitmap_value_insert_into_set (PA_OUT,
2372 expression_for_id (i));
2373 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2375 bitmap_set_t pa_in = bitmap_set_new ();
2376 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2377 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2378 bitmap_value_insert_into_set (PA_OUT,
2379 expression_for_id (i));
2380 bitmap_set_free (pa_in);
2382 else
2383 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2384 bitmap_value_insert_into_set (PA_OUT,
2385 expression_for_id (i));
2390 /* Prune expressions that are clobbered in block and thus become
2391 invalid if translated from PA_OUT to PA_IN. */
2392 prune_clobbered_mems (PA_OUT, block);
2394 /* PA_IN starts with PA_OUT - TMP_GEN.
2395 Then we subtract things from ANTIC_IN. */
2396 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2398 /* For partial antic, we want to put back in the phi results, since
2399 we will properly avoid making them partially antic over backedges. */
2400 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2401 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2403 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2404 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2406 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2408 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2410 changed = true;
2411 bitmap_set_bit (changed_blocks, block->index);
2412 FOR_EACH_EDGE (e, ei, block->preds)
2413 bitmap_set_bit (changed_blocks, e->src->index);
2415 else
2416 bitmap_clear_bit (changed_blocks, block->index);
2418 maybe_dump_sets:
2419 if (dump_file && (dump_flags & TDF_DETAILS))
2421 if (PA_OUT)
2422 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2424 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2426 if (old_PA_IN)
2427 bitmap_set_free (old_PA_IN);
2428 if (PA_OUT)
2429 bitmap_set_free (PA_OUT);
2430 return changed;
2433 /* Compute ANTIC and partial ANTIC sets. */
2435 static void
2436 compute_antic (void)
2438 bool changed = true;
2439 int num_iterations = 0;
2440 basic_block block;
2441 int i;
2443 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2444 We pre-build the map of blocks with incoming abnormal edges here. */
2445 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2446 bitmap_clear (has_abnormal_preds);
2448 FOR_ALL_BB (block)
2450 edge_iterator ei;
2451 edge e;
2453 FOR_EACH_EDGE (e, ei, block->preds)
2455 e->flags &= ~EDGE_DFS_BACK;
2456 if (e->flags & EDGE_ABNORMAL)
2458 bitmap_set_bit (has_abnormal_preds, block->index);
2459 break;
2463 BB_VISITED (block) = 0;
2464 BB_DEFERRED (block) = 0;
2466 /* While we are here, give empty ANTIC_IN sets to each block. */
2467 ANTIC_IN (block) = bitmap_set_new ();
2468 PA_IN (block) = bitmap_set_new ();
2471 /* At the exit block we anticipate nothing. */
2472 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun)) = 1;
2474 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2475 bitmap_ones (changed_blocks);
2476 while (changed)
2478 if (dump_file && (dump_flags & TDF_DETAILS))
2479 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2480 /* ??? We need to clear our PHI translation cache here as the
2481 ANTIC sets shrink and we restrict valid translations to
2482 those having operands with leaders in ANTIC. Same below
2483 for PA ANTIC computation. */
2484 num_iterations++;
2485 changed = false;
2486 for (i = postorder_num - 1; i >= 0; i--)
2488 if (bitmap_bit_p (changed_blocks, postorder[i]))
2490 basic_block block = BASIC_BLOCK (postorder[i]);
2491 changed |= compute_antic_aux (block,
2492 bitmap_bit_p (has_abnormal_preds,
2493 block->index));
2496 /* Theoretically possible, but *highly* unlikely. */
2497 gcc_checking_assert (num_iterations < 500);
2500 statistics_histogram_event (cfun, "compute_antic iterations",
2501 num_iterations);
2503 if (do_partial_partial)
2505 bitmap_ones (changed_blocks);
2506 mark_dfs_back_edges ();
2507 num_iterations = 0;
2508 changed = true;
2509 while (changed)
2511 if (dump_file && (dump_flags & TDF_DETAILS))
2512 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2513 num_iterations++;
2514 changed = false;
2515 for (i = postorder_num - 1 ; i >= 0; i--)
2517 if (bitmap_bit_p (changed_blocks, postorder[i]))
2519 basic_block block = BASIC_BLOCK (postorder[i]);
2520 changed
2521 |= compute_partial_antic_aux (block,
2522 bitmap_bit_p (has_abnormal_preds,
2523 block->index));
2526 /* Theoretically possible, but *highly* unlikely. */
2527 gcc_checking_assert (num_iterations < 500);
2529 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2530 num_iterations);
2532 sbitmap_free (has_abnormal_preds);
2533 sbitmap_free (changed_blocks);
2537 /* Inserted expressions are placed onto this worklist, which is used
2538 for performing quick dead code elimination of insertions we made
2539 that didn't turn out to be necessary. */
2540 static bitmap inserted_exprs;
2542 /* The actual worker for create_component_ref_by_pieces. */
2544 static tree
2545 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2546 unsigned int *operand, gimple_seq *stmts)
2548 vn_reference_op_t currop = &ref->operands[*operand];
2549 tree genop;
2550 ++*operand;
2551 switch (currop->opcode)
2553 case CALL_EXPR:
2555 tree folded, sc = NULL_TREE;
2556 unsigned int nargs = 0;
2557 tree fn, *args;
2558 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2559 fn = currop->op0;
2560 else
2561 fn = find_or_generate_expression (block, currop->op0, stmts);
2562 if (!fn)
2563 return NULL_TREE;
2564 if (currop->op1)
2566 sc = find_or_generate_expression (block, currop->op1, stmts);
2567 if (!sc)
2568 return NULL_TREE;
2570 args = XNEWVEC (tree, ref->operands.length () - 1);
2571 while (*operand < ref->operands.length ())
2573 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2574 operand, stmts);
2575 if (!args[nargs])
2576 return NULL_TREE;
2577 nargs++;
2579 folded = build_call_array (currop->type,
2580 (TREE_CODE (fn) == FUNCTION_DECL
2581 ? build_fold_addr_expr (fn) : fn),
2582 nargs, args);
2583 free (args);
2584 if (sc)
2585 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2586 return folded;
2589 case MEM_REF:
2591 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2592 stmts);
2593 if (!baseop)
2594 return NULL_TREE;
2595 tree offset = currop->op0;
2596 if (TREE_CODE (baseop) == ADDR_EXPR
2597 && handled_component_p (TREE_OPERAND (baseop, 0)))
2599 HOST_WIDE_INT off;
2600 tree base;
2601 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2602 &off);
2603 gcc_assert (base);
2604 offset = int_const_binop (PLUS_EXPR, offset,
2605 build_int_cst (TREE_TYPE (offset),
2606 off));
2607 baseop = build_fold_addr_expr (base);
2609 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2612 case TARGET_MEM_REF:
2614 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2615 vn_reference_op_t nextop = &ref->operands[++*operand];
2616 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2617 stmts);
2618 if (!baseop)
2619 return NULL_TREE;
2620 if (currop->op0)
2622 genop0 = find_or_generate_expression (block, currop->op0, stmts);
2623 if (!genop0)
2624 return NULL_TREE;
2626 if (nextop->op0)
2628 genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2629 if (!genop1)
2630 return NULL_TREE;
2632 return build5 (TARGET_MEM_REF, currop->type,
2633 baseop, currop->op2, genop0, currop->op1, genop1);
2636 case ADDR_EXPR:
2637 if (currop->op0)
2639 gcc_assert (is_gimple_min_invariant (currop->op0));
2640 return currop->op0;
2642 /* Fallthrough. */
2643 case REALPART_EXPR:
2644 case IMAGPART_EXPR:
2645 case VIEW_CONVERT_EXPR:
2647 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2648 stmts);
2649 if (!genop0)
2650 return NULL_TREE;
2651 return fold_build1 (currop->opcode, currop->type, genop0);
2654 case WITH_SIZE_EXPR:
2656 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2657 stmts);
2658 if (!genop0)
2659 return NULL_TREE;
2660 tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2661 if (!genop1)
2662 return NULL_TREE;
2663 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2666 case BIT_FIELD_REF:
2668 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2669 stmts);
2670 if (!genop0)
2671 return NULL_TREE;
2672 tree op1 = currop->op0;
2673 tree op2 = currop->op1;
2674 return fold_build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2677 /* For array ref vn_reference_op's, operand 1 of the array ref
2678 is op0 of the reference op and operand 3 of the array ref is
2679 op1. */
2680 case ARRAY_RANGE_REF:
2681 case ARRAY_REF:
2683 tree genop0;
2684 tree genop1 = currop->op0;
2685 tree genop2 = currop->op1;
2686 tree genop3 = currop->op2;
2687 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2688 stmts);
2689 if (!genop0)
2690 return NULL_TREE;
2691 genop1 = find_or_generate_expression (block, genop1, stmts);
2692 if (!genop1)
2693 return NULL_TREE;
2694 if (genop2)
2696 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2697 /* Drop zero minimum index if redundant. */
2698 if (integer_zerop (genop2)
2699 && (!domain_type
2700 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2701 genop2 = NULL_TREE;
2702 else
2704 genop2 = find_or_generate_expression (block, genop2, stmts);
2705 if (!genop2)
2706 return NULL_TREE;
2709 if (genop3)
2711 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2712 /* We can't always put a size in units of the element alignment
2713 here as the element alignment may be not visible. See
2714 PR43783. Simply drop the element size for constant
2715 sizes. */
2716 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2717 genop3 = NULL_TREE;
2718 else
2720 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2721 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2722 genop3 = find_or_generate_expression (block, genop3, stmts);
2723 if (!genop3)
2724 return NULL_TREE;
2727 return build4 (currop->opcode, currop->type, genop0, genop1,
2728 genop2, genop3);
2730 case COMPONENT_REF:
2732 tree op0;
2733 tree op1;
2734 tree genop2 = currop->op1;
2735 op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2736 if (!op0)
2737 return NULL_TREE;
2738 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2739 op1 = currop->op0;
2740 if (genop2)
2742 genop2 = find_or_generate_expression (block, genop2, stmts);
2743 if (!genop2)
2744 return NULL_TREE;
2746 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2749 case SSA_NAME:
2751 genop = find_or_generate_expression (block, currop->op0, stmts);
2752 return genop;
2754 case STRING_CST:
2755 case INTEGER_CST:
2756 case COMPLEX_CST:
2757 case VECTOR_CST:
2758 case REAL_CST:
2759 case CONSTRUCTOR:
2760 case VAR_DECL:
2761 case PARM_DECL:
2762 case CONST_DECL:
2763 case RESULT_DECL:
2764 case FUNCTION_DECL:
2765 return currop->op0;
2767 default:
2768 gcc_unreachable ();
2772 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2773 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2774 trying to rename aggregates into ssa form directly, which is a no no.
