* ipa-utils.h (ipa_polymorphic_call_context): Turn into class; add ctors.
[official-gcc.git] / gcc / tree-ssa-pre.c
blobf191ec05af3cbf909e8ec91d2a9d50c9298ed39b
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 "basic-block.h"
28 #include "gimple-pretty-print.h"
29 #include "tree-inline.h"
30 #include "inchash.h"
31 #include "hash-table.h"
32 #include "tree-ssa-alias.h"
33 #include "internal-fn.h"
34 #include "gimple-fold.h"
35 #include "tree-eh.h"
36 #include "gimple-expr.h"
37 #include "is-a.h"
38 #include "gimple.h"
39 #include "gimplify.h"
40 #include "gimple-iterator.h"
41 #include "gimplify-me.h"
42 #include "gimple-ssa.h"
43 #include "tree-cfg.h"
44 #include "tree-phinodes.h"
45 #include "ssa-iterators.h"
46 #include "stringpool.h"
47 #include "tree-ssanames.h"
48 #include "tree-ssa-loop.h"
49 #include "tree-into-ssa.h"
50 #include "expr.h"
51 #include "tree-dfa.h"
52 #include "tree-ssa.h"
53 #include "tree-iterator.h"
54 #include "alloc-pool.h"
55 #include "obstack.h"
56 #include "tree-pass.h"
57 #include "flags.h"
58 #include "langhooks.h"
59 #include "cfgloop.h"
60 #include "tree-ssa-sccvn.h"
61 #include "tree-scalar-evolution.h"
62 #include "params.h"
63 #include "dbgcnt.h"
64 #include "domwalk.h"
65 #include "ipa-prop.h"
66 #include "tree-ssa-propagate.h"
67 #include "ipa-utils.h"
69 /* TODO:
71 1. Avail sets can be shared by making an avail_find_leader that
72 walks up the dominator tree and looks in those avail sets.
73 This might affect code optimality, it's unclear right now.
74 2. Strength reduction can be performed by anticipating expressions
75 we can repair later on.
76 3. We can do back-substitution or smarter value numbering to catch
77 commutative expressions split up over multiple statements.
80 /* For ease of terminology, "expression node" in the below refers to
81 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
82 represent the actual statement containing the expressions we care about,
83 and we cache the value number by putting it in the expression. */
85 /* Basic algorithm
87 First we walk the statements to generate the AVAIL sets, the
88 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
89 generation of values/expressions by a given block. We use them
90 when computing the ANTIC sets. The AVAIL sets consist of
91 SSA_NAME's that represent values, so we know what values are
92 available in what blocks. AVAIL is a forward dataflow problem. In
93 SSA, values are never killed, so we don't need a kill set, or a
94 fixpoint iteration, in order to calculate the AVAIL sets. In
95 traditional parlance, AVAIL sets tell us the downsafety of the
96 expressions/values.
98 Next, we generate the ANTIC sets. These sets represent the
99 anticipatable expressions. ANTIC is a backwards dataflow
100 problem. An expression is anticipatable in a given block if it could
101 be generated in that block. This means that if we had to perform
102 an insertion in that block, of the value of that expression, we
103 could. Calculating the ANTIC sets requires phi translation of
104 expressions, because the flow goes backwards through phis. We must
105 iterate to a fixpoint of the ANTIC sets, because we have a kill
106 set. Even in SSA form, values are not live over the entire
107 function, only from their definition point onwards. So we have to
108 remove values from the ANTIC set once we go past the definition
109 point of the leaders that make them up.
110 compute_antic/compute_antic_aux performs this computation.
112 Third, we perform insertions to make partially redundant
113 expressions fully redundant.
115 An expression is partially redundant (excluding partial
116 anticipation) if:
118 1. It is AVAIL in some, but not all, of the predecessors of a
119 given block.
120 2. It is ANTIC in all the predecessors.
122 In order to make it fully redundant, we insert the expression into
123 the predecessors where it is not available, but is ANTIC.
125 For the partial anticipation case, we only perform insertion if it
126 is partially anticipated in some block, and fully available in all
127 of the predecessors.
129 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
130 performs these steps.
132 Fourth, we eliminate fully redundant expressions.
133 This is a simple statement walk that replaces redundant
134 calculations with the now available values. */
136 /* Representations of value numbers:
138 Value numbers are represented by a representative SSA_NAME. We
139 will create fake SSA_NAME's in situations where we need a
140 representative but do not have one (because it is a complex
141 expression). In order to facilitate storing the value numbers in
142 bitmaps, and keep the number of wasted SSA_NAME's down, we also
143 associate a value_id with each value number, and create full blown
144 ssa_name's only where we actually need them (IE in operands of
145 existing expressions).
147 Theoretically you could replace all the value_id's with
148 SSA_NAME_VERSION, but this would allocate a large number of
149 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
150 It would also require an additional indirection at each point we
151 use the value id. */
153 /* Representation of expressions on value numbers:
155 Expressions consisting of value numbers are represented the same
156 way as our VN internally represents them, with an additional
157 "pre_expr" wrapping around them in order to facilitate storing all
158 of the expressions in the same sets. */
160 /* Representation of sets:
162 The dataflow sets do not need to be sorted in any particular order
163 for the majority of their lifetime, are simply represented as two
164 bitmaps, one that keeps track of values present in the set, and one
165 that keeps track of expressions present in the set.
167 When we need them in topological order, we produce it on demand by
168 transforming the bitmap into an array and sorting it into topo
169 order. */
171 /* Type of expression, used to know which member of the PRE_EXPR union
172 is valid. */
174 enum pre_expr_kind
176 NAME,
177 NARY,
178 REFERENCE,
179 CONSTANT
182 typedef union pre_expr_union_d
184 tree name;
185 tree constant;
186 vn_nary_op_t nary;
187 vn_reference_t reference;
188 } pre_expr_union;
190 typedef struct pre_expr_d : typed_noop_remove <pre_expr_d>
192 enum pre_expr_kind kind;
193 unsigned int id;
194 pre_expr_union u;
196 /* hash_table support. */
197 typedef pre_expr_d value_type;
198 typedef pre_expr_d compare_type;
199 static inline hashval_t hash (const pre_expr_d *);
200 static inline int equal (const pre_expr_d *, const pre_expr_d *);
201 } *pre_expr;
203 #define PRE_EXPR_NAME(e) (e)->u.name
204 #define PRE_EXPR_NARY(e) (e)->u.nary
205 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
206 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
208 /* Compare E1 and E1 for equality. */
210 inline int
211 pre_expr_d::equal (const value_type *e1, const compare_type *e2)
213 if (e1->kind != e2->kind)
214 return false;
216 switch (e1->kind)
218 case CONSTANT:
219 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
220 PRE_EXPR_CONSTANT (e2));
221 case NAME:
222 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
223 case NARY:
224 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
225 case REFERENCE:
226 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
227 PRE_EXPR_REFERENCE (e2));
228 default:
229 gcc_unreachable ();
233 /* Hash E. */
235 inline hashval_t
236 pre_expr_d::hash (const value_type *e)
238 switch (e->kind)
240 case CONSTANT:
241 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
242 case NAME:
243 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
244 case NARY:
245 return PRE_EXPR_NARY (e)->hashcode;
246 case REFERENCE:
247 return PRE_EXPR_REFERENCE (e)->hashcode;
248 default:
249 gcc_unreachable ();
253 /* Next global expression id number. */
254 static unsigned int next_expression_id;
256 /* Mapping from expression to id number we can use in bitmap sets. */
257 static vec<pre_expr> expressions;
258 static hash_table<pre_expr_d> *expression_to_id;
259 static vec<unsigned> name_to_id;
261 /* Allocate an expression id for EXPR. */
263 static inline unsigned int
264 alloc_expression_id (pre_expr expr)
266 struct pre_expr_d **slot;
267 /* Make sure we won't overflow. */
268 gcc_assert (next_expression_id + 1 > next_expression_id);
269 expr->id = next_expression_id++;
270 expressions.safe_push (expr);
271 if (expr->kind == NAME)
273 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
274 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
275 re-allocations by using vec::reserve upfront. There is no
276 vec::quick_grow_cleared unfortunately. */
277 unsigned old_len = name_to_id.length ();
278 name_to_id.reserve (num_ssa_names - old_len);
279 name_to_id.safe_grow_cleared (num_ssa_names);
280 gcc_assert (name_to_id[version] == 0);
281 name_to_id[version] = expr->id;
283 else
285 slot = expression_to_id->find_slot (expr, INSERT);
286 gcc_assert (!*slot);
287 *slot = expr;
289 return next_expression_id - 1;
292 /* Return the expression id for tree EXPR. */
294 static inline unsigned int
295 get_expression_id (const pre_expr expr)
297 return expr->id;
300 static inline unsigned int
301 lookup_expression_id (const pre_expr expr)
303 struct pre_expr_d **slot;
305 if (expr->kind == NAME)
307 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
308 if (name_to_id.length () <= version)
309 return 0;
310 return name_to_id[version];
312 else
314 slot = expression_to_id->find_slot (expr, NO_INSERT);
315 if (!slot)
316 return 0;
317 return ((pre_expr)*slot)->id;
321 /* Return the existing expression id for EXPR, or create one if one
322 does not exist yet. */
324 static inline unsigned int
325 get_or_alloc_expression_id (pre_expr expr)
327 unsigned int id = lookup_expression_id (expr);
328 if (id == 0)
329 return alloc_expression_id (expr);
330 return expr->id = id;
333 /* Return the expression that has expression id ID */
335 static inline pre_expr
336 expression_for_id (unsigned int id)
338 return expressions[id];
341 /* Free the expression id field in all of our expressions,
342 and then destroy the expressions array. */
344 static void
345 clear_expression_ids (void)
347 expressions.release ();
350 static alloc_pool pre_expr_pool;
352 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
354 static pre_expr
355 get_or_alloc_expr_for_name (tree name)
357 struct pre_expr_d expr;
358 pre_expr result;
359 unsigned int result_id;
361 expr.kind = NAME;
362 expr.id = 0;
363 PRE_EXPR_NAME (&expr) = name;
364 result_id = lookup_expression_id (&expr);
365 if (result_id != 0)
366 return expression_for_id (result_id);
368 result = (pre_expr) pool_alloc (pre_expr_pool);
369 result->kind = NAME;
370 PRE_EXPR_NAME (result) = name;
371 alloc_expression_id (result);
372 return result;
375 /* An unordered bitmap set. One bitmap tracks values, the other,
376 expressions. */
377 typedef struct bitmap_set
379 bitmap_head expressions;
380 bitmap_head values;
381 } *bitmap_set_t;
383 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
384 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
386 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
387 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
389 /* Mapping from value id to expressions with that value_id. */
390 static vec<bitmap> value_expressions;
392 /* Sets that we need to keep track of. */
393 typedef struct bb_bitmap_sets
395 /* The EXP_GEN set, which represents expressions/values generated in
396 a basic block. */
397 bitmap_set_t exp_gen;
399 /* The PHI_GEN set, which represents PHI results generated in a
400 basic block. */
401 bitmap_set_t phi_gen;
403 /* The TMP_GEN set, which represents results/temporaries generated
404 in a basic block. IE the LHS of an expression. */
405 bitmap_set_t tmp_gen;
407 /* The AVAIL_OUT set, which represents which values are available in
408 a given basic block. */
409 bitmap_set_t avail_out;
411 /* The ANTIC_IN set, which represents which values are anticipatable
412 in a given basic block. */
413 bitmap_set_t antic_in;
415 /* The PA_IN set, which represents which values are
416 partially anticipatable in a given basic block. */
417 bitmap_set_t pa_in;
419 /* The NEW_SETS set, which is used during insertion to augment the
420 AVAIL_OUT set of blocks with the new insertions performed during
421 the current iteration. */
422 bitmap_set_t new_sets;
424 /* A cache for value_dies_in_block_x. */
425 bitmap expr_dies;
427 /* True if we have visited this block during ANTIC calculation. */
428 unsigned int visited : 1;
430 /* True we have deferred processing this block during ANTIC
431 calculation until its successor is processed. */
432 unsigned int deferred : 1;
434 /* True when the block contains a call that might not return. */
435 unsigned int contains_may_not_return_call : 1;
436 } *bb_value_sets_t;
438 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
439 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
440 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
