Merge from trunk @217148.
[official-gcc.git] / gcc / tree-ssa-pre.c
blob24e820adb97e1e34aae631e20914149d61aa1902
1 /* SSA-PRE for trees.
2 Copyright (C) 2001-2014 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 "predict.h"
28 #include "vec.h"
29 #include "hashtab.h"
30 #include "hash-set.h"
31 #include "machmode.h"
32 #include "hard-reg-set.h"
33 #include "input.h"
34 #include "function.h"
35 #include "dominance.h"
36 #include "cfg.h"
37 #include "cfganal.h"
38 #include "basic-block.h"
39 #include "gimple-pretty-print.h"
40 #include "tree-inline.h"
41 #include "inchash.h"
42 #include "hash-table.h"
43 #include "tree-ssa-alias.h"
44 #include "internal-fn.h"
45 #include "gimple-fold.h"
46 #include "tree-eh.h"
47 #include "gimple-expr.h"
48 #include "is-a.h"
49 #include "gimple.h"
50 #include "gimplify.h"
51 #include "gimple-iterator.h"
52 #include "gimplify-me.h"
53 #include "gimple-ssa.h"
54 #include "tree-cfg.h"
55 #include "tree-phinodes.h"
56 #include "ssa-iterators.h"
57 #include "stringpool.h"
58 #include "tree-ssanames.h"
59 #include "tree-ssa-loop.h"
60 #include "tree-into-ssa.h"
61 #include "expr.h"
62 #include "tree-dfa.h"
63 #include "tree-ssa.h"
64 #include "tree-iterator.h"
65 #include "alloc-pool.h"
66 #include "obstack.h"
67 #include "tree-pass.h"
68 #include "flags.h"
69 #include "langhooks.h"
70 #include "cfgloop.h"
71 #include "tree-ssa-sccvn.h"
72 #include "tree-scalar-evolution.h"
73 #include "params.h"
74 #include "dbgcnt.h"
75 #include "domwalk.h"
76 #include "hash-map.h"
77 #include "plugin-api.h"
78 #include "ipa-ref.h"
79 #include "cgraph.h"
80 #include "ipa-prop.h"
81 #include "tree-ssa-propagate.h"
82 #include "ipa-utils.h"
84 /* TODO:
86 1. Avail sets can be shared by making an avail_find_leader that
87 walks up the dominator tree and looks in those avail sets.
88 This might affect code optimality, it's unclear right now.
89 2. Strength reduction can be performed by anticipating expressions
90 we can repair later on.
91 3. We can do back-substitution or smarter value numbering to catch
92 commutative expressions split up over multiple statements.
95 /* For ease of terminology, "expression node" in the below refers to
96 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
97 represent the actual statement containing the expressions we care about,
98 and we cache the value number by putting it in the expression. */
100 /* Basic algorithm
102 First we walk the statements to generate the AVAIL sets, the
103 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
104 generation of values/expressions by a given block. We use them
105 when computing the ANTIC sets. The AVAIL sets consist of
106 SSA_NAME's that represent values, so we know what values are
107 available in what blocks. AVAIL is a forward dataflow problem. In
108 SSA, values are never killed, so we don't need a kill set, or a
109 fixpoint iteration, in order to calculate the AVAIL sets. In
110 traditional parlance, AVAIL sets tell us the downsafety of the
111 expressions/values.
113 Next, we generate the ANTIC sets. These sets represent the
114 anticipatable expressions. ANTIC is a backwards dataflow
115 problem. An expression is anticipatable in a given block if it could
116 be generated in that block. This means that if we had to perform
117 an insertion in that block, of the value of that expression, we
118 could. Calculating the ANTIC sets requires phi translation of
119 expressions, because the flow goes backwards through phis. We must
120 iterate to a fixpoint of the ANTIC sets, because we have a kill
121 set. Even in SSA form, values are not live over the entire
122 function, only from their definition point onwards. So we have to
123 remove values from the ANTIC set once we go past the definition
124 point of the leaders that make them up.
125 compute_antic/compute_antic_aux performs this computation.
127 Third, we perform insertions to make partially redundant
128 expressions fully redundant.
130 An expression is partially redundant (excluding partial
131 anticipation) if:
133 1. It is AVAIL in some, but not all, of the predecessors of a
134 given block.
135 2. It is ANTIC in all the predecessors.
137 In order to make it fully redundant, we insert the expression into
138 the predecessors where it is not available, but is ANTIC.
140 For the partial anticipation case, we only perform insertion if it
141 is partially anticipated in some block, and fully available in all
142 of the predecessors.
144 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
145 performs these steps.
147 Fourth, we eliminate fully redundant expressions.
148 This is a simple statement walk that replaces redundant
149 calculations with the now available values. */
151 /* Representations of value numbers:
153 Value numbers are represented by a representative SSA_NAME. We
154 will create fake SSA_NAME's in situations where we need a
155 representative but do not have one (because it is a complex
156 expression). In order to facilitate storing the value numbers in
157 bitmaps, and keep the number of wasted SSA_NAME's down, we also
158 associate a value_id with each value number, and create full blown
159 ssa_name's only where we actually need them (IE in operands of
160 existing expressions).
162 Theoretically you could replace all the value_id's with
163 SSA_NAME_VERSION, but this would allocate a large number of
164 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
165 It would also require an additional indirection at each point we
166 use the value id. */
168 /* Representation of expressions on value numbers:
170 Expressions consisting of value numbers are represented the same
171 way as our VN internally represents them, with an additional
172 "pre_expr" wrapping around them in order to facilitate storing all
173 of the expressions in the same sets. */
175 /* Representation of sets:
177 The dataflow sets do not need to be sorted in any particular order
178 for the majority of their lifetime, are simply represented as two
179 bitmaps, one that keeps track of values present in the set, and one
180 that keeps track of expressions present in the set.
182 When we need them in topological order, we produce it on demand by
183 transforming the bitmap into an array and sorting it into topo
184 order. */
186 /* Type of expression, used to know which member of the PRE_EXPR union
187 is valid. */
189 enum pre_expr_kind
191 NAME,
192 NARY,
193 REFERENCE,
194 CONSTANT
197 typedef union pre_expr_union_d
199 tree name;
200 tree constant;
201 vn_nary_op_t nary;
202 vn_reference_t reference;
203 } pre_expr_union;
205 typedef struct pre_expr_d : typed_noop_remove <pre_expr_d>
207 enum pre_expr_kind kind;
208 unsigned int id;
209 pre_expr_union u;
211 /* hash_table support. */
212 typedef pre_expr_d value_type;
213 typedef pre_expr_d compare_type;
214 static inline hashval_t hash (const pre_expr_d *);
215 static inline int equal (const pre_expr_d *, const pre_expr_d *);
216 } *pre_expr;
218 #define PRE_EXPR_NAME(e) (e)->u.name
219 #define PRE_EXPR_NARY(e) (e)->u.nary
220 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
221 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
223 /* Compare E1 and E1 for equality. */
225 inline int
226 pre_expr_d::equal (const value_type *e1, const compare_type *e2)
228 if (e1->kind != e2->kind)
229 return false;
231 switch (e1->kind)
233 case CONSTANT:
234 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
235 PRE_EXPR_CONSTANT (e2));
236 case NAME:
237 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
238 case NARY:
239 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
240 case REFERENCE:
241 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
242 PRE_EXPR_REFERENCE (e2));
243 default:
244 gcc_unreachable ();
248 /* Hash E. */
250 inline hashval_t
251 pre_expr_d::hash (const value_type *e)
253 switch (e->kind)
255 case CONSTANT:
256 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
257 case NAME:
258 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
259 case NARY:
260 return PRE_EXPR_NARY (e)->hashcode;
261 case REFERENCE:
262 return PRE_EXPR_REFERENCE (e)->hashcode;
263 default:
264 gcc_unreachable ();
268 /* Next global expression id number. */
269 static unsigned int next_expression_id;
271 /* Mapping from expression to id number we can use in bitmap sets. */
272 static vec<pre_expr> expressions;
273 static hash_table<pre_expr_d> *expression_to_id;
274 static vec<unsigned> name_to_id;
276 /* Allocate an expression id for EXPR. */
278 static inline unsigned int
279 alloc_expression_id (pre_expr expr)
281 struct pre_expr_d **slot;
282 /* Make sure we won't overflow. */
283 gcc_assert (next_expression_id + 1 > next_expression_id);
284 expr->id = next_expression_id++;
285 expressions.safe_push (expr);
286 if (expr->kind == NAME)
288 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
289 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
290 re-allocations by using vec::reserve upfront. */
291 unsigned old_len = name_to_id.length ();
292 name_to_id.reserve (num_ssa_names - old_len);
293 name_to_id.quick_grow_cleared (num_ssa_names);
294 gcc_assert (name_to_id[version] == 0);
295 name_to_id[version] = expr->id;
297 else
299 slot = expression_to_id->find_slot (expr, INSERT);
300 gcc_assert (!*slot);
301 *slot = expr;
303 return next_expression_id - 1;
306 /* Return the expression id for tree EXPR. */
308 static inline unsigned int
309 get_expression_id (const pre_expr expr)
311 return expr->id;
314 static inline unsigned int
315 lookup_expression_id (const pre_expr expr)
317 struct pre_expr_d **slot;
319 if (expr->kind == NAME)
321 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
322 if (name_to_id.length () <= version)
323 return 0;
324 return name_to_id[version];
326 else
328 slot = expression_to_id->find_slot (expr, NO_INSERT);
329 if (!slot)
330 return 0;
331 return ((pre_expr)*slot)->id;
335 /* Return the existing expression id for EXPR, or create one if one
336 does not exist yet. */
338 static inline unsigned int
339 get_or_alloc_expression_id (pre_expr expr)
341 unsigned int id = lookup_expression_id (expr);
342 if (id == 0)
343 return alloc_expression_id (expr);
344 return expr->id = id;
347 /* Return the expression that has expression id ID */
349 static inline pre_expr
350 expression_for_id (unsigned int id)
352 return expressions[id];
355 /* Free the expression id field in all of our expressions,
356 and then destroy the expressions array. */
358 static void
359 clear_expression_ids (void)
361 expressions.release ();
364 static alloc_pool pre_expr_pool;
366 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
368 static pre_expr
369 get_or_alloc_expr_for_name (tree name)
371 struct pre_expr_d expr;
372 pre_expr result;
373 unsigned int result_id;
375 expr.kind = NAME;
376 expr.id = 0;
377 PRE_EXPR_NAME (&expr) = name;
378 result_id = lookup_expression_id (&expr);
379 if (result_id != 0)
380 return expression_for_id (result_id);
382 result = (pre_expr) pool_alloc (pre_expr_pool);
383 result->kind = NAME;
384 PRE_EXPR_NAME (result) = name;
385 alloc_expression_id (result);
386 return result;
389 /* An unordered bitmap set. One bitmap tracks values, the other,
390 expressions. */
391 typedef struct bitmap_set
393 bitmap_head expressions;
394 bitmap_head values;
395 } *bitmap_set_t;
397 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
398 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
400 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
401 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
403 /* Mapping from value id to expressions with that value_id. */
404 static vec<bitmap> value_expressions;
406 /* Sets that we need to keep track of. */
407 typedef struct bb_bitmap_sets
409 /* The EXP_GEN set, which represents expressions/values generated in
410 a basic block. */
411 bitmap_set_t exp_gen;
413 /* The PHI_GEN set, which represents PHI results generated in a
414 basic block. */
415 bitmap_set_t phi_gen;
417 /* The TMP_GEN set, which represents results/temporaries generated
418 in a basic block. IE the LHS of an expression. */
419 bitmap_set_t tmp_gen;
421 /* The AVAIL_OUT set, which represents which values are available in
422 a given basic block. */
423 bitmap_set_t avail_out;
425 /* The ANTIC_IN set, which represents which values are anticipatable
426 in a given basic block. */
427 bitmap_set_t antic_in;
429 /* The PA_IN set, which represents which values are
430 partially anticipatable in a given basic block. */
431 bitmap_set_t pa_in;
433 /* The NEW_SETS set, which is used during insertion to augment the
434 AVAIL_OUT set of blocks with the new insertions performed during
435 the current iteration. */
436 bitmap_set_t new_sets;
438 /* A cache for value_dies_in_block_x. */
439 bitmap expr_dies;
441 /* The live virtual operand on successor edges. */
442 tree vop_on_exit;
444 /* True if we have visited this block during ANTIC calculation. */
445 unsigned int visited : 1;
447 /* True when the block contains a call that might not return. */
448 unsigned int contains_may_not_return_call : 1;
449 } *bb_value_sets_t;
451 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
452 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
453 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
454 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
455 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
456 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
457 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
458 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
459 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
460 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
461 #define BB_LIVE_VOP_ON_EXIT(BB) ((bb_value_sets_t) ((BB)->aux))->vop_on_exit
464 /* Basic block list in postorder. */
465 static int *postorder;
466 static int postorder_num;
468 /* This structure is used to keep track of statistics on what
469 optimization PRE was able to perform. */
470 static struct
472 /* The number of RHS computations eliminated by PRE. */
473 int eliminations;
475 /* The number of new expressions/temporaries generated by PRE. */
476 int insertions;
478 /* The number of inserts found due to partial anticipation */
479 int pa_insert;
481 /* The number of new PHI nodes added by PRE. */
482 int phis;
483 } pre_stats;
485 static bool do_partial_partial;
486 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int);
487 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
488 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
489 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
490 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
491 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
492 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
493 unsigned int, bool);
494 static bitmap_set_t bitmap_set_new (void);
495 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
496 tree);
497 static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
498 static unsigned int get_expr_value_id (pre_expr);
500 /* We can add and remove elements and entries to and from sets
501 and hash tables, so we use alloc pools for them. */
503 static alloc_pool bitmap_set_pool;
504 static bitmap_obstack grand_bitmap_obstack;
506 /* Set of blocks with statements that have had their EH properties changed. */
507 static bitmap need_eh_cleanup;
509 /* Set of blocks with statements that have had their AB properties changed. */
510 static bitmap need_ab_cleanup;
512 /* A three tuple {e, pred, v} used to cache phi translations in the
513 phi_translate_table. */
515 typedef struct expr_pred_trans_d : typed_free_remove<expr_pred_trans_d>
517 /* The expression. */
518 pre_expr e;
520 /* The predecessor block along which we translated the expression. */
521 basic_block pred;
523 /* The value that resulted from the translation. */
524 pre_expr v;
526 /* The hashcode for the expression, pred pair. This is cached for
527 speed reasons. */
528 hashval_t hashcode;
530 /* hash_table support. */
531 typedef expr_pred_trans_d value_type;
532 typedef expr_pred_trans_d compare_type;
533 static inline hashval_t hash (const value_type *);
534 static inline int equal (const value_type *, const compare_type *);
535 } *expr_pred_trans_t;
536 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
538 inline hashval_t
539 expr_pred_trans_d::hash (const expr_pred_trans_d *e)
541 return e->hashcode;
544 inline int
545 expr_pred_trans_d::equal (const value_type *ve1,
546 const compare_type *ve2)
548 basic_block b1 = ve1->pred;
549 basic_block b2 = ve2->pred;
551 /* If they are not translations for the same basic block, they can't
552 be equal. */
553 if (b1 != b2)
554 return false;
555 return pre_expr_d::equal (ve1->e, ve2->e);
558 /* The phi_translate_table caches phi translations for a given
559 expression and predecessor. */
560 static hash_table<expr_pred_trans_d> *phi_translate_table;
562 /* Add the tuple mapping from {expression E, basic block PRED} to
563 the phi translation table and return whether it pre-existed. */
565 static inline bool
566 phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred)
568 expr_pred_trans_t *slot;
569 expr_pred_trans_d tem;
570 hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e),
571 pred->index);
572 tem.e = e;
573 tem.pred = pred;
574 tem.hashcode = hash;
575 slot = phi_translate_table->find_slot_with_hash (&tem, hash, INSERT);
576 if (*slot)
578 *entry = *slot;
579 return true;
582 *entry = *slot = XNEW (struct expr_pred_trans_d);
583 (*entry)->e = e;
584 (*entry)->pred = pred;
585 (*entry)->hashcode = hash;
586 return false;
590 /* Add expression E to the expression set of value id V. */
592 static void
593 add_to_value (unsigned int v, pre_expr e)
595 bitmap set;
597 gcc_checking_assert (get_expr_value_id (e) == v);
599 if (v >= value_expressions.length ())
601 value_expressions.safe_grow_cleared (v + 1);
604 set = value_expressions[v];
605 if (!set)
607 set = BITMAP_ALLOC (&grand_bitmap_obstack);
608 value_expressions[v] = set;
611 bitmap_set_bit (set, get_or_alloc_expression_id (e));
614 /* Create a new bitmap set and return it. */
616 static bitmap_set_t
617 bitmap_set_new (void)
619 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
620 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
621 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
622 return ret;
625 /* Return the value id for a PRE expression EXPR. */
627 static unsigned int
628 get_expr_value_id (pre_expr expr)
630 unsigned int id;
631 switch (expr->kind)
633 case CONSTANT:
634 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
635 break;
636 case NAME:
637 id = VN_INFO (PRE_EXPR_NAME (expr))->value_id;
638 break;
639 case NARY:
640 id = PRE_EXPR_NARY (expr)->value_id;
641 break;
642 case REFERENCE:
643 id = PRE_EXPR_REFERENCE (expr)->value_id;
644 break;
645 default:
646 gcc_unreachable ();
648 /* ??? We cannot assert that expr has a value-id (it can be 0), because
649 we assign value-ids only to expressions that have a result
650 in set_hashtable_value_ids. */
651 return id;
654 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
656 static tree
657 sccvn_valnum_from_value_id (unsigned int val)
659 bitmap_iterator bi;
660 unsigned int i;
661 bitmap exprset = value_expressions[val];
662 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
664 pre_expr vexpr = expression_for_id (i);
665 if (vexpr->kind == NAME)
666 return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
667 else if (vexpr->kind == CONSTANT)
668 return PRE_EXPR_CONSTANT (vexpr);
670 return NULL_TREE;
673 /* Remove an expression EXPR from a bitmapped set. */
675 static void
676 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
678 unsigned int val = get_expr_value_id (expr);
679 if (!value_id_constant_p (val))
681 bitmap_clear_bit (&set->values, val);
682 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
686 static void
687 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
688 unsigned int val, bool allow_constants)
690 if (allow_constants || !value_id_constant_p (val))
692 /* We specifically expect this and only this function to be able to
693 insert constants into a set. */
694 bitmap_set_bit (&set->values, val);
695 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
699 /* Insert an expression EXPR into a bitmapped set. */
701 static void
702 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
704 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
707 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
709 static void
710 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
712 bitmap_copy (&dest->expressions, &orig->expressions);
713 bitmap_copy (&dest->values, &orig->values);
717 /* Free memory used up by SET. */
718 static void
719 bitmap_set_free (bitmap_set_t set)
721 bitmap_clear (&set->expressions);
722 bitmap_clear (&set->values);
726 /* Generate an topological-ordered array of bitmap set SET. */
728 static vec<pre_expr>
729 sorted_array_from_bitmap_set (bitmap_set_t set)
731 unsigned int i, j;
732 bitmap_iterator bi, bj;
733 vec<pre_expr> result;
735 /* Pre-allocate enough space for the array. */
736 result.create (bitmap_count_bits (&set->expressions));
738 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
740 /* The number of expressions having a given value is usually
741 relatively small. Thus, rather than making a vector of all
742 the expressions and sorting it by value-id, we walk the values
743 and check in the reverse mapping that tells us what expressions
744 have a given value, to filter those in our set. As a result,
745 the expressions are inserted in value-id order, which means
746 topological order.