2776 Thus, this routine doesn't create temporaries, it just builds a
2777 single access expression for the array, calling
2778 find_or_generate_expression to build the innermost pieces.
2780 This function is a subroutine of create_expression_by_pieces, and
2781 should not be called on it's own unless you really know what you
2782 are doing. */
2784 static tree
2785 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2786 gimple_seq *stmts)
2788 unsigned int op = 0;
2789 return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2792 /* Find a simple leader for an expression, or generate one using
2793 create_expression_by_pieces from a NARY expression for the value.
2794 BLOCK is the basic_block we are looking for leaders in.
2795 OP is the tree expression to find a leader for or generate.
2796 Returns the leader or NULL_TREE on failure. */
2798 static tree
2799 find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2801 pre_expr expr = get_or_alloc_expr_for (op);
2802 unsigned int lookfor = get_expr_value_id (expr);
2803 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2804 if (leader)
2806 if (leader->kind == NAME)
2807 return PRE_EXPR_NAME (leader);
2808 else if (leader->kind == CONSTANT)
2809 return PRE_EXPR_CONSTANT (leader);
2811 /* Defer. */
2812 return NULL_TREE;
2815 /* It must be a complex expression, so generate it recursively. Note
2816 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2817 where the insert algorithm fails to insert a required expression. */
2818 bitmap exprset = value_expressions[lookfor];
2819 bitmap_iterator bi;
2820 unsigned int i;
2821 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2823 pre_expr temp = expression_for_id (i);
2824 /* We cannot insert random REFERENCE expressions at arbitrary
2825 places. We can insert NARYs which eventually re-materializes
2826 its operand values. */
2827 if (temp->kind == NARY)
2828 return create_expression_by_pieces (block, temp, stmts,
2829 get_expr_type (expr));
2832 /* Defer. */
2833 return NULL_TREE;
2836 #define NECESSARY GF_PLF_1
2838 /* Create an expression in pieces, so that we can handle very complex
2839 expressions that may be ANTIC, but not necessary GIMPLE.
2840 BLOCK is the basic block the expression will be inserted into,
2841 EXPR is the expression to insert (in value form)
2842 STMTS is a statement list to append the necessary insertions into.
2844 This function will die if we hit some value that shouldn't be
2845 ANTIC but is (IE there is no leader for it, or its components).
2846 The function returns NULL_TREE in case a different antic expression
2847 has to be inserted first.
2848 This function may also generate expressions that are themselves
2849 partially or fully redundant. Those that are will be either made
2850 fully redundant during the next iteration of insert (for partially
2851 redundant ones), or eliminated by eliminate (for fully redundant
2852 ones). */
2854 static tree
2855 create_expression_by_pieces (basic_block block, pre_expr expr,
2856 gimple_seq *stmts, tree type)
2858 tree name;
2859 tree folded;
2860 gimple_seq forced_stmts = NULL;
2861 unsigned int value_id;
2862 gimple_stmt_iterator gsi;
2863 tree exprtype = type ? type : get_expr_type (expr);
2864 pre_expr nameexpr;
2865 gimple newstmt;
2867 switch (expr->kind)
2869 /* We may hit the NAME/CONSTANT case if we have to convert types
2870 that value numbering saw through. */
2871 case NAME:
2872 folded = PRE_EXPR_NAME (expr);
2873 break;
2874 case CONSTANT:
2875 folded = PRE_EXPR_CONSTANT (expr);
2876 break;
2877 case REFERENCE:
2879 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2880 folded = create_component_ref_by_pieces (block, ref, stmts);
2881 if (!folded)
2882 return NULL_TREE;
2884 break;
2885 case NARY:
2887 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2888 tree *genop = XALLOCAVEC (tree, nary->length);
2889 unsigned i;
2890 for (i = 0; i < nary->length; ++i)
2892 genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2893 if (!genop[i])
2894 return NULL_TREE;
2895 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2896 may have conversions stripped. */
2897 if (nary->opcode == POINTER_PLUS_EXPR)
2899 if (i == 0)
2900 genop[i] = fold_convert (nary->type, genop[i]);
2901 else if (i == 1)
2902 genop[i] = convert_to_ptrofftype (genop[i]);
2904 else
2905 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]);
2907 if (nary->opcode == CONSTRUCTOR)
2909 vec<constructor_elt, va_gc> *elts = NULL;
2910 for (i = 0; i < nary->length; ++i)
2911 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2912 folded = build_constructor (nary->type, elts);
2914 else
2916 switch (nary->length)
2918 case 1:
2919 folded = fold_build1 (nary->opcode, nary->type,
2920 genop[0]);
2921 break;
2922 case 2:
2923 folded = fold_build2 (nary->opcode, nary->type,
2924 genop[0], genop[1]);
2925 break;
2926 case 3:
2927 folded = fold_build3 (nary->opcode, nary->type,
2928 genop[0], genop[1], genop[2]);
2929 break;
2930 default:
2931 gcc_unreachable ();
2935 break;
2936 default:
2937 gcc_unreachable ();
2940 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2941 folded = fold_convert (exprtype, folded);
2943 /* Force the generated expression to be a sequence of GIMPLE
2944 statements.
2945 We have to call unshare_expr because force_gimple_operand may
2946 modify the tree we pass to it. */
2947 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
2948 false, NULL);
2950 /* If we have any intermediate expressions to the value sets, add them
2951 to the value sets and chain them in the instruction stream. */
2952 if (forced_stmts)
2954 gsi = gsi_start (forced_stmts);
2955 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2957 gimple stmt = gsi_stmt (gsi);
2958 tree forcedname = gimple_get_lhs (stmt);
2959 pre_expr nameexpr;
2961 if (TREE_CODE (forcedname) == SSA_NAME)
2963 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
2964 VN_INFO_GET (forcedname)->valnum = forcedname;
2965 VN_INFO (forcedname)->value_id = get_next_value_id ();
2966 nameexpr = get_or_alloc_expr_for_name (forcedname);
2967 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2968 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2969 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2972 gimple_seq_add_seq (stmts, forced_stmts);
2975 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2976 newstmt = gimple_build_assign (name, folded);
2977 gimple_set_plf (newstmt, NECESSARY, false);
2979 gimple_seq_add_stmt (stmts, newstmt);
2980 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
2982 /* Fold the last statement. */
2983 gsi = gsi_last (*stmts);
2984 if (fold_stmt_inplace (&gsi))
2985 update_stmt (gsi_stmt (gsi));
2987 /* Add a value number to the temporary.
2988 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2989 we are creating the expression by pieces, and this particular piece of
2990 the expression may have been represented. There is no harm in replacing
2991 here. */
2992 value_id = get_expr_value_id (expr);
2993 VN_INFO_GET (name)->value_id = value_id;
2994 VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id);
2995 if (VN_INFO (name)->valnum == NULL_TREE)
2996 VN_INFO (name)->valnum = name;
2997 gcc_assert (VN_INFO (name)->valnum != NULL_TREE);
2998 nameexpr = get_or_alloc_expr_for_name (name);
2999 add_to_value (value_id, nameexpr);
3000 if (NEW_SETS (block))
3001 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3002 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3004 pre_stats.insertions++;
3005 if (dump_file && (dump_flags & TDF_DETAILS))
3007 fprintf (dump_file, "Inserted ");
3008 print_gimple_stmt (dump_file, newstmt, 0, 0);
3009 fprintf (dump_file, " in predecessor %d (%04d)\n",
3010 block->index, value_id);
3013 return name;
3017 /* Returns true if we want to inhibit the insertions of PHI nodes
3018 for the given EXPR for basic block BB (a member of a loop).