441 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
442 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
443 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
444 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
445 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
446 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
447 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
448 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
451 /* Basic block list in postorder. */
452 static int *postorder;
453 static int postorder_num;
455 /* This structure is used to keep track of statistics on what
456 optimization PRE was able to perform. */
457 static struct
459 /* The number of RHS computations eliminated by PRE. */
460 int eliminations;
462 /* The number of new expressions/temporaries generated by PRE. */
463 int insertions;
465 /* The number of inserts found due to partial anticipation */
466 int pa_insert;
468 /* The number of new PHI nodes added by PRE. */
469 int phis;
470 } pre_stats;
472 static bool do_partial_partial;
473 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int);
474 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
475 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
476 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
477 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
478 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
479 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
480 unsigned int, bool);
481 static bitmap_set_t bitmap_set_new (void);
482 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
483 tree);
484 static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
485 static unsigned int get_expr_value_id (pre_expr);
487 /* We can add and remove elements and entries to and from sets
488 and hash tables, so we use alloc pools for them. */
490 static alloc_pool bitmap_set_pool;
491 static bitmap_obstack grand_bitmap_obstack;
493 /* Set of blocks with statements that have had their EH properties changed. */
494 static bitmap need_eh_cleanup;
496 /* Set of blocks with statements that have had their AB properties changed. */
497 static bitmap need_ab_cleanup;
499 /* A three tuple {e, pred, v} used to cache phi translations in the
500 phi_translate_table. */
502 typedef struct expr_pred_trans_d : typed_free_remove<expr_pred_trans_d>
504 /* The expression. */
505 pre_expr e;
507 /* The predecessor block along which we translated the expression. */
508 basic_block pred;
510 /* The value that resulted from the translation. */
511 pre_expr v;
513 /* The hashcode for the expression, pred pair. This is cached for
514 speed reasons. */
515 hashval_t hashcode;
517 /* hash_table support. */
518 typedef expr_pred_trans_d value_type;
519 typedef expr_pred_trans_d compare_type;
520 static inline hashval_t hash (const value_type *);
521 static inline int equal (const value_type *, const compare_type *);
522 } *expr_pred_trans_t;
523 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
525 inline hashval_t
526 expr_pred_trans_d::hash (const expr_pred_trans_d *e)
528 return e->hashcode;
531 inline int
532 expr_pred_trans_d::equal (const value_type *ve1,
533 const compare_type *ve2)
535 basic_block b1 = ve1->pred;
536 basic_block b2 = ve2->pred;
538 /* If they are not translations for the same basic block, they can't
539 be equal. */
540 if (b1 != b2)
541 return false;
542 return pre_expr_d::equal (ve1->e, ve2->e);
545 /* The phi_translate_table caches phi translations for a given
546 expression and predecessor. */
547 static hash_table<expr_pred_trans_d> *phi_translate_table;
549 /* Add the tuple mapping from {expression E, basic block PRED} to
550 the phi translation table and return whether it pre-existed. */
552 static inline bool
553 phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred)
555 expr_pred_trans_t *slot;
556 expr_pred_trans_d tem;
557 hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e),
558 pred->index);
559 tem.e = e;
560 tem.pred = pred;
561 tem.hashcode = hash;
562 slot = phi_translate_table->find_slot_with_hash (&tem, hash, INSERT);
563 if (*slot)
565 *entry = *slot;
566 return true;
569 *entry = *slot = XNEW (struct expr_pred_trans_d);
570 (*entry)->e = e;
571 (*entry)->pred = pred;
572 (*entry)->hashcode = hash;
573 return false;
577 /* Add expression E to the expression set of value id V. */
579 static void
580 add_to_value (unsigned int v, pre_expr e)
582 bitmap set;
584 gcc_checking_assert (get_expr_value_id (e) == v);
586 if (v >= value_expressions.length ())
588 value_expressions.safe_grow_cleared (v + 1);
591 set = value_expressions[v];
592 if (!set)
594 set = BITMAP_ALLOC (&grand_bitmap_obstack);
595 value_expressions[v] = set;
598 bitmap_set_bit (set, get_or_alloc_expression_id (e));
601 /* Create a new bitmap set and return it. */
603 static bitmap_set_t
604 bitmap_set_new (void)
606 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
607 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
608 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
609 return ret;
612 /* Return the value id for a PRE expression EXPR. */
614 static unsigned int
615 get_expr_value_id (pre_expr expr)
617 unsigned int id;
618 switch (expr->kind)
620 case CONSTANT:
621 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
622 break;
623 case NAME:
624 id = VN_INFO (PRE_EXPR_NAME (expr))->value_id;
625 break;
626 case NARY:
627 id = PRE_EXPR_NARY (expr)->value_id;
628 break;
629 case REFERENCE:
630 id = PRE_EXPR_REFERENCE (expr)->value_id;
631 break;
632 default:
633 gcc_unreachable ();
635 /* ??? We cannot assert that expr has a value-id (it can be 0), because
636 we assign value-ids only to expressions that have a result
637 in set_hashtable_value_ids. */
638 return id;
641 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
643 static tree
644 sccvn_valnum_from_value_id (unsigned int val)
646 bitmap_iterator bi;
647 unsigned int i;
648 bitmap exprset = value_expressions[val];
649 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
651 pre_expr vexpr = expression_for_id (i);
652 if (vexpr->kind == NAME)
653 return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
654 else if (vexpr->kind == CONSTANT)
655 return PRE_EXPR_CONSTANT (vexpr);
657 return NULL_TREE;
660 /* Remove an expression EXPR from a bitmapped set. */
662 static void
663 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
665 unsigned int val = get_expr_value_id (expr);
666 if (!value_id_constant_p (val))
668 bitmap_clear_bit (&set->values, val);
669 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
673 static void
674 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
675 unsigned int val, bool allow_constants)
677 if (allow_constants || !value_id_constant_p (val))
679 /* We specifically expect this and only this function to be able to
680 insert constants into a set. */
681 bitmap_set_bit (&set->values, val);
682 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
686 /* Insert an expression EXPR into a bitmapped set. */
688 static void
689 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
691 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
694 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
696 static void
697 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
699 bitmap_copy (&dest->expressions, &orig->expressions);
700 bitmap_copy (&dest->values, &orig->values);
704 /* Free memory used up by SET. */
705 static void
706 bitmap_set_free (bitmap_set_t set)
708 bitmap_clear (&set->expressions);
709 bitmap_clear (&set->values);
713 /* Generate an topological-ordered array of bitmap set SET. */
715 static vec<pre_expr>
716 sorted_array_from_bitmap_set (bitmap_set_t set)
718 unsigned int i, j;
719 bitmap_iterator bi, bj;
720 vec<pre_expr> result;
722 /* Pre-allocate roughly enough space for the array. */
723 result.create (bitmap_count_bits (&set->values));
725 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
727 /* The number of expressions having a given value is usually
728 relatively small. Thus, rather than making a vector of all
729 the expressions and sorting it by value-id, we walk the values
730 and check in the reverse mapping that tells us what expressions
731 have a given value, to filter those in our set. As a result,
732 the expressions are inserted in value-id order, which means
733 topological order.
735 If this is somehow a significant lose for some cases, we can
736 choose which set to walk based on the set size. */
737 bitmap exprset = value_expressions[i];
738 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj)
740 if (bitmap_bit_p (&set->expressions, j))
741 result.safe_push (expression_for_id (j));
745 return result;
748 /* Perform bitmapped set operation DEST &= ORIG. */
750 static void
751 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
753 bitmap_iterator bi;
754 unsigned int i;
756 if (dest != orig)
758 bitmap_head temp;
759 bitmap_initialize (&temp, &grand_bitmap_obstack);
761 bitmap_and_into (&dest->values, &orig->values);
762 bitmap_copy (&temp, &dest->expressions);
763 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
765 pre_expr expr = expression_for_id (i);
766 unsigned int value_id = get_expr_value_id (expr);
767 if (!bitmap_bit_p (&dest->values, value_id))
768 bitmap_clear_bit (&dest->expressions, i);
770 bitmap_clear (&temp);
774 /* Subtract all values and expressions contained in ORIG from DEST. */
776 static bitmap_set_t
777 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
779 bitmap_set_t result = bitmap_set_new ();
780 bitmap_iterator bi;
781 unsigned int i;
783 bitmap_and_compl (&result->expressions, &dest->expressions,
784 &orig->expressions);
786 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
788 pre_expr expr = expression_for_id (i);
789 unsigned int value_id = get_expr_value_id (expr);
790 bitmap_set_bit (&result->values, value_id);
793 return result;
796 /* Subtract all the values in bitmap set B from bitmap set A. */
798 static void
799 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
801 unsigned int i;
802 bitmap_iterator bi;
803 bitmap_head temp;
805 bitmap_initialize (&temp, &grand_bitmap_obstack);
807 bitmap_copy (&temp, &a->expressions);
808 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
810 pre_expr expr = expression_for_id (i);
811 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
812 bitmap_remove_from_set (a, expr);
814 bitmap_clear (&temp);
818 /* Return true if bitmapped set SET contains the value VALUE_ID. */
820 static bool
821 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
823 if (value_id_constant_p (value_id))
824 return true;
826 if (!set || bitmap_empty_p (&set->expressions))
827 return false;
829 return bitmap_bit_p (&set->values, value_id);
832 static inline bool
833 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
835 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
838 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
840 static void
841 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
842 const pre_expr expr)
844 bitmap exprset;
845 unsigned int i;
846 bitmap_iterator bi;
848 if (value_id_constant_p (lookfor))
849 return;
851 if (!bitmap_set_contains_value (set, lookfor))
852 return;
854 /* The number of expressions having a given value is usually
855 significantly less than the total number of expressions in SET.
856 Thus, rather than check, for each expression in SET, whether it
857 has the value LOOKFOR, we walk the reverse mapping that tells us
858 what expressions have a given value, and see if any of those
859 expressions are in our set. For large testcases, this is about
860 5-10x faster than walking the bitmap. If this is somehow a
861 significant lose for some cases, we can choose which set to walk
862 based on the set size. */
863 exprset = value_expressions[lookfor];
864 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
866 if (bitmap_clear_bit (&set->expressions, i))
868 bitmap_set_bit (&set->expressions, get_expression_id (expr));
869 return;
873 gcc_unreachable ();
876 /* Return true if two bitmap sets are equal. */
878 static bool
879 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
881 return bitmap_equal_p (&a->values, &b->values);
884 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
885 and add it otherwise. */
887 static void
888 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
890 unsigned int val = get_expr_value_id (expr);
892 if (bitmap_set_contains_value (set, val))
893 bitmap_set_replace_value (set, val, expr);
894 else
895 bitmap_insert_into_set (set, expr);
898 /* Insert EXPR into SET if EXPR's value is not already present in
899 SET. */
901 static void
902 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
904 unsigned int val = get_expr_value_id (expr);
906 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
908 /* Constant values are always considered to be part of the set. */
909 if (value_id_constant_p (val))
910 return;
912 /* If the value membership changed, add the expression. */
913 if (bitmap_set_bit (&set->values, val))