748 If this is somehow a significant lose for some cases, we can
749 choose which set to walk based on the set size. */
750 bitmap exprset = value_expressions[i];
751 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj)
753 if (bitmap_bit_p (&set->expressions, j))
754 result.quick_push (expression_for_id (j));
758 return result;
761 /* Perform bitmapped set operation DEST &= ORIG. */
763 static void
764 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
766 bitmap_iterator bi;
767 unsigned int i;
769 if (dest != orig)
771 bitmap_head temp;
772 bitmap_initialize (&temp, &grand_bitmap_obstack);
774 bitmap_and_into (&dest->values, &orig->values);
775 bitmap_copy (&temp, &dest->expressions);
776 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
778 pre_expr expr = expression_for_id (i);
779 unsigned int value_id = get_expr_value_id (expr);
780 if (!bitmap_bit_p (&dest->values, value_id))
781 bitmap_clear_bit (&dest->expressions, i);
783 bitmap_clear (&temp);
787 /* Subtract all values and expressions contained in ORIG from DEST. */
789 static bitmap_set_t
790 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
792 bitmap_set_t result = bitmap_set_new ();
793 bitmap_iterator bi;
794 unsigned int i;
796 bitmap_and_compl (&result->expressions, &dest->expressions,
797 &orig->expressions);
799 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
801 pre_expr expr = expression_for_id (i);
802 unsigned int value_id = get_expr_value_id (expr);
803 bitmap_set_bit (&result->values, value_id);
806 return result;
809 /* Subtract all the values in bitmap set B from bitmap set A. */
811 static void
812 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
814 unsigned int i;
815 bitmap_iterator bi;
816 bitmap_head temp;
818 bitmap_initialize (&temp, &grand_bitmap_obstack);
820 bitmap_copy (&temp, &a->expressions);
821 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
823 pre_expr expr = expression_for_id (i);
824 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
825 bitmap_remove_from_set (a, expr);
827 bitmap_clear (&temp);
831 /* Return true if bitmapped set SET contains the value VALUE_ID. */
833 static bool
834 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
836 if (value_id_constant_p (value_id))
837 return true;
839 if (!set || bitmap_empty_p (&set->expressions))
840 return false;
842 return bitmap_bit_p (&set->values, value_id);
845 static inline bool
846 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
848 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
851 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
853 static void
854 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
855 const pre_expr expr)
857 bitmap exprset;
858 unsigned int i;
859 bitmap_iterator bi;
861 if (value_id_constant_p (lookfor))
862 return;
864 if (!bitmap_set_contains_value (set, lookfor))
865 return;
867 /* The number of expressions having a given value is usually
868 significantly less than the total number of expressions in SET.
869 Thus, rather than check, for each expression in SET, whether it
870 has the value LOOKFOR, we walk the reverse mapping that tells us
871 what expressions have a given value, and see if any of those
872 expressions are in our set. For large testcases, this is about
873 5-10x faster than walking the bitmap. If this is somehow a
874 significant lose for some cases, we can choose which set to walk
875 based on the set size. */
876 exprset = value_expressions[lookfor];
877 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
879 if (bitmap_clear_bit (&set->expressions, i))
881 bitmap_set_bit (&set->expressions, get_expression_id (expr));
882 return;
886 gcc_unreachable ();
889 /* Return true if two bitmap sets are equal. */
891 static bool
892 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
894 return bitmap_equal_p (&a->values, &b->values);
897 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
898 and add it otherwise. */
900 static void
901 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
903 unsigned int val = get_expr_value_id (expr);
905 if (bitmap_set_contains_value (set, val))
906 bitmap_set_replace_value (set, val, expr);
907 else
908 bitmap_insert_into_set (set, expr);
911 /* Insert EXPR into SET if EXPR's value is not already present in
912 SET. */
914 static void
915 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
917 unsigned int val = get_expr_value_id (expr);
919 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
921 /* Constant values are always considered to be part of the set. */
922 if (value_id_constant_p (val))
923 return;
925 /* If the value membership changed, add the expression. */
926 if (bitmap_set_bit (&set->values, val))
927 bitmap_set_bit (&set->expressions, expr->id);
930 /* Print out EXPR to outfile. */
932 static void
933 print_pre_expr (FILE *outfile, const pre_expr expr)
935 switch (expr->kind)
937 case CONSTANT:
938 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
939 break;
940 case NAME:
941 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
942 break;
943 case NARY:
945 unsigned int i;
946 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
947 fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode));
948 for (i = 0; i < nary->length; i++)
950 print_generic_expr (outfile, nary->op[i], 0);
951 if (i != (unsigned) nary->length - 1)
952 fprintf (outfile, ",");
954 fprintf (outfile, "}");
956 break;
958 case REFERENCE:
960 vn_reference_op_t vro;
961 unsigned int i;
962 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
963 fprintf (outfile, "{");
964 for (i = 0;
965 ref->operands.iterate (i, &vro);
966 i++)
968 bool closebrace = false;
969 if (vro->opcode != SSA_NAME
970 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
972 fprintf (outfile, "%s", get_tree_code_name (vro->opcode));
973 if (vro->op0)
975 fprintf (outfile, "<");
976 closebrace = true;
979 if (vro->op0)
981 print_generic_expr (outfile, vro->op0, 0);
982 if (vro->op1)
984 fprintf (outfile, ",");
985 print_generic_expr (outfile, vro->op1, 0);
987 if (vro->op2)
989 fprintf (outfile, ",");
990 print_generic_expr (outfile, vro->op2, 0);
993 if (closebrace)
994 fprintf (outfile, ">");
995 if (i != ref->operands.length () - 1)
996 fprintf (outfile, ",");
998 fprintf (outfile, "}");
999 if (ref->vuse)
1001 fprintf (outfile, "@");
1002 print_generic_expr (outfile, ref->vuse, 0);
1005 break;
1008 void debug_pre_expr (pre_expr);
1010 /* Like print_pre_expr but always prints to stderr. */
1011 DEBUG_FUNCTION void
1012 debug_pre_expr (pre_expr e)
1014 print_pre_expr (stderr, e);
1015 fprintf (stderr, "\n");
1018 /* Print out SET to OUTFILE. */
1020 static void
1021 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1022 const char *setname, int blockindex)
1024 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1025 if (set)
1027 bool first = true;
1028 unsigned i;
1029 bitmap_iterator bi;
1031 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1033 const pre_expr expr = expression_for_id (i);
1035 if (!first)
1036 fprintf (outfile, ", ");
1037 first = false;
1038 print_pre_expr (outfile, expr);
1040 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1043 fprintf (outfile, " }\n");
1046 void debug_bitmap_set (bitmap_set_t);
1048 DEBUG_FUNCTION void
1049 debug_bitmap_set (bitmap_set_t set)
1051 print_bitmap_set (stderr, set, "debug", 0);
1054 void debug_bitmap_sets_for (basic_block);
1056 DEBUG_FUNCTION void
1057 debug_bitmap_sets_for (basic_block bb)
1059 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1060 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1061 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1062 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1063 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1064 if (do_partial_partial)
1065 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1066 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1069 /* Print out the expressions that have VAL to OUTFILE. */
1071 static void
1072 print_value_expressions (FILE *outfile, unsigned int val)
1074 bitmap set = value_expressions[val];
1075 if (set)
1077 bitmap_set x;
1078 char s[10];
1079 sprintf (s, "%04d", val);
1080 x.expressions = *set;
1081 print_bitmap_set (outfile, &x, s, 0);
1086 DEBUG_FUNCTION void
1087 debug_value_expressions (unsigned int val)
1089 print_value_expressions (stderr, val);
1092 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1093 represent it. */
1095 static pre_expr
1096 get_or_alloc_expr_for_constant (tree constant)
1098 unsigned int result_id;
1099 unsigned int value_id;
1100 struct pre_expr_d expr;
1101 pre_expr newexpr;
1103 expr.kind = CONSTANT;
1104 PRE_EXPR_CONSTANT (&expr) = constant;
1105 result_id = lookup_expression_id (&expr);
1106 if (result_id != 0)
1107 return expression_for_id (result_id);
1109 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1110 newexpr->kind = CONSTANT;
1111 PRE_EXPR_CONSTANT (newexpr) = constant;
1112 alloc_expression_id (newexpr);
1113 value_id = get_or_alloc_constant_value_id (constant);
1114 add_to_value (value_id, newexpr);
1115 return newexpr;
1118 /* Given a value id V, find the actual tree representing the constant
1119 value if there is one, and return it. Return NULL if we can't find
1120 a constant. */
1122 static tree
1123 get_constant_for_value_id (unsigned int v)
1125 if (value_id_constant_p (v))
1127 unsigned int i;
1128 bitmap_iterator bi;
1129 bitmap exprset = value_expressions[v];
1131 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1133 pre_expr expr = expression_for_id (i);
1134 if (expr->kind == CONSTANT)
1135 return PRE_EXPR_CONSTANT (expr);
1138 return NULL;
1141 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1142 Currently only supports constants and SSA_NAMES. */
1143 static pre_expr
1144 get_or_alloc_expr_for (tree t)
1146 if (TREE_CODE (t) == SSA_NAME)
1147 return get_or_alloc_expr_for_name (t);
1148 else if (is_gimple_min_invariant (t))
1149 return get_or_alloc_expr_for_constant (t);
1150 else
1152 /* More complex expressions can result from SCCVN expression
1153 simplification that inserts values for them. As they all
1154 do not have VOPs the get handled by the nary ops struct. */
1155 vn_nary_op_t result;
1156 unsigned int result_id;
1157 vn_nary_op_lookup (t, &result);
1158 if (result != NULL)
1160 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1161 e->kind = NARY;
1162 PRE_EXPR_NARY (e) = result;
1163 result_id = lookup_expression_id (e);
1164 if (result_id != 0)
1166 pool_free (pre_expr_pool, e);
1167 e = expression_for_id (result_id);
1168 return e;
1170 alloc_expression_id (e);
1171 return e;
1174 return NULL;
1177 /* Return the folded version of T if T, when folded, is a gimple
1178 min_invariant. Otherwise, return T. */
1180 static pre_expr
1181 fully_constant_expression (pre_expr e)
1183 switch (e->kind)
1185 case CONSTANT:
1186 return e;
1187 case NARY:
1189 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1190 switch (TREE_CODE_CLASS (nary->opcode))
1192 case tcc_binary:
1193 case tcc_comparison:
1195 /* We have to go from trees to pre exprs to value ids to
1196 constants. */
1197 tree naryop0 = nary->op[0];
1198 tree naryop1 = nary->op[1];
1199 tree result;
1200 if (!is_gimple_min_invariant (naryop0))
1202 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1203 unsigned int vrep0 = get_expr_value_id (rep0);
1204 tree const0 = get_constant_for_value_id (vrep0);
1205 if (const0)
1206 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1208 if (!is_gimple_min_invariant (naryop1))
1210 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1211 unsigned int vrep1 = get_expr_value_id (rep1);
1212 tree const1 = get_constant_for_value_id (vrep1);
1213 if (const1)
1214 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1216 result = fold_binary (nary->opcode, nary->type,
1217 naryop0, naryop1);
1218 if (result && is_gimple_min_invariant (result))
1219 return get_or_alloc_expr_for_constant (result);
1220 /* We might have simplified the expression to a
1221 SSA_NAME for example from x_1 * 1. But we cannot
1222 insert a PHI for x_1 unconditionally as x_1 might
1223 not be available readily. */
1224 return e;
1226 case tcc_reference:
1227 if (nary->opcode != REALPART_EXPR
1228 && nary->opcode != IMAGPART_EXPR
1229 && nary->opcode != VIEW_CONVERT_EXPR)
1230 return e;
1231 /* Fallthrough. */
1232 case tcc_unary:
1234 /* We have to go from trees to pre exprs to value ids to
1235 constants. */
1236 tree naryop0 = nary->op[0];
1237 tree const0, result;
1238 if (is_gimple_min_invariant (naryop0))
1239 const0 = naryop0;
1240 else
1242 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1243 unsigned int vrep0 = get_expr_value_id (rep0);
1244 const0 = get_constant_for_value_id (vrep0);
1246 result = NULL;
1247 if (const0)
1249 tree type1 = TREE_TYPE (nary->op[0]);
1250 const0 = fold_convert (type1, const0);
1251 result = fold_unary (nary->opcode, nary->type, const0);
1253 if (result && is_gimple_min_invariant (result))
1254 return get_or_alloc_expr_for_constant (result);
1255 return e;
1257 default:
1258 return e;
1261 case REFERENCE:
1263 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1264 tree folded;
1265 if ((folded = fully_constant_vn_reference_p (ref)))
1266 return get_or_alloc_expr_for_constant (folded);
1267 return e;
1269 default:
1270 return e;
1272 return e;
1275 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1276 it has the value it would have in BLOCK. Set *SAME_VALID to true
1277 in case the new vuse doesn't change the value id of the OPERANDS. */
1279 static tree
1280 translate_vuse_through_block (vec<vn_reference_op_s> operands,
1281 alias_set_type set, tree type, tree vuse,
1282 basic_block phiblock,
1283 basic_block block, bool *same_valid)
1285 gimple phi = SSA_NAME_DEF_STMT (vuse);
1286 ao_ref ref;
1287 edge e = NULL;
1288 bool use_oracle;
1290 *same_valid = true;
1292 if (gimple_bb (phi) != phiblock)
1293 return vuse;
1295 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1297 /* Use the alias-oracle to find either the PHI node in this block,
1298 the first VUSE used in this block that is equivalent to vuse or
1299 the first VUSE which definition in this block kills the value. */
1300 if (gimple_code (phi) == GIMPLE_PHI)
1301 e = find_edge (block, phiblock);
1302 else if (use_oracle)
1303 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1305 vuse = gimple_vuse (phi);
1306 phi = SSA_NAME_DEF_STMT (vuse);
1307 if (gimple_bb (phi) != phiblock)
1308 return vuse;
1309 if (gimple_code (phi) == GIMPLE_PHI)
1311 e = find_edge (block, phiblock);
1312 break;
1315 else
1316 return NULL_TREE;
1318 if (e)
1320 if (use_oracle)
1322 bitmap visited = NULL;
1323 unsigned int cnt;
1324 /* Try to find a vuse that dominates this phi node by skipping
1325 non-clobbering statements. */
1326 vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false,
1327 NULL, NULL);
1328 if (visited)
1329 BITMAP_FREE (visited);
1331 else
1332 vuse = NULL_TREE;
1333 if (!vuse)
1335 /* If we didn't find any, the value ID can't stay the same,
1336 but return the translated vuse. */
1337 *same_valid = false;
1338 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1340 /* ??? We would like to return vuse here as this is the canonical
1341 upmost vdef that this reference is associated with. But during
1342 insertion of the references into the hash tables we only ever
1343 directly insert with their direct gimple_vuse, hence returning
1344 something else would make us not find the other expression. */
1345 return PHI_ARG_DEF (phi, e->dest_idx);
1348 return NULL_TREE;
1351 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1352 SET2. This is used to avoid making a set consisting of the union
1353 of PA_IN and ANTIC_IN during insert. */
1355 static inline pre_expr
1356 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1358 pre_expr result;
1360 result = bitmap_find_leader (set1, val);
1361 if (!result && set2)
1362 result = bitmap_find_leader (set2, val);
1363 return result;
1366 /* Get the tree type for our PRE expression e. */
1368 static tree
1369 get_expr_type (const pre_expr e)
1371 switch (e->kind)
1373 case NAME:
1374 return TREE_TYPE (PRE_EXPR_NAME (e));
1375 case CONSTANT:
1376 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1377 case REFERENCE:
1378 return PRE_EXPR_REFERENCE (e)->type;
1379 case NARY:
1380 return PRE_EXPR_NARY (e)->type;
1382 gcc_unreachable ();
1385 /* Get a representative SSA_NAME for a given expression.