3019 We want to do this, when we fear that the induction variable we
3020 create might inhibit vectorization. */
3022 static bool
3023 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3025 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3026 vec<vn_reference_op_s> ops = vr->operands;
3027 vn_reference_op_t op;
3028 unsigned i;
3030 /* If we aren't going to vectorize we don't inhibit anything. */
3031 if (!flag_tree_loop_vectorize)
3032 return false;
3034 /* Otherwise we inhibit the insertion when the address of the
3035 memory reference is a simple induction variable. In other
3036 cases the vectorizer won't do anything anyway (either it's
3037 loop invariant or a complicated expression). */
3038 FOR_EACH_VEC_ELT (ops, i, op)
3040 switch (op->opcode)
3042 case CALL_EXPR:
3043 /* Calls are not a problem. */
3044 return false;
3046 case ARRAY_REF:
3047 case ARRAY_RANGE_REF:
3048 if (TREE_CODE (op->op0) != SSA_NAME)
3049 break;
3050 /* Fallthru. */
3051 case SSA_NAME:
3053 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3054 affine_iv iv;
3055 /* Default defs are loop invariant. */
3056 if (!defbb)
3057 break;
3058 /* Defined outside this loop, also loop invariant. */
3059 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3060 break;
3061 /* If it's a simple induction variable inhibit insertion,
3062 the vectorizer might be interested in this one. */
3063 if (simple_iv (bb->loop_father, bb->loop_father,
3064 op->op0, &iv, true))
3065 return true;
3066 /* No simple IV, vectorizer can't do anything, hence no
3067 reason to inhibit the transformation for this operand. */
3068 break;
3070 default:
3071 break;
3074 return false;
3077 /* Insert the to-be-made-available values of expression EXPRNUM for each
3078 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3079 merge the result with a phi node, given the same value number as
3080 NODE. Return true if we have inserted new stuff. */
3082 static bool
3083 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3084 vec<pre_expr> avail)
3086 pre_expr expr = expression_for_id (exprnum);
3087 pre_expr newphi;
3088 unsigned int val = get_expr_value_id (expr);
3089 edge pred;
3090 bool insertions = false;
3091 bool nophi = false;
3092 basic_block bprime;
3093 pre_expr eprime;
3094 edge_iterator ei;
3095 tree type = get_expr_type (expr);
3096 tree temp;
3097 gimple phi;
3099 /* Make sure we aren't creating an induction variable. */
3100 if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
3102 bool firstinsideloop = false;
3103 bool secondinsideloop = false;
3104 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3105 EDGE_PRED (block, 0)->src);
3106 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3107 EDGE_PRED (block, 1)->src);
3108 /* Induction variables only have one edge inside the loop. */
3109 if ((firstinsideloop ^ secondinsideloop)
3110 && (expr->kind != REFERENCE
3111 || inhibit_phi_insertion (block, expr)))
3113 if (dump_file && (dump_flags & TDF_DETAILS))
3114 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3115 nophi = true;
3119 /* Make the necessary insertions. */
3120 FOR_EACH_EDGE (pred, ei, block->preds)
3122 gimple_seq stmts = NULL;
3123 tree builtexpr;
3124 bprime = pred->src;
3125 eprime = avail[pred->dest_idx];
3127 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3129 builtexpr = create_expression_by_pieces (bprime, eprime,
3130 &stmts, type);
3131 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3132 gsi_insert_seq_on_edge (pred, stmts);
3133 if (!builtexpr)
3135 /* We cannot insert a PHI node if we failed to insert
3136 on one edge. */
3137 nophi = true;
3138 continue;
3140 avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3141 insertions = true;
3143 else if (eprime->kind == CONSTANT)
3145 /* Constants may not have the right type, fold_convert
3146 should give us back a constant with the right type. */
3147 tree constant = PRE_EXPR_CONSTANT (eprime);
3148 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3150 tree builtexpr = fold_convert (type, constant);
3151 if (!is_gimple_min_invariant (builtexpr))
3153 tree forcedexpr = force_gimple_operand (builtexpr,
3154 &stmts, true,
3155 NULL);
3156 if (!is_gimple_min_invariant (forcedexpr))
3158 if (forcedexpr != builtexpr)
3160 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3161 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3163 if (stmts)
3165 gimple_stmt_iterator gsi;
3166 gsi = gsi_start (stmts);
3167 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3169 gimple stmt = gsi_stmt (gsi);
3170 tree lhs = gimple_get_lhs (stmt);
3171 if (TREE_CODE (lhs) == SSA_NAME)
3172 bitmap_set_bit (inserted_exprs,
3173 SSA_NAME_VERSION (lhs));
3174 gimple_set_plf (stmt, NECESSARY, false);
3176 gsi_insert_seq_on_edge (pred, stmts);
3178 avail[pred->dest_idx]
3179 = get_or_alloc_expr_for_name (forcedexpr);
3182 else
3183 avail[pred->dest_idx]
3184 = get_or_alloc_expr_for_constant (builtexpr);
3187 else if (eprime->kind == NAME)
3189 /* We may have to do a conversion because our value
3190 numbering can look through types in certain cases, but
3191 our IL requires all operands of a phi node have the same
3192 type. */
3193 tree name = PRE_EXPR_NAME (eprime);
3194 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3196 tree builtexpr;
3197 tree forcedexpr;
3198 builtexpr = fold_convert (type, name);
3199 forcedexpr = force_gimple_operand (builtexpr,
3200 &stmts, true,
3201 NULL);
3203 if (forcedexpr != name)
3205 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3206 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3209 if (stmts)
3211 gimple_stmt_iterator gsi;
3212 gsi = gsi_start (stmts);
3213 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3215 gimple stmt = gsi_stmt (gsi);
3216 tree lhs = gimple_get_lhs (stmt);
3217 if (TREE_CODE (lhs) == SSA_NAME)
3218 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3219 gimple_set_plf (stmt, NECESSARY, false);
3221 gsi_insert_seq_on_edge (pred, stmts);
3223 avail[pred->dest_idx] = get_or_alloc_expr_for_name (forcedexpr);
3227 /* If we didn't want a phi node, and we made insertions, we still have
3228 inserted new stuff, and thus return true. If we didn't want a phi node,
3229 and didn't make insertions, we haven't added anything new, so return
3230 false. */
3231 if (nophi && insertions)
3232 return true;
3233 else if (nophi && !insertions)
3234 return false;
3236 /* Now build a phi for the new variable. */
3237 temp = make_temp_ssa_name (type, NULL, "prephitmp");
3238 phi = create_phi_node (temp, block);
3240 gimple_set_plf (phi, NECESSARY, false);
3241 VN_INFO_GET (temp)->value_id = val;
3242 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3243 if (VN_INFO (temp)->valnum == NULL_TREE)
3244 VN_INFO (temp)->valnum = temp;
3245 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3246 FOR_EACH_EDGE (pred, ei, block->preds)
3248 pre_expr ae = avail[pred->dest_idx];
3249 gcc_assert (get_expr_type (ae) == type
3250 || useless_type_conversion_p (type, get_expr_type (ae)));
3251 if (ae->kind == CONSTANT)
3252 add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3253 pred, UNKNOWN_LOCATION);
3254 else
3255 add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3258 newphi = get_or_alloc_expr_for_name (temp);
3259 add_to_value (val, newphi);
3261 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3262 this insertion, since we test for the existence of this value in PHI_GEN
3263 before proceeding with the partial redundancy checks in insert_aux.
3265 The value may exist in AVAIL_OUT, in particular, it could be represented
3266 by the expression we are trying to eliminate, in which case we want the
3267 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3268 inserted there.
3270 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3271 this block, because if it did, it would have existed in our dominator's
3272 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3275 bitmap_insert_into_set (PHI_GEN (block), newphi);
3276 bitmap_value_replace_in_set (AVAIL_OUT (block),
3277 newphi);
3278 bitmap_insert_into_set (NEW_SETS (block),
3279 newphi);
3281 if (dump_file && (dump_flags & TDF_DETAILS))
3283 fprintf (dump_file, "Created phi ");
3284 print_gimple_stmt (dump_file, phi, 0, 0);
3285 fprintf (dump_file, " in block %d (%04d)\n", block->index, val);
3287 pre_stats.phis++;
3288 return true;
3293 /* Perform insertion of partially redundant values.
3294 For BLOCK, do the following:
3295 1. Propagate the NEW_SETS of the dominator into the current block.
3296 If the block has multiple predecessors,
3297 2a. Iterate over the ANTIC expressions for the block to see if
3298 any of them are partially redundant.
3299 2b. If so, insert them into the necessary predecessors to make
3300 the expression fully redundant.
3301 2c. Insert a new PHI merging the values of the predecessors.
3302 2d. Insert the new PHI, and the new expressions, into the
3303 NEW_SETS set.
3304 3. Recursively call ourselves on the dominator children of BLOCK.
3306 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3307 do_regular_insertion and do_partial_insertion.