914 bitmap_set_bit (&set->expressions, expr->id);
917 /* Print out EXPR to outfile. */
919 static void
920 print_pre_expr (FILE *outfile, const pre_expr expr)
922 switch (expr->kind)
924 case CONSTANT:
925 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
926 break;
927 case NAME:
928 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
929 break;
930 case NARY:
932 unsigned int i;
933 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
934 fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode));
935 for (i = 0; i < nary->length; i++)
937 print_generic_expr (outfile, nary->op[i], 0);
938 if (i != (unsigned) nary->length - 1)
939 fprintf (outfile, ",");
941 fprintf (outfile, "}");
943 break;
945 case REFERENCE:
947 vn_reference_op_t vro;
948 unsigned int i;
949 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
950 fprintf (outfile, "{");
951 for (i = 0;
952 ref->operands.iterate (i, &vro);
953 i++)
955 bool closebrace = false;
956 if (vro->opcode != SSA_NAME
957 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
959 fprintf (outfile, "%s", get_tree_code_name (vro->opcode));
960 if (vro->op0)
962 fprintf (outfile, "<");
963 closebrace = true;
966 if (vro->op0)
968 print_generic_expr (outfile, vro->op0, 0);
969 if (vro->op1)
971 fprintf (outfile, ",");
972 print_generic_expr (outfile, vro->op1, 0);
974 if (vro->op2)
976 fprintf (outfile, ",");
977 print_generic_expr (outfile, vro->op2, 0);
980 if (closebrace)
981 fprintf (outfile, ">");
982 if (i != ref->operands.length () - 1)
983 fprintf (outfile, ",");
985 fprintf (outfile, "}");
986 if (ref->vuse)
988 fprintf (outfile, "@");
989 print_generic_expr (outfile, ref->vuse, 0);
992 break;
995 void debug_pre_expr (pre_expr);
997 /* Like print_pre_expr but always prints to stderr. */
998 DEBUG_FUNCTION void
999 debug_pre_expr (pre_expr e)
1001 print_pre_expr (stderr, e);
1002 fprintf (stderr, "\n");
1005 /* Print out SET to OUTFILE. */
1007 static void
1008 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1009 const char *setname, int blockindex)
1011 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1012 if (set)
1014 bool first = true;
1015 unsigned i;
1016 bitmap_iterator bi;
1018 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1020 const pre_expr expr = expression_for_id (i);
1022 if (!first)
1023 fprintf (outfile, ", ");
1024 first = false;
1025 print_pre_expr (outfile, expr);
1027 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1030 fprintf (outfile, " }\n");
1033 void debug_bitmap_set (bitmap_set_t);
1035 DEBUG_FUNCTION void
1036 debug_bitmap_set (bitmap_set_t set)
1038 print_bitmap_set (stderr, set, "debug", 0);
1041 void debug_bitmap_sets_for (basic_block);
1043 DEBUG_FUNCTION void
1044 debug_bitmap_sets_for (basic_block bb)
1046 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1047 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1048 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1049 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1050 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1051 if (do_partial_partial)
1052 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1053 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1056 /* Print out the expressions that have VAL to OUTFILE. */
1058 static void
1059 print_value_expressions (FILE *outfile, unsigned int val)
1061 bitmap set = value_expressions[val];
1062 if (set)
1064 bitmap_set x;
1065 char s[10];
1066 sprintf (s, "%04d", val);
1067 x.expressions = *set;
1068 print_bitmap_set (outfile, &x, s, 0);
1073 DEBUG_FUNCTION void
1074 debug_value_expressions (unsigned int val)
1076 print_value_expressions (stderr, val);
1079 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1080 represent it. */
1082 static pre_expr
1083 get_or_alloc_expr_for_constant (tree constant)
1085 unsigned int result_id;
1086 unsigned int value_id;
1087 struct pre_expr_d expr;
1088 pre_expr newexpr;
1090 expr.kind = CONSTANT;
1091 PRE_EXPR_CONSTANT (&expr) = constant;
1092 result_id = lookup_expression_id (&expr);
1093 if (result_id != 0)
1094 return expression_for_id (result_id);
1096 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1097 newexpr->kind = CONSTANT;
1098 PRE_EXPR_CONSTANT (newexpr) = constant;
1099 alloc_expression_id (newexpr);
1100 value_id = get_or_alloc_constant_value_id (constant);
1101 add_to_value (value_id, newexpr);
1102 return newexpr;
1105 /* Given a value id V, find the actual tree representing the constant
1106 value if there is one, and return it. Return NULL if we can't find
1107 a constant. */
1109 static tree
1110 get_constant_for_value_id (unsigned int v)
1112 if (value_id_constant_p (v))
1114 unsigned int i;
1115 bitmap_iterator bi;
1116 bitmap exprset = value_expressions[v];
1118 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1120 pre_expr expr = expression_for_id (i);
1121 if (expr->kind == CONSTANT)
1122 return PRE_EXPR_CONSTANT (expr);
1125 return NULL;
1128 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1129 Currently only supports constants and SSA_NAMES. */
1130 static pre_expr
1131 get_or_alloc_expr_for (tree t)
1133 if (TREE_CODE (t) == SSA_NAME)
1134 return get_or_alloc_expr_for_name (t);
1135 else if (is_gimple_min_invariant (t))
1136 return get_or_alloc_expr_for_constant (t);
1137 else
1139 /* More complex expressions can result from SCCVN expression
1140 simplification that inserts values for them. As they all
1141 do not have VOPs the get handled by the nary ops struct. */
1142 vn_nary_op_t result;
1143 unsigned int result_id;
1144 vn_nary_op_lookup (t, &result);
1145 if (result != NULL)
1147 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1148 e->kind = NARY;
1149 PRE_EXPR_NARY (e) = result;
1150 result_id = lookup_expression_id (e);
1151 if (result_id != 0)
1153 pool_free (pre_expr_pool, e);
1154 e = expression_for_id (result_id);
1155 return e;
1157 alloc_expression_id (e);
1158 return e;
1161 return NULL;
1164 /* Return the folded version of T if T, when folded, is a gimple
1165 min_invariant. Otherwise, return T. */
1167 static pre_expr
1168 fully_constant_expression (pre_expr e)
1170 switch (e->kind)
1172 case CONSTANT:
1173 return e;
1174 case NARY:
1176 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1177 switch (TREE_CODE_CLASS (nary->opcode))
1179 case tcc_binary:
1180 case tcc_comparison:
1182 /* We have to go from trees to pre exprs to value ids to
1183 constants. */
1184 tree naryop0 = nary->op[0];
1185 tree naryop1 = nary->op[1];
1186 tree result;
1187 if (!is_gimple_min_invariant (naryop0))
1189 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1190 unsigned int vrep0 = get_expr_value_id (rep0);
1191 tree const0 = get_constant_for_value_id (vrep0);
1192 if (const0)
1193 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1195 if (!is_gimple_min_invariant (naryop1))
1197 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1198 unsigned int vrep1 = get_expr_value_id (rep1);
1199 tree const1 = get_constant_for_value_id (vrep1);
1200 if (const1)
1201 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1203 result = fold_binary (nary->opcode, nary->type,
1204 naryop0, naryop1);
1205 if (result && is_gimple_min_invariant (result))
1206 return get_or_alloc_expr_for_constant (result);
1207 /* We might have simplified the expression to a
1208 SSA_NAME for example from x_1 * 1. But we cannot
1209 insert a PHI for x_1 unconditionally as x_1 might
1210 not be available readily. */
1211 return e;
1213 case tcc_reference:
1214 if (nary->opcode != REALPART_EXPR
1215 && nary->opcode != IMAGPART_EXPR
1216 && nary->opcode != VIEW_CONVERT_EXPR)
1217 return e;
1218 /* Fallthrough. */
1219 case tcc_unary:
1221 /* We have to go from trees to pre exprs to value ids to
1222 constants. */
1223 tree naryop0 = nary->op[0];
1224 tree const0, result;
1225 if (is_gimple_min_invariant (naryop0))
1226 const0 = naryop0;
1227 else
1229 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1230 unsigned int vrep0 = get_expr_value_id (rep0);
1231 const0 = get_constant_for_value_id (vrep0);
1233 result = NULL;
1234 if (const0)
1236 tree type1 = TREE_TYPE (nary->op[0]);
1237 const0 = fold_convert (type1, const0);
1238 result = fold_unary (nary->opcode, nary->type, const0);
1240 if (result && is_gimple_min_invariant (result))
1241 return get_or_alloc_expr_for_constant (result);
1242 return e;
1244 default:
1245 return e;
1248 case REFERENCE:
1250 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1251 tree folded;
1252 if ((folded = fully_constant_vn_reference_p (ref)))
1253 return get_or_alloc_expr_for_constant (folded);
1254 return e;
1256 default:
1257 return e;
1259 return e;
1262 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1263 it has the value it would have in BLOCK. Set *SAME_VALID to true
1264 in case the new vuse doesn't change the value id of the OPERANDS. */
1266 static tree
1267 translate_vuse_through_block (vec<vn_reference_op_s> operands,
1268 alias_set_type set, tree type, tree vuse,
1269 basic_block phiblock,
1270 basic_block block, bool *same_valid)
1272 gimple phi = SSA_NAME_DEF_STMT (vuse);
1273 ao_ref ref;
1274 edge e = NULL;
1275 bool use_oracle;
1277 *same_valid = true;
1279 if (gimple_bb (phi) != phiblock)
1280 return vuse;
1282 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1284 /* Use the alias-oracle to find either the PHI node in this block,
1285 the first VUSE used in this block that is equivalent to vuse or
1286 the first VUSE which definition in this block kills the value. */
1287 if (gimple_code (phi) == GIMPLE_PHI)
1288 e = find_edge (block, phiblock);
1289 else if (use_oracle)
1290 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1292 vuse = gimple_vuse (phi);
1293 phi = SSA_NAME_DEF_STMT (vuse);
1294 if (gimple_bb (phi) != phiblock)
1295 return vuse;
1296 if (gimple_code (phi) == GIMPLE_PHI)
1298 e = find_edge (block, phiblock);
1299 break;
1302 else
1303 return NULL_TREE;
1305 if (e)
1307 if (use_oracle)
1309 bitmap visited = NULL;
1310 unsigned int cnt;
1311 /* Try to find a vuse that dominates this phi node by skipping
1312 non-clobbering statements. */
1313 vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false,
1314 NULL, NULL);
1315 if (visited)
1316 BITMAP_FREE (visited);
1318 else
1319 vuse = NULL_TREE;
1320 if (!vuse)
1322 /* If we didn't find any, the value ID can't stay the same,
1323 but return the translated vuse. */
1324 *same_valid = false;
1325 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1327 /* ??? We would like to return vuse here as this is the canonical
1328 upmost vdef that this reference is associated with. But during
1329 insertion of the references into the hash tables we only ever
1330 directly insert with their direct gimple_vuse, hence returning
1331 something else would make us not find the other expression. */
1332 return PHI_ARG_DEF (phi, e->dest_idx);
1335 return NULL_TREE;
1338 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1339 SET2. This is used to avoid making a set consisting of the union
1340 of PA_IN and ANTIC_IN during insert. */
1342 static inline pre_expr
1343 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1345 pre_expr result;
1347 result = bitmap_find_leader (set1, val);
1348 if (!result && set2)
1349 result = bitmap_find_leader (set2, val);
1350 return result;
1353 /* Get the tree type for our PRE expression e. */
1355 static tree
1356 get_expr_type (const pre_expr e)
1358 switch (e->kind)
1360 case NAME:
1361 return TREE_TYPE (PRE_EXPR_NAME (e));
1362 case CONSTANT:
1363 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1364 case REFERENCE:
1365 return PRE_EXPR_REFERENCE (e)->type;
1366 case NARY:
1367 return PRE_EXPR_NARY (e)->type;
1369 gcc_unreachable ();
1372 /* Get a representative SSA_NAME for a given expression.
1373 Since all of our sub-expressions are treated as values, we require
1374 them to be SSA_NAME's for simplicity.
1375 Prior versions of GVNPRE used to use "value handles" here, so that
1376 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1377 either case, the operands are really values (IE we do not expect
1378 them to be usable without finding leaders). */
1380 static tree
1381 get_representative_for (const pre_expr e)
1383 tree name;
1384 unsigned int value_id = get_expr_value_id (e);
1386 switch (e->kind)
1388 case NAME:
1389 return PRE_EXPR_NAME (e);
1390 case CONSTANT:
1391 return PRE_EXPR_CONSTANT (e);
1392 case NARY:
1393 case REFERENCE:
1395 /* Go through all of the expressions representing this value
1396 and pick out an SSA_NAME. */
1397 unsigned int i;
1398 bitmap_iterator bi;
1399 bitmap exprs = value_expressions[value_id];
1400 EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1402 pre_expr rep = expression_for_id (i);
1403 if (rep->kind == NAME)
1404 return PRE_EXPR_NAME (rep);
1405 else if (rep->kind == CONSTANT)
1406 return PRE_EXPR_CONSTANT (rep);
1409 break;
1412 /* If we reached here we couldn't find an SSA_NAME. This can
1413 happen when we've discovered a value that has never appeared in
1414 the program as set to an SSA_NAME, as the result of phi translation.