1386 Since all of our sub-expressions are treated as values, we require
1387 them to be SSA_NAME's for simplicity.
1388 Prior versions of GVNPRE used to use "value handles" here, so that
1389 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1390 either case, the operands are really values (IE we do not expect
1391 them to be usable without finding leaders). */
1393 static tree
1394 get_representative_for (const pre_expr e)
1396 tree name;
1397 unsigned int value_id = get_expr_value_id (e);
1399 switch (e->kind)
1401 case NAME:
1402 return PRE_EXPR_NAME (e);
1403 case CONSTANT:
1404 return PRE_EXPR_CONSTANT (e);
1405 case NARY:
1406 case REFERENCE:
1408 /* Go through all of the expressions representing this value
1409 and pick out an SSA_NAME. */
1410 unsigned int i;
1411 bitmap_iterator bi;
1412 bitmap exprs = value_expressions[value_id];
1413 EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1415 pre_expr rep = expression_for_id (i);
1416 if (rep->kind == NAME)
1417 return PRE_EXPR_NAME (rep);
1418 else if (rep->kind == CONSTANT)
1419 return PRE_EXPR_CONSTANT (rep);
1422 break;
1425 /* If we reached here we couldn't find an SSA_NAME. This can
1426 happen when we've discovered a value that has never appeared in
1427 the program as set to an SSA_NAME, as the result of phi translation.
1428 Create one here.
1429 ??? We should be able to re-use this when we insert the statement
1430 to compute it. */
1431 name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1432 VN_INFO_GET (name)->value_id = value_id;
1433 VN_INFO (name)->valnum = name;
1434 /* ??? For now mark this SSA name for release by SCCVN. */
1435 VN_INFO (name)->needs_insertion = true;
1436 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1437 if (dump_file && (dump_flags & TDF_DETAILS))
1439 fprintf (dump_file, "Created SSA_NAME representative ");
1440 print_generic_expr (dump_file, name, 0);
1441 fprintf (dump_file, " for expression:");
1442 print_pre_expr (dump_file, e);
1443 fprintf (dump_file, " (%04d)\n", value_id);
1446 return name;
1451 static pre_expr
1452 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1453 basic_block pred, basic_block phiblock);
1455 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1456 the phis in PRED. Return NULL if we can't find a leader for each part
1457 of the translated expression. */
1459 static pre_expr
1460 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1461 basic_block pred, basic_block phiblock)
1463 switch (expr->kind)
1465 case NARY:
1467 unsigned int i;
1468 bool changed = false;
1469 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1470 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1471 sizeof_vn_nary_op (nary->length));
1472 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1474 for (i = 0; i < newnary->length; i++)
1476 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1477 continue;
1478 else
1480 pre_expr leader, result;
1481 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1482 leader = find_leader_in_sets (op_val_id, set1, set2);
1483 result = phi_translate (leader, set1, set2, pred, phiblock);
1484 if (result && result != leader)
1486 tree name = get_representative_for (result);
1487 if (!name)
1488 return NULL;
1489 newnary->op[i] = name;
1491 else if (!result)
1492 return NULL;
1494 changed |= newnary->op[i] != nary->op[i];
1497 if (changed)
1499 pre_expr constant;
1500 unsigned int new_val_id;
1502 tree result = vn_nary_op_lookup_pieces (newnary->length,
1503 newnary->opcode,
1504 newnary->type,
1505 &newnary->op[0],
1506 &nary);
1507 if (result && is_gimple_min_invariant (result))
1508 return get_or_alloc_expr_for_constant (result);
1510 expr = (pre_expr) pool_alloc (pre_expr_pool);
1511 expr->kind = NARY;
1512 expr->id = 0;
1513 if (nary)
1515 PRE_EXPR_NARY (expr) = nary;
1516 constant = fully_constant_expression (expr);
1517 if (constant != expr)
1518 return constant;
1520 new_val_id = nary->value_id;
1521 get_or_alloc_expression_id (expr);
1523 else
1525 new_val_id = get_next_value_id ();
1526 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
1527 nary = vn_nary_op_insert_pieces (newnary->length,
1528 newnary->opcode,
1529 newnary->type,
1530 &newnary->op[0],
1531 result, new_val_id);
1532 PRE_EXPR_NARY (expr) = nary;
1533 constant = fully_constant_expression (expr);
1534 if (constant != expr)
1535 return constant;
1536 get_or_alloc_expression_id (expr);
1538 add_to_value (new_val_id, expr);
1540 return expr;
1542 break;
1544 case REFERENCE:
1546 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1547 vec<vn_reference_op_s> operands = ref->operands;
1548 tree vuse = ref->vuse;
1549 tree newvuse = vuse;
1550 vec<vn_reference_op_s> newoperands = vNULL;
1551 bool changed = false, same_valid = true;
1552 unsigned int i, n;
1553 vn_reference_op_t operand;
1554 vn_reference_t newref;
1556 for (i = 0; operands.iterate (i, &operand); i++)
1558 pre_expr opresult;
1559 pre_expr leader;
1560 tree op[3];
1561 tree type = operand->type;
1562 vn_reference_op_s newop = *operand;
1563 op[0] = operand->op0;
1564 op[1] = operand->op1;
1565 op[2] = operand->op2;
1566 for (n = 0; n < 3; ++n)
1568 unsigned int op_val_id;
1569 if (!op[n])
1570 continue;
1571 if (TREE_CODE (op[n]) != SSA_NAME)
1573 /* We can't possibly insert these. */
1574 if (n != 0
1575 && !is_gimple_min_invariant (op[n]))
1576 break;
1577 continue;
1579 op_val_id = VN_INFO (op[n])->value_id;
1580 leader = find_leader_in_sets (op_val_id, set1, set2);
1581 if (!leader)
1582 break;
1583 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1584 if (!opresult)
1585 break;
1586 if (opresult != leader)
1588 tree name = get_representative_for (opresult);
1589 if (!name)
1590 break;
1591 changed |= name != op[n];
1592 op[n] = name;
1595 if (n != 3)
1597 newoperands.release ();
1598 return NULL;
1600 if (!changed)
1601 continue;
1602 if (!newoperands.exists ())
1603 newoperands = operands.copy ();
1604 /* We may have changed from an SSA_NAME to a constant */
1605 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1606 newop.opcode = TREE_CODE (op[0]);
1607 newop.type = type;
1608 newop.op0 = op[0];
1609 newop.op1 = op[1];
1610 newop.op2 = op[2];
1611 newoperands[i] = newop;
1613 gcc_checking_assert (i == operands.length ());
1615 if (vuse)
1617 newvuse = translate_vuse_through_block (newoperands.exists ()
1618 ? newoperands : operands,
1619 ref->set, ref->type,
1620 vuse, phiblock, pred,
1621 &same_valid);
1622 if (newvuse == NULL_TREE)
1624 newoperands.release ();
1625 return NULL;
1629 if (changed || newvuse != vuse)
1631 unsigned int new_val_id;
1632 pre_expr constant;
1634 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1635 ref->type,
1636 newoperands.exists ()
1637 ? newoperands : operands,
1638 &newref, VN_WALK);
1639 if (result)
1640 newoperands.release ();
1642 /* We can always insert constants, so if we have a partial
1643 redundant constant load of another type try to translate it
1644 to a constant of appropriate type. */
1645 if (result && is_gimple_min_invariant (result))
1647 tree tem = result;
1648 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1650 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1651 if (tem && !is_gimple_min_invariant (tem))
1652 tem = NULL_TREE;
1654 if (tem)
1655 return get_or_alloc_expr_for_constant (tem);
1658 /* If we'd have to convert things we would need to validate
1659 if we can insert the translated expression. So fail
1660 here for now - we cannot insert an alias with a different
1661 type in the VN tables either, as that would assert. */
1662 if (result
1663 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1664 return NULL;
1665 else if (!result && newref
1666 && !useless_type_conversion_p (ref->type, newref->type))
1668 newoperands.release ();
1669 return NULL;
1672 expr = (pre_expr) pool_alloc (pre_expr_pool);
1673 expr->kind = REFERENCE;
1674 expr->id = 0;
1676 if (newref)
1678 PRE_EXPR_REFERENCE (expr) = newref;
1679 constant = fully_constant_expression (expr);
1680 if (constant != expr)
1681 return constant;
1683 new_val_id = newref->value_id;
1684 get_or_alloc_expression_id (expr);
1686 else
1688 if (changed || !same_valid)
1690 new_val_id = get_next_value_id ();
1691 value_expressions.safe_grow_cleared
1692 (get_max_value_id () + 1);
1694 else
1695 new_val_id = ref->value_id;
1696 if (!newoperands.exists ())
1697 newoperands = operands.copy ();
1698 newref = vn_reference_insert_pieces (newvuse, ref->set,
1699 ref->type,
1700 newoperands,
1701 result, new_val_id);
1702 newoperands = vNULL;
1703 PRE_EXPR_REFERENCE (expr) = newref;
1704 constant = fully_constant_expression (expr);
1705 if (constant != expr)
1706 return constant;
1707 get_or_alloc_expression_id (expr);
1709 add_to_value (new_val_id, expr);
1711 newoperands.release ();
1712 return expr;
1714 break;
1716 case NAME:
1718 tree name = PRE_EXPR_NAME (expr);
1719 gimple def_stmt = SSA_NAME_DEF_STMT (name);
1720 /* If the SSA name is defined by a PHI node in this block,
1721 translate it. */
1722 if (gimple_code (def_stmt) == GIMPLE_PHI
1723 && gimple_bb (def_stmt) == phiblock)
1725 edge e = find_edge (pred, gimple_bb (def_stmt));
1726 tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1728 /* Handle constant. */
1729 if (is_gimple_min_invariant (def))
1730 return get_or_alloc_expr_for_constant (def);
1732 return get_or_alloc_expr_for_name (def);
1734 /* Otherwise return it unchanged - it will get removed if its
1735 value is not available in PREDs AVAIL_OUT set of expressions
1736 by the subtraction of TMP_GEN. */
1737 return expr;
1740 default:
1741 gcc_unreachable ();
1745 /* Wrapper around phi_translate_1 providing caching functionality. */
1747 static pre_expr
1748 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1749 basic_block pred, basic_block phiblock)
1751 expr_pred_trans_t slot = NULL;
1752 pre_expr phitrans;
1754 if (!expr)
1755 return NULL;
1757 /* Constants contain no values that need translation. */
1758 if (expr->kind == CONSTANT)
1759 return expr;
1761 if (value_id_constant_p (get_expr_value_id (expr)))
1762 return expr;
1764 /* Don't add translations of NAMEs as those are cheap to translate. */
1765 if (expr->kind != NAME)
1767 if (phi_trans_add (&slot, expr, pred))
1768 return slot->v;
1769 /* Store NULL for the value we want to return in the case of
1770 recursing. */
1771 slot->v = NULL;
1774 /* Translate. */
1775 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1777 if (slot)
1779 if (phitrans)
1780 slot->v = phitrans;
1781 else
1782 /* Remove failed translations again, they cause insert
1783 iteration to not pick up new opportunities reliably. */
1784 phi_translate_table->remove_elt_with_hash (slot, slot->hashcode);
1787 return phitrans;
1791 /* For each expression in SET, translate the values through phi nodes
1792 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1793 expressions in DEST. */
1795 static void
1796 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1797 basic_block phiblock)
1799 vec<pre_expr> exprs;
1800 pre_expr expr;
1801 int i;
1803 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1805 bitmap_set_copy (dest, set);
1806 return;
1809 exprs = sorted_array_from_bitmap_set (set);
1810 FOR_EACH_VEC_ELT (exprs, i, expr)
1812 pre_expr translated;
1813 translated = phi_translate (expr, set, NULL, pred, phiblock);
1814 if (!translated)
1815 continue;
1817 /* We might end up with multiple expressions from SET being
1818 translated to the same value. In this case we do not want
1819 to retain the NARY or REFERENCE expression but prefer a NAME
1820 which would be the leader. */
1821 if (translated->kind == NAME)
1822 bitmap_value_replace_in_set (dest, translated);
1823 else
1824 bitmap_value_insert_into_set (dest, translated);
1826 exprs.release ();
1829 /* Find the leader for a value (i.e., the name representing that
1830 value) in a given set, and return it. Return NULL if no leader
1831 is found. */
1833 static pre_expr
1834 bitmap_find_leader (bitmap_set_t set, unsigned int val)
1836 if (value_id_constant_p (val))
1838 unsigned int i;
1839 bitmap_iterator bi;
1840 bitmap exprset = value_expressions[val];
1842 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1844 pre_expr expr = expression_for_id (i);
1845 if (expr->kind == CONSTANT)
1846 return expr;
1849 if (bitmap_set_contains_value (set, val))
1851 /* Rather than walk the entire bitmap of expressions, and see
1852 whether any of them has the value we are looking for, we look
1853 at the reverse mapping, which tells us the set of expressions
1854 that have a given value (IE value->expressions with that
1855 value) and see if any of those expressions are in our set.
1856 The number of expressions per value is usually significantly
1857 less than the number of expressions in the set. In fact, for
1858 large testcases, doing it this way is roughly 5-10x faster
1859 than walking the bitmap.
1860 If this is somehow a significant lose for some cases, we can
1861 choose which set to walk based on which set is smaller. */
1862 unsigned int i;
1863 bitmap_iterator bi;
1864 bitmap exprset = value_expressions[val];
1866 EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1867 return expression_for_id (i);
1869 return NULL;
1872 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1873 BLOCK by seeing if it is not killed in the block. Note that we are
1874 only determining whether there is a store that kills it. Because
1875 of the order in which clean iterates over values, we are guaranteed
1876 that altered operands will have caused us to be eliminated from the
1877 ANTIC_IN set already. */
1879 static bool
1880 value_dies_in_block_x (pre_expr expr, basic_block block)
1882 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1883 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1884 gimple def;
1885 gimple_stmt_iterator gsi;
1886 unsigned id = get_expression_id (expr);
1887 bool res = false;
1888 ao_ref ref;
1890 if (!vuse)
1891 return false;
1893 /* Lookup a previously calculated result. */
1894 if (EXPR_DIES (block)
1895 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1896 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1898 /* A memory expression {e, VUSE} dies in the block if there is a
1899 statement that may clobber e. If, starting statement walk from the
1900 top of the basic block, a statement uses VUSE there can be no kill
1901 inbetween that use and the original statement that loaded {e, VUSE},
1902 so we can stop walking. */
1903 ref.base = NULL_TREE;
1904 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1906 tree def_vuse, def_vdef;
1907 def = gsi_stmt (gsi);
1908 def_vuse = gimple_vuse (def);
1909 def_vdef = gimple_vdef (def);
1911 /* Not a memory statement. */
1912 if (!def_vuse)
1913 continue;
1915 /* Not a may-def. */
1916 if (!def_vdef)
1918 /* A load with the same VUSE, we're done. */
1919 if (def_vuse == vuse)
1920 break;
1922 continue;
1925 /* Init ref only if we really need it. */
1926 if (ref.base == NULL_TREE
1927 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1928 refx->operands))
1930 res = true;
1931 break;
1933 /* If the statement may clobber expr, it dies. */
1934 if (stmt_may_clobber_ref_p_1 (def, &ref))
1936 res = true;
1937 break;
1941 /* Remember the result. */
1942 if (!EXPR_DIES (block))
1943 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1944 bitmap_set_bit (EXPR_DIES (block), id * 2);
1945 if (res)
1946 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1948 return res;
1952 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1953 contains its value-id. */
1955 static bool
1956 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1958 if (op && TREE_CODE (op) == SSA_NAME)
1960 unsigned int value_id = VN_INFO (op)->value_id;
1961 if (!(bitmap_set_contains_value (set1, value_id)
1962 || (set2 && bitmap_set_contains_value (set2, value_id))))
1963 return false;
1965 return true;
1968 /* Determine if the expression EXPR is valid in SET1 U SET2.
1969 ONLY SET2 CAN BE NULL.
1970 This means that we have a leader for each part of the expression
1971 (if it consists of values), or the expression is an SSA_NAME.