3311 static bool
3312 do_regular_insertion (basic_block block, basic_block dom)
3314 bool new_stuff = false;
3315 vec<pre_expr> exprs;
3316 pre_expr expr;
3317 vec<pre_expr> avail = vNULL;
3318 int i;
3320 exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3321 avail.safe_grow (EDGE_COUNT (block->preds));
3323 FOR_EACH_VEC_ELT (exprs, i, expr)
3325 if (expr->kind == NARY
3326 || expr->kind == REFERENCE)
3328 unsigned int val;
3329 bool by_some = false;
3330 bool cant_insert = false;
3331 bool all_same = true;
3332 pre_expr first_s = NULL;
3333 edge pred;
3334 basic_block bprime;
3335 pre_expr eprime = NULL;
3336 edge_iterator ei;
3337 pre_expr edoubleprime = NULL;
3338 bool do_insertion = false;
3340 val = get_expr_value_id (expr);
3341 if (bitmap_set_contains_value (PHI_GEN (block), val))
3342 continue;
3343 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3345 if (dump_file && (dump_flags & TDF_DETAILS))
3347 fprintf (dump_file, "Found fully redundant value: ");
3348 print_pre_expr (dump_file, expr);
3349 fprintf (dump_file, "\n");
3351 continue;
3354 FOR_EACH_EDGE (pred, ei, block->preds)
3356 unsigned int vprime;
3358 /* We should never run insertion for the exit block
3359 and so not come across fake pred edges. */
3360 gcc_assert (!(pred->flags & EDGE_FAKE));
3361 bprime = pred->src;
3362 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3363 bprime, block);
3365 /* eprime will generally only be NULL if the
3366 value of the expression, translated
3367 through the PHI for this predecessor, is
3368 undefined. If that is the case, we can't
3369 make the expression fully redundant,
3370 because its value is undefined along a
3371 predecessor path. We can thus break out
3372 early because it doesn't matter what the
3373 rest of the results are. */
3374 if (eprime == NULL)
3376 avail[pred->dest_idx] = NULL;
3377 cant_insert = true;
3378 break;
3381 eprime = fully_constant_expression (eprime);
3382 vprime = get_expr_value_id (eprime);
3383 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3384 vprime);
3385 if (edoubleprime == NULL)
3387 avail[pred->dest_idx] = eprime;
3388 all_same = false;
3390 else
3392 avail[pred->dest_idx] = edoubleprime;
3393 by_some = true;
3394 /* We want to perform insertions to remove a redundancy on
3395 a path in the CFG we want to optimize for speed. */
3396 if (optimize_edge_for_speed_p (pred))
3397 do_insertion = true;
3398 if (first_s == NULL)
3399 first_s = edoubleprime;
3400 else if (!pre_expr_d::equal (first_s, edoubleprime))
3401 all_same = false;
3404 /* If we can insert it, it's not the same value
3405 already existing along every predecessor, and
3406 it's defined by some predecessor, it is
3407 partially redundant. */
3408 if (!cant_insert && !all_same && by_some)
3410 if (!do_insertion)
3412 if (dump_file && (dump_flags & TDF_DETAILS))
3414 fprintf (dump_file, "Skipping partial redundancy for "
3415 "expression ");
3416 print_pre_expr (dump_file, expr);
3417 fprintf (dump_file, " (%04d), no redundancy on to be "
3418 "optimized for speed edge\n", val);
3421 else if (dbg_cnt (treepre_insert))
3423 if (dump_file && (dump_flags & TDF_DETAILS))
3425 fprintf (dump_file, "Found partial redundancy for "
3426 "expression ");
3427 print_pre_expr (dump_file, expr);
3428 fprintf (dump_file, " (%04d)\n",
3429 get_expr_value_id (expr));
3431 if (insert_into_preds_of_block (block,
3432 get_expression_id (expr),
3433 avail))
3434 new_stuff = true;
3437 /* If all edges produce the same value and that value is
3438 an invariant, then the PHI has the same value on all
3439 edges. Note this. */
3440 else if (!cant_insert && all_same)
3442 gcc_assert (edoubleprime->kind == CONSTANT
3443 || edoubleprime->kind == NAME);
3445 tree temp = make_temp_ssa_name (get_expr_type (expr),
3446 NULL, "pretmp");
3447 gimple assign = gimple_build_assign (temp,
3448 edoubleprime->kind == CONSTANT ? PRE_EXPR_CONSTANT (edoubleprime) : PRE_EXPR_NAME (edoubleprime));
3449 gimple_stmt_iterator gsi = gsi_after_labels (block);
3450 gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3452 gimple_set_plf (assign, NECESSARY, false);
3453 VN_INFO_GET (temp)->value_id = val;
3454 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3455 if (VN_INFO (temp)->valnum == NULL_TREE)
3456 VN_INFO (temp)->valnum = temp;
3457 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3458 pre_expr newe = get_or_alloc_expr_for_name (temp);
3459 add_to_value (val, newe);
3460 bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3461 bitmap_insert_into_set (NEW_SETS (block), newe);
3466 exprs.release ();
3467 return new_stuff;
3471 /* Perform insertion for partially anticipatable expressions. There
3472 is only one case we will perform insertion for these. This case is
3473 if the expression is partially anticipatable, and fully available.
3474 In this case, we know that putting it earlier will enable us to
3475 remove the later computation. */
3478 static bool
3479 do_partial_partial_insertion (basic_block block, basic_block dom)
3481 bool new_stuff = false;
3482 vec<pre_expr> exprs;
3483 pre_expr expr;
3484 auto_vec<pre_expr> avail;
3485 int i;
3487 exprs = sorted_array_from_bitmap_set (PA_IN (block));
3488 avail.safe_grow (EDGE_COUNT (block->preds));
3490 FOR_EACH_VEC_ELT (exprs, i, expr)
3492 if (expr->kind == NARY
3493 || expr->kind == REFERENCE)
3495 unsigned int val;
3496 bool by_all = true;
3497 bool cant_insert = false;
3498 edge pred;
3499 basic_block bprime;
3500 pre_expr eprime = NULL;
3501 edge_iterator ei;
3503 val = get_expr_value_id (expr);
3504 if (bitmap_set_contains_value (PHI_GEN (block), val))
3505 continue;
3506 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3507 continue;
3509 FOR_EACH_EDGE (pred, ei, block->preds)
3511 unsigned int vprime;
3512 pre_expr edoubleprime;
3514 /* We should never run insertion for the exit block
3515 and so not come across fake pred edges. */
3516 gcc_assert (!(pred->flags & EDGE_FAKE));
3517 bprime = pred->src;
3518 eprime = phi_translate (expr, ANTIC_IN (block),
3519 PA_IN (block),
3520 bprime, block);
3522 /* eprime will generally only be NULL if the
3523 value of the expression, translated
3524 through the PHI for this predecessor, is
3525 undefined. If that is the case, we can't
3526 make the expression fully redundant,
3527 because its value is undefined along a
3528 predecessor path. We can thus break out
3529 early because it doesn't matter what the
3530 rest of the results are. */
3531 if (eprime == NULL)
3533 avail[pred->dest_idx] = NULL;
3534 cant_insert = true;
3535 break;
3538 eprime = fully_constant_expression (eprime);
3539 vprime = get_expr_value_id (eprime);
3540 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3541 avail[pred->dest_idx] = edoubleprime;
3542 if (edoubleprime == NULL)
3544 by_all = false;
3545 break;
3549 /* If we can insert it, it's not the same value
3550 already existing along every predecessor, and
3551 it's defined by some predecessor, it is
3552 partially redundant. */
3553 if (!cant_insert && by_all)
3555 edge succ;
3556 bool do_insertion = false;
3558 /* Insert only if we can remove a later expression on a path
3559 that we want to optimize for speed.
3560 The phi node that we will be inserting in BLOCK is not free,
3561 and inserting it for the sake of !optimize_for_speed successor
3562 may cause regressions on the speed path. */
3563 FOR_EACH_EDGE (succ, ei, block->succs)
3565 if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3566 || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3568 if (optimize_edge_for_speed_p (succ))
3569 do_insertion = true;
3573 if (!do_insertion)
3575 if (dump_file && (dump_flags & TDF_DETAILS))
3577 fprintf (dump_file, "Skipping partial partial redundancy "
3578 "for expression ");
3579 print_pre_expr (dump_file, expr);
3580 fprintf (dump_file, " (%04d), not (partially) anticipated "
3581 "on any to be optimized for speed edges\n", val);
3584 else if (dbg_cnt (treepre_insert))
3586 pre_stats.pa_insert++;
3587 if (dump_file && (dump_flags & TDF_DETAILS))
3589 fprintf (dump_file, "Found partial partial redundancy "
3590 "for expression ");
3591 print_pre_expr (dump_file, expr);
3592 fprintf (dump_file, " (%04d)\n",
3593 get_expr_value_id (expr));
3595 if (insert_into_preds_of_block (block,
3596 get_expression_id (expr),
3597 avail))
3598 new_stuff = true;
3604 exprs.release ();
3605 return new_stuff;
3608 static bool
3609 insert_aux (basic_block block)
3611 basic_block son;
3612 bool new_stuff = false;
3614 if (block)
3616 basic_block dom;
3617 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3618 if (dom)
3620 unsigned i;
3621 bitmap_iterator bi;
3622 bitmap_set_t newset = NEW_SETS (dom);
3623 if (newset)
3625 /* Note that we need to value_replace both NEW_SETS, and
3626 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3627 represented by some non-simple expression here that we want
3628 to replace it with. */
3629 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3631 pre_expr expr = expression_for_id (i);
3632 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3633 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3636 if (!single_pred_p (block))
3638 new_stuff |= do_regular_insertion (block, dom);
3639 if (do_partial_partial)
3640 new_stuff |= do_partial_partial_insertion (block, dom);
3644 for (son = first_dom_son (CDI_DOMINATORS, block);
3645 son;
3646 son = next_dom_son (CDI_DOMINATORS, son))
3648 new_stuff |= insert_aux (son);
3651 return new_stuff;
3654 /* Perform insertion of partially redundant values. */
3656 static void
3657 insert (void)
3659 bool new_stuff = true;
3660 basic_block bb;
3661 int num_iterations = 0;
3663 FOR_ALL_BB (bb)
3664 NEW_SETS (bb) = bitmap_set_new ();
3666 while (new_stuff)
3668 num_iterations++;
3669 if (dump_file && dump_flags & TDF_DETAILS)
3670 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3671 new_stuff = insert_aux (ENTRY_BLOCK_PTR_FOR_FN (cfun));
3673 /* Clear the NEW sets before the next iteration. We have already
3674 fully propagated its contents. */
3675 if (new_stuff)
3676 FOR_ALL_BB (bb)
3677 bitmap_set_free (NEW_SETS (bb));
3679 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3683 /* Compute the AVAIL set for all basic blocks.