1415 Create one here.
1416 ??? We should be able to re-use this when we insert the statement
1417 to compute it. */
1418 name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1419 VN_INFO_GET (name)->value_id = value_id;
1420 VN_INFO (name)->valnum = name;
1421 /* ??? For now mark this SSA name for release by SCCVN. */
1422 VN_INFO (name)->needs_insertion = true;
1423 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1424 if (dump_file && (dump_flags & TDF_DETAILS))
1426 fprintf (dump_file, "Created SSA_NAME representative ");
1427 print_generic_expr (dump_file, name, 0);
1428 fprintf (dump_file, " for expression:");
1429 print_pre_expr (dump_file, e);
1430 fprintf (dump_file, " (%04d)\n", value_id);
1433 return name;
1438 static pre_expr
1439 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1440 basic_block pred, basic_block phiblock);
1442 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1443 the phis in PRED. Return NULL if we can't find a leader for each part
1444 of the translated expression. */
1446 static pre_expr
1447 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1448 basic_block pred, basic_block phiblock)
1450 switch (expr->kind)
1452 case NARY:
1454 unsigned int i;
1455 bool changed = false;
1456 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1457 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1458 sizeof_vn_nary_op (nary->length));
1459 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1461 for (i = 0; i < newnary->length; i++)
1463 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1464 continue;
1465 else
1467 pre_expr leader, result;
1468 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1469 leader = find_leader_in_sets (op_val_id, set1, set2);
1470 result = phi_translate (leader, set1, set2, pred, phiblock);
1471 if (result && result != leader)
1473 tree name = get_representative_for (result);
1474 if (!name)
1475 return NULL;
1476 newnary->op[i] = name;
1478 else if (!result)
1479 return NULL;
1481 changed |= newnary->op[i] != nary->op[i];
1484 if (changed)
1486 pre_expr constant;
1487 unsigned int new_val_id;
1489 tree result = vn_nary_op_lookup_pieces (newnary->length,
1490 newnary->opcode,
1491 newnary->type,
1492 &newnary->op[0],
1493 &nary);
1494 if (result && is_gimple_min_invariant (result))
1495 return get_or_alloc_expr_for_constant (result);
1497 expr = (pre_expr) pool_alloc (pre_expr_pool);
1498 expr->kind = NARY;
1499 expr->id = 0;
1500 if (nary)
1502 PRE_EXPR_NARY (expr) = nary;
1503 constant = fully_constant_expression (expr);
1504 if (constant != expr)
1505 return constant;
1507 new_val_id = nary->value_id;
1508 get_or_alloc_expression_id (expr);
1510 else
1512 new_val_id = get_next_value_id ();
1513 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
1514 nary = vn_nary_op_insert_pieces (newnary->length,
1515 newnary->opcode,
1516 newnary->type,
1517 &newnary->op[0],
1518 result, new_val_id);
1519 PRE_EXPR_NARY (expr) = nary;
1520 constant = fully_constant_expression (expr);
1521 if (constant != expr)
1522 return constant;
1523 get_or_alloc_expression_id (expr);
1525 add_to_value (new_val_id, expr);
1527 return expr;
1529 break;
1531 case REFERENCE:
1533 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1534 vec<vn_reference_op_s> operands = ref->operands;
1535 tree vuse = ref->vuse;
1536 tree newvuse = vuse;
1537 vec<vn_reference_op_s> newoperands = vNULL;
1538 bool changed = false, same_valid = true;
1539 unsigned int i, j, n;
1540 vn_reference_op_t operand;
1541 vn_reference_t newref;
1543 for (i = 0, j = 0;
1544 operands.iterate (i, &operand); i++, j++)
1546 pre_expr opresult;
1547 pre_expr leader;
1548 tree op[3];
1549 tree type = operand->type;
1550 vn_reference_op_s newop = *operand;
1551 op[0] = operand->op0;
1552 op[1] = operand->op1;
1553 op[2] = operand->op2;
1554 for (n = 0; n < 3; ++n)
1556 unsigned int op_val_id;
1557 if (!op[n])
1558 continue;
1559 if (TREE_CODE (op[n]) != SSA_NAME)
1561 /* We can't possibly insert these. */
1562 if (n != 0
1563 && !is_gimple_min_invariant (op[n]))
1564 break;
1565 continue;
1567 op_val_id = VN_INFO (op[n])->value_id;
1568 leader = find_leader_in_sets (op_val_id, set1, set2);
1569 if (!leader)
1570 break;
1571 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1572 if (!opresult)
1573 break;
1574 if (opresult != leader)
1576 tree name = get_representative_for (opresult);
1577 if (!name)
1578 break;
1579 changed |= name != op[n];
1580 op[n] = name;
1583 if (n != 3)
1585 newoperands.release ();
1586 return NULL;
1588 if (!newoperands.exists ())
1589 newoperands = operands.copy ();
1590 /* We may have changed from an SSA_NAME to a constant */
1591 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1592 newop.opcode = TREE_CODE (op[0]);
1593 newop.type = type;
1594 newop.op0 = op[0];
1595 newop.op1 = op[1];
1596 newop.op2 = op[2];
1597 /* If it transforms a non-constant ARRAY_REF into a constant
1598 one, adjust the constant offset. */
1599 if (newop.opcode == ARRAY_REF
1600 && newop.off == -1
1601 && TREE_CODE (op[0]) == INTEGER_CST
1602 && TREE_CODE (op[1]) == INTEGER_CST
1603 && TREE_CODE (op[2]) == INTEGER_CST)
1605 offset_int off = ((wi::to_offset (op[0])
1606 - wi::to_offset (op[1]))
1607 * wi::to_offset (op[2]));
1608 if (wi::fits_shwi_p (off))
1609 newop.off = off.to_shwi ();
1611 newoperands[j] = newop;
1612 /* If it transforms from an SSA_NAME to an address, fold with
1613 a preceding indirect reference. */
1614 if (j > 0 && op[0] && TREE_CODE (op[0]) == ADDR_EXPR
1615 && newoperands[j - 1].opcode == MEM_REF)
1616 vn_reference_fold_indirect (&newoperands, &j);
1618 if (i != operands.length ())
1620 newoperands.release ();
1621 return NULL;
1624 if (vuse)
1626 newvuse = translate_vuse_through_block (newoperands,
1627 ref->set, ref->type,
1628 vuse, phiblock, pred,
1629 &same_valid);
1630 if (newvuse == NULL_TREE)
1632 newoperands.release ();
1633 return NULL;
1637 if (changed || newvuse != vuse)
1639 unsigned int new_val_id;
1640 pre_expr constant;
1642 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1643 ref->type,
1644 newoperands,
1645 &newref, VN_WALK);
1646 if (result)
1647 newoperands.release ();
1649 /* We can always insert constants, so if we have a partial
1650 redundant constant load of another type try to translate it
1651 to a constant of appropriate type. */
1652 if (result && is_gimple_min_invariant (result))
1654 tree tem = result;
1655 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1657 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1658 if (tem && !is_gimple_min_invariant (tem))
1659 tem = NULL_TREE;
1661 if (tem)
1662 return get_or_alloc_expr_for_constant (tem);
1665 /* If we'd have to convert things we would need to validate
1666 if we can insert the translated expression. So fail
1667 here for now - we cannot insert an alias with a different
1668 type in the VN tables either, as that would assert. */
1669 if (result
1670 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1671 return NULL;
1672 else if (!result && newref
1673 && !useless_type_conversion_p (ref->type, newref->type))
1675 newoperands.release ();
1676 return NULL;
1679 expr = (pre_expr) pool_alloc (pre_expr_pool);
1680 expr->kind = REFERENCE;
1681 expr->id = 0;
1683 if (newref)
1685 PRE_EXPR_REFERENCE (expr) = newref;
1686 constant = fully_constant_expression (expr);
1687 if (constant != expr)
1688 return constant;
1690 new_val_id = newref->value_id;
1691 get_or_alloc_expression_id (expr);
1693 else
1695 if (changed || !same_valid)
1697 new_val_id = get_next_value_id ();
1698 value_expressions.safe_grow_cleared
1699 (get_max_value_id () + 1);
1701 else
1702 new_val_id = ref->value_id;
1703 newref = vn_reference_insert_pieces (newvuse, ref->set,
1704 ref->type,
1705 newoperands,
1706 result, new_val_id);
1707 newoperands.create (0);
1708 PRE_EXPR_REFERENCE (expr) = newref;
1709 constant = fully_constant_expression (expr);
1710 if (constant != expr)
1711 return constant;
1712 get_or_alloc_expression_id (expr);
1714 add_to_value (new_val_id, expr);
1716 newoperands.release ();
1717 return expr;
1719 break;
1721 case NAME:
1723 tree name = PRE_EXPR_NAME (expr);
1724 gimple def_stmt = SSA_NAME_DEF_STMT (name);
1725 /* If the SSA name is defined by a PHI node in this block,
1726 translate it. */
1727 if (gimple_code (def_stmt) == GIMPLE_PHI
1728 && gimple_bb (def_stmt) == phiblock)
1730 edge e = find_edge (pred, gimple_bb (def_stmt));
1731 tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1733 /* Handle constant. */
1734 if (is_gimple_min_invariant (def))
1735 return get_or_alloc_expr_for_constant (def);
1737 return get_or_alloc_expr_for_name (def);
1739 /* Otherwise return it unchanged - it will get cleaned if its
1740 value is not available in PREDs AVAIL_OUT set of expressions. */
1741 return expr;
1744 default:
1745 gcc_unreachable ();
1749 /* Wrapper around phi_translate_1 providing caching functionality. */
1751 static pre_expr
1752 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1753 basic_block pred, basic_block phiblock)
1755 expr_pred_trans_t slot = NULL;
1756 pre_expr phitrans;
1758 if (!expr)
1759 return NULL;
1761 /* Constants contain no values that need translation. */
1762 if (expr->kind == CONSTANT)
1763 return expr;
1765 if (value_id_constant_p (get_expr_value_id (expr)))
1766 return expr;
1768 /* Don't add translations of NAMEs as those are cheap to translate. */
1769 if (expr->kind != NAME)
1771 if (phi_trans_add (&slot, expr, pred))
1772 return slot->v;
1773 /* Store NULL for the value we want to return in the case of
1774 recursing. */
1775 slot->v = NULL;
1778 /* Translate. */
1779 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1781 if (slot)
1783 if (phitrans)
1784 slot->v = phitrans;
1785 else
1786 /* Remove failed translations again, they cause insert
1787 iteration to not pick up new opportunities reliably. */
1788 phi_translate_table->remove_elt_with_hash (slot, slot->hashcode);
1791 return phitrans;
1795 /* For each expression in SET, translate the values through phi nodes
1796 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1797 expressions in DEST. */
1799 static void
1800 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1801 basic_block phiblock)
1803 vec<pre_expr> exprs;
1804 pre_expr expr;
1805 int i;
1807 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1809 bitmap_set_copy (dest, set);
1810 return;
1813 exprs = sorted_array_from_bitmap_set (set);
1814 FOR_EACH_VEC_ELT (exprs, i, expr)
1816 pre_expr translated;
1817 translated = phi_translate (expr, set, NULL, pred, phiblock);
1818 if (!translated)
1819 continue;
1821 /* We might end up with multiple expressions from SET being
1822 translated to the same value. In this case we do not want
1823 to retain the NARY or REFERENCE expression but prefer a NAME
1824 which would be the leader. */
1825 if (translated->kind == NAME)
1826 bitmap_value_replace_in_set (dest, translated);
1827 else
1828 bitmap_value_insert_into_set (dest, translated);
1830 exprs.release ();
1833 /* Find the leader for a value (i.e., the name representing that
1834 value) in a given set, and return it. Return NULL if no leader
1835 is found. */
1837 static pre_expr
1838 bitmap_find_leader (bitmap_set_t set, unsigned int val)
1840 if (value_id_constant_p (val))
1842 unsigned int i;
1843 bitmap_iterator bi;
1844 bitmap exprset = value_expressions[val];
1846 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1848 pre_expr expr = expression_for_id (i);
1849 if (expr->kind == CONSTANT)
1850 return expr;
1853 if (bitmap_set_contains_value (set, val))
1855 /* Rather than walk the entire bitmap of expressions, and see
1856 whether any of them has the value we are looking for, we look
1857 at the reverse mapping, which tells us the set of expressions
1858 that have a given value (IE value->expressions with that
1859 value) and see if any of those expressions are in our set.
1860 The number of expressions per value is usually significantly
1861 less than the number of expressions in the set. In fact, for
1862 large testcases, doing it this way is roughly 5-10x faster
1863 than walking the bitmap.
1864 If this is somehow a significant lose for some cases, we can
1865 choose which set to walk based on which set is smaller. */
1866 unsigned int i;
1867 bitmap_iterator bi;
1868 bitmap exprset = value_expressions[val];
1870 EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1871 return expression_for_id (i);
1873 return NULL;
1876 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1877 BLOCK by seeing if it is not killed in the block. Note that we are
1878 only determining whether there is a store that kills it. Because
1879 of the order in which clean iterates over values, we are guaranteed
1880 that altered operands will have caused us to be eliminated from the
1881 ANTIC_IN set already. */
1883 static bool
1884 value_dies_in_block_x (pre_expr expr, basic_block block)
1886 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1887 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1888 gimple def;
1889 gimple_stmt_iterator gsi;
1890 unsigned id = get_expression_id (expr);
1891 bool res = false;
1892 ao_ref ref;
1894 if (!vuse)
1895 return false;
1897 /* Lookup a previously calculated result. */
1898 if (EXPR_DIES (block)
1899 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1900 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1902 /* A memory expression {e, VUSE} dies in the block if there is a
1903 statement that may clobber e. If, starting statement walk from the
1904 top of the basic block, a statement uses VUSE there can be no kill
1905 inbetween that use and the original statement that loaded {e, VUSE},
1906 so we can stop walking. */
1907 ref.base = NULL_TREE;
1908 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1910 tree def_vuse, def_vdef;
1911 def = gsi_stmt (gsi);
1912 def_vuse = gimple_vuse (def);
1913 def_vdef = gimple_vdef (def);
1915 /* Not a memory statement. */
1916 if (!def_vuse)
1917 continue;
1919 /* Not a may-def. */
1920 if (!def_vdef)
1922 /* A load with the same VUSE, we're done. */
1923 if (def_vuse == vuse)
1924 break;
1926 continue;
1929 /* Init ref only if we really need it. */
1930 if (ref.base == NULL_TREE
1931 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1932 refx->operands))
1934 res = true;
1935 break;
1937 /* If the statement may clobber expr, it dies. */
1938 if (stmt_may_clobber_ref_p_1 (def, &ref))
1940 res = true;
1941 break;
1945 /* Remember the result. */
1946 if (!EXPR_DIES (block))
1947 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1948 bitmap_set_bit (EXPR_DIES (block), id * 2);
1949 if (res)
1950 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1952 return res;
1956 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1957 contains its value-id. */
1959 static bool
1960 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1962 if (op && TREE_CODE (op) == SSA_NAME)
1964 unsigned int value_id = VN_INFO (op)->value_id;
1965 if (!(bitmap_set_contains_value (set1, value_id)
1966 || (set2 && bitmap_set_contains_value (set2, value_id))))
1967 return false;
1969 return true;
1972 /* Determine if the expression EXPR is valid in SET1 U SET2.
1973 ONLY SET2 CAN BE NULL.
1974 This means that we have a leader for each part of the expression
1975 (if it consists of values), or the expression is an SSA_NAME.