1972 For loads/calls, we also see if the vuse is killed in this block. */
1974 static bool
1975 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr)
1977 switch (expr->kind)
1979 case NAME:
1980 /* By construction all NAMEs are available. Non-available
1981 NAMEs are removed by subtracting TMP_GEN from the sets. */
1982 return true;
1983 case NARY:
1985 unsigned int i;
1986 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1987 for (i = 0; i < nary->length; i++)
1988 if (!op_valid_in_sets (set1, set2, nary->op[i]))
1989 return false;
1990 return true;
1992 break;
1993 case REFERENCE:
1995 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1996 vn_reference_op_t vro;
1997 unsigned int i;
1999 FOR_EACH_VEC_ELT (ref->operands, i, vro)
2001 if (!op_valid_in_sets (set1, set2, vro->op0)
2002 || !op_valid_in_sets (set1, set2, vro->op1)
2003 || !op_valid_in_sets (set1, set2, vro->op2))
2004 return false;
2006 return true;
2008 default:
2009 gcc_unreachable ();
2013 /* Clean the set of expressions that are no longer valid in SET1 or
2014 SET2. This means expressions that are made up of values we have no
2015 leaders for in SET1 or SET2. This version is used for partial
2016 anticipation, which means it is not valid in either ANTIC_IN or
2017 PA_IN. */
2019 static void
2020 dependent_clean (bitmap_set_t set1, bitmap_set_t set2)
2022 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
2023 pre_expr expr;
2024 int i;
2026 FOR_EACH_VEC_ELT (exprs, i, expr)
2028 if (!valid_in_sets (set1, set2, expr))
2029 bitmap_remove_from_set (set1, expr);
2031 exprs.release ();
2034 /* Clean the set of expressions that are no longer valid in SET. This
2035 means expressions that are made up of values we have no leaders for
2036 in SET. */
2038 static void
2039 clean (bitmap_set_t set)
2041 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set);
2042 pre_expr expr;
2043 int i;
2045 FOR_EACH_VEC_ELT (exprs, i, expr)
2047 if (!valid_in_sets (set, NULL, expr))
2048 bitmap_remove_from_set (set, expr);
2050 exprs.release ();
2053 /* Clean the set of expressions that are no longer valid in SET because
2054 they are clobbered in BLOCK or because they trap and may not be executed. */
2056 static void
2057 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2059 bitmap_iterator bi;
2060 unsigned i;
2062 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2064 pre_expr expr = expression_for_id (i);
2065 if (expr->kind == REFERENCE)
2067 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2068 if (ref->vuse)
2070 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2071 if (!gimple_nop_p (def_stmt)
2072 && ((gimple_bb (def_stmt) != block
2073 && !dominated_by_p (CDI_DOMINATORS,
2074 block, gimple_bb (def_stmt)))
2075 || (gimple_bb (def_stmt) == block
2076 && value_dies_in_block_x (expr, block))))
2077 bitmap_remove_from_set (set, expr);
2080 else if (expr->kind == NARY)
2082 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2083 /* If the NARY may trap make sure the block does not contain
2084 a possible exit point.
2085 ??? This is overly conservative if we translate AVAIL_OUT
2086 as the available expression might be after the exit point. */
2087 if (BB_MAY_NOTRETURN (block)
2088 && vn_nary_may_trap (nary))
2089 bitmap_remove_from_set (set, expr);
2094 static sbitmap has_abnormal_preds;
2096 /* List of blocks that may have changed during ANTIC computation and
2097 thus need to be iterated over. */
2099 static sbitmap changed_blocks;
2101 /* Compute the ANTIC set for BLOCK.
2103 If succs(BLOCK) > 1 then
2104 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2105 else if succs(BLOCK) == 1 then
2106 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2108 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2111 static bool
2112 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2114 bool changed = false;
2115 bitmap_set_t S, old, ANTIC_OUT;
2116 bitmap_iterator bi;
2117 unsigned int bii;
2118 edge e;
2119 edge_iterator ei;
2121 old = ANTIC_OUT = S = NULL;
2122 BB_VISITED (block) = 1;
2124 /* If any edges from predecessors are abnormal, antic_in is empty,
2125 so do nothing. */
2126 if (block_has_abnormal_pred_edge)
2127 goto maybe_dump_sets;
2129 old = ANTIC_IN (block);
2130 ANTIC_OUT = bitmap_set_new ();
2132 /* If the block has no successors, ANTIC_OUT is empty. */
2133 if (EDGE_COUNT (block->succs) == 0)
2135 /* If we have one successor, we could have some phi nodes to
2136 translate through. */
2137 else if (single_succ_p (block))
2139 basic_block succ_bb = single_succ (block);
2140 gcc_assert (BB_VISITED (succ_bb));
2141 phi_translate_set (ANTIC_OUT, ANTIC_IN (succ_bb), block, succ_bb);
2143 /* If we have multiple successors, we take the intersection of all of
2144 them. Note that in the case of loop exit phi nodes, we may have
2145 phis to translate through. */
2146 else
2148 size_t i;
2149 basic_block bprime, first = NULL;
2151 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2152 FOR_EACH_EDGE (e, ei, block->succs)
2154 if (!first
2155 && BB_VISITED (e->dest))
2156 first = e->dest;
2157 else if (BB_VISITED (e->dest))
2158 worklist.quick_push (e->dest);
2161 /* Of multiple successors we have to have visited one already
2162 which is guaranteed by iteration order. */
2163 gcc_assert (first != NULL);
2165 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2167 FOR_EACH_VEC_ELT (worklist, i, bprime)
2169 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2171 bitmap_set_t tmp = bitmap_set_new ();
2172 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2173 bitmap_set_and (ANTIC_OUT, tmp);
2174 bitmap_set_free (tmp);
2176 else
2177 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2181 /* Prune expressions that are clobbered in block and thus become
2182 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2183 prune_clobbered_mems (ANTIC_OUT, block);
2185 /* Generate ANTIC_OUT - TMP_GEN. */
2186 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2188 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2189 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2190 TMP_GEN (block));
2192 /* Then union in the ANTIC_OUT - TMP_GEN values,
2193 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2194 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2195 bitmap_value_insert_into_set (ANTIC_IN (block),
2196 expression_for_id (bii));
2198 clean (ANTIC_IN (block));
2200 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2202 changed = true;
2203 bitmap_set_bit (changed_blocks, block->index);
2204 FOR_EACH_EDGE (e, ei, block->preds)
2205 bitmap_set_bit (changed_blocks, e->src->index);
2207 else
2208 bitmap_clear_bit (changed_blocks, block->index);
2210 maybe_dump_sets:
2211 if (dump_file && (dump_flags & TDF_DETAILS))
2213 if (ANTIC_OUT)
2214 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2216 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2217 block->index);
2219 if (S)
2220 print_bitmap_set (dump_file, S, "S", block->index);
2222 if (old)
2223 bitmap_set_free (old);
2224 if (S)
2225 bitmap_set_free (S);
2226 if (ANTIC_OUT)
2227 bitmap_set_free (ANTIC_OUT);
2228 return changed;
2231 /* Compute PARTIAL_ANTIC for BLOCK.
2233 If succs(BLOCK) > 1 then
2234 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2235 in ANTIC_OUT for all succ(BLOCK)
2236 else if succs(BLOCK) == 1 then
2237 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2239 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2240 - ANTIC_IN[BLOCK])
2243 static bool
2244 compute_partial_antic_aux (basic_block block,
2245 bool block_has_abnormal_pred_edge)
2247 bool changed = false;
2248 bitmap_set_t old_PA_IN;
2249 bitmap_set_t PA_OUT;
2250 edge e;
2251 edge_iterator ei;
2252 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2254 old_PA_IN = PA_OUT = NULL;
2256 /* If any edges from predecessors are abnormal, antic_in is empty,
2257 so do nothing. */
2258 if (block_has_abnormal_pred_edge)
2259 goto maybe_dump_sets;
2261 /* If there are too many partially anticipatable values in the
2262 block, phi_translate_set can take an exponential time: stop
2263 before the translation starts. */
2264 if (max_pa
2265 && single_succ_p (block)
2266 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2267 goto maybe_dump_sets;
2269 old_PA_IN = PA_IN (block);
2270 PA_OUT = bitmap_set_new ();
2272 /* If the block has no successors, ANTIC_OUT is empty. */
2273 if (EDGE_COUNT (block->succs) == 0)
2275 /* If we have one successor, we could have some phi nodes to
2276 translate through. Note that we can't phi translate across DFS
2277 back edges in partial antic, because it uses a union operation on
2278 the successors. For recurrences like IV's, we will end up
2279 generating a new value in the set on each go around (i + 3 (VH.1)
2280 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2281 else if (single_succ_p (block))
2283 basic_block succ = single_succ (block);
2284 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2285 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2287 /* If we have multiple successors, we take the union of all of
2288 them. */
2289 else
2291 size_t i;
2292 basic_block bprime;
2294 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2295 FOR_EACH_EDGE (e, ei, block->succs)
2297 if (e->flags & EDGE_DFS_BACK)
2298 continue;
2299 worklist.quick_push (e->dest);
2301 if (worklist.length () > 0)
2303 FOR_EACH_VEC_ELT (worklist, i, bprime)
2305 unsigned int i;
2306 bitmap_iterator bi;
2308 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2309 bitmap_value_insert_into_set (PA_OUT,
2310 expression_for_id (i));
2311 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2313 bitmap_set_t pa_in = bitmap_set_new ();
2314 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2315 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2316 bitmap_value_insert_into_set (PA_OUT,
2317 expression_for_id (i));
2318 bitmap_set_free (pa_in);
2320 else
2321 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2322 bitmap_value_insert_into_set (PA_OUT,
2323 expression_for_id (i));
2328 /* Prune expressions that are clobbered in block and thus become
2329 invalid if translated from PA_OUT to PA_IN. */
2330 prune_clobbered_mems (PA_OUT, block);
2332 /* PA_IN starts with PA_OUT - TMP_GEN.
2333 Then we subtract things from ANTIC_IN. */
2334 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2336 /* For partial antic, we want to put back in the phi results, since
2337 we will properly avoid making them partially antic over backedges. */
2338 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2339 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2341 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2342 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2344 dependent_clean (PA_IN (block), ANTIC_IN (block));
2346 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2348 changed = true;
2349 bitmap_set_bit (changed_blocks, block->index);
2350 FOR_EACH_EDGE (e, ei, block->preds)
2351 bitmap_set_bit (changed_blocks, e->src->index);
2353 else
2354 bitmap_clear_bit (changed_blocks, block->index);
2356 maybe_dump_sets:
2357 if (dump_file && (dump_flags & TDF_DETAILS))
2359 if (PA_OUT)
2360 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2362 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2364 if (old_PA_IN)
2365 bitmap_set_free (old_PA_IN);
2366 if (PA_OUT)
2367 bitmap_set_free (PA_OUT);
2368 return changed;
2371 /* Compute ANTIC and partial ANTIC sets. */
2373 static void
2374 compute_antic (void)
2376 bool changed = true;
2377 int num_iterations = 0;
2378 basic_block block;
2379 int i;
2381 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2382 We pre-build the map of blocks with incoming abnormal edges here. */
2383 has_abnormal_preds = sbitmap_alloc (last_basic_block_for_fn (cfun));
2384 bitmap_clear (has_abnormal_preds);
2386 FOR_ALL_BB_FN (block, cfun)
2388 edge_iterator ei;
2389 edge e;
2391 FOR_EACH_EDGE (e, ei, block->preds)
2393 e->flags &= ~EDGE_DFS_BACK;
2394 if (e->flags & EDGE_ABNORMAL)
2396 bitmap_set_bit (has_abnormal_preds, block->index);
2397 break;
2401 BB_VISITED (block) = 0;
2403 /* While we are here, give empty ANTIC_IN sets to each block. */
2404 ANTIC_IN (block) = bitmap_set_new ();
2405 PA_IN (block) = bitmap_set_new ();
2408 /* At the exit block we anticipate nothing. */
2409 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun)) = 1;
2411 changed_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun) + 1);
2412 bitmap_ones (changed_blocks);
2413 while (changed)
2415 if (dump_file && (dump_flags & TDF_DETAILS))
2416 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2417 /* ??? We need to clear our PHI translation cache here as the
2418 ANTIC sets shrink and we restrict valid translations to
2419 those having operands with leaders in ANTIC. Same below
2420 for PA ANTIC computation. */
2421 num_iterations++;
2422 changed = false;
2423 for (i = postorder_num - 1; i >= 0; i--)
2425 if (bitmap_bit_p (changed_blocks, postorder[i]))
2427 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2428 changed |= compute_antic_aux (block,
2429 bitmap_bit_p (has_abnormal_preds,
2430 block->index));
2433 /* Theoretically possible, but *highly* unlikely. */
2434 gcc_checking_assert (num_iterations < 500);
2437 statistics_histogram_event (cfun, "compute_antic iterations",
2438 num_iterations);
2440 if (do_partial_partial)
2442 bitmap_ones (changed_blocks);
2443 mark_dfs_back_edges ();
2444 num_iterations = 0;
2445 changed = true;
2446 while (changed)
2448 if (dump_file && (dump_flags & TDF_DETAILS))
2449 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2450 num_iterations++;
2451 changed = false;
2452 for (i = postorder_num - 1 ; i >= 0; i--)
2454 if (bitmap_bit_p (changed_blocks, postorder[i]))
2456 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2457 changed
2458 |= compute_partial_antic_aux (block,
2459 bitmap_bit_p (has_abnormal_preds,
2460 block->index));
2463 /* Theoretically possible, but *highly* unlikely. */
2464 gcc_checking_assert (num_iterations < 500);
2466 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2467 num_iterations);
2469 sbitmap_free (has_abnormal_preds);
2470 sbitmap_free (changed_blocks);
2474 /* Inserted expressions are placed onto this worklist, which is used
2475 for performing quick dead code elimination of insertions we made
2476 that didn't turn out to be necessary. */
2477 static bitmap inserted_exprs;
2479 /* The actual worker for create_component_ref_by_pieces. */
2481 static tree
2482 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2483 unsigned int *operand, gimple_seq *stmts)
2485 vn_reference_op_t currop = &ref->operands[*operand];
2486 tree genop;
2487 ++*operand;
2488 switch (currop->opcode)
2490 case CALL_EXPR:
2492 tree folded, sc = NULL_TREE;
2493 unsigned int nargs = 0;
2494 tree fn, *args;
2495 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2496 fn = currop->op0;
2497 else
2498 fn = find_or_generate_expression (block, currop->op0, stmts);
2499 if (!fn)
2500 return NULL_TREE;
2501 if (currop->op1)
2503 sc = find_or_generate_expression (block, currop->op1, stmts);
2504 if (!sc)
2505 return NULL_TREE;
2507 args = XNEWVEC (tree, ref->operands.length () - 1);
2508 while (*operand < ref->operands.length ())
2510 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2511 operand, stmts);
2512 if (!args[nargs])
2513 return NULL_TREE;
2514 nargs++;
2516 folded = build_call_array (currop->type,
2517 (TREE_CODE (fn) == FUNCTION_DECL
2518 ? build_fold_addr_expr (fn) : fn),
2519 nargs, args);
2520 if (currop->with_bounds)
2521 CALL_WITH_BOUNDS_P (folded) = true;
2522 free (args);
2523 if (sc)
2524 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2525 return folded;
2528 case MEM_REF:
2530 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2531 stmts);
2532 if (!baseop)
2533 return NULL_TREE;
2534 tree offset = currop->op0;
2535 if (TREE_CODE (baseop) == ADDR_EXPR
2536 && handled_component_p (TREE_OPERAND (baseop, 0)))
2538 HOST_WIDE_INT off;
2539 tree base;
2540 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2541 &off);
2542 gcc_assert (base);
2543 offset = int_const_binop (PLUS_EXPR, offset,
2544 build_int_cst (TREE_TYPE (offset),
2545 off));
2546 baseop = build_fold_addr_expr (base);
2548 tree t = fold_build2 (MEM_REF, currop->type, baseop, offset);
2549 REF_REVERSE_STORAGE_ORDER (t) = currop->reverse;
2550 return t;
2553 case TARGET_MEM_REF:
2555 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2556 vn_reference_op_t nextop = &ref->operands[++*operand];
2557 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2558 stmts);
2559 if (!baseop)
2560 return NULL_TREE;
2561 if (currop->op0)
2563 genop0 = find_or_generate_expression (block, currop->op0, stmts);
2564 if (!genop0)
2565 return NULL_TREE;
2567 if (nextop->op0)
2569 genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2570 if (!genop1)
2571 return NULL_TREE;
2573 return build5 (TARGET_MEM_REF, currop->type,
2574 baseop, currop->op2, genop0, currop->op1, genop1);
2577 case ADDR_EXPR:
2578 if (currop->op0)
2580 gcc_assert (is_gimple_min_invariant (currop->op0));
2581 return currop->op0;
2583 /* Fallthrough. */
2584 case REALPART_EXPR:
2585 case IMAGPART_EXPR:
2586 case VIEW_CONVERT_EXPR:
2588 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2589 stmts);
2590 if (!genop0)
2591 return NULL_TREE;
2592 return fold_build1 (currop->opcode, currop->type, genop0);
2595 case WITH_SIZE_EXPR:
2597 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2598 stmts);
2599 if (!genop0)
2600 return NULL_TREE;
2601 tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2602 if (!genop1)
2603 return NULL_TREE;
2604 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2607 case BIT_FIELD_REF:
2609 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2610 stmts);
2611 if (!genop0)
2612 return NULL_TREE;
2613 tree op1 = currop->op0;
2614 tree op2 = currop->op1;
2615 tree t = build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2616 REF_REVERSE_STORAGE_ORDER (t) = currop->reverse;
2617 return fold (t);
2620 /* For array ref vn_reference_op's, operand 1 of the array ref
2621 is op0 of the reference op and operand 3 of the array ref is
2622 op1. */
2623 case ARRAY_RANGE_REF:
2624 case ARRAY_REF:
2626 tree genop0;
2627 tree genop1 = currop->op0;
2628 tree genop2 = currop->op1;
2629 tree genop3 = currop->op2;
2630 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2631 stmts);
2632 if (!genop0)
2633 return NULL_TREE;
2634 genop1 = find_or_generate_expression (block, genop1, stmts);
2635 if (!genop1)
2636 return NULL_TREE;
2637 if (genop2)
2639 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2640 /* Drop zero minimum index if redundant. */
2641 if (integer_zerop (genop2)
2642 && (!domain_type
2643 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2644 genop2 = NULL_TREE;
2645 else
2647 genop2 = find_or_generate_expression (block, genop2, stmts);
2648 if (!genop2)
2649 return NULL_TREE;
2652 if (genop3)
2654 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2655 /* We can't always put a size in units of the element alignment
2656 here as the element alignment may be not visible. See
2657 PR43783. Simply drop the element size for constant
2658 sizes. */
2659 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2660 genop3 = NULL_TREE;
2661 else
2663 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2664 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2665 genop3 = find_or_generate_expression (block, genop3, stmts);
2666 if (!genop3)
2667 return NULL_TREE;
2670 return build4 (currop->opcode, currop->type, genop0, genop1,
2671 genop2, genop3);
2673 case COMPONENT_REF:
2675 tree op0;
2676 tree op1;
2677 tree genop2 = currop->op1;
2678 op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2679 if (!op0)
2680 return NULL_TREE;
2681 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2682 op1 = currop->op0;
2683 if (genop2)
2685 genop2 = find_or_generate_expression (block, genop2, stmts);
2686 if (!genop2)
2687 return NULL_TREE;
2689 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2692 case SSA_NAME:
2694 genop = find_or_generate_expression (block, currop->op0, stmts);
2695 return genop;
2697 case STRING_CST:
2698 case INTEGER_CST:
2699 case COMPLEX_CST:
2700 case VECTOR_CST:
2701 case REAL_CST:
2702 case CONSTRUCTOR:
2703 case VAR_DECL:
2704 case PARM_DECL:
2705 case CONST_DECL:
2706 case RESULT_DECL:
2707 case FUNCTION_DECL:
2708 return currop->op0;
2710 default:
2711 gcc_unreachable ();
2715 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2716 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2717 trying to rename aggregates into ssa form directly, which is a no no.