3685 This function performs value numbering of the statements in each basic
3686 block. The AVAIL sets are built from information we glean while doing
3687 this value numbering, since the AVAIL sets contain only one entry per
3688 value.
3690 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3691 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3693 static void
3694 compute_avail (void)
3697 basic_block block, son;
3698 basic_block *worklist;
3699 size_t sp = 0;
3700 unsigned i;
3702 /* We pretend that default definitions are defined in the entry block.
3703 This includes function arguments and the static chain decl. */
3704 for (i = 1; i < num_ssa_names; ++i)
3706 tree name = ssa_name (i);
3707 pre_expr e;
3708 if (!name
3709 || !SSA_NAME_IS_DEFAULT_DEF (name)
3710 || has_zero_uses (name)
3711 || virtual_operand_p (name))
3712 continue;
3714 e = get_or_alloc_expr_for_name (name);
3715 add_to_value (get_expr_value_id (e), e);
3716 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)), e);
3717 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3721 if (dump_file && (dump_flags & TDF_DETAILS))
3723 print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3724 "tmp_gen", ENTRY_BLOCK);
3725 print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3726 "avail_out", ENTRY_BLOCK);
3729 /* Allocate the worklist. */
3730 worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
3732 /* Seed the algorithm by putting the dominator children of the entry
3733 block on the worklist. */
3734 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR_FOR_FN (cfun));
3735 son;
3736 son = next_dom_son (CDI_DOMINATORS, son))
3737 worklist[sp++] = son;
3739 /* Loop until the worklist is empty. */
3740 while (sp)
3742 gimple_stmt_iterator gsi;
3743 gimple stmt;
3744 basic_block dom;
3746 /* Pick a block from the worklist. */
3747 block = worklist[--sp];
3749 /* Initially, the set of available values in BLOCK is that of
3750 its immediate dominator. */
3751 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3752 if (dom)
3753 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3755 /* Generate values for PHI nodes. */
3756 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3758 tree result = gimple_phi_result (gsi_stmt (gsi));
3760 /* We have no need for virtual phis, as they don't represent
3761 actual computations. */
3762 if (virtual_operand_p (result))
3763 continue;
3765 pre_expr e = get_or_alloc_expr_for_name (result);
3766 add_to_value (get_expr_value_id (e), e);
3767 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3768 bitmap_insert_into_set (PHI_GEN (block), e);
3771 BB_MAY_NOTRETURN (block) = 0;
3773 /* Now compute value numbers and populate value sets with all
3774 the expressions computed in BLOCK. */
3775 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3777 ssa_op_iter iter;
3778 tree op;
3780 stmt = gsi_stmt (gsi);
3782 /* Cache whether the basic-block has any non-visible side-effect
3783 or control flow.
3784 If this isn't a call or it is the last stmt in the
3785 basic-block then the CFG represents things correctly. */
3786 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3788 /* Non-looping const functions always return normally.
3789 Otherwise the call might not return or have side-effects
3790 that forbids hoisting possibly trapping expressions
3791 before it. */
3792 int flags = gimple_call_flags (stmt);
3793 if (!(flags & ECF_CONST)
3794 || (flags & ECF_LOOPING_CONST_OR_PURE))
3795 BB_MAY_NOTRETURN (block) = 1;
3798 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3800 pre_expr e = get_or_alloc_expr_for_name (op);
3802 add_to_value (get_expr_value_id (e), e);
3803 bitmap_insert_into_set (TMP_GEN (block), e);
3804 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3807 if (gimple_has_side_effects (stmt)
3808 || stmt_could_throw_p (stmt)
3809 || is_gimple_debug (stmt))
3810 continue;
3812 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3814 if (ssa_undefined_value_p (op))
3815 continue;
3816 pre_expr e = get_or_alloc_expr_for_name (op);
3817 bitmap_value_insert_into_set (EXP_GEN (block), e);
3820 switch (gimple_code (stmt))
3822 case GIMPLE_RETURN:
3823 continue;
3825 case GIMPLE_CALL:
3827 vn_reference_t ref;
3828 pre_expr result = NULL;
3829 auto_vec<vn_reference_op_s> ops;
3831 /* We can value number only calls to real functions. */
3832 if (gimple_call_internal_p (stmt))
3833 continue;
3835 copy_reference_ops_from_call (stmt, &ops);
3836 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3837 gimple_expr_type (stmt),
3838 ops, &ref, VN_NOWALK);
3839 if (!ref)
3840 continue;
3842 /* If the value of the call is not invalidated in
3843 this block until it is computed, add the expression
3844 to EXP_GEN. */
3845 if (!gimple_vuse (stmt)
3846 || gimple_code
3847 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
3848 || gimple_bb (SSA_NAME_DEF_STMT
3849 (gimple_vuse (stmt))) != block)
3851 result = (pre_expr) pool_alloc (pre_expr_pool);
3852 result->kind = REFERENCE;
3853 result->id = 0;
3854 PRE_EXPR_REFERENCE (result) = ref;
3856 get_or_alloc_expression_id (result);
3857 add_to_value (get_expr_value_id (result), result);
3858 bitmap_value_insert_into_set (EXP_GEN (block), result);
3860 continue;
3863 case GIMPLE_ASSIGN:
3865 pre_expr result = NULL;
3866 switch (vn_get_stmt_kind (stmt))
3868 case VN_NARY:
3870 enum tree_code code = gimple_assign_rhs_code (stmt);
3871 vn_nary_op_t nary;
3873 /* COND_EXPR and VEC_COND_EXPR are awkward in
3874 that they contain an embedded complex expression.
3875 Don't even try to shove those through PRE. */
3876 if (code == COND_EXPR
3877 || code == VEC_COND_EXPR)
3878 continue;
3880 vn_nary_op_lookup_stmt (stmt, &nary);
3881 if (!nary)
3882 continue;
3884 /* If the NARY traps and there was a preceding
3885 point in the block that might not return avoid
3886 adding the nary to EXP_GEN. */
3887 if (BB_MAY_NOTRETURN (block)
3888 && vn_nary_may_trap (nary))
3889 continue;
3891 result = (pre_expr) pool_alloc (pre_expr_pool);
3892 result->kind = NARY;
3893 result->id = 0;
3894 PRE_EXPR_NARY (result) = nary;
3895 break;
3898 case VN_REFERENCE:
3900 vn_reference_t ref;
3901 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3902 gimple_vuse (stmt),
3903 VN_WALK, &ref);
3904 if (!ref)
3905 continue;
3907 /* If the value of the reference is not invalidated in
3908 this block until it is computed, add the expression
3909 to EXP_GEN. */
3910 if (gimple_vuse (stmt))
3912 gimple def_stmt;
3913 bool ok = true;
3914 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
3915 while (!gimple_nop_p (def_stmt)
3916 && gimple_code (def_stmt) != GIMPLE_PHI
3917 && gimple_bb (def_stmt) == block)
3919 if (stmt_may_clobber_ref_p
3920 (def_stmt, gimple_assign_rhs1 (stmt)))
3922 ok = false;
3923 break;
3925 def_stmt
3926 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
3928 if (!ok)
3929 continue;
3932 result = (pre_expr) pool_alloc (pre_expr_pool);
3933 result->kind = REFERENCE;
3934 result->id = 0;
3935 PRE_EXPR_REFERENCE (result) = ref;
3936 break;
3939 default:
3940 continue;
3943 get_or_alloc_expression_id (result);
3944 add_to_value (get_expr_value_id (result), result);
3945 bitmap_value_insert_into_set (EXP_GEN (block), result);
3946 continue;
3948 default:
3949 break;
3953 if (dump_file && (dump_flags & TDF_DETAILS))
3955 print_bitmap_set (dump_file, EXP_GEN (block),
3956 "exp_gen", block->index);
3957 print_bitmap_set (dump_file, PHI_GEN (block),
3958 "phi_gen", block->index);
3959 print_bitmap_set (dump_file, TMP_GEN (block),
3960 "tmp_gen", block->index);
3961 print_bitmap_set (dump_file, AVAIL_OUT (block),
3962 "avail_out", block->index);
3965 /* Put the dominator children of BLOCK on the worklist of blocks
3966 to compute available sets for. */
3967 for (son = first_dom_son (CDI_DOMINATORS, block);
3968 son;
3969 son = next_dom_son (CDI_DOMINATORS, son))
3970 worklist[sp++] = son;
3973 free (worklist);
3977 /* Local state for the eliminate domwalk. */
3978 static vec<gimple> el_to_remove;
3979 static vec<gimple> el_to_update;
3980 static unsigned int el_todo;
3981 static vec<tree> el_avail;
3982 static vec<tree> el_avail_stack;
3984 /* Return a leader for OP that is available at the current point of the
3985 eliminate domwalk. */
3987 static tree
3988 eliminate_avail (tree op)
3990 tree valnum = VN_INFO (op)->valnum;
3991 if (TREE_CODE (valnum) == SSA_NAME)
3993 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
3994 return valnum;
3995 if (el_avail.length () > SSA_NAME_VERSION (valnum))
3996 return el_avail[SSA_NAME_VERSION (valnum)];
3998 else if (is_gimple_min_invariant (valnum))
3999 return valnum;
4000 return NULL_TREE;
4003 /* At the current point of the eliminate domwalk make OP available. */
4005 static void
4006 eliminate_push_avail (tree op)
4008 tree valnum = VN_INFO (op)->valnum;
4009 if (TREE_CODE (valnum) == SSA_NAME)
4011 if (el_avail.length () <= SSA_NAME_VERSION (valnum))
4012 el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1);
4013 el_avail[SSA_NAME_VERSION (valnum)] = op;
4014 el_avail_stack.safe_push (op);
4018 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
4019 the leader for the expression if insertion was successful. */
4021 static tree
4022 eliminate_insert (gimple_stmt_iterator *gsi, tree val)
4024 tree expr = vn_get_expr_for (val);
4025 if (!CONVERT_EXPR_P (expr)
4026 && TREE_CODE (expr) != VIEW_CONVERT_EXPR)
4027 return NULL_TREE;
4029 tree op = TREE_OPERAND (expr, 0);
4030 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op;
4031 if (!leader)
4032 return NULL_TREE;
4034 tree res = make_temp_ssa_name (TREE_TYPE (val), NULL, "pretmp");
4035 gimple tem = gimple_build_assign (res,
4036 fold_build1 (TREE_CODE (expr),
4037 TREE_TYPE (expr), leader));
4038 gsi_insert_before (gsi, tem, GSI_SAME_STMT);
4039 VN_INFO_GET (res)->valnum = val;
4041 if (TREE_CODE (leader) == SSA_NAME)
4042 gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true);
4044 pre_stats.insertions++;
4045 if (dump_file && (dump_flags & TDF_DETAILS))
4047 fprintf (dump_file, "Inserted ");
4048 print_gimple_stmt (dump_file, tem, 0, 0);
4051 return res;
4054 class eliminate_dom_walker : public dom_walker
4056 public:
4057 eliminate_dom_walker (cdi_direction direction) : dom_walker (direction) {}
4059 virtual void before_dom_children (basic_block);
4060 virtual void after_dom_children (basic_block);
4063 /* Perform elimination for the basic-block B during the domwalk. */
4065 void
4066 eliminate_dom_walker::before_dom_children (basic_block b)
4068 gimple_stmt_iterator gsi;
4069 gimple stmt;
4071 /* Mark new bb. */
4072 el_avail_stack.safe_push (NULL_TREE);
4074 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4076 gimple stmt, phi = gsi_stmt (gsi);
4077 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4078 gimple_stmt_iterator gsi2;
4080 /* We want to perform redundant PHI elimination. Do so by
4081 replacing the PHI with a single copy if possible.