1976 For loads/calls, we also see if the vuse is killed in this block. */
1978 static bool
1979 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
1980 basic_block block)
1982 switch (expr->kind)
1984 case NAME:
1985 return bitmap_find_leader (AVAIL_OUT (block),
1986 get_expr_value_id (expr)) != NULL;
1987 case NARY:
1989 unsigned int i;
1990 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1991 for (i = 0; i < nary->length; i++)
1992 if (!op_valid_in_sets (set1, set2, nary->op[i]))
1993 return false;
1994 return true;
1996 break;
1997 case REFERENCE:
1999 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2000 vn_reference_op_t vro;
2001 unsigned int i;
2003 FOR_EACH_VEC_ELT (ref->operands, i, vro)
2005 if (!op_valid_in_sets (set1, set2, vro->op0)
2006 || !op_valid_in_sets (set1, set2, vro->op1)
2007 || !op_valid_in_sets (set1, set2, vro->op2))
2008 return false;
2010 return true;
2012 default:
2013 gcc_unreachable ();
2017 /* Clean the set of expressions that are no longer valid in SET1 or
2018 SET2. This means expressions that are made up of values we have no
2019 leaders for in SET1 or SET2. This version is used for partial
2020 anticipation, which means it is not valid in either ANTIC_IN or
2021 PA_IN. */
2023 static void
2024 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2026 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
2027 pre_expr expr;
2028 int i;
2030 FOR_EACH_VEC_ELT (exprs, i, expr)
2032 if (!valid_in_sets (set1, set2, expr, block))
2033 bitmap_remove_from_set (set1, expr);
2035 exprs.release ();
2038 /* Clean the set of expressions that are no longer valid in SET. This
2039 means expressions that are made up of values we have no leaders for
2040 in SET. */
2042 static void
2043 clean (bitmap_set_t set, basic_block block)
2045 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set);
2046 pre_expr expr;
2047 int i;
2049 FOR_EACH_VEC_ELT (exprs, i, expr)
2051 if (!valid_in_sets (set, NULL, expr, block))
2052 bitmap_remove_from_set (set, expr);
2054 exprs.release ();
2057 /* Clean the set of expressions that are no longer valid in SET because
2058 they are clobbered in BLOCK or because they trap and may not be executed. */
2060 static void
2061 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2063 bitmap_iterator bi;
2064 unsigned i;
2066 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2068 pre_expr expr = expression_for_id (i);
2069 if (expr->kind == REFERENCE)
2071 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2072 if (ref->vuse)
2074 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2075 if (!gimple_nop_p (def_stmt)
2076 && ((gimple_bb (def_stmt) != block
2077 && !dominated_by_p (CDI_DOMINATORS,
2078 block, gimple_bb (def_stmt)))
2079 || (gimple_bb (def_stmt) == block
2080 && value_dies_in_block_x (expr, block))))
2081 bitmap_remove_from_set (set, expr);
2084 else if (expr->kind == NARY)
2086 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2087 /* If the NARY may trap make sure the block does not contain
2088 a possible exit point.
2089 ??? This is overly conservative if we translate AVAIL_OUT
2090 as the available expression might be after the exit point. */
2091 if (BB_MAY_NOTRETURN (block)
2092 && vn_nary_may_trap (nary))
2093 bitmap_remove_from_set (set, expr);
2098 static sbitmap has_abnormal_preds;
2100 /* List of blocks that may have changed during ANTIC computation and
2101 thus need to be iterated over. */
2103 static sbitmap changed_blocks;
2105 /* Decide whether to defer a block for a later iteration, or PHI
2106 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2107 should defer the block, and true if we processed it. */
2109 static bool
2110 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2111 basic_block block, basic_block phiblock)
2113 if (!BB_VISITED (phiblock))
2115 bitmap_set_bit (changed_blocks, block->index);
2116 BB_VISITED (block) = 0;
2117 BB_DEFERRED (block) = 1;
2118 return false;
2120 else
2121 phi_translate_set (dest, source, block, phiblock);
2122 return true;
2125 /* Compute the ANTIC set for BLOCK.
2127 If succs(BLOCK) > 1 then
2128 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2129 else if succs(BLOCK) == 1 then
2130 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2132 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2135 static bool
2136 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2138 bool changed = false;
2139 bitmap_set_t S, old, ANTIC_OUT;
2140 bitmap_iterator bi;
2141 unsigned int bii;
2142 edge e;
2143 edge_iterator ei;
2145 old = ANTIC_OUT = S = NULL;
2146 BB_VISITED (block) = 1;
2148 /* If any edges from predecessors are abnormal, antic_in is empty,
2149 so do nothing. */
2150 if (block_has_abnormal_pred_edge)
2151 goto maybe_dump_sets;
2153 old = ANTIC_IN (block);
2154 ANTIC_OUT = bitmap_set_new ();
2156 /* If the block has no successors, ANTIC_OUT is empty. */
2157 if (EDGE_COUNT (block->succs) == 0)
2159 /* If we have one successor, we could have some phi nodes to
2160 translate through. */
2161 else if (single_succ_p (block))
2163 basic_block succ_bb = single_succ (block);
2165 /* We trade iterations of the dataflow equations for having to
2166 phi translate the maximal set, which is incredibly slow
2167 (since the maximal set often has 300+ members, even when you
2168 have a small number of blocks).
2169 Basically, we defer the computation of ANTIC for this block
2170 until we have processed it's successor, which will inevitably
2171 have a *much* smaller set of values to phi translate once
2172 clean has been run on it.
2173 The cost of doing this is that we technically perform more
2174 iterations, however, they are lower cost iterations.
2176 Timings for PRE on tramp3d-v4:
2177 without maximal set fix: 11 seconds
2178 with maximal set fix/without deferring: 26 seconds
2179 with maximal set fix/with deferring: 11 seconds
2182 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2183 block, succ_bb))
2185 changed = true;
2186 goto maybe_dump_sets;
2189 /* If we have multiple successors, we take the intersection of all of
2190 them. Note that in the case of loop exit phi nodes, we may have
2191 phis to translate through. */
2192 else
2194 size_t i;
2195 basic_block bprime, first = NULL;
2197 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2198 FOR_EACH_EDGE (e, ei, block->succs)
2200 if (!first
2201 && BB_VISITED (e->dest))
2202 first = e->dest;
2203 else if (BB_VISITED (e->dest))
2204 worklist.quick_push (e->dest);
2207 /* Of multiple successors we have to have visited one already. */
2208 if (!first)
2210 bitmap_set_bit (changed_blocks, block->index);
2211 BB_VISITED (block) = 0;
2212 BB_DEFERRED (block) = 1;
2213 changed = true;
2214 goto maybe_dump_sets;
2217 if (!gimple_seq_empty_p (phi_nodes (first)))
2218 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2219 else
2220 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2222 FOR_EACH_VEC_ELT (worklist, i, bprime)
2224 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2226 bitmap_set_t tmp = bitmap_set_new ();
2227 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2228 bitmap_set_and (ANTIC_OUT, tmp);
2229 bitmap_set_free (tmp);
2231 else
2232 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2236 /* Prune expressions that are clobbered in block and thus become
2237 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2238 prune_clobbered_mems (ANTIC_OUT, block);
2240 /* Generate ANTIC_OUT - TMP_GEN. */
2241 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2243 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2244 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2245 TMP_GEN (block));
2247 /* Then union in the ANTIC_OUT - TMP_GEN values,
2248 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2249 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2250 bitmap_value_insert_into_set (ANTIC_IN (block),
2251 expression_for_id (bii));
2253 clean (ANTIC_IN (block), block);
2255 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2257 changed = true;
2258 bitmap_set_bit (changed_blocks, block->index);
2259 FOR_EACH_EDGE (e, ei, block->preds)
2260 bitmap_set_bit (changed_blocks, e->src->index);
2262 else
2263 bitmap_clear_bit (changed_blocks, block->index);
2265 maybe_dump_sets:
2266 if (dump_file && (dump_flags & TDF_DETAILS))
2268 if (!BB_DEFERRED (block) || BB_VISITED (block))
2270 if (ANTIC_OUT)
2271 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2273 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2274 block->index);
2276 if (S)
2277 print_bitmap_set (dump_file, S, "S", block->index);
2279 else
2281 fprintf (dump_file,
2282 "Block %d was deferred for a future iteration.\n",
2283 block->index);
2286 if (old)
2287 bitmap_set_free (old);
2288 if (S)
2289 bitmap_set_free (S);
2290 if (ANTIC_OUT)
2291 bitmap_set_free (ANTIC_OUT);
2292 return changed;
2295 /* Compute PARTIAL_ANTIC for BLOCK.
2297 If succs(BLOCK) > 1 then
2298 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2299 in ANTIC_OUT for all succ(BLOCK)
2300 else if succs(BLOCK) == 1 then
2301 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2303 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2304 - ANTIC_IN[BLOCK])
2307 static bool
2308 compute_partial_antic_aux (basic_block block,
2309 bool block_has_abnormal_pred_edge)
2311 bool changed = false;
2312 bitmap_set_t old_PA_IN;
2313 bitmap_set_t PA_OUT;
2314 edge e;
2315 edge_iterator ei;
2316 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2318 old_PA_IN = PA_OUT = NULL;
2320 /* If any edges from predecessors are abnormal, antic_in is empty,
2321 so do nothing. */
2322 if (block_has_abnormal_pred_edge)
2323 goto maybe_dump_sets;
2325 /* If there are too many partially anticipatable values in the
2326 block, phi_translate_set can take an exponential time: stop
2327 before the translation starts. */
2328 if (max_pa
2329 && single_succ_p (block)
2330 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2331 goto maybe_dump_sets;
2333 old_PA_IN = PA_IN (block);
2334 PA_OUT = bitmap_set_new ();
2336 /* If the block has no successors, ANTIC_OUT is empty. */
2337 if (EDGE_COUNT (block->succs) == 0)
2339 /* If we have one successor, we could have some phi nodes to
2340 translate through. Note that we can't phi translate across DFS
2341 back edges in partial antic, because it uses a union operation on
2342 the successors. For recurrences like IV's, we will end up
2343 generating a new value in the set on each go around (i + 3 (VH.1)
2344 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2345 else if (single_succ_p (block))
2347 basic_block succ = single_succ (block);
2348 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2349 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2351 /* If we have multiple successors, we take the union of all of
2352 them. */
2353 else
2355 size_t i;
2356 basic_block bprime;
2358 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2359 FOR_EACH_EDGE (e, ei, block->succs)
2361 if (e->flags & EDGE_DFS_BACK)
2362 continue;
2363 worklist.quick_push (e->dest);
2365 if (worklist.length () > 0)
2367 FOR_EACH_VEC_ELT (worklist, i, bprime)
2369 unsigned int i;
2370 bitmap_iterator bi;
2372 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2373 bitmap_value_insert_into_set (PA_OUT,
2374 expression_for_id (i));
2375 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2377 bitmap_set_t pa_in = bitmap_set_new ();
2378 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2379 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2380 bitmap_value_insert_into_set (PA_OUT,
2381 expression_for_id (i));
2382 bitmap_set_free (pa_in);
2384 else
2385 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2386 bitmap_value_insert_into_set (PA_OUT,
2387 expression_for_id (i));
2392 /* Prune expressions that are clobbered in block and thus become
2393 invalid if translated from PA_OUT to PA_IN. */
2394 prune_clobbered_mems (PA_OUT, block);
2396 /* PA_IN starts with PA_OUT - TMP_GEN.