2719 Thus, this routine doesn't create temporaries, it just builds a
2720 single access expression for the array, calling
2721 find_or_generate_expression to build the innermost pieces.
2723 This function is a subroutine of create_expression_by_pieces, and
2724 should not be called on it's own unless you really know what you
2725 are doing. */
2727 static tree
2728 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2729 gimple_seq *stmts)
2731 unsigned int op = 0;
2732 return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2735 /* Find a simple leader for an expression, or generate one using
2736 create_expression_by_pieces from a NARY expression for the value.
2737 BLOCK is the basic_block we are looking for leaders in.
2738 OP is the tree expression to find a leader for or generate.
2739 Returns the leader or NULL_TREE on failure. */
2741 static tree
2742 find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2744 pre_expr expr = get_or_alloc_expr_for (op);
2745 unsigned int lookfor = get_expr_value_id (expr);
2746 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2747 if (leader)
2749 if (leader->kind == NAME)
2750 return PRE_EXPR_NAME (leader);
2751 else if (leader->kind == CONSTANT)
2752 return PRE_EXPR_CONSTANT (leader);
2754 /* Defer. */
2755 return NULL_TREE;
2758 /* It must be a complex expression, so generate it recursively. Note
2759 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2760 where the insert algorithm fails to insert a required expression. */
2761 bitmap exprset = value_expressions[lookfor];
2762 bitmap_iterator bi;
2763 unsigned int i;
2764 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2766 pre_expr temp = expression_for_id (i);
2767 /* We cannot insert random REFERENCE expressions at arbitrary
2768 places. We can insert NARYs which eventually re-materializes
2769 its operand values. */
2770 if (temp->kind == NARY)
2771 return create_expression_by_pieces (block, temp, stmts,
2772 get_expr_type (expr));
2775 /* Defer. */
2776 return NULL_TREE;
2779 #define NECESSARY GF_PLF_1
2781 /* Create an expression in pieces, so that we can handle very complex
2782 expressions that may be ANTIC, but not necessary GIMPLE.
2783 BLOCK is the basic block the expression will be inserted into,
2784 EXPR is the expression to insert (in value form)
2785 STMTS is a statement list to append the necessary insertions into.
2787 This function will die if we hit some value that shouldn't be
2788 ANTIC but is (IE there is no leader for it, or its components).
2789 The function returns NULL_TREE in case a different antic expression
2790 has to be inserted first.
2791 This function may also generate expressions that are themselves
2792 partially or fully redundant. Those that are will be either made
2793 fully redundant during the next iteration of insert (for partially
2794 redundant ones), or eliminated by eliminate (for fully redundant
2795 ones). */
2797 static tree
2798 create_expression_by_pieces (basic_block block, pre_expr expr,
2799 gimple_seq *stmts, tree type)
2801 tree name;
2802 tree folded;
2803 gimple_seq forced_stmts = NULL;
2804 unsigned int value_id;
2805 gimple_stmt_iterator gsi;
2806 tree exprtype = type ? type : get_expr_type (expr);
2807 pre_expr nameexpr;
2808 gimple newstmt;
2810 switch (expr->kind)
2812 /* We may hit the NAME/CONSTANT case if we have to convert types
2813 that value numbering saw through. */
2814 case NAME:
2815 folded = PRE_EXPR_NAME (expr);
2816 break;
2817 case CONSTANT:
2818 folded = PRE_EXPR_CONSTANT (expr);
2819 break;
2820 case REFERENCE:
2822 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2823 folded = create_component_ref_by_pieces (block, ref, stmts);
2824 if (!folded)
2825 return NULL_TREE;
2827 break;
2828 case NARY:
2830 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2831 tree *genop = XALLOCAVEC (tree, nary->length);
2832 unsigned i;
2833 for (i = 0; i < nary->length; ++i)
2835 genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2836 if (!genop[i])
2837 return NULL_TREE;
2838 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2839 may have conversions stripped. */
2840 if (nary->opcode == POINTER_PLUS_EXPR)
2842 if (i == 0)
2843 genop[i] = gimple_convert (&forced_stmts,
2844 nary->type, genop[i]);
2845 else if (i == 1)
2846 genop[i] = gimple_convert (&forced_stmts,
2847 sizetype, genop[i]);
2849 else
2850 genop[i] = gimple_convert (&forced_stmts,
2851 TREE_TYPE (nary->op[i]), genop[i]);
2853 if (nary->opcode == CONSTRUCTOR)
2855 vec<constructor_elt, va_gc> *elts = NULL;
2856 for (i = 0; i < nary->length; ++i)
2857 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2858 folded = build_constructor (nary->type, elts);
2860 else
2862 switch (nary->length)
2864 case 1:
2865 folded = fold_build1 (nary->opcode, nary->type,
2866 genop[0]);
2867 break;
2868 case 2:
2869 folded = fold_build2 (nary->opcode, nary->type,
2870 genop[0], genop[1]);
2871 break;
2872 case 3:
2873 folded = fold_build3 (nary->opcode, nary->type,
2874 genop[0], genop[1], genop[2]);
2875 break;
2876 default:
2877 gcc_unreachable ();
2881 break;
2882 default:
2883 gcc_unreachable ();
2886 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2887 folded = fold_convert (exprtype, folded);
2889 /* Force the generated expression to be a sequence of GIMPLE
2890 statements.
2891 We have to call unshare_expr because force_gimple_operand may
2892 modify the tree we pass to it. */
2893 gimple_seq tem = NULL;
2894 folded = force_gimple_operand (unshare_expr (folded), &tem,
2895 false, NULL);
2896 gimple_seq_add_seq_without_update (&forced_stmts, tem);
2898 /* If we have any intermediate expressions to the value sets, add them
2899 to the value sets and chain them in the instruction stream. */
2900 if (forced_stmts)
2902 gsi = gsi_start (forced_stmts);
2903 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2905 gimple stmt = gsi_stmt (gsi);
2906 tree forcedname = gimple_get_lhs (stmt);
2907 pre_expr nameexpr;
2909 if (TREE_CODE (forcedname) == SSA_NAME)
2911 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
2912 VN_INFO_GET (forcedname)->valnum = forcedname;
2913 VN_INFO (forcedname)->value_id = get_next_value_id ();
2914 nameexpr = get_or_alloc_expr_for_name (forcedname);
2915 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2916 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2917 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2920 gimple_set_vuse (stmt, BB_LIVE_VOP_ON_EXIT (block));
2921 gimple_set_modified (stmt, true);
2923 gimple_seq_add_seq (stmts, forced_stmts);
2926 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2927 newstmt = gimple_build_assign (name, folded);
2928 gimple_set_vuse (newstmt, BB_LIVE_VOP_ON_EXIT (block));
2929 gimple_set_modified (newstmt, true);
2930 gimple_set_plf (newstmt, NECESSARY, false);
2932 gimple_seq_add_stmt (stmts, newstmt);
2933 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
2935 /* Fold the last statement. */
2936 gsi = gsi_last (*stmts);
2937 if (fold_stmt_inplace (&gsi))
2938 update_stmt (gsi_stmt (gsi));
2940 /* Add a value number to the temporary.
2941 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2942 we are creating the expression by pieces, and this particular piece of
2943 the expression may have been represented. There is no harm in replacing
2944 here. */
2945 value_id = get_expr_value_id (expr);
2946 VN_INFO_GET (name)->value_id = value_id;
2947 VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id);
2948 if (VN_INFO (name)->valnum == NULL_TREE)
2949 VN_INFO (name)->valnum = name;
2950 gcc_assert (VN_INFO (name)->valnum != NULL_TREE);
2951 nameexpr = get_or_alloc_expr_for_name (name);
2952 add_to_value (value_id, nameexpr);
2953 if (NEW_SETS (block))
2954 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2955 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2957 pre_stats.insertions++;
2958 if (dump_file && (dump_flags & TDF_DETAILS))
2960 fprintf (dump_file, "Inserted ");
2961 print_gimple_stmt (dump_file, newstmt, 0, 0);
2962 fprintf (dump_file, " in predecessor %d (%04d)\n",
2963 block->index, value_id);
2966 return name;
2970 /* Insert the to-be-made-available values of expression EXPRNUM for each
2971 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
2972 merge the result with a phi node, given the same value number as
2973 NODE. Return true if we have inserted new stuff. */
2975 static bool
2976 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
2977 vec<pre_expr> avail)
2979 pre_expr expr = expression_for_id (exprnum);
2980 pre_expr newphi;
2981 unsigned int val = get_expr_value_id (expr);
2982 edge pred;
2983 bool insertions = false;
2984 bool nophi = false;
2985 basic_block bprime;
2986 pre_expr eprime;
2987 edge_iterator ei;
2988 tree type = get_expr_type (expr);
2989 tree temp;
2990 gimple phi;
2992 /* Make sure we aren't creating an induction variable. */
2993 if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
2995 bool firstinsideloop = false;
2996 bool secondinsideloop = false;
2997 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
2998 EDGE_PRED (block, 0)->src);
2999 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3000 EDGE_PRED (block, 1)->src);
3001 /* Induction variables only have one edge inside the loop. */
3002 if ((firstinsideloop ^ secondinsideloop)
3003 && expr->kind != REFERENCE)
3005 if (dump_file && (dump_flags & TDF_DETAILS))
3006 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3007 nophi = true;
3011 /* Make the necessary insertions. */
3012 FOR_EACH_EDGE (pred, ei, block->preds)
3014 gimple_seq stmts = NULL;
3015 tree builtexpr;
3016 bprime = pred->src;
3017 eprime = avail[pred->dest_idx];
3019 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3021 builtexpr = create_expression_by_pieces (bprime, eprime,
3022 &stmts, type);
3023 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3024 gsi_insert_seq_on_edge (pred, stmts);
3025 if (!builtexpr)
3027 /* We cannot insert a PHI node if we failed to insert
3028 on one edge. */
3029 nophi = true;
3030 continue;
3032 avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3033 insertions = true;
3035 else if (eprime->kind == CONSTANT)
3037 /* Constants may not have the right type, fold_convert
3038 should give us back a constant with the right type. */
3039 tree constant = PRE_EXPR_CONSTANT (eprime);
3040 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3042 tree builtexpr = fold_convert (type, constant);
3043 if (!is_gimple_min_invariant (builtexpr))
3045 tree forcedexpr = force_gimple_operand (builtexpr,
3046 &stmts, true,
3047 NULL);
3048 if (!is_gimple_min_invariant (forcedexpr))
3050 if (forcedexpr != builtexpr)
3052 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3053 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3055 if (stmts)
3057 gimple_stmt_iterator gsi;
3058 gsi = gsi_start (stmts);
3059 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3061 gimple stmt = gsi_stmt (gsi);
3062 tree lhs = gimple_get_lhs (stmt);
3063 if (TREE_CODE (lhs) == SSA_NAME)
3064 bitmap_set_bit (inserted_exprs,
3065 SSA_NAME_VERSION (lhs));
3066 gimple_set_plf (stmt, NECESSARY, false);
3068 gsi_insert_seq_on_edge (pred, stmts);
3070 avail[pred->dest_idx]
3071 = get_or_alloc_expr_for_name (forcedexpr);
3074 else
3075 avail[pred->dest_idx]
3076 = get_or_alloc_expr_for_constant (builtexpr);
3079 else if (eprime->kind == NAME)
3081 /* We may have to do a conversion because our value
3082 numbering can look through types in certain cases, but
3083 our IL requires all operands of a phi node have the same
3084 type. */
3085 tree name = PRE_EXPR_NAME (eprime);
3086 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3088 tree builtexpr;
3089 tree forcedexpr;
3090 builtexpr = fold_convert (type, name);
3091 forcedexpr = force_gimple_operand (builtexpr,
3092 &stmts, true,
3093 NULL);
3095 if (forcedexpr != name)
3097 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3098 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3101 if (stmts)
3103 gimple_stmt_iterator gsi;
3104 gsi = gsi_start (stmts);
3105 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3107 gimple stmt = gsi_stmt (gsi);
3108 tree lhs = gimple_get_lhs (stmt);
3109 if (TREE_CODE (lhs) == SSA_NAME)
3110 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3111 gimple_set_plf (stmt, NECESSARY, false);
3113 gsi_insert_seq_on_edge (pred, stmts);
3115 avail[pred->dest_idx] = get_or_alloc_expr_for_name (forcedexpr);
3119 /* If we didn't want a phi node, and we made insertions, we still have
3120 inserted new stuff, and thus return true. If we didn't want a phi node,
3121 and didn't make insertions, we haven't added anything new, so return
3122 false. */
3123 if (nophi && insertions)
3124 return true;
3125 else if (nophi && !insertions)
3126 return false;
3128 /* Now build a phi for the new variable. */
3129 temp = make_temp_ssa_name (type, NULL, "prephitmp");
3130 phi = create_phi_node (temp, block);
3132 gimple_set_plf (phi, NECESSARY, false);
3133 VN_INFO_GET (temp)->value_id = val;
3134 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3135 if (VN_INFO (temp)->valnum == NULL_TREE)
3136 VN_INFO (temp)->valnum = temp;
3137 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3138 FOR_EACH_EDGE (pred, ei, block->preds)
3140 pre_expr ae = avail[pred->dest_idx];
3141 gcc_assert (get_expr_type (ae) == type
3142 || useless_type_conversion_p (type, get_expr_type (ae)));
3143 if (ae->kind == CONSTANT)
3144 add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3145 pred, UNKNOWN_LOCATION);
3146 else
3147 add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3150 newphi = get_or_alloc_expr_for_name (temp);
3151 add_to_value (val, newphi);
3153 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3154 this insertion, since we test for the existence of this value in PHI_GEN
3155 before proceeding with the partial redundancy checks in insert_aux.
3157 The value may exist in AVAIL_OUT, in particular, it could be represented
3158 by the expression we are trying to eliminate, in which case we want the
3159 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3160 inserted there.
3162 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3163 this block, because if it did, it would have existed in our dominator's
3164 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3167 bitmap_insert_into_set (PHI_GEN (block), newphi);
3168 bitmap_value_replace_in_set (AVAIL_OUT (block),
3169 newphi);
3170 bitmap_insert_into_set (NEW_SETS (block),
3171 newphi);
3173 if (dump_file && (dump_flags & TDF_DETAILS))
3175 fprintf (dump_file, "Created phi ");
3176 print_gimple_stmt (dump_file, phi, 0, 0);
3177 fprintf (dump_file, " in block %d (%04d)\n", block->index, val);
3179 pre_stats.phis++;
3180 return true;
3185 /* Perform insertion of partially redundant values.
3186 For BLOCK, do the following:
3187 1. Propagate the NEW_SETS of the dominator into the current block.
3188 If the block has multiple predecessors,
3189 2a. Iterate over the ANTIC expressions for the block to see if
3190 any of them are partially redundant.
3191 2b. If so, insert them into the necessary predecessors to make
3192 the expression fully redundant.
3193 2c. Insert a new PHI merging the values of the predecessors.
3194 2d. Insert the new PHI, and the new expressions, into the
3195 NEW_SETS set.
3196 3. Recursively call ourselves on the dominator children of BLOCK.
3198 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3199 do_regular_insertion and do_partial_insertion.