4082 Do not touch inserted, single-argument or virtual PHIs. */
4083 if (gimple_phi_num_args (phi) == 1
4084 || virtual_operand_p (res))
4086 gsi_next (&gsi);
4087 continue;
4090 sprime = eliminate_avail (res);
4091 if (!sprime
4092 || sprime == res)
4094 eliminate_push_avail (res);
4095 gsi_next (&gsi);
4096 continue;
4098 else if (is_gimple_min_invariant (sprime))
4100 if (!useless_type_conversion_p (TREE_TYPE (res),
4101 TREE_TYPE (sprime)))
4102 sprime = fold_convert (TREE_TYPE (res), sprime);
4105 if (dump_file && (dump_flags & TDF_DETAILS))
4107 fprintf (dump_file, "Replaced redundant PHI node defining ");
4108 print_generic_expr (dump_file, res, 0);
4109 fprintf (dump_file, " with ");
4110 print_generic_expr (dump_file, sprime, 0);
4111 fprintf (dump_file, "\n");
4114 remove_phi_node (&gsi, false);
4116 if (inserted_exprs
4117 && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4118 && TREE_CODE (sprime) == SSA_NAME)
4119 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4121 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4122 sprime = fold_convert (TREE_TYPE (res), sprime);
4123 stmt = gimple_build_assign (res, sprime);
4124 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4126 gsi2 = gsi_after_labels (b);
4127 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4128 /* Queue the copy for eventual removal. */
4129 el_to_remove.safe_push (stmt);
4130 /* If we inserted this PHI node ourself, it's not an elimination. */
4131 if (inserted_exprs
4132 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4133 pre_stats.phis--;
4134 else
4135 pre_stats.eliminations++;
4138 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4140 tree lhs = NULL_TREE;
4141 tree rhs = NULL_TREE;
4143 stmt = gsi_stmt (gsi);
4145 if (gimple_has_lhs (stmt))
4146 lhs = gimple_get_lhs (stmt);
4148 if (gimple_assign_single_p (stmt))
4149 rhs = gimple_assign_rhs1 (stmt);
4151 /* Lookup the RHS of the expression, see if we have an
4152 available computation for it. If so, replace the RHS with
4153 the available computation. */
4154 if (gimple_has_lhs (stmt)
4155 && TREE_CODE (lhs) == SSA_NAME
4156 && !gimple_has_volatile_ops (stmt))
4158 tree sprime;
4159 gimple orig_stmt = stmt;
4161 sprime = eliminate_avail (lhs);
4162 /* If there is no usable leader mark lhs as leader for its value. */
4163 if (!sprime)
4164 eliminate_push_avail (lhs);
4166 /* See PR43491. Do not replace a global register variable when
4167 it is a the RHS of an assignment. Do replace local register
4168 variables since gcc does not guarantee a local variable will
4169 be allocated in register.
4170 Do not perform copy propagation or undo constant propagation. */
4171 if (gimple_assign_single_p (stmt)
4172 && (TREE_CODE (rhs) == SSA_NAME
4173 || is_gimple_min_invariant (rhs)
4174 || (TREE_CODE (rhs) == VAR_DECL
4175 && is_global_var (rhs)
4176 && DECL_HARD_REGISTER (rhs))))
4177 continue;
4179 if (!sprime)
4181 /* If there is no existing usable leader but SCCVN thinks
4182 it has an expression it wants to use as replacement,
4183 insert that. */
4184 tree val = VN_INFO (lhs)->valnum;
4185 if (val != VN_TOP
4186 && TREE_CODE (val) == SSA_NAME
4187 && VN_INFO (val)->needs_insertion
4188 && VN_INFO (val)->expr != NULL_TREE
4189 && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE)
4190 eliminate_push_avail (sprime);
4192 else if (is_gimple_min_invariant (sprime))
4194 /* If there is no existing leader but SCCVN knows this
4195 value is constant, use that constant. */
4196 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4197 TREE_TYPE (sprime)))
4198 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4200 if (dump_file && (dump_flags & TDF_DETAILS))
4202 fprintf (dump_file, "Replaced ");
4203 print_gimple_expr (dump_file, stmt, 0, 0);
4204 fprintf (dump_file, " with ");
4205 print_generic_expr (dump_file, sprime, 0);
4206 fprintf (dump_file, " in ");
4207 print_gimple_stmt (dump_file, stmt, 0, 0);
4209 pre_stats.eliminations++;
4210 propagate_tree_value_into_stmt (&gsi, sprime);
4211 stmt = gsi_stmt (gsi);
4212 update_stmt (stmt);
4214 /* If we removed EH side-effects from the statement, clean
4215 its EH information. */
4216 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4218 bitmap_set_bit (need_eh_cleanup,
4219 gimple_bb (stmt)->index);
4220 if (dump_file && (dump_flags & TDF_DETAILS))
4221 fprintf (dump_file, " Removed EH side-effects.\n");
4223 continue;
4226 if (sprime
4227 && sprime != lhs
4228 && (rhs == NULL_TREE
4229 || TREE_CODE (rhs) != SSA_NAME
4230 || may_propagate_copy (rhs, sprime)))
4232 bool can_make_abnormal_goto
4233 = is_gimple_call (stmt)
4234 && stmt_can_make_abnormal_goto (stmt);
4236 gcc_assert (sprime != rhs);
4238 if (dump_file && (dump_flags & TDF_DETAILS))
4240 fprintf (dump_file, "Replaced ");
4241 print_gimple_expr (dump_file, stmt, 0, 0);
4242 fprintf (dump_file, " with ");
4243 print_generic_expr (dump_file, sprime, 0);
4244 fprintf (dump_file, " in ");
4245 print_gimple_stmt (dump_file, stmt, 0, 0);
4248 if (TREE_CODE (sprime) == SSA_NAME)
4249 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4250 NECESSARY, true);
4251 /* We need to make sure the new and old types actually match,
4252 which may require adding a simple cast, which fold_convert
4253 will do for us. */
4254 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4255 && !useless_type_conversion_p (gimple_expr_type (stmt),
4256 TREE_TYPE (sprime)))
4257 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4259 pre_stats.eliminations++;
4260 propagate_tree_value_into_stmt (&gsi, sprime);
4261 stmt = gsi_stmt (gsi);
4262 update_stmt (stmt);
4264 /* If we removed EH side-effects from the statement, clean
4265 its EH information. */
4266 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4268 bitmap_set_bit (need_eh_cleanup,
4269 gimple_bb (stmt)->index);
4270 if (dump_file && (dump_flags & TDF_DETAILS))
4271 fprintf (dump_file, " Removed EH side-effects.\n");
4274 /* Likewise for AB side-effects. */
4275 if (can_make_abnormal_goto
4276 && !stmt_can_make_abnormal_goto (stmt))
4278 bitmap_set_bit (need_ab_cleanup,
4279 gimple_bb (stmt)->index);
4280 if (dump_file && (dump_flags & TDF_DETAILS))
4281 fprintf (dump_file, " Removed AB side-effects.\n");
4285 /* If the statement is a scalar store, see if the expression
4286 has the same value number as its rhs. If so, the store is
4287 dead. */
4288 else if (gimple_assign_single_p (stmt)
4289 && !gimple_has_volatile_ops (stmt)
4290 && !is_gimple_reg (gimple_assign_lhs (stmt))