2397 Then we subtract things from ANTIC_IN. */
2398 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2400 /* For partial antic, we want to put back in the phi results, since
2401 we will properly avoid making them partially antic over backedges. */
2402 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2403 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2405 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2406 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2408 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2410 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2412 changed = true;
2413 bitmap_set_bit (changed_blocks, block->index);
2414 FOR_EACH_EDGE (e, ei, block->preds)
2415 bitmap_set_bit (changed_blocks, e->src->index);
2417 else
2418 bitmap_clear_bit (changed_blocks, block->index);
2420 maybe_dump_sets:
2421 if (dump_file && (dump_flags & TDF_DETAILS))
2423 if (PA_OUT)
2424 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2426 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2428 if (old_PA_IN)
2429 bitmap_set_free (old_PA_IN);
2430 if (PA_OUT)
2431 bitmap_set_free (PA_OUT);
2432 return changed;
2435 /* Compute ANTIC and partial ANTIC sets. */
2437 static void
2438 compute_antic (void)
2440 bool changed = true;
2441 int num_iterations = 0;
2442 basic_block block;
2443 int i;
2445 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2446 We pre-build the map of blocks with incoming abnormal edges here. */
2447 has_abnormal_preds = sbitmap_alloc (last_basic_block_for_fn (cfun));
2448 bitmap_clear (has_abnormal_preds);
2450 FOR_ALL_BB_FN (block, cfun)
2452 edge_iterator ei;
2453 edge e;
2455 FOR_EACH_EDGE (e, ei, block->preds)
2457 e->flags &= ~EDGE_DFS_BACK;
2458 if (e->flags & EDGE_ABNORMAL)
2460 bitmap_set_bit (has_abnormal_preds, block->index);
2461 break;
2465 BB_VISITED (block) = 0;
2466 BB_DEFERRED (block) = 0;
2468 /* While we are here, give empty ANTIC_IN sets to each block. */
2469 ANTIC_IN (block) = bitmap_set_new ();
2470 PA_IN (block) = bitmap_set_new ();
2473 /* At the exit block we anticipate nothing. */
2474 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun)) = 1;
2476 changed_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun) + 1);
2477 bitmap_ones (changed_blocks);
2478 while (changed)
2480 if (dump_file && (dump_flags & TDF_DETAILS))
2481 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2482 /* ??? We need to clear our PHI translation cache here as the
2483 ANTIC sets shrink and we restrict valid translations to
2484 those having operands with leaders in ANTIC. Same below
2485 for PA ANTIC computation. */
2486 num_iterations++;
2487 changed = false;
2488 for (i = postorder_num - 1; i >= 0; i--)
2490 if (bitmap_bit_p (changed_blocks, postorder[i]))
2492 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2493 changed |= compute_antic_aux (block,
2494 bitmap_bit_p (has_abnormal_preds,
2495 block->index));
2498 /* Theoretically possible, but *highly* unlikely. */
2499 gcc_checking_assert (num_iterations < 500);
2502 statistics_histogram_event (cfun, "compute_antic iterations",
2503 num_iterations);
2505 if (do_partial_partial)
2507 bitmap_ones (changed_blocks);
2508 mark_dfs_back_edges ();
2509 num_iterations = 0;
2510 changed = true;
2511 while (changed)
2513 if (dump_file && (dump_flags & TDF_DETAILS))
2514 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2515 num_iterations++;
2516 changed = false;
2517 for (i = postorder_num - 1 ; i >= 0; i--)
2519 if (bitmap_bit_p (changed_blocks, postorder[i]))
2521 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2522 changed
2523 |= compute_partial_antic_aux (block,
2524 bitmap_bit_p (has_abnormal_preds,
2525 block->index));
2528 /* Theoretically possible, but *highly* unlikely. */
2529 gcc_checking_assert (num_iterations < 500);
2531 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2532 num_iterations);
2534 sbitmap_free (has_abnormal_preds);
2535 sbitmap_free (changed_blocks);
2539 /* Inserted expressions are placed onto this worklist, which is used
2540 for performing quick dead code elimination of insertions we made
2541 that didn't turn out to be necessary. */
2542 static bitmap inserted_exprs;
2544 /* The actual worker for create_component_ref_by_pieces. */
2546 static tree
2547 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2548 unsigned int *operand, gimple_seq *stmts)
2550 vn_reference_op_t currop = &ref->operands[*operand];
2551 tree genop;
2552 ++*operand;
2553 switch (currop->opcode)
2555 case CALL_EXPR:
2557 tree folded, sc = NULL_TREE;
2558 unsigned int nargs = 0;
2559 tree fn, *args;
2560 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2561 fn = currop->op0;
2562 else
2563 fn = find_or_generate_expression (block, currop->op0, stmts);
2564 if (!fn)
2565 return NULL_TREE;
2566 if (currop->op1)
2568 sc = find_or_generate_expression (block, currop->op1, stmts);
2569 if (!sc)
2570 return NULL_TREE;
2572 args = XNEWVEC (tree, ref->operands.length () - 1);
2573 while (*operand < ref->operands.length ())
2575 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2576 operand, stmts);
2577 if (!args[nargs])
2578 return NULL_TREE;
2579 nargs++;
2581 folded = build_call_array (currop->type,
2582 (TREE_CODE (fn) == FUNCTION_DECL
2583 ? build_fold_addr_expr (fn) : fn),
2584 nargs, args);
2585 free (args);
2586 if (sc)
2587 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2588 return folded;
2591 case MEM_REF:
2593 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2594 stmts);
2595 if (!baseop)
2596 return NULL_TREE;
2597 tree offset = currop->op0;
2598 if (TREE_CODE (baseop) == ADDR_EXPR
2599 && handled_component_p (TREE_OPERAND (baseop, 0)))
2601 HOST_WIDE_INT off;
2602 tree base;
2603 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2604 &off);
2605 gcc_assert (base);
2606 offset = int_const_binop (PLUS_EXPR, offset,
2607 build_int_cst (TREE_TYPE (offset),
2608 off));
2609 baseop = build_fold_addr_expr (base);
2611 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2614 case TARGET_MEM_REF:
2616 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2617 vn_reference_op_t nextop = &ref->operands[++*operand];
2618 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2619 stmts);
2620 if (!baseop)
2621 return NULL_TREE;
2622 if (currop->op0)
2624 genop0 = find_or_generate_expression (block, currop->op0, stmts);
2625 if (!genop0)
2626 return NULL_TREE;
2628 if (nextop->op0)
2630 genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2631 if (!genop1)
2632 return NULL_TREE;
2634 return build5 (TARGET_MEM_REF, currop->type,
2635 baseop, currop->op2, genop0, currop->op1, genop1);
2638 case ADDR_EXPR:
2639 if (currop->op0)
2641 gcc_assert (is_gimple_min_invariant (currop->op0));
2642 return currop->op0;
2644 /* Fallthrough. */
2645 case REALPART_EXPR:
2646 case IMAGPART_EXPR:
2647 case VIEW_CONVERT_EXPR:
2649 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2650 stmts);
2651 if (!genop0)
2652 return NULL_TREE;
2653 return fold_build1 (currop->opcode, currop->type, genop0);
2656 case WITH_SIZE_EXPR:
2658 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2659 stmts);
2660 if (!genop0)
2661 return NULL_TREE;
2662 tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2663 if (!genop1)
2664 return NULL_TREE;
2665 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2668 case BIT_FIELD_REF:
2670 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2671 stmts);
2672 if (!genop0)
2673 return NULL_TREE;
2674 tree op1 = currop->op0;
2675 tree op2 = currop->op1;
2676 return fold_build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2679 /* For array ref vn_reference_op's, operand 1 of the array ref
2680 is op0 of the reference op and operand 3 of the array ref is
2681 op1. */
2682 case ARRAY_RANGE_REF:
2683 case ARRAY_REF:
2685 tree genop0;
2686 tree genop1 = currop->op0;
2687 tree genop2 = currop->op1;
2688 tree genop3 = currop->op2;
2689 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2690 stmts);
2691 if (!genop0)
2692 return NULL_TREE;
2693 genop1 = find_or_generate_expression (block, genop1, stmts);
2694 if (!genop1)
2695 return NULL_TREE;
2696 if (genop2)
2698 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2699 /* Drop zero minimum index if redundant. */
2700 if (integer_zerop (genop2)
2701 && (!domain_type
2702 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2703 genop2 = NULL_TREE;
2704 else
2706 genop2 = find_or_generate_expression (block, genop2, stmts);
2707 if (!genop2)
2708 return NULL_TREE;
2711 if (genop3)
2713 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2714 /* We can't always put a size in units of the element alignment
2715 here as the element alignment may be not visible. See
2716 PR43783. Simply drop the element size for constant
2717 sizes. */
2718 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2719 genop3 = NULL_TREE;
2720 else
2722 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2723 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2724 genop3 = find_or_generate_expression (block, genop3, stmts);
2725 if (!genop3)
2726 return NULL_TREE;
2729 return build4 (currop->opcode, currop->type, genop0, genop1,
2730 genop2, genop3);
2732 case COMPONENT_REF:
2734 tree op0;
2735 tree op1;
2736 tree genop2 = currop->op1;
2737 op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2738 if (!op0)
2739 return NULL_TREE;
2740 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2741 op1 = currop->op0;
2742 if (genop2)
2744 genop2 = find_or_generate_expression (block, genop2, stmts);
2745 if (!genop2)
2746 return NULL_TREE;
2748 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2751 case SSA_NAME:
2753 genop = find_or_generate_expression (block, currop->op0, stmts);
2754 return genop;
2756 case STRING_CST:
2757 case INTEGER_CST:
2758 case COMPLEX_CST:
2759 case VECTOR_CST:
2760 case REAL_CST:
2761 case CONSTRUCTOR:
2762 case VAR_DECL:
2763 case PARM_DECL:
2764 case CONST_DECL:
2765 case RESULT_DECL:
2766 case FUNCTION_DECL:
2767 return currop->op0;
2769 default:
2770 gcc_unreachable ();
2774 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2775 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2776 trying to rename aggregates into ssa form directly, which is a no no.
2778 Thus, this routine doesn't create temporaries, it just builds a
2779 single access expression for the array, calling
2780 find_or_generate_expression to build the innermost pieces.
2782 This function is a subroutine of create_expression_by_pieces, and
2783 should not be called on it's own unless you really know what you
2784 are doing. */
2786 static tree
2787 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2788 gimple_seq *stmts)
2790 unsigned int op = 0;
2791 return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2794 /* Find a simple leader for an expression, or generate one using
2795 create_expression_by_pieces from a NARY expression for the value.
2796 BLOCK is the basic_block we are looking for leaders in.
2797 OP is the tree expression to find a leader for or generate.
2798 Returns the leader or NULL_TREE on failure. */
2800 static tree
2801 find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2803 pre_expr expr = get_or_alloc_expr_for (op);
2804 unsigned int lookfor = get_expr_value_id (expr);
2805 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2806 if (leader)
2808 if (leader->kind == NAME)
2809 return PRE_EXPR_NAME (leader);
2810 else if (leader->kind == CONSTANT)
2811 return PRE_EXPR_CONSTANT (leader);
2813 /* Defer. */
2814 return NULL_TREE;
2817 /* It must be a complex expression, so generate it recursively. Note
2818 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2819 where the insert algorithm fails to insert a required expression. */
2820 bitmap exprset = value_expressions[lookfor];
2821 bitmap_iterator bi;
2822 unsigned int i;
2823 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2825 pre_expr temp = expression_for_id (i);
2826 /* We cannot insert random REFERENCE expressions at arbitrary
2827 places. We can insert NARYs which eventually re-materializes
2828 its operand values. */
2829 if (temp->kind == NARY)
2830 return create_expression_by_pieces (block, temp, stmts,
2831 get_expr_type (expr));
2834 /* Defer. */
2835 return NULL_TREE;
2838 #define NECESSARY GF_PLF_1
2840 /* Create an expression in pieces, so that we can handle very complex
2841 expressions that may be ANTIC, but not necessary GIMPLE.
2842 BLOCK is the basic block the expression will be inserted into,
2843 EXPR is the expression to insert (in value form)
2844 STMTS is a statement list to append the necessary insertions into.
2846 This function will die if we hit some value that shouldn't be
2847 ANTIC but is (IE there is no leader for it, or its components).
2848 The function returns NULL_TREE in case a different antic expression
2849 has to be inserted first.
2850 This function may also generate expressions that are themselves
2851 partially or fully redundant. Those that are will be either made
2852 fully redundant during the next iteration of insert (for partially
2853 redundant ones), or eliminated by eliminate (for fully redundant
2854 ones). */
2856 static tree
2857 create_expression_by_pieces (basic_block block, pre_expr expr,
2858 gimple_seq *stmts, tree type)
2860 tree name;
2861 tree folded;
2862 gimple_seq forced_stmts = NULL;
2863 unsigned int value_id;
2864 gimple_stmt_iterator gsi;
2865 tree exprtype = type ? type : get_expr_type (expr);
2866 pre_expr nameexpr;
2867 gimple newstmt;
2869 switch (expr->kind)
2871 /* We may hit the NAME/CONSTANT case if we have to convert types
2872 that value numbering saw through. */
2873 case NAME:
2874 folded = PRE_EXPR_NAME (expr);
2875 break;
2876 case CONSTANT:
2877 folded = PRE_EXPR_CONSTANT (expr);
2878 break;
2879 case REFERENCE:
2881 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2882 folded = create_component_ref_by_pieces (block, ref, stmts);
2883 if (!folded)
2884 return NULL_TREE;
2886 break;
2887 case NARY:
2889 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2890 tree *genop = XALLOCAVEC (tree, nary->length);
2891 unsigned i;
2892 for (i = 0; i < nary->length; ++i)
2894 genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2895 if (!genop[i])
2896 return NULL_TREE;
2897 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2898 may have conversions stripped. */
2899 if (nary->opcode == POINTER_PLUS_EXPR)
2901 if (i == 0)
2902 genop[i] = fold_convert (nary->type, genop[i]);
2903 else if (i == 1)
2904 genop[i] = convert_to_ptrofftype (genop[i]);
2906 else
2907 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]);
2909 if (nary->opcode == CONSTRUCTOR)
2911 vec<constructor_elt, va_gc> *elts = NULL;
2912 for (i = 0; i < nary->length; ++i)
2913 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2914 folded = build_constructor (nary->type, elts);
2916 else
2918 switch (nary->length)
2920 case 1:
2921 folded = fold_build1 (nary->opcode, nary->type,
2922 genop[0]);
2923 break;
2924 case 2:
2925 folded = fold_build2 (nary->opcode, nary->type,
2926 genop[0], genop[1]);
2927 break;
2928 case 3:
2929 folded = fold_build3 (nary->opcode, nary->type,
2930 genop[0], genop[1], genop[2]);
2931 break;
2932 default:
2933 gcc_unreachable ();
2937 break;
2938 default:
2939 gcc_unreachable ();
2942 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2943 folded = fold_convert (exprtype, folded);
2945 /* Force the generated expression to be a sequence of GIMPLE
2946 statements.
2947 We have to call unshare_expr because force_gimple_operand may
2948 modify the tree we pass to it. */
2949 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
2950 false, NULL);
2952 /* If we have any intermediate expressions to the value sets, add them
2953 to the value sets and chain them in the instruction stream. */
2954 if (forced_stmts)
2956 gsi = gsi_start (forced_stmts);
2957 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2959 gimple stmt = gsi_stmt (gsi);
2960 tree forcedname = gimple_get_lhs (stmt);
2961 pre_expr nameexpr;
2963 if (TREE_CODE (forcedname) == SSA_NAME)
2965 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
2966 VN_INFO_GET (forcedname)->valnum = forcedname;
2967 VN_INFO (forcedname)->value_id = get_next_value_id ();
2968 nameexpr = get_or_alloc_expr_for_name (forcedname);
2969 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2970 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2971 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2974 gimple_seq_add_seq (stmts, forced_stmts);
2977 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2978 newstmt = gimple_build_assign (name, folded);
2979 gimple_set_plf (newstmt, NECESSARY, false);
2981 gimple_seq_add_stmt (stmts, newstmt);
2982 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
2984 /* Fold the last statement. */
2985 gsi = gsi_last (*stmts);
2986 if (fold_stmt_inplace (&gsi))
2987 update_stmt (gsi_stmt (gsi));
2989 /* Add a value number to the temporary.