3203 static bool
3204 do_regular_insertion (basic_block block, basic_block dom)
3206 bool new_stuff = false;
3207 vec<pre_expr> exprs;
3208 pre_expr expr;
3209 vec<pre_expr> avail = vNULL;
3210 int i;
3212 exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3213 avail.safe_grow (EDGE_COUNT (block->preds));
3215 FOR_EACH_VEC_ELT (exprs, i, expr)
3217 if (expr->kind == NARY
3218 || expr->kind == REFERENCE)
3220 unsigned int val;
3221 bool by_some = false;
3222 bool cant_insert = false;
3223 bool all_same = true;
3224 pre_expr first_s = NULL;
3225 edge pred;
3226 basic_block bprime;
3227 pre_expr eprime = NULL;
3228 edge_iterator ei;
3229 pre_expr edoubleprime = NULL;
3230 bool do_insertion = false;
3232 val = get_expr_value_id (expr);
3233 if (bitmap_set_contains_value (PHI_GEN (block), val))
3234 continue;
3235 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3237 if (dump_file && (dump_flags & TDF_DETAILS))
3239 fprintf (dump_file, "Found fully redundant value: ");
3240 print_pre_expr (dump_file, expr);
3241 fprintf (dump_file, "\n");
3243 continue;
3246 FOR_EACH_EDGE (pred, ei, block->preds)
3248 unsigned int vprime;
3250 /* We should never run insertion for the exit block
3251 and so not come across fake pred edges. */
3252 gcc_assert (!(pred->flags & EDGE_FAKE));
3253 bprime = pred->src;
3254 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3255 bprime, block);
3257 /* eprime will generally only be NULL if the
3258 value of the expression, translated
3259 through the PHI for this predecessor, is
3260 undefined. If that is the case, we can't
3261 make the expression fully redundant,
3262 because its value is undefined along a
3263 predecessor path. We can thus break out
3264 early because it doesn't matter what the
3265 rest of the results are. */
3266 if (eprime == NULL)
3268 avail[pred->dest_idx] = NULL;
3269 cant_insert = true;
3270 break;
3273 eprime = fully_constant_expression (eprime);
3274 vprime = get_expr_value_id (eprime);
3275 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3276 vprime);
3277 if (edoubleprime == NULL)
3279 avail[pred->dest_idx] = eprime;
3280 all_same = false;
3282 else
3284 avail[pred->dest_idx] = edoubleprime;
3285 by_some = true;
3286 /* We want to perform insertions to remove a redundancy on
3287 a path in the CFG we want to optimize for speed. */
3288 if (optimize_edge_for_speed_p (pred))
3289 do_insertion = true;
3290 if (first_s == NULL)
3291 first_s = edoubleprime;
3292 else if (!pre_expr_d::equal (first_s, edoubleprime))
3293 all_same = false;
3296 /* If we can insert it, it's not the same value
3297 already existing along every predecessor, and
3298 it's defined by some predecessor, it is
3299 partially redundant. */
3300 if (!cant_insert && !all_same && by_some)
3302 if (!do_insertion)
3304 if (dump_file && (dump_flags & TDF_DETAILS))
3306 fprintf (dump_file, "Skipping partial redundancy for "
3307 "expression ");
3308 print_pre_expr (dump_file, expr);
3309 fprintf (dump_file, " (%04d), no redundancy on to be "
3310 "optimized for speed edge\n", val);
3313 else if (dbg_cnt (treepre_insert))
3315 if (dump_file && (dump_flags & TDF_DETAILS))
3317 fprintf (dump_file, "Found partial redundancy for "
3318 "expression ");
3319 print_pre_expr (dump_file, expr);
3320 fprintf (dump_file, " (%04d)\n",
3321 get_expr_value_id (expr));
3323 if (insert_into_preds_of_block (block,
3324 get_expression_id (expr),
3325 avail))
3326 new_stuff = true;
3329 /* If all edges produce the same value and that value is
3330 an invariant, then the PHI has the same value on all
3331 edges. Note this. */
3332 else if (!cant_insert && all_same)
3334 gcc_assert (edoubleprime->kind == CONSTANT
3335 || edoubleprime->kind == NAME);
3337 tree temp = make_temp_ssa_name (get_expr_type (expr),
3338 NULL, "pretmp");
3339 gimple assign = gimple_build_assign (temp,
3340 edoubleprime->kind == CONSTANT ? PRE_EXPR_CONSTANT (edoubleprime) : PRE_EXPR_NAME (edoubleprime));
3341 gimple_stmt_iterator gsi = gsi_after_labels (block);
3342 gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3344 gimple_set_plf (assign, NECESSARY, false);
3345 VN_INFO_GET (temp)->value_id = val;
3346 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3347 if (VN_INFO (temp)->valnum == NULL_TREE)
3348 VN_INFO (temp)->valnum = temp;
3349 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3350 pre_expr newe = get_or_alloc_expr_for_name (temp);
3351 add_to_value (val, newe);
3352 bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3353 bitmap_insert_into_set (NEW_SETS (block), newe);
3358 exprs.release ();
3359 return new_stuff;
3363 /* Perform insertion for partially anticipatable expressions. There
3364 is only one case we will perform insertion for these. This case is
3365 if the expression is partially anticipatable, and fully available.
3366 In this case, we know that putting it earlier will enable us to
3367 remove the later computation. */
3370 static bool
3371 do_partial_partial_insertion (basic_block block, basic_block dom)
3373 bool new_stuff = false;
3374 vec<pre_expr> exprs;
3375 pre_expr expr;
3376 auto_vec<pre_expr> avail;
3377 int i;
3379 exprs = sorted_array_from_bitmap_set (PA_IN (block));
3380 avail.safe_grow (EDGE_COUNT (block->preds));
3382 FOR_EACH_VEC_ELT (exprs, i, expr)
3384 if (expr->kind == NARY
3385 || expr->kind == REFERENCE)
3387 unsigned int val;
3388 bool by_all = true;
3389 bool cant_insert = false;
3390 edge pred;
3391 basic_block bprime;
3392 pre_expr eprime = NULL;
3393 edge_iterator ei;
3395 val = get_expr_value_id (expr);
3396 if (bitmap_set_contains_value (PHI_GEN (block), val))
3397 continue;
3398 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3399 continue;
3401 FOR_EACH_EDGE (pred, ei, block->preds)
3403 unsigned int vprime;
3404 pre_expr edoubleprime;
3406 /* We should never run insertion for the exit block
3407 and so not come across fake pred edges. */
3408 gcc_assert (!(pred->flags & EDGE_FAKE));
3409 bprime = pred->src;
3410 eprime = phi_translate (expr, ANTIC_IN (block),
3411 PA_IN (block),
3412 bprime, block);
3414 /* eprime will generally only be NULL if the
3415 value of the expression, translated
3416 through the PHI for this predecessor, is
3417 undefined. If that is the case, we can't
3418 make the expression fully redundant,
3419 because its value is undefined along a
3420 predecessor path. We can thus break out
3421 early because it doesn't matter what the
3422 rest of the results are. */
3423 if (eprime == NULL)
3425 avail[pred->dest_idx] = NULL;
3426 cant_insert = true;
3427 break;
3430 eprime = fully_constant_expression (eprime);
3431 vprime = get_expr_value_id (eprime);
3432 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3433 avail[pred->dest_idx] = edoubleprime;
3434 if (edoubleprime == NULL)
3436 by_all = false;
3437 break;
3441 /* If we can insert it, it's not the same value
3442 already existing along every predecessor, and
3443 it's defined by some predecessor, it is
3444 partially redundant. */
3445 if (!cant_insert && by_all)
3447 edge succ;
3448 bool do_insertion = false;
3450 /* Insert only if we can remove a later expression on a path
3451 that we want to optimize for speed.
3452 The phi node that we will be inserting in BLOCK is not free,
3453 and inserting it for the sake of !optimize_for_speed successor
3454 may cause regressions on the speed path. */
3455 FOR_EACH_EDGE (succ, ei, block->succs)
3457 if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3458 || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3460 if (optimize_edge_for_speed_p (succ))
3461 do_insertion = true;
3465 if (!do_insertion)
3467 if (dump_file && (dump_flags & TDF_DETAILS))
3469 fprintf (dump_file, "Skipping partial partial redundancy "
3470 "for expression ");
3471 print_pre_expr (dump_file, expr);
3472 fprintf (dump_file, " (%04d), not (partially) anticipated "
3473 "on any to be optimized for speed edges\n", val);
3476 else if (dbg_cnt (treepre_insert))
3478 pre_stats.pa_insert++;
3479 if (dump_file && (dump_flags & TDF_DETAILS))
3481 fprintf (dump_file, "Found partial partial redundancy "
3482 "for expression ");
3483 print_pre_expr (dump_file, expr);
3484 fprintf (dump_file, " (%04d)\n",
3485 get_expr_value_id (expr));
3487 if (insert_into_preds_of_block (block,
3488 get_expression_id (expr),
3489 avail))
3490 new_stuff = true;
3496 exprs.release ();
3497 return new_stuff;
3500 static bool
3501 insert_aux (basic_block block)
3503 basic_block son;
3504 bool new_stuff = false;
3506 if (block)
3508 basic_block dom;
3509 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3510 if (dom)
3512 unsigned i;
3513 bitmap_iterator bi;
3514 bitmap_set_t newset = NEW_SETS (dom);
3515 if (newset)
3517 /* Note that we need to value_replace both NEW_SETS, and
3518 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3519 represented by some non-simple expression here that we want
3520 to replace it with. */
3521 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3523 pre_expr expr = expression_for_id (i);
3524 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3525 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3528 if (!single_pred_p (block))
3530 new_stuff |= do_regular_insertion (block, dom);
3531 if (do_partial_partial)
3532 new_stuff |= do_partial_partial_insertion (block, dom);
3536 for (son = first_dom_son (CDI_DOMINATORS, block);
3537 son;
3538 son = next_dom_son (CDI_DOMINATORS, son))
3540 new_stuff |= insert_aux (son);
3543 return new_stuff;
3546 /* Perform insertion of partially redundant values. */
3548 static void
3549 insert (void)
3551 bool new_stuff = true;
3552 basic_block bb;
3553 int num_iterations = 0;
3555 FOR_ALL_BB_FN (bb, cfun)
3556 NEW_SETS (bb) = bitmap_set_new ();
3558 while (new_stuff)
3560 num_iterations++;
3561 if (dump_file && dump_flags & TDF_DETAILS)
3562 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3563 new_stuff = insert_aux (ENTRY_BLOCK_PTR_FOR_FN (cfun));
3565 /* Clear the NEW sets before the next iteration. We have already
3566 fully propagated its contents. */
3567 if (new_stuff)
3568 FOR_ALL_BB_FN (bb, cfun)
3569 bitmap_set_free (NEW_SETS (bb));
3571 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3575 /* Compute the AVAIL set for all basic blocks.
3577 This function performs value numbering of the statements in each basic
3578 block. The AVAIL sets are built from information we glean while doing
3579 this value numbering, since the AVAIL sets contain only one entry per
3580 value.
3582 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3583 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3585 static void
3586 compute_avail (void)
3589 basic_block block, son;
3590 basic_block *worklist;
3591 size_t sp = 0;
3592 unsigned i;
3594 /* We pretend that default definitions are defined in the entry block.
3595 This includes function arguments and the static chain decl. */
3596 for (i = 1; i < num_ssa_names; ++i)
3598 tree name = ssa_name (i);
3599 pre_expr e;
3600 if (!name
3601 || !SSA_NAME_IS_DEFAULT_DEF (name)
3602 || has_zero_uses (name)
3603 || virtual_operand_p (name))
3604 continue;
3606 e = get_or_alloc_expr_for_name (name);
3607 add_to_value (get_expr_value_id (e), e);
3608 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)), e);
3609 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3613 if (dump_file && (dump_flags & TDF_DETAILS))
3615 print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3616 "tmp_gen", ENTRY_BLOCK);
3617 print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3618 "avail_out", ENTRY_BLOCK);
3621 /* Allocate the worklist. */
3622 worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
3624 /* Seed the algorithm by putting the dominator children of the entry
3625 block on the worklist. */
3626 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR_FOR_FN (cfun));
3627 son;
3628 son = next_dom_son (CDI_DOMINATORS, son))
3629 worklist[sp++] = son;
3631 BB_LIVE_VOP_ON_EXIT (ENTRY_BLOCK_PTR_FOR_FN (cfun))
3632 = ssa_default_def (cfun, gimple_vop (cfun));
3634 /* Loop until the worklist is empty. */
3635 while (sp)
3637 gimple_stmt_iterator gsi;
3638 gimple stmt;
3639 basic_block dom;
3641 /* Pick a block from the worklist. */
3642 block = worklist[--sp];
3644 /* Initially, the set of available values in BLOCK is that of
3645 its immediate dominator. */
3646 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3647 if (dom)
3649 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3650 BB_LIVE_VOP_ON_EXIT (block) = BB_LIVE_VOP_ON_EXIT (dom);
3653 /* Generate values for PHI nodes. */
3654 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3656 tree result = gimple_phi_result (gsi_stmt (gsi));
3658 /* We have no need for virtual phis, as they don't represent
3659 actual computations. */
3660 if (virtual_operand_p (result))
3662 BB_LIVE_VOP_ON_EXIT (block) = result;
3663 continue;
3666 pre_expr e = get_or_alloc_expr_for_name (result);
3667 add_to_value (get_expr_value_id (e), e);
3668 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3669 bitmap_insert_into_set (PHI_GEN (block), e);
3672 BB_MAY_NOTRETURN (block) = 0;
3674 /* Now compute value numbers and populate value sets with all
3675 the expressions computed in BLOCK. */
3676 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3678 ssa_op_iter iter;
3679 tree op;
3681 stmt = gsi_stmt (gsi);
3683 /* Cache whether the basic-block has any non-visible side-effect
3684 or control flow.
3685 If this isn't a call or it is the last stmt in the
3686 basic-block then the CFG represents things correctly. */
3687 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3689 /* Non-looping const functions always return normally.
3690 Otherwise the call might not return or have side-effects
3691 that forbids hoisting possibly trapping expressions
3692 before it. */
3693 int flags = gimple_call_flags (stmt);
3694 if (!(flags & ECF_CONST)
3695 || (flags & ECF_LOOPING_CONST_OR_PURE))
3696 BB_MAY_NOTRETURN (block) = 1;
3699 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3701 pre_expr e = get_or_alloc_expr_for_name (op);
3703 add_to_value (get_expr_value_id (e), e);
3704 bitmap_insert_into_set (TMP_GEN (block), e);
3705 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3708 if (gimple_vdef (stmt))
3709 BB_LIVE_VOP_ON_EXIT (block) = gimple_vdef (stmt);
3711 if (gimple_has_side_effects (stmt)
3712 || stmt_could_throw_p (stmt)
3713 || is_gimple_debug (stmt))
3714 continue;
3716 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3718 if (ssa_undefined_value_p (op))
3719 continue;
3720 pre_expr e = get_or_alloc_expr_for_name (op);
3721 bitmap_value_insert_into_set (EXP_GEN (block), e);
3724 switch (gimple_code (stmt))
3726 case GIMPLE_RETURN:
3727 continue;
3729 case GIMPLE_CALL:
3731 vn_reference_t ref;
3732 vn_reference_s ref1;
3733 pre_expr result = NULL;
3735 /* We can value number only calls to real functions. */
3736 if (gimple_call_internal_p (stmt))
3737 continue;
3739 vn_reference_lookup_call (stmt, &ref, &ref1);
3740 if (!ref)
3741 continue;
3743 /* If the value of the call is not invalidated in
3744 this block until it is computed, add the expression
3745 to EXP_GEN. */
3746 if (!gimple_vuse (stmt)
3747 || gimple_code
3748 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
3749 || gimple_bb (SSA_NAME_DEF_STMT
3750 (gimple_vuse (stmt))) != block)
3752 result = (pre_expr) pool_alloc (pre_expr_pool);
3753 result->kind = REFERENCE;
3754 result->id = 0;
3755 PRE_EXPR_REFERENCE (result) = ref;
3757 get_or_alloc_expression_id (result);
3758 add_to_value (get_expr_value_id (result), result);
3759 bitmap_value_insert_into_set (EXP_GEN (block), result);
3761 continue;
3764 case GIMPLE_ASSIGN:
3766 pre_expr result = NULL;
3767 switch (vn_get_stmt_kind (stmt))
3769 case VN_NARY:
3771 enum tree_code code = gimple_assign_rhs_code (stmt);
3772 vn_nary_op_t nary;
3774 /* COND_EXPR and VEC_COND_EXPR are awkward in
3775 that they contain an embedded complex expression.