4291 && (TREE_CODE (rhs) == SSA_NAME
4292 || is_gimple_min_invariant (rhs)))
4294 tree val;
4295 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4296 gimple_vuse (stmt), VN_WALK, NULL);
4297 if (TREE_CODE (rhs) == SSA_NAME)
4298 rhs = VN_INFO (rhs)->valnum;
4299 if (val
4300 && operand_equal_p (val, rhs, 0))
4302 if (dump_file && (dump_flags & TDF_DETAILS))
4304 fprintf (dump_file, "Deleted redundant store ");
4305 print_gimple_stmt (dump_file, stmt, 0, 0);
4308 /* Queue stmt for removal. */
4309 el_to_remove.safe_push (stmt);
4312 /* Visit COND_EXPRs and fold the comparison with the
4313 available value-numbers. */
4314 else if (gimple_code (stmt) == GIMPLE_COND)
4316 tree op0 = gimple_cond_lhs (stmt);
4317 tree op1 = gimple_cond_rhs (stmt);
4318 tree result;
4320 if (TREE_CODE (op0) == SSA_NAME)
4321 op0 = VN_INFO (op0)->valnum;
4322 if (TREE_CODE (op1) == SSA_NAME)
4323 op1 = VN_INFO (op1)->valnum;
4324 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4325 op0, op1);
4326 if (result && TREE_CODE (result) == INTEGER_CST)
4328 if (integer_zerop (result))
4329 gimple_cond_make_false (stmt);
4330 else
4331 gimple_cond_make_true (stmt);
4332 update_stmt (stmt);
4333 el_todo = TODO_cleanup_cfg;
4336 /* Visit indirect calls and turn them into direct calls if
4337 possible. */
4338 if (is_gimple_call (stmt))
4340 tree orig_fn = gimple_call_fn (stmt);
4341 tree fn;
4342 if (!orig_fn)
4343 continue;
4344 if (TREE_CODE (orig_fn) == SSA_NAME)
4345 fn = VN_INFO (orig_fn)->valnum;
4346 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4347 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4349 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4350 if (!gimple_call_addr_fndecl (fn))
4352 fn = ipa_intraprocedural_devirtualization (stmt);
4353 if (fn)
4354 fn = build_fold_addr_expr (fn);
4357 else
4358 continue;
4359 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4360 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4361 TREE_TYPE (fn)))
4363 bool can_make_abnormal_goto
4364 = stmt_can_make_abnormal_goto (stmt);
4365 bool was_noreturn = gimple_call_noreturn_p (stmt);
4367 if (dump_file && (dump_flags & TDF_DETAILS))
4369 fprintf (dump_file, "Replacing call target with ");
4370 print_generic_expr (dump_file, fn, 0);
4371 fprintf (dump_file, " in ");
4372 print_gimple_stmt (dump_file, stmt, 0, 0);
4375 gimple_call_set_fn (stmt, fn);
4376 el_to_update.safe_push (stmt);
4378 /* When changing a call into a noreturn call, cfg cleanup
4379 is needed to fix up the noreturn call. */
4380 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4381 el_todo |= TODO_cleanup_cfg;
4383 /* If we removed EH side-effects from the statement, clean
4384 its EH information. */
4385 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4387 bitmap_set_bit (need_eh_cleanup,
4388 gimple_bb (stmt)->index);
4389 if (dump_file && (dump_flags & TDF_DETAILS))
4390 fprintf (dump_file, " Removed EH side-effects.\n");
4393 /* Likewise for AB side-effects. */
4394 if (can_make_abnormal_goto
4395 && !stmt_can_make_abnormal_goto (stmt))
4397 bitmap_set_bit (need_ab_cleanup,
4398 gimple_bb (stmt)->index);
4399 if (dump_file && (dump_flags & TDF_DETAILS))
4400 fprintf (dump_file, " Removed AB side-effects.\n");
4403 /* Changing an indirect call to a direct call may
4404 have exposed different semantics. This may
4405 require an SSA update. */
4406 el_todo |= TODO_update_ssa_only_virtuals;
4412 /* Make no longer available leaders no longer available. */
4414 void
4415 eliminate_dom_walker::after_dom_children (basic_block)
4417 tree entry;
4418 while ((entry = el_avail_stack.pop ()) != NULL_TREE)
4419 el_avail[SSA_NAME_VERSION (VN_INFO (entry)->valnum)] = NULL_TREE;
4422 /* Eliminate fully redundant computations. */
4424 static unsigned int
4425 eliminate (void)
4427 gimple_stmt_iterator gsi;
4428 gimple stmt;
4429 unsigned i;
4431 need_eh_cleanup = BITMAP_ALLOC (NULL);
4432 need_ab_cleanup = BITMAP_ALLOC (NULL);
4434 el_to_remove.create (0);
4435 el_to_update.create (0);
4436 el_todo = 0;
4437 el_avail.create (0);
4438 el_avail_stack.create (0);
4440 eliminate_dom_walker (CDI_DOMINATORS).walk (cfun->cfg->x_entry_block_ptr);
4442 el_avail.release ();
4443 el_avail_stack.release ();
4445 /* We cannot remove stmts during BB walk, especially not release SSA
4446 names there as this confuses the VN machinery. The stmts ending
4447 up in el_to_remove are either stores or simple copies. */
4448 FOR_EACH_VEC_ELT (el_to_remove, i, stmt)
4450 tree lhs = gimple_assign_lhs (stmt);
4451 tree rhs = gimple_assign_rhs1 (stmt);
4452 use_operand_p use_p;
4453 gimple use_stmt;
4455 /* If there is a single use only, propagate the equivalency
4456 instead of keeping the copy. */
4457 if (TREE_CODE (lhs) == SSA_NAME
4458 && TREE_CODE (rhs) == SSA_NAME
4459 && single_imm_use (lhs, &use_p, &use_stmt)
4460 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4462 SET_USE (use_p, rhs);
4463 update_stmt (use_stmt);
4464 if (inserted_exprs
4465 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4466 && TREE_CODE (rhs) == SSA_NAME)
4467 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4470 /* If this is a store or a now unused copy, remove it. */
4471 if (TREE_CODE (lhs) != SSA_NAME
4472 || has_zero_uses (lhs))
4474 basic_block bb = gimple_bb (stmt);
4475 gsi = gsi_for_stmt (stmt);
4476 unlink_stmt_vdef (stmt);
4477 if (gsi_remove (&gsi, true))
4478 bitmap_set_bit (need_eh_cleanup, bb->index);
4479 if (inserted_exprs
4480 && TREE_CODE (lhs) == SSA_NAME)
4481 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4482 release_defs (stmt);
4485 el_to_remove.release ();
4487 /* We cannot update call statements with virtual operands during
4488 SSA walk. This might remove them which in turn makes our
4489 VN lattice invalid. */
4490 FOR_EACH_VEC_ELT (el_to_update, i, stmt)
4491 update_stmt (stmt);
4492 el_to_update.release ();
4494 return el_todo;
4497 /* Perform CFG cleanups made necessary by elimination. */
4499 static unsigned
4500 fini_eliminate (void)
4502 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4503 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4505 if (do_eh_cleanup)
4506 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4508 if (do_ab_cleanup)
4509 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4511 BITMAP_FREE (need_eh_cleanup);
4512 BITMAP_FREE (need_ab_cleanup);
4514 if (do_eh_cleanup || do_ab_cleanup)
4515 return TODO_cleanup_cfg;
4516 return 0;
4519 /* Borrow a bit of tree-ssa-dce.c for the moment.