2990 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2991 we are creating the expression by pieces, and this particular piece of
2992 the expression may have been represented. There is no harm in replacing
2993 here. */
2994 value_id = get_expr_value_id (expr);
2995 VN_INFO_GET (name)->value_id = value_id;
2996 VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id);
2997 if (VN_INFO (name)->valnum == NULL_TREE)
2998 VN_INFO (name)->valnum = name;
2999 gcc_assert (VN_INFO (name)->valnum != NULL_TREE);
3000 nameexpr = get_or_alloc_expr_for_name (name);
3001 add_to_value (value_id, nameexpr);
3002 if (NEW_SETS (block))
3003 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3004 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3006 pre_stats.insertions++;
3007 if (dump_file && (dump_flags & TDF_DETAILS))
3009 fprintf (dump_file, "Inserted ");
3010 print_gimple_stmt (dump_file, newstmt, 0, 0);
3011 fprintf (dump_file, " in predecessor %d (%04d)\n",
3012 block->index, value_id);
3015 return name;
3019 /* Insert the to-be-made-available values of expression EXPRNUM for each
3020 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3021 merge the result with a phi node, given the same value number as
3022 NODE. Return true if we have inserted new stuff. */
3024 static bool
3025 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3026 vec<pre_expr> avail)
3028 pre_expr expr = expression_for_id (exprnum);
3029 pre_expr newphi;
3030 unsigned int val = get_expr_value_id (expr);
3031 edge pred;
3032 bool insertions = false;
3033 bool nophi = false;
3034 basic_block bprime;
3035 pre_expr eprime;
3036 edge_iterator ei;
3037 tree type = get_expr_type (expr);
3038 tree temp;
3039 gimple phi;
3041 /* Make sure we aren't creating an induction variable. */
3042 if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
3044 bool firstinsideloop = false;
3045 bool secondinsideloop = false;
3046 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3047 EDGE_PRED (block, 0)->src);
3048 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3049 EDGE_PRED (block, 1)->src);
3050 /* Induction variables only have one edge inside the loop. */
3051 if ((firstinsideloop ^ secondinsideloop)
3052 && expr->kind != REFERENCE)
3054 if (dump_file && (dump_flags & TDF_DETAILS))
3055 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3056 nophi = true;
3060 /* Make the necessary insertions. */
3061 FOR_EACH_EDGE (pred, ei, block->preds)
3063 gimple_seq stmts = NULL;
3064 tree builtexpr;
3065 bprime = pred->src;
3066 eprime = avail[pred->dest_idx];
3068 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3070 builtexpr = create_expression_by_pieces (bprime, eprime,
3071 &stmts, type);
3072 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3073 gsi_insert_seq_on_edge (pred, stmts);
3074 if (!builtexpr)
3076 /* We cannot insert a PHI node if we failed to insert
3077 on one edge. */
3078 nophi = true;
3079 continue;
3081 avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3082 insertions = true;
3084 else if (eprime->kind == CONSTANT)
3086 /* Constants may not have the right type, fold_convert
3087 should give us back a constant with the right type. */
3088 tree constant = PRE_EXPR_CONSTANT (eprime);
3089 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3091 tree builtexpr = fold_convert (type, constant);
3092 if (!is_gimple_min_invariant (builtexpr))
3094 tree forcedexpr = force_gimple_operand (builtexpr,
3095 &stmts, true,
3096 NULL);
3097 if (!is_gimple_min_invariant (forcedexpr))
3099 if (forcedexpr != builtexpr)
3101 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3102 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3104 if (stmts)
3106 gimple_stmt_iterator gsi;
3107 gsi = gsi_start (stmts);
3108 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3110 gimple stmt = gsi_stmt (gsi);
3111 tree lhs = gimple_get_lhs (stmt);
3112 if (TREE_CODE (lhs) == SSA_NAME)
3113 bitmap_set_bit (inserted_exprs,
3114 SSA_NAME_VERSION (lhs));
3115 gimple_set_plf (stmt, NECESSARY, false);
3117 gsi_insert_seq_on_edge (pred, stmts);
3119 avail[pred->dest_idx]
3120 = get_or_alloc_expr_for_name (forcedexpr);
3123 else
3124 avail[pred->dest_idx]
3125 = get_or_alloc_expr_for_constant (builtexpr);
3128 else if (eprime->kind == NAME)
3130 /* We may have to do a conversion because our value
3131 numbering can look through types in certain cases, but
3132 our IL requires all operands of a phi node have the same
3133 type. */
3134 tree name = PRE_EXPR_NAME (eprime);
3135 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3137 tree builtexpr;
3138 tree forcedexpr;
3139 builtexpr = fold_convert (type, name);
3140 forcedexpr = force_gimple_operand (builtexpr,
3141 &stmts, true,
3142 NULL);
3144 if (forcedexpr != name)
3146 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3147 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3150 if (stmts)
3152 gimple_stmt_iterator gsi;
3153 gsi = gsi_start (stmts);
3154 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3156 gimple stmt = gsi_stmt (gsi);
3157 tree lhs = gimple_get_lhs (stmt);
3158 if (TREE_CODE (lhs) == SSA_NAME)
3159 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3160 gimple_set_plf (stmt, NECESSARY, false);
3162 gsi_insert_seq_on_edge (pred, stmts);
3164 avail[pred->dest_idx] = get_or_alloc_expr_for_name (forcedexpr);
3168 /* If we didn't want a phi node, and we made insertions, we still have
3169 inserted new stuff, and thus return true. If we didn't want a phi node,
3170 and didn't make insertions, we haven't added anything new, so return
3171 false. */
3172 if (nophi && insertions)
3173 return true;
3174 else if (nophi && !insertions)
3175 return false;
3177 /* Now build a phi for the new variable. */
3178 temp = make_temp_ssa_name (type, NULL, "prephitmp");
3179 phi = create_phi_node (temp, block);
3181 gimple_set_plf (phi, NECESSARY, false);
3182 VN_INFO_GET (temp)->value_id = val;
3183 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3184 if (VN_INFO (temp)->valnum == NULL_TREE)
3185 VN_INFO (temp)->valnum = temp;
3186 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3187 FOR_EACH_EDGE (pred, ei, block->preds)
3189 pre_expr ae = avail[pred->dest_idx];
3190 gcc_assert (get_expr_type (ae) == type
3191 || useless_type_conversion_p (type, get_expr_type (ae)));
3192 if (ae->kind == CONSTANT)
3193 add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3194 pred, UNKNOWN_LOCATION);
3195 else
3196 add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3199 newphi = get_or_alloc_expr_for_name (temp);
3200 add_to_value (val, newphi);
3202 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3203 this insertion, since we test for the existence of this value in PHI_GEN
3204 before proceeding with the partial redundancy checks in insert_aux.
3206 The value may exist in AVAIL_OUT, in particular, it could be represented
3207 by the expression we are trying to eliminate, in which case we want the
3208 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3209 inserted there.
3211 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3212 this block, because if it did, it would have existed in our dominator's
3213 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3216 bitmap_insert_into_set (PHI_GEN (block), newphi);
3217 bitmap_value_replace_in_set (AVAIL_OUT (block),
3218 newphi);
3219 bitmap_insert_into_set (NEW_SETS (block),
3220 newphi);
3222 if (dump_file && (dump_flags & TDF_DETAILS))
3224 fprintf (dump_file, "Created phi ");
3225 print_gimple_stmt (dump_file, phi, 0, 0);
3226 fprintf (dump_file, " in block %d (%04d)\n", block->index, val);
3228 pre_stats.phis++;
3229 return true;
3234 /* Perform insertion of partially redundant values.
3235 For BLOCK, do the following:
3236 1. Propagate the NEW_SETS of the dominator into the current block.
3237 If the block has multiple predecessors,
3238 2a. Iterate over the ANTIC expressions for the block to see if
3239 any of them are partially redundant.
3240 2b. If so, insert them into the necessary predecessors to make
3241 the expression fully redundant.
3242 2c. Insert a new PHI merging the values of the predecessors.
3243 2d. Insert the new PHI, and the new expressions, into the
3244 NEW_SETS set.
3245 3. Recursively call ourselves on the dominator children of BLOCK.
3247 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3248 do_regular_insertion and do_partial_insertion.
3252 static bool
3253 do_regular_insertion (basic_block block, basic_block dom)
3255 bool new_stuff = false;
3256 vec<pre_expr> exprs;
3257 pre_expr expr;
3258 vec<pre_expr> avail = vNULL;
3259 int i;
3261 exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3262 avail.safe_grow (EDGE_COUNT (block->preds));
3264 FOR_EACH_VEC_ELT (exprs, i, expr)
3266 if (expr->kind == NARY
3267 || expr->kind == REFERENCE)
3269 unsigned int val;
3270 bool by_some = false;
3271 bool cant_insert = false;
3272 bool all_same = true;
3273 pre_expr first_s = NULL;
3274 edge pred;
3275 basic_block bprime;
3276 pre_expr eprime = NULL;
3277 edge_iterator ei;
3278 pre_expr edoubleprime = NULL;
3279 bool do_insertion = false;
3281 val = get_expr_value_id (expr);
3282 if (bitmap_set_contains_value (PHI_GEN (block), val))
3283 continue;
3284 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3286 if (dump_file && (dump_flags & TDF_DETAILS))
3288 fprintf (dump_file, "Found fully redundant value: ");
3289 print_pre_expr (dump_file, expr);
3290 fprintf (dump_file, "\n");
3292 continue;
3295 FOR_EACH_EDGE (pred, ei, block->preds)
3297 unsigned int vprime;
3299 /* We should never run insertion for the exit block
3300 and so not come across fake pred edges. */
3301 gcc_assert (!(pred->flags & EDGE_FAKE));
3302 bprime = pred->src;
3303 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3304 bprime, block);
3306 /* eprime will generally only be NULL if the
3307 value of the expression, translated
3308 through the PHI for this predecessor, is
3309 undefined. If that is the case, we can't
3310 make the expression fully redundant,
3311 because its value is undefined along a
3312 predecessor path. We can thus break out
3313 early because it doesn't matter what the
3314 rest of the results are. */
3315 if (eprime == NULL)
3317 avail[pred->dest_idx] = NULL;
3318 cant_insert = true;
3319 break;
3322 eprime = fully_constant_expression (eprime);
3323 vprime = get_expr_value_id (eprime);
3324 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3325 vprime);
3326 if (edoubleprime == NULL)
3328 avail[pred->dest_idx] = eprime;
3329 all_same = false;
3331 else
3333 avail[pred->dest_idx] = edoubleprime;
3334 by_some = true;
3335 /* We want to perform insertions to remove a redundancy on
3336 a path in the CFG we want to optimize for speed. */
3337 if (optimize_edge_for_speed_p (pred))
3338 do_insertion = true;
3339 if (first_s == NULL)
3340 first_s = edoubleprime;
3341 else if (!pre_expr_d::equal (first_s, edoubleprime))
3342 all_same = false;
3345 /* If we can insert it, it's not the same value
3346 already existing along every predecessor, and
3347 it's defined by some predecessor, it is
3348 partially redundant. */
3349 if (!cant_insert && !all_same && by_some)
3351 if (!do_insertion)
3353 if (dump_file && (dump_flags & TDF_DETAILS))
3355 fprintf (dump_file, "Skipping partial redundancy for "
3356 "expression ");
3357 print_pre_expr (dump_file, expr);
3358 fprintf (dump_file, " (%04d), no redundancy on to be "
3359 "optimized for speed edge\n", val);
3362 else if (dbg_cnt (treepre_insert))
3364 if (dump_file && (dump_flags & TDF_DETAILS))
3366 fprintf (dump_file, "Found partial redundancy for "
3367 "expression ");
3368 print_pre_expr (dump_file, expr);
3369 fprintf (dump_file, " (%04d)\n",
3370 get_expr_value_id (expr));
3372 if (insert_into_preds_of_block (block,
3373 get_expression_id (expr),
3374 avail))
3375 new_stuff = true;
3378 /* If all edges produce the same value and that value is
3379 an invariant, then the PHI has the same value on all
3380 edges. Note this. */
3381 else if (!cant_insert && all_same)
3383 gcc_assert (edoubleprime->kind == CONSTANT
3384 || edoubleprime->kind == NAME);
3386 tree temp = make_temp_ssa_name (get_expr_type (expr),
3387 NULL, "pretmp");
3388 gimple assign = gimple_build_assign (temp,
3389 edoubleprime->kind == CONSTANT ? PRE_EXPR_CONSTANT (edoubleprime) : PRE_EXPR_NAME (edoubleprime));
3390 gimple_stmt_iterator gsi = gsi_after_labels (block);
3391 gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3393 gimple_set_plf (assign, NECESSARY, false);
3394 VN_INFO_GET (temp)->value_id = val;
3395 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3396 if (VN_INFO (temp)->valnum == NULL_TREE)
3397 VN_INFO (temp)->valnum = temp;
3398 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3399 pre_expr newe = get_or_alloc_expr_for_name (temp);
3400 add_to_value (val, newe);
3401 bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3402 bitmap_insert_into_set (NEW_SETS (block), newe);
3407 exprs.release ();
3408 return new_stuff;
3412 /* Perform insertion for partially anticipatable expressions. There
3413 is only one case we will perform insertion for these. This case is
3414 if the expression is partially anticipatable, and fully available.