3776 Don't even try to shove those through PRE. */
3777 if (code == COND_EXPR
3778 || code == VEC_COND_EXPR)
3779 continue;
3781 vn_nary_op_lookup_stmt (stmt, &nary);
3782 if (!nary)
3783 continue;
3785 /* If the NARY traps and there was a preceding
3786 point in the block that might not return avoid
3787 adding the nary to EXP_GEN. */
3788 if (BB_MAY_NOTRETURN (block)
3789 && vn_nary_may_trap (nary))
3790 continue;
3792 result = (pre_expr) pool_alloc (pre_expr_pool);
3793 result->kind = NARY;
3794 result->id = 0;
3795 PRE_EXPR_NARY (result) = nary;
3796 break;
3799 case VN_REFERENCE:
3801 vn_reference_t ref;
3802 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3803 gimple_vuse (stmt),
3804 VN_WALK, &ref);
3805 if (!ref)
3806 continue;
3808 /* If the value of the reference is not invalidated in
3809 this block until it is computed, add the expression
3810 to EXP_GEN. */
3811 if (gimple_vuse (stmt))
3813 gimple def_stmt;
3814 bool ok = true;
3815 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
3816 while (!gimple_nop_p (def_stmt)
3817 && gimple_code (def_stmt) != GIMPLE_PHI
3818 && gimple_bb (def_stmt) == block)
3820 if (stmt_may_clobber_ref_p
3821 (def_stmt, gimple_assign_rhs1 (stmt)))
3823 ok = false;
3824 break;
3826 def_stmt
3827 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
3829 if (!ok)
3830 continue;
3833 result = (pre_expr) pool_alloc (pre_expr_pool);
3834 result->kind = REFERENCE;
3835 result->id = 0;
3836 PRE_EXPR_REFERENCE (result) = ref;
3837 break;
3840 default:
3841 continue;
3844 get_or_alloc_expression_id (result);
3845 add_to_value (get_expr_value_id (result), result);
3846 bitmap_value_insert_into_set (EXP_GEN (block), result);
3847 continue;
3849 default:
3850 break;
3854 if (dump_file && (dump_flags & TDF_DETAILS))
3856 print_bitmap_set (dump_file, EXP_GEN (block),
3857 "exp_gen", block->index);
3858 print_bitmap_set (dump_file, PHI_GEN (block),
3859 "phi_gen", block->index);
3860 print_bitmap_set (dump_file, TMP_GEN (block),
3861 "tmp_gen", block->index);
3862 print_bitmap_set (dump_file, AVAIL_OUT (block),
3863 "avail_out", block->index);
3866 /* Put the dominator children of BLOCK on the worklist of blocks
3867 to compute available sets for. */
3868 for (son = first_dom_son (CDI_DOMINATORS, block);
3869 son;
3870 son = next_dom_son (CDI_DOMINATORS, son))
3871 worklist[sp++] = son;
3874 free (worklist);
3878 /* Local state for the eliminate domwalk. */
3879 static vec<gimple> el_to_remove;
3880 static unsigned int el_todo;
3881 static vec<tree> el_avail;
3882 static vec<tree> el_avail_stack;
3884 /* Return a leader for OP that is available at the current point of the
3885 eliminate domwalk. */
3887 static tree
3888 eliminate_avail (tree op)
3890 tree valnum = VN_INFO (op)->valnum;
3891 if (TREE_CODE (valnum) == SSA_NAME)
3893 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
3894 return valnum;
3895 if (el_avail.length () > SSA_NAME_VERSION (valnum))
3896 return el_avail[SSA_NAME_VERSION (valnum)];
3898 else if (is_gimple_min_invariant (valnum))
3899 return valnum;
3900 return NULL_TREE;
3903 /* At the current point of the eliminate domwalk make OP available. */
3905 static void
3906 eliminate_push_avail (tree op)
3908 tree valnum = VN_INFO (op)->valnum;
3909 if (TREE_CODE (valnum) == SSA_NAME)
3911 if (el_avail.length () <= SSA_NAME_VERSION (valnum))
3912 el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1);
3913 el_avail[SSA_NAME_VERSION (valnum)] = op;
3914 el_avail_stack.safe_push (op);
3918 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
3919 the leader for the expression if insertion was successful. */
3921 static tree
3922 eliminate_insert (gimple_stmt_iterator *gsi, tree val)
3924 tree expr = vn_get_expr_for (val);
3925 if (!CONVERT_EXPR_P (expr)
3926 && TREE_CODE (expr) != VIEW_CONVERT_EXPR)
3927 return NULL_TREE;
3929 tree op = TREE_OPERAND (expr, 0);
3930 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op;
3931 if (!leader)
3932 return NULL_TREE;
3934 tree res = make_temp_ssa_name (TREE_TYPE (val), NULL, "pretmp");
3935 gimple tem = gimple_build_assign (res,
3936 fold_build1 (TREE_CODE (expr),
3937 TREE_TYPE (expr), leader));
3938 gsi_insert_before (gsi, tem, GSI_SAME_STMT);
3939 VN_INFO_GET (res)->valnum = val;
3941 if (TREE_CODE (leader) == SSA_NAME)
3942 gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true);
3944 pre_stats.insertions++;
3945 if (dump_file && (dump_flags & TDF_DETAILS))
3947 fprintf (dump_file, "Inserted ");
3948 print_gimple_stmt (dump_file, tem, 0, 0);
3951 return res;
3954 class eliminate_dom_walker : public dom_walker
3956 public:
3957 eliminate_dom_walker (cdi_direction direction, bool do_pre_)
3958 : dom_walker (direction), do_pre (do_pre_) {}
3960 virtual void before_dom_children (basic_block);
3961 virtual void after_dom_children (basic_block);
3963 bool do_pre;
3966 /* Perform elimination for the basic-block B during the domwalk. */
3968 void
3969 eliminate_dom_walker::before_dom_children (basic_block b)
3971 gimple_stmt_iterator gsi;
3972 gimple stmt;
3974 /* Mark new bb. */
3975 el_avail_stack.safe_push (NULL_TREE);
3977 /* ??? If we do nothing for unreachable blocks then this will confuse
3978 tailmerging. Eventually we can reduce its reliance on SCCVN now
3979 that we fully copy/constant-propagate (most) things. */
3981 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
3983 gimple phi = gsi_stmt (gsi);
3984 tree res = PHI_RESULT (phi);
3986 if (virtual_operand_p (res))
3988 gsi_next (&gsi);
3989 continue;
3992 tree sprime = eliminate_avail (res);
3993 if (sprime
3994 && sprime != res)
3996 if (dump_file && (dump_flags & TDF_DETAILS))
3998 fprintf (dump_file, "Replaced redundant PHI node defining ");
3999 print_generic_expr (dump_file, res, 0);
4000 fprintf (dump_file, " with ");
4001 print_generic_expr (dump_file, sprime, 0);
4002 fprintf (dump_file, "\n");
4005 /* If we inserted this PHI node ourself, it's not an elimination. */
4006 if (inserted_exprs
4007 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4008 pre_stats.phis--;
4009 else
4010 pre_stats.eliminations++;
4012 /* If we will propagate into all uses don't bother to do
4013 anything. */
4014 if (may_propagate_copy (res, sprime))
4016 /* Mark the PHI for removal. */
4017 el_to_remove.safe_push (phi);
4018 gsi_next (&gsi);
4019 continue;
4022 remove_phi_node (&gsi, false);
4024 if (inserted_exprs
4025 && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4026 && TREE_CODE (sprime) == SSA_NAME)
4027 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4029 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4030 sprime = fold_convert (TREE_TYPE (res), sprime);
4031 gimple stmt = gimple_build_assign (res, sprime);
4032 /* ??? It cannot yet be necessary (DOM walk). */
4033 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4035 gimple_stmt_iterator gsi2 = gsi_after_labels (b);
4036 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4037 continue;
4040 eliminate_push_avail (res);
4041 gsi_next (&gsi);
4044 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4046 tree sprime = NULL_TREE;
4047 stmt = gsi_stmt (gsi);
4048 tree lhs = gimple_get_lhs (stmt);
4049 if (lhs && TREE_CODE (lhs) == SSA_NAME
4050 && !gimple_has_volatile_ops (stmt)
4051 /* See PR43491. Do not replace a global register variable when
4052 it is a the RHS of an assignment. Do replace local register
4053 variables since gcc does not guarantee a local variable will
4054 be allocated in register.
4055 ??? The fix isn't effective here. This should instead
4056 be ensured by not value-numbering them the same but treating
4057 them like volatiles? */
4058 && !(gimple_assign_single_p (stmt)
4059 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL
4060 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt))
4061 && is_global_var (gimple_assign_rhs1 (stmt)))))
4063 sprime = eliminate_avail (lhs);
4064 if (!sprime)
4066 /* If there is no existing usable leader but SCCVN thinks
4067 it has an expression it wants to use as replacement,
4068 insert that. */
4069 tree val = VN_INFO (lhs)->valnum;
4070 if (val != VN_TOP
4071 && TREE_CODE (val) == SSA_NAME
4072 && VN_INFO (val)->needs_insertion
4073 && VN_INFO (val)->expr != NULL_TREE
4074 && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE)
4075 eliminate_push_avail (sprime);
4078 /* If this now constitutes a copy duplicate points-to
4079 and range info appropriately. This is especially
4080 important for inserted code. See tree-ssa-copy.c
4081 for similar code. */
4082 if (sprime
4083 && TREE_CODE (sprime) == SSA_NAME)
4085 basic_block sprime_b = gimple_bb (SSA_NAME_DEF_STMT (sprime));
4086 if (POINTER_TYPE_P (TREE_TYPE (lhs))
4087 && SSA_NAME_PTR_INFO (lhs)
4088 && !SSA_NAME_PTR_INFO (sprime))
4090 duplicate_ssa_name_ptr_info (sprime,
4091 SSA_NAME_PTR_INFO (lhs));
4092 if (b != sprime_b)
4093 mark_ptr_info_alignment_unknown
4094 (SSA_NAME_PTR_INFO (sprime));
4096 else if (!POINTER_TYPE_P (TREE_TYPE (lhs))
4097 && SSA_NAME_RANGE_INFO (lhs)
4098 && !SSA_NAME_RANGE_INFO (sprime)
4099 && b == sprime_b)
4100 duplicate_ssa_name_range_info (sprime,
4101 SSA_NAME_RANGE_TYPE (lhs),
4102 SSA_NAME_RANGE_INFO (lhs));
4105 /* Inhibit the use of an inserted PHI on a loop header when
4106 the address of the memory reference is a simple induction
4107 variable. In other cases the vectorizer won't do anything
4108 anyway (either it's loop invariant or a complicated
4109 expression). */
4110 if (sprime
4111 && TREE_CODE (sprime) == SSA_NAME
4112 && do_pre
4113 && flag_tree_loop_vectorize
4114 && loop_outer (b->loop_father)
4115 && has_zero_uses (sprime)
4116 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))
4117 && gimple_assign_load_p (stmt))
4119 gimple def_stmt = SSA_NAME_DEF_STMT (sprime);
4120 basic_block def_bb = gimple_bb (def_stmt);
4121 if (gimple_code (def_stmt) == GIMPLE_PHI
4122 && b->loop_father->header == def_bb)
4124 ssa_op_iter iter;
4125 tree op;
4126 bool found = false;
4127 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4129 affine_iv iv;
4130 def_bb = gimple_bb (SSA_NAME_DEF_STMT (op));
4131 if (def_bb
4132 && flow_bb_inside_loop_p (b->loop_father, def_bb)
4133 && simple_iv (b->loop_father,
4134 b->loop_father, op, &iv, true))
4136 found = true;
4137 break;
4140 if (found)
4142 if (dump_file && (dump_flags & TDF_DETAILS))
4144 fprintf (dump_file, "Not replacing ");
4145 print_gimple_expr (dump_file, stmt, 0, 0);
4146 fprintf (dump_file, " with ");
4147 print_generic_expr (dump_file, sprime, 0);
4148 fprintf (dump_file, " which would add a loop"
4149 " carried dependence to loop %d\n",
4150 b->loop_father->num);
4152 /* Don't keep sprime available. */
4153 sprime = NULL_TREE;
4158 if (sprime)
4160 /* If we can propagate the value computed for LHS into
4161 all uses don't bother doing anything with this stmt. */
4162 if (may_propagate_copy (lhs, sprime))
4164 /* Mark it for removal. */
4165 el_to_remove.safe_push (stmt);
4167 /* ??? Don't count copy/constant propagations. */
4168 if (gimple_assign_single_p (stmt)
4169 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4170 || gimple_assign_rhs1 (stmt) == sprime))
4171 continue;
4173 if (dump_file && (dump_flags & TDF_DETAILS))
4175 fprintf (dump_file, "Replaced ");
4176 print_gimple_expr (dump_file, stmt, 0, 0);
4177 fprintf (dump_file, " with ");
4178 print_generic_expr (dump_file, sprime, 0);
4179 fprintf (dump_file, " in all uses of ");
4180 print_gimple_stmt (dump_file, stmt, 0, 0);
4183 pre_stats.eliminations++;
4184 continue;
4187 /* If this is an assignment from our leader (which
4188 happens in the case the value-number is a constant)
4189 then there is nothing to do. */
4190 if (gimple_assign_single_p (stmt)
4191 && sprime == gimple_assign_rhs1 (stmt))
4192 continue;
4194 /* Else replace its RHS. */
4195 bool can_make_abnormal_goto
4196 = is_gimple_call (stmt)
4197 && stmt_can_make_abnormal_goto (stmt);
4199 if (dump_file && (dump_flags & TDF_DETAILS))
4201 fprintf (dump_file, "Replaced ");
4202 print_gimple_expr (dump_file, stmt, 0, 0);
4203 fprintf (dump_file, " with ");
4204 print_generic_expr (dump_file, sprime, 0);
4205 fprintf (dump_file, " in ");
4206 print_gimple_stmt (dump_file, stmt, 0, 0);
4209 if (TREE_CODE (sprime) == SSA_NAME)
4210 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4211 NECESSARY, true);
4213 pre_stats.eliminations++;
4214 gimple orig_stmt = stmt;
4215 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4216 TREE_TYPE (sprime)))
4217 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4218 tree vdef = gimple_vdef (stmt);
4219 tree vuse = gimple_vuse (stmt);
4220 propagate_tree_value_into_stmt (&gsi, sprime);
4221 stmt = gsi_stmt (gsi);
4222 update_stmt (stmt);
4223 if (vdef != gimple_vdef (stmt))
4224 VN_INFO (vdef)->valnum = vuse;
4226 /* If we removed EH side-effects from the statement, clean
4227 its EH information. */
4228 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4230 bitmap_set_bit (need_eh_cleanup,
4231 gimple_bb (stmt)->index);
4232 if (dump_file && (dump_flags & TDF_DETAILS))
4233 fprintf (dump_file, " Removed EH side-effects.\n");
4236 /* Likewise for AB side-effects. */
4237 if (can_make_abnormal_goto
4238 && !stmt_can_make_abnormal_goto (stmt))
4240 bitmap_set_bit (need_ab_cleanup,
4241 gimple_bb (stmt)->index);
4242 if (dump_file && (dump_flags & TDF_DETAILS))
4243 fprintf (dump_file, " Removed AB side-effects.\n");
4246 continue;
4250 /* If the statement is a scalar store, see if the expression
4251 has the same value number as its rhs. If so, the store is
4252 dead. */
4253 if (gimple_assign_single_p (stmt)
4254 && !gimple_has_volatile_ops (stmt)
4255 && !is_gimple_reg (gimple_assign_lhs (stmt))
4256 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4257 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4259 tree val;
4260 tree rhs = gimple_assign_rhs1 (stmt);
4261 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4262 gimple_vuse (stmt), VN_WALK, NULL);
4263 if (TREE_CODE (rhs) == SSA_NAME)
4264 rhs = VN_INFO (rhs)->valnum;
4265 if (val
4266 && operand_equal_p (val, rhs, 0))
4268 if (dump_file && (dump_flags & TDF_DETAILS))
4270 fprintf (dump_file, "Deleted redundant store ");
4271 print_gimple_stmt (dump_file, stmt, 0, 0);
4274 /* Queue stmt for removal. */
4275 el_to_remove.safe_push (stmt);
4276 continue;
4280 bool can_make_abnormal_goto = stmt_can_make_abnormal_goto (stmt);
4281 bool was_noreturn = (is_gimple_call (stmt)
4282 && gimple_call_noreturn_p (stmt));
4283 tree vdef = gimple_vdef (stmt);
4284 tree vuse = gimple_vuse (stmt);
4286 /* If we didn't replace the whole stmt (or propagate the result
4287 into all uses), replace all uses on this stmt with their
4288 leaders. */
4289 use_operand_p use_p;
4290 ssa_op_iter iter;
4291 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
4293 tree use = USE_FROM_PTR (use_p);
4294 /* ??? The call code above leaves stmt operands un-updated. */
4295 if (TREE_CODE (use) != SSA_NAME)
4296 continue;
4297 tree sprime = eliminate_avail (use);
4298 if (sprime && sprime != use
4299 && may_propagate_copy (use, sprime)
4300 /* We substitute into debug stmts to avoid excessive
4301 debug temporaries created by removed stmts, but we need
4302 to avoid doing so for inserted sprimes as we never want
4303 to create debug temporaries for them. */
4304 && (!inserted_exprs
4305 || TREE_CODE (sprime) != SSA_NAME
4306 || !is_gimple_debug (stmt)
4307 || !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))))
4309 propagate_value (use_p, sprime);
4310 gimple_set_modified (stmt, true);
4311 if (TREE_CODE (sprime) == SSA_NAME
4312 && !is_gimple_debug (stmt))
4313 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4314 NECESSARY, true);
4318 /* Visit indirect calls and turn them into direct calls if
4319 possible using the devirtualization machinery. */
4320 if (is_gimple_call (stmt))
4322 tree fn = gimple_call_fn (stmt);
4323 if (fn
4324 && flag_devirtualize
4325 && virtual_method_call_p (fn))
4327 tree otr_type = obj_type_ref_class (fn);
4328 tree instance;
4329 ipa_polymorphic_call_context context (current_function_decl, fn, stmt, &instance);
4330 bool final;
4332 context.get_dynamic_type (instance, OBJ_TYPE_REF_OBJECT (fn), otr_type, stmt);
4334 vec <cgraph_node *>targets
4335 = possible_polymorphic_call_targets (obj_type_ref_class (fn),
4336 tree_to_uhwi
4337 (OBJ_TYPE_REF_TOKEN (fn)),
4338 context,
4339 &final);
4340 if (dump_enabled_p ())
4341 dump_possible_polymorphic_call_targets (dump_file,
4342 obj_type_ref_class (fn),
4343 tree_to_uhwi
4344 (OBJ_TYPE_REF_TOKEN (fn)),
4345 context);
4346 if (final && targets.length () <= 1 && dbg_cnt (devirt))
4348 tree fn;
4349 if (targets.length () == 1)
4350 fn = targets[0]->decl;
4351 else
4352 fn = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
4353 if (dump_enabled_p ())
4355 location_t loc = gimple_location_safe (stmt);
4356 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc,
4357 "converting indirect call to "
4358 "function %s\n",
4359 cgraph_node::get (fn)->name ());
4361 gimple_call_set_fndecl (stmt, fn);
4362 gimple_set_modified (stmt, true);
4367 if (gimple_modified_p (stmt))
4369 /* If a formerly non-invariant ADDR_EXPR is turned into an
4370 invariant one it was on a separate stmt. */
4371 if (gimple_assign_single_p (stmt)
4372 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR)
4373 recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
4374 gimple old_stmt = stmt;
4375 if (is_gimple_call (stmt))
4377 /* ??? Only fold calls inplace for now, this may create new
4378 SSA names which in turn will confuse free_scc_vn SSA name
4379 release code. */
4380 fold_stmt_inplace (&gsi);
4381 /* When changing a call into a noreturn call, cfg cleanup
4382 is needed to fix up the noreturn call. */
4383 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4384 el_todo |= TODO_cleanup_cfg;
4386 else
4388 fold_stmt (&gsi);
4389 stmt = gsi_stmt (gsi);
4390 if ((gimple_code (stmt) == GIMPLE_COND
4391 && (gimple_cond_true_p (stmt)
4392 || gimple_cond_false_p (stmt)))
4393 || (gimple_code (stmt) == GIMPLE_SWITCH
4394 && TREE_CODE (gimple_switch_index (stmt)) == INTEGER_CST))
4395 el_todo |= TODO_cleanup_cfg;
4397 /* If we removed EH side-effects from the statement, clean
4398 its EH information. */
4399 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
4401 bitmap_set_bit (need_eh_cleanup,
4402 gimple_bb (stmt)->index);
4403 if (dump_file && (dump_flags & TDF_DETAILS))
4404 fprintf (dump_file, " Removed EH side-effects.\n");
4406 /* Likewise for AB side-effects. */
4407 if (can_make_abnormal_goto
4408 && !stmt_can_make_abnormal_goto (stmt))
4410 bitmap_set_bit (need_ab_cleanup,
4411 gimple_bb (stmt)->index);
4412 if (dump_file && (dump_flags & TDF_DETAILS))
4413 fprintf (dump_file, " Removed AB side-effects.\n");
4415 update_stmt (stmt);
4416 if (vdef != gimple_vdef (stmt))
4417 VN_INFO (vdef)->valnum = vuse;
4420 /* Make new values available - for fully redundant LHS we
4421 continue with the next stmt above and skip this. */
4422 def_operand_p defp;
4423 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_DEF)
4424 eliminate_push_avail (DEF_FROM_PTR (defp));
4427 /* Replace destination PHI arguments. */
4428 edge_iterator ei;
4429 edge e;
4430 FOR_EACH_EDGE (e, ei, b->succs)
4432 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
4434 gimple phi = gsi_stmt (gsi);
4435 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
4436 tree arg = USE_FROM_PTR (use_p);
4437 if (TREE_CODE (arg) != SSA_NAME
4438 || virtual_operand_p (arg))
4439 continue;
4440 tree sprime = eliminate_avail (arg);
4441 if (sprime && may_propagate_copy (arg, sprime))
4443 propagate_value (use_p, sprime);
4444 if (TREE_CODE (sprime) == SSA_NAME)
4445 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4451 /* Make no longer available leaders no longer available. */
4453 void
4454 eliminate_dom_walker::after_dom_children (basic_block)
4456 tree entry;
4457 while ((entry = el_avail_stack.pop ()) != NULL_TREE)
4458 el_avail[SSA_NAME_VERSION (VN_INFO (entry)->valnum)] = NULL_TREE;
4461 /* Eliminate fully redundant computations. */
4463 static unsigned int
4464 eliminate (bool do_pre)
4466 gimple_stmt_iterator gsi;
4467 gimple stmt;
4469 need_eh_cleanup = BITMAP_ALLOC (NULL);
4470 need_ab_cleanup = BITMAP_ALLOC (NULL);
4472 el_to_remove.create (0);
4473 el_todo = 0;
4474 el_avail.create (num_ssa_names);
4475 el_avail_stack.create (0);
4477 eliminate_dom_walker (CDI_DOMINATORS,
4478 do_pre).walk (cfun->cfg->x_entry_block_ptr);
4480 el_avail.release ();
4481 el_avail_stack.release ();
4483 /* We cannot remove stmts during BB walk, especially not release SSA
4484 names there as this confuses the VN machinery. The stmts ending
4485 up in el_to_remove are either stores or simple copies.