4520 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4521 this may be a bit faster, and we may want critical edges kept split. */
4523 /* If OP's defining statement has not already been determined to be necessary,
4524 mark that statement necessary. Return the stmt, if it is newly
4525 necessary. */
4527 static inline gimple
4528 mark_operand_necessary (tree op)
4530 gimple stmt;
4532 gcc_assert (op);
4534 if (TREE_CODE (op) != SSA_NAME)
4535 return NULL;
4537 stmt = SSA_NAME_DEF_STMT (op);
4538 gcc_assert (stmt);
4540 if (gimple_plf (stmt, NECESSARY)
4541 || gimple_nop_p (stmt))
4542 return NULL;
4544 gimple_set_plf (stmt, NECESSARY, true);
4545 return stmt;
4548 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4549 to insert PHI nodes sometimes, and because value numbering of casts isn't
4550 perfect, we sometimes end up inserting dead code. This simple DCE-like
4551 pass removes any insertions we made that weren't actually used. */
4553 static void
4554 remove_dead_inserted_code (void)
4556 bitmap worklist;
4557 unsigned i;
4558 bitmap_iterator bi;
4559 gimple t;
4561 worklist = BITMAP_ALLOC (NULL);
4562 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4564 t = SSA_NAME_DEF_STMT (ssa_name (i));
4565 if (gimple_plf (t, NECESSARY))
4566 bitmap_set_bit (worklist, i);
4568 while (!bitmap_empty_p (worklist))
4570 i = bitmap_first_set_bit (worklist);
4571 bitmap_clear_bit (worklist, i);
4572 t = SSA_NAME_DEF_STMT (ssa_name (i));
4574 /* PHI nodes are somewhat special in that each PHI alternative has
4575 data and control dependencies. All the statements feeding the
4576 PHI node's arguments are always necessary. */
4577 if (gimple_code (t) == GIMPLE_PHI)
4579 unsigned k;
4581 for (k = 0; k < gimple_phi_num_args (t); k++)
4583 tree arg = PHI_ARG_DEF (t, k);
4584 if (TREE_CODE (arg) == SSA_NAME)
4586 gimple n = mark_operand_necessary (arg);
4587 if (n)
4588 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4592 else
4594 /* Propagate through the operands. Examine all the USE, VUSE and
4595 VDEF operands in this statement. Mark all the statements
4596 which feed this statement's uses as necessary. */
4597 ssa_op_iter iter;
4598 tree use;
4600 /* The operands of VDEF expressions are also needed as they
4601 represent potential definitions that may reach this
4602 statement (VDEF operands allow us to follow def-def
4603 links). */
4605 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4607 gimple n = mark_operand_necessary (use);
4608 if (n)
4609 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4614 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4616 t = SSA_NAME_DEF_STMT (ssa_name (i));
4617 if (!gimple_plf (t, NECESSARY))
4619 gimple_stmt_iterator gsi;
4621 if (dump_file && (dump_flags & TDF_DETAILS))
4623 fprintf (dump_file, "Removing unnecessary insertion:");
4624 print_gimple_stmt (dump_file, t, 0, 0);
4627 gsi = gsi_for_stmt (t);
4628 if (gimple_code (t) == GIMPLE_PHI)
4629 remove_phi_node (&gsi, true);
4630 else
4632 gsi_remove (&gsi, true);
4633 release_defs (t);
4637 BITMAP_FREE (worklist);
4641 /* Initialize data structures used by PRE. */
4643 static void
4644 init_pre (void)
4646 basic_block bb;
4648 next_expression_id = 1;
4649 expressions.create (0);
4650 expressions.safe_push (NULL);
4651 value_expressions.create (get_max_value_id () + 1);
4652 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
4653 name_to_id.create (0);
4655 inserted_exprs = BITMAP_ALLOC (NULL);
4657 connect_infinite_loops_to_exit ();
4658 memset (&pre_stats, 0, sizeof (pre_stats));
4660 postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
4661 postorder_num = inverted_post_order_compute (postorder);
4663 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4665 calculate_dominance_info (CDI_POST_DOMINATORS);
4666 calculate_dominance_info (CDI_DOMINATORS);
4668 bitmap_obstack_initialize (&grand_bitmap_obstack);
4669 phi_translate_table.create (5110);
4670 expression_to_id.create (num_ssa_names * 3);
4671 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4672 sizeof (struct bitmap_set), 30);
4673 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4674 sizeof (struct pre_expr_d), 30);
4675 FOR_ALL_BB (bb)
4677 EXP_GEN (bb) = bitmap_set_new ();
4678 PHI_GEN (bb) = bitmap_set_new ();
4679 TMP_GEN (bb) = bitmap_set_new ();
4680 AVAIL_OUT (bb) = bitmap_set_new ();
4685 /* Deallocate data structures used by PRE. */
4687 static void
4688 fini_pre ()
4690 free (postorder);
4691 value_expressions.release ();
4692 BITMAP_FREE (inserted_exprs);
4693 bitmap_obstack_release (&grand_bitmap_obstack);
4694 free_alloc_pool (bitmap_set_pool);
4695 free_alloc_pool (pre_expr_pool);
4696 phi_translate_table.dispose ();
4697 expression_to_id.dispose ();
4698 name_to_id.release ();
4700 free_aux_for_blocks ();
4702 free_dominance_info (CDI_POST_DOMINATORS);
4705 /* Gate and execute functions for PRE. */
4707 static unsigned int
4708 do_pre (void)
4710 unsigned int todo = 0;
4712 do_partial_partial =
4713 flag_tree_partial_pre && optimize_function_for_speed_p (cfun);
4715 /* This has to happen before SCCVN runs because
4716 loop_optimizer_init may create new phis, etc. */
4717 loop_optimizer_init (LOOPS_NORMAL);
4719 if (!run_scc_vn (VN_WALK))
4721 loop_optimizer_finalize ();
4722 return 0;
4725 init_pre ();
4726 scev_initialize ();
4728 /* Collect and value number expressions computed in each basic block. */
4729 compute_avail ();
4731 /* Insert can get quite slow on an incredibly large number of basic
4732 blocks due to some quadratic behavior. Until this behavior is
4733 fixed, don't run it when he have an incredibly large number of
4734 bb's. If we aren't going to run insert, there is no point in
4735 computing ANTIC, either, even though it's plenty fast. */
4736 if (n_basic_blocks_for_fn (cfun) < 4000)
4738 compute_antic ();
4739 insert ();
4742 /* Make sure to remove fake edges before committing our inserts.
4743 This makes sure we don't end up with extra critical edges that
4744 we would need to split. */
4745 remove_fake_exit_edges ();
4746 gsi_commit_edge_inserts ();
4748 /* Remove all the redundant expressions. */
4749 todo |= eliminate ();
4751 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4752 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4753 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4754 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4756 clear_expression_ids ();
4757 remove_dead_inserted_code ();
4758 todo |= TODO_verify_flow;
4760 scev_finalize ();
4761 fini_pre ();
4762 todo |= fini_eliminate ();
4763 loop_optimizer_finalize ();
4765 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4766 case we can merge the block with the remaining predecessor of the block.
4767 It should either:
4768 - call merge_blocks after each tail merge iteration
4769 - call merge_blocks after all tail merge iterations
4770 - mark TODO_cleanup_cfg when necessary
4771 - share the cfg cleanup with fini_pre. */
4772 todo |= tail_merge_optimize (todo);
4774 free_scc_vn ();
4776 /* Tail merging invalidates the virtual SSA web, together with
4777 cfg-cleanup opportunities exposed by PRE this will wreck the
4778 SSA updating machinery. So make sure to run update-ssa
4779 manually, before eventually scheduling cfg-cleanup as part of
4780 the todo. */
4781 update_ssa (TODO_update_ssa_only_virtuals);
4783 return todo;
4786 static bool
4787 gate_pre (void)
4789 return flag_tree_pre != 0;
4792 namespace {
4794 const pass_data pass_data_pre =
4796 GIMPLE_PASS, /* type */
4797 "pre", /* name */
4798 OPTGROUP_NONE, /* optinfo_flags */
4799 true, /* has_gate */
4800 true, /* has_execute */
4801 TV_TREE_PRE, /* tv_id */
4802 ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */
4803 0, /* properties_provided */
4804 0, /* properties_destroyed */
4805 TODO_rebuild_alias, /* todo_flags_start */
4806 TODO_verify_ssa, /* todo_flags_finish */
4809 class pass_pre : public gimple_opt_pass
4811 public:
4812 pass_pre (gcc::context *ctxt)
4813 : gimple_opt_pass (pass_data_pre, ctxt)
4816 /* opt_pass methods: */
4817 bool gate () { return gate_pre (); }
4818 unsigned int execute () { return do_pre (); }
4820 }; // class pass_pre
4822 } // anon namespace
4824 gimple_opt_pass *
4825 make_pass_pre (gcc::context *ctxt)
4827 return new pass_pre (ctxt);
4831 /* Gate and execute functions for FRE. */
4833 static unsigned int
4834 execute_fre (void)
4836 unsigned int todo = 0;
4838 if (!run_scc_vn (VN_WALKREWRITE))
4839 return 0;
4841 memset (&pre_stats, 0, sizeof (pre_stats));
4843 /* Remove all the redundant expressions. */
4844 todo |= eliminate ();
4846 todo |= fini_eliminate ();
4848 free_scc_vn ();
4850 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4851 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4853 return todo;
4856 static bool
4857 gate_fre (void)
4859 return flag_tree_fre != 0;
4862 namespace {
4864 const pass_data pass_data_fre =
4866 GIMPLE_PASS, /* type */
4867 "fre", /* name */
4868 OPTGROUP_NONE, /* optinfo_flags */
4869 true, /* has_gate */
4870 true, /* has_execute */
4871 TV_TREE_FRE, /* tv_id */
4872 ( PROP_cfg | PROP_ssa ), /* properties_required */
4873 0, /* properties_provided */
4874 0, /* properties_destroyed */
4875 0, /* todo_flags_start */
4876 TODO_verify_ssa, /* todo_flags_finish */
4879 class pass_fre : public gimple_opt_pass
4881 public:
4882 pass_fre (gcc::context *ctxt)
4883 : gimple_opt_pass (pass_data_fre, ctxt)
4886 /* opt_pass methods: */
4887 opt_pass * clone () { return new pass_fre (m_ctxt); }
4888 bool gate () { return gate_fre (); }
4889 unsigned int execute () { return execute_fre (); }
4891 }; // class pass_fre
4893 } // anon namespace
4895 gimple_opt_pass *
4896 make_pass_fre (gcc::context *ctxt)
4898 return new pass_fre (ctxt);