4486 Remove stmts in reverse order to make debug stmt creation possible. */
4487 while (!el_to_remove.is_empty ())
4489 stmt = el_to_remove.pop ();
4491 if (dump_file && (dump_flags & TDF_DETAILS))
4493 fprintf (dump_file, "Removing dead stmt ");
4494 print_gimple_stmt (dump_file, stmt, 0, 0);
4497 tree lhs;
4498 if (gimple_code (stmt) == GIMPLE_PHI)
4499 lhs = gimple_phi_result (stmt);
4500 else
4501 lhs = gimple_get_lhs (stmt);
4503 if (inserted_exprs
4504 && TREE_CODE (lhs) == SSA_NAME)
4505 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4507 gsi = gsi_for_stmt (stmt);
4508 if (gimple_code (stmt) == GIMPLE_PHI)
4509 remove_phi_node (&gsi, true);
4510 else
4512 basic_block bb = gimple_bb (stmt);
4513 unlink_stmt_vdef (stmt);
4514 if (gsi_remove (&gsi, true))
4515 bitmap_set_bit (need_eh_cleanup, bb->index);
4516 release_defs (stmt);
4519 /* Removing a stmt may expose a forwarder block. */
4520 el_todo |= TODO_cleanup_cfg;
4522 el_to_remove.release ();
4524 return el_todo;
4527 /* Perform CFG cleanups made necessary by elimination. */
4529 static unsigned
4530 fini_eliminate (void)
4532 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4533 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4535 if (do_eh_cleanup)
4536 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4538 if (do_ab_cleanup)
4539 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4541 BITMAP_FREE (need_eh_cleanup);
4542 BITMAP_FREE (need_ab_cleanup);
4544 if (do_eh_cleanup || do_ab_cleanup)
4545 return TODO_cleanup_cfg;
4546 return 0;
4549 /* Borrow a bit of tree-ssa-dce.c for the moment.
4550 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4551 this may be a bit faster, and we may want critical edges kept split. */
4553 /* If OP's defining statement has not already been determined to be necessary,
4554 mark that statement necessary. Return the stmt, if it is newly
4555 necessary. */
4557 static inline gimple
4558 mark_operand_necessary (tree op)
4560 gimple stmt;
4562 gcc_assert (op);
4564 if (TREE_CODE (op) != SSA_NAME)
4565 return NULL;
4567 stmt = SSA_NAME_DEF_STMT (op);
4568 gcc_assert (stmt);
4570 if (gimple_plf (stmt, NECESSARY)
4571 || gimple_nop_p (stmt))
4572 return NULL;
4574 gimple_set_plf (stmt, NECESSARY, true);
4575 return stmt;
4578 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4579 to insert PHI nodes sometimes, and because value numbering of casts isn't
4580 perfect, we sometimes end up inserting dead code. This simple DCE-like
4581 pass removes any insertions we made that weren't actually used. */
4583 static void
4584 remove_dead_inserted_code (void)
4586 bitmap worklist;
4587 unsigned i;
4588 bitmap_iterator bi;
4589 gimple t;
4591 worklist = BITMAP_ALLOC (NULL);
4592 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4594 t = SSA_NAME_DEF_STMT (ssa_name (i));
4595 if (gimple_plf (t, NECESSARY))
4596 bitmap_set_bit (worklist, i);
4598 while (!bitmap_empty_p (worklist))
4600 i = bitmap_first_set_bit (worklist);
4601 bitmap_clear_bit (worklist, i);
4602 t = SSA_NAME_DEF_STMT (ssa_name (i));
4604 /* PHI nodes are somewhat special in that each PHI alternative has
4605 data and control dependencies. All the statements feeding the
4606 PHI node's arguments are always necessary. */
4607 if (gimple_code (t) == GIMPLE_PHI)
4609 unsigned k;
4611 for (k = 0; k < gimple_phi_num_args (t); k++)
4613 tree arg = PHI_ARG_DEF (t, k);
4614 if (TREE_CODE (arg) == SSA_NAME)
4616 gimple n = mark_operand_necessary (arg);
4617 if (n)
4618 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4622 else
4624 /* Propagate through the operands. Examine all the USE, VUSE and
4625 VDEF operands in this statement. Mark all the statements
4626 which feed this statement's uses as necessary. */
4627 ssa_op_iter iter;
4628 tree use;
4630 /* The operands of VDEF expressions are also needed as they
4631 represent potential definitions that may reach this
4632 statement (VDEF operands allow us to follow def-def
4633 links). */
4635 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4637 gimple n = mark_operand_necessary (use);
4638 if (n)
4639 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4644 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4646 t = SSA_NAME_DEF_STMT (ssa_name (i));
4647 if (!gimple_plf (t, NECESSARY))
4649 gimple_stmt_iterator gsi;
4651 if (dump_file && (dump_flags & TDF_DETAILS))
4653 fprintf (dump_file, "Removing unnecessary insertion:");
4654 print_gimple_stmt (dump_file, t, 0, 0);
4657 gsi = gsi_for_stmt (t);
4658 if (gimple_code (t) == GIMPLE_PHI)
4659 remove_phi_node (&gsi, true);
4660 else
4662 gsi_remove (&gsi, true);
4663 release_defs (t);
4667 BITMAP_FREE (worklist);
4671 /* Initialize data structures used by PRE. */
4673 static void
4674 init_pre (void)
4676 basic_block bb;
4678 next_expression_id = 1;
4679 expressions.create (0);
4680 expressions.safe_push (NULL);
4681 value_expressions.create (get_max_value_id () + 1);
4682 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
4683 name_to_id.create (0);
4685 inserted_exprs = BITMAP_ALLOC (NULL);
4687 connect_infinite_loops_to_exit ();
4688 memset (&pre_stats, 0, sizeof (pre_stats));
4690 postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
4691 postorder_num = inverted_post_order_compute (postorder);
4693 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4695 calculate_dominance_info (CDI_POST_DOMINATORS);
4696 calculate_dominance_info (CDI_DOMINATORS);
4698 bitmap_obstack_initialize (&grand_bitmap_obstack);
4699 phi_translate_table = new hash_table<expr_pred_trans_d> (5110);
4700 expression_to_id = new hash_table<pre_expr_d> (num_ssa_names * 3);
4701 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4702 sizeof (struct bitmap_set), 30);
4703 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4704 sizeof (struct pre_expr_d), 30);
4705 FOR_ALL_BB_FN (bb, cfun)
4707 EXP_GEN (bb) = bitmap_set_new ();
4708 PHI_GEN (bb) = bitmap_set_new ();
4709 TMP_GEN (bb) = bitmap_set_new ();
4710 AVAIL_OUT (bb) = bitmap_set_new ();
4715 /* Deallocate data structures used by PRE. */
4717 static void
4718 fini_pre ()
4720 free (postorder);
4721 value_expressions.release ();
4722 BITMAP_FREE (inserted_exprs);
4723 bitmap_obstack_release (&grand_bitmap_obstack);
4724 free_alloc_pool (bitmap_set_pool);
4725 free_alloc_pool (pre_expr_pool);
4726 delete phi_translate_table;
4727 phi_translate_table = NULL;
4728 delete expression_to_id;
4729 expression_to_id = NULL;
4730 name_to_id.release ();
4732 free_aux_for_blocks ();
4734 free_dominance_info (CDI_POST_DOMINATORS);
4737 namespace {
4739 const pass_data pass_data_pre =
4741 GIMPLE_PASS, /* type */
4742 "pre", /* name */
4743 OPTGROUP_NONE, /* optinfo_flags */
4744 TV_TREE_PRE, /* tv_id */
4745 /* PROP_no_crit_edges is ensured by placing pass_split_crit_edges before
4746 pass_pre. */
4747 ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */
4748 0, /* properties_provided */
4749 PROP_no_crit_edges, /* properties_destroyed */
4750 TODO_rebuild_alias, /* todo_flags_start */
4751 0, /* todo_flags_finish */
4754 class pass_pre : public gimple_opt_pass
4756 public:
4757 pass_pre (gcc::context *ctxt)
4758 : gimple_opt_pass (pass_data_pre, ctxt)
4761 /* opt_pass methods: */
4762 virtual bool gate (function *) { return flag_tree_pre != 0; }
4763 virtual unsigned int execute (function *);
4765 }; // class pass_pre
4767 unsigned int
4768 pass_pre::execute (function *fun)
4770 unsigned int todo = 0;
4772 do_partial_partial =
4773 flag_tree_partial_pre && optimize_function_for_speed_p (fun);
4775 /* This has to happen before SCCVN runs because
4776 loop_optimizer_init may create new phis, etc. */
4777 loop_optimizer_init (LOOPS_NORMAL);
4779 if (!run_scc_vn (VN_WALK))
4781 loop_optimizer_finalize ();
4782 return 0;
4785 init_pre ();
4786 scev_initialize ();
4788 /* Collect and value number expressions computed in each basic block. */
4789 compute_avail ();
4791 /* Insert can get quite slow on an incredibly large number of basic
4792 blocks due to some quadratic behavior. Until this behavior is
4793 fixed, don't run it when he have an incredibly large number of
4794 bb's. If we aren't going to run insert, there is no point in
4795 computing ANTIC, either, even though it's plenty fast. */
4796 if (n_basic_blocks_for_fn (fun) < 4000)
4798 compute_antic ();
4799 insert ();
4802 /* Make sure to remove fake edges before committing our inserts.
4803 This makes sure we don't end up with extra critical edges that
4804 we would need to split. */
4805 remove_fake_exit_edges ();
4806 gsi_commit_edge_inserts ();
4808 /* Eliminate folds statements which might (should not...) end up
4809 not keeping virtual operands up-to-date. */
4810 gcc_assert (!need_ssa_update_p (fun));
4812 /* Remove all the redundant expressions. */
4813 todo |= eliminate (true);
4815 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
4816 statistics_counter_event (fun, "PA inserted", pre_stats.pa_insert);
4817 statistics_counter_event (fun, "New PHIs", pre_stats.phis);
4818 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
4820 clear_expression_ids ();
4821 remove_dead_inserted_code ();
4823 scev_finalize ();
4824 fini_pre ();
4825 todo |= fini_eliminate ();
4826 loop_optimizer_finalize ();
4828 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4829 case we can merge the block with the remaining predecessor of the block.
4830 It should either:
4831 - call merge_blocks after each tail merge iteration
4832 - call merge_blocks after all tail merge iterations
4833 - mark TODO_cleanup_cfg when necessary
4834 - share the cfg cleanup with fini_pre. */
4835 todo |= tail_merge_optimize (todo);
4837 free_scc_vn ();
4839 /* Tail merging invalidates the virtual SSA web, together with
4840 cfg-cleanup opportunities exposed by PRE this will wreck the
4841 SSA updating machinery. So make sure to run update-ssa
4842 manually, before eventually scheduling cfg-cleanup as part of
4843 the todo. */
4844 update_ssa (TODO_update_ssa_only_virtuals);
4846 return todo;
4849 } // anon namespace
4851 gimple_opt_pass *
4852 make_pass_pre (gcc::context *ctxt)
4854 return new pass_pre (ctxt);
4857 namespace {
4859 const pass_data pass_data_fre =
4861 GIMPLE_PASS, /* type */
4862 "fre", /* name */
4863 OPTGROUP_NONE, /* optinfo_flags */
4864 TV_TREE_FRE, /* tv_id */
4865 ( PROP_cfg | PROP_ssa ), /* properties_required */
4866 0, /* properties_provided */
4867 0, /* properties_destroyed */
4868 0, /* todo_flags_start */
4869 0, /* todo_flags_finish */
4872 class pass_fre : public gimple_opt_pass
4874 public:
4875 pass_fre (gcc::context *ctxt)
4876 : gimple_opt_pass (pass_data_fre, ctxt)
4879 /* opt_pass methods: */
4880 opt_pass * clone () { return new pass_fre (m_ctxt); }
4881 virtual bool gate (function *) { return flag_tree_fre != 0; }
4882 virtual unsigned int execute (function *);
4884 }; // class pass_fre
4886 unsigned int
4887 pass_fre::execute (function *fun)
4889 unsigned int todo = 0;
4891 if (!run_scc_vn (VN_WALKREWRITE))
4892 return 0;
4894 memset (&pre_stats, 0, sizeof (pre_stats));
4896 /* Remove all the redundant expressions. */
4897 todo |= eliminate (false);
4899 todo |= fini_eliminate ();
4901 free_scc_vn ();
4903 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
4904 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
4906 return todo;
4909 } // anon namespace
4911 gimple_opt_pass *
4912 make_pass_fre (gcc::context *ctxt)
4914 return new pass_fre (ctxt);