3415 In this case, we know that putting it earlier will enable us to
3416 remove the later computation. */
3419 static bool
3420 do_partial_partial_insertion (basic_block block, basic_block dom)
3422 bool new_stuff = false;
3423 vec<pre_expr> exprs;
3424 pre_expr expr;
3425 auto_vec<pre_expr> avail;
3426 int i;
3428 exprs = sorted_array_from_bitmap_set (PA_IN (block));
3429 avail.safe_grow (EDGE_COUNT (block->preds));
3431 FOR_EACH_VEC_ELT (exprs, i, expr)
3433 if (expr->kind == NARY
3434 || expr->kind == REFERENCE)
3436 unsigned int val;
3437 bool by_all = true;
3438 bool cant_insert = false;
3439 edge pred;
3440 basic_block bprime;
3441 pre_expr eprime = NULL;
3442 edge_iterator ei;
3444 val = get_expr_value_id (expr);
3445 if (bitmap_set_contains_value (PHI_GEN (block), val))
3446 continue;
3447 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3448 continue;
3450 FOR_EACH_EDGE (pred, ei, block->preds)
3452 unsigned int vprime;
3453 pre_expr edoubleprime;
3455 /* We should never run insertion for the exit block
3456 and so not come across fake pred edges. */
3457 gcc_assert (!(pred->flags & EDGE_FAKE));
3458 bprime = pred->src;
3459 eprime = phi_translate (expr, ANTIC_IN (block),
3460 PA_IN (block),
3461 bprime, block);
3463 /* eprime will generally only be NULL if the
3464 value of the expression, translated
3465 through the PHI for this predecessor, is
3466 undefined. If that is the case, we can't
3467 make the expression fully redundant,
3468 because its value is undefined along a
3469 predecessor path. We can thus break out
3470 early because it doesn't matter what the
3471 rest of the results are. */
3472 if (eprime == NULL)
3474 avail[pred->dest_idx] = NULL;
3475 cant_insert = true;
3476 break;
3479 eprime = fully_constant_expression (eprime);
3480 vprime = get_expr_value_id (eprime);
3481 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3482 avail[pred->dest_idx] = edoubleprime;
3483 if (edoubleprime == NULL)
3485 by_all = false;
3486 break;
3490 /* If we can insert it, it's not the same value
3491 already existing along every predecessor, and
3492 it's defined by some predecessor, it is
3493 partially redundant. */
3494 if (!cant_insert && by_all)
3496 edge succ;
3497 bool do_insertion = false;
3499 /* Insert only if we can remove a later expression on a path
3500 that we want to optimize for speed.
3501 The phi node that we will be inserting in BLOCK is not free,
3502 and inserting it for the sake of !optimize_for_speed successor
3503 may cause regressions on the speed path. */
3504 FOR_EACH_EDGE (succ, ei, block->succs)
3506 if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3507 || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3509 if (optimize_edge_for_speed_p (succ))
3510 do_insertion = true;
3514 if (!do_insertion)
3516 if (dump_file && (dump_flags & TDF_DETAILS))
3518 fprintf (dump_file, "Skipping partial partial redundancy "
3519 "for expression ");
3520 print_pre_expr (dump_file, expr);
3521 fprintf (dump_file, " (%04d), not (partially) anticipated "
3522 "on any to be optimized for speed edges\n", val);
3525 else if (dbg_cnt (treepre_insert))
3527 pre_stats.pa_insert++;
3528 if (dump_file && (dump_flags & TDF_DETAILS))
3530 fprintf (dump_file, "Found partial partial redundancy "
3531 "for expression ");
3532 print_pre_expr (dump_file, expr);
3533 fprintf (dump_file, " (%04d)\n",
3534 get_expr_value_id (expr));
3536 if (insert_into_preds_of_block (block,
3537 get_expression_id (expr),
3538 avail))
3539 new_stuff = true;
3545 exprs.release ();
3546 return new_stuff;
3549 static bool
3550 insert_aux (basic_block block)
3552 basic_block son;
3553 bool new_stuff = false;
3555 if (block)
3557 basic_block dom;
3558 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3559 if (dom)
3561 unsigned i;
3562 bitmap_iterator bi;
3563 bitmap_set_t newset = NEW_SETS (dom);
3564 if (newset)
3566 /* Note that we need to value_replace both NEW_SETS, and
3567 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3568 represented by some non-simple expression here that we want
3569 to replace it with. */
3570 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3572 pre_expr expr = expression_for_id (i);
3573 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3574 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3577 if (!single_pred_p (block))
3579 new_stuff |= do_regular_insertion (block, dom);
3580 if (do_partial_partial)
3581 new_stuff |= do_partial_partial_insertion (block, dom);
3585 for (son = first_dom_son (CDI_DOMINATORS, block);
3586 son;
3587 son = next_dom_son (CDI_DOMINATORS, son))
3589 new_stuff |= insert_aux (son);
3592 return new_stuff;
3595 /* Perform insertion of partially redundant values. */
3597 static void
3598 insert (void)
3600 bool new_stuff = true;
3601 basic_block bb;
3602 int num_iterations = 0;
3604 FOR_ALL_BB_FN (bb, cfun)
3605 NEW_SETS (bb) = bitmap_set_new ();
3607 while (new_stuff)
3609 num_iterations++;
3610 if (dump_file && dump_flags & TDF_DETAILS)
3611 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3612 new_stuff = insert_aux (ENTRY_BLOCK_PTR_FOR_FN (cfun));
3614 /* Clear the NEW sets before the next iteration. We have already
3615 fully propagated its contents. */
3616 if (new_stuff)
3617 FOR_ALL_BB_FN (bb, cfun)
3618 bitmap_set_free (NEW_SETS (bb));
3620 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3624 /* Compute the AVAIL set for all basic blocks.
3626 This function performs value numbering of the statements in each basic
3627 block. The AVAIL sets are built from information we glean while doing
3628 this value numbering, since the AVAIL sets contain only one entry per
3629 value.
3631 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3632 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3634 static void
3635 compute_avail (void)
3638 basic_block block, son;
3639 basic_block *worklist;
3640 size_t sp = 0;
3641 unsigned i;
3643 /* We pretend that default definitions are defined in the entry block.
3644 This includes function arguments and the static chain decl. */
3645 for (i = 1; i < num_ssa_names; ++i)
3647 tree name = ssa_name (i);
3648 pre_expr e;
3649 if (!name
3650 || !SSA_NAME_IS_DEFAULT_DEF (name)
3651 || has_zero_uses (name)
3652 || virtual_operand_p (name))
3653 continue;
3655 e = get_or_alloc_expr_for_name (name);
3656 add_to_value (get_expr_value_id (e), e);
3657 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)), e);
3658 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3662 if (dump_file && (dump_flags & TDF_DETAILS))
3664 print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3665 "tmp_gen", ENTRY_BLOCK);
3666 print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3667 "avail_out", ENTRY_BLOCK);
3670 /* Allocate the worklist. */
3671 worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
3673 /* Seed the algorithm by putting the dominator children of the entry
3674 block on the worklist. */
3675 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR_FOR_FN (cfun));
3676 son;
3677 son = next_dom_son (CDI_DOMINATORS, son))
3678 worklist[sp++] = son;
3680 /* Loop until the worklist is empty. */
3681 while (sp)
3683 gimple_stmt_iterator gsi;
3684 gimple stmt;
3685 basic_block dom;
3687 /* Pick a block from the worklist. */
3688 block = worklist[--sp];
3690 /* Initially, the set of available values in BLOCK is that of
3691 its immediate dominator. */
3692 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3693 if (dom)
3694 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3696 /* Generate values for PHI nodes. */
3697 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3699 tree result = gimple_phi_result (gsi_stmt (gsi));
3701 /* We have no need for virtual phis, as they don't represent
3702 actual computations. */
3703 if (virtual_operand_p (result))
3704 continue;
3706 pre_expr e = get_or_alloc_expr_for_name (result);
3707 add_to_value (get_expr_value_id (e), e);
3708 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3709 bitmap_insert_into_set (PHI_GEN (block), e);
3712 BB_MAY_NOTRETURN (block) = 0;
3714 /* Now compute value numbers and populate value sets with all
3715 the expressions computed in BLOCK. */
3716 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3718 ssa_op_iter iter;
3719 tree op;
3721 stmt = gsi_stmt (gsi);
3723 /* Cache whether the basic-block has any non-visible side-effect
3724 or control flow.
3725 If this isn't a call or it is the last stmt in the
3726 basic-block then the CFG represents things correctly. */
3727 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3729 /* Non-looping const functions always return normally.
3730 Otherwise the call might not return or have side-effects
3731 that forbids hoisting possibly trapping expressions
3732 before it. */
3733 int flags = gimple_call_flags (stmt);
3734 if (!(flags & ECF_CONST)
3735 || (flags & ECF_LOOPING_CONST_OR_PURE))
3736 BB_MAY_NOTRETURN (block) = 1;
3739 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3741 pre_expr e = get_or_alloc_expr_for_name (op);
3743 add_to_value (get_expr_value_id (e), e);
3744 bitmap_insert_into_set (TMP_GEN (block), e);
3745 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3748 if (gimple_has_side_effects (stmt)
3749 || stmt_could_throw_p (stmt)
3750 || is_gimple_debug (stmt))
3751 continue;
3753 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3755 if (ssa_undefined_value_p (op))
3756 continue;
3757 pre_expr e = get_or_alloc_expr_for_name (op);
3758 bitmap_value_insert_into_set (EXP_GEN (block), e);
3761 switch (gimple_code (stmt))
3763 case GIMPLE_RETURN:
3764 continue;
3766 case GIMPLE_CALL:
3768 vn_reference_t ref;
3769 pre_expr result = NULL;
3770 auto_vec<vn_reference_op_s> ops;
3772 /* We can value number only calls to real functions. */
3773 if (gimple_call_internal_p (stmt))
3774 continue;
3776 copy_reference_ops_from_call (stmt, &ops);
3777 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3778 gimple_expr_type (stmt),
3779 ops, &ref, VN_NOWALK);
3780 if (!ref)
3781 continue;
3783 /* If the value of the call is not invalidated in
3784 this block until it is computed, add the expression
3785 to EXP_GEN. */
3786 if (!gimple_vuse (stmt)
3787 || gimple_code
3788 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
3789 || gimple_bb (SSA_NAME_DEF_STMT
3790 (gimple_vuse (stmt))) != block)
3792 result = (pre_expr) pool_alloc (pre_expr_pool);
3793 result->kind = REFERENCE;
3794 result->id = 0;
3795 PRE_EXPR_REFERENCE (result) = ref;
3797 get_or_alloc_expression_id (result);
3798 add_to_value (get_expr_value_id (result), result);
3799 bitmap_value_insert_into_set (EXP_GEN (block), result);
3801 continue;
3804 case GIMPLE_ASSIGN:
3806 pre_expr result = NULL;
3807 switch (vn_get_stmt_kind (stmt))
3809 case VN_NARY:
3811 enum tree_code code = gimple_assign_rhs_code (stmt);
3812 vn_nary_op_t nary;
3814 /* COND_EXPR and VEC_COND_EXPR are awkward in
3815 that they contain an embedded complex expression.
3816 Don't even try to shove those through PRE. */
3817 if (code == COND_EXPR
3818 || code == VEC_COND_EXPR)
3819 continue;
3821 vn_nary_op_lookup_stmt (stmt, &nary);
3822 if (!nary)
3823 continue;
3825 /* If the NARY traps and there was a preceding
3826 point in the block that might not return avoid
3827 adding the nary to EXP_GEN. */
3828 if (BB_MAY_NOTRETURN (block)
3829 && vn_nary_may_trap (nary))
3830 continue;
3832 result = (pre_expr) pool_alloc (pre_expr_pool);
3833 result->kind = NARY;
3834 result->id = 0;
3835 PRE_EXPR_NARY (result) = nary;
3836 break;
3839 case VN_REFERENCE:
3841 vn_reference_t ref;
3842 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3843 gimple_vuse (stmt),
3844 VN_WALK, &ref);
3845 if (!ref)
3846 continue;
3848 /* If the value of the reference is not invalidated in
3849 this block until it is computed, add the expression
3850 to EXP_GEN. */
3851 if (gimple_vuse (stmt))
3853 gimple def_stmt;
3854 bool ok = true;
3855 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
3856 while (!gimple_nop_p (def_stmt)
3857 && gimple_code (def_stmt) != GIMPLE_PHI
3858 && gimple_bb (def_stmt) == block)
3860 if (stmt_may_clobber_ref_p
3861 (def_stmt, gimple_assign_rhs1 (stmt)))
3863 ok = false;
3864 break;
3866 def_stmt
3867 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
3869 if (!ok)
3870 continue;
3873 result = (pre_expr) pool_alloc (pre_expr_pool);
3874 result->kind = REFERENCE;
3875 result->id = 0;
3876 PRE_EXPR_REFERENCE (result) = ref;
3877 break;
3880 default:
3881 continue;
3884 get_or_alloc_expression_id (result);
3885 add_to_value (get_expr_value_id (result), result);
3886 bitmap_value_insert_into_set (EXP_GEN (block), result);
3887 continue;
3889 default:
3890 break;
3894 if (dump_file && (dump_flags & TDF_DETAILS))
3896 print_bitmap_set (dump_file, EXP_GEN (block),
3897 "exp_gen", block->index);
3898 print_bitmap_set (dump_file, PHI_GEN (block),
3899 "phi_gen", block->index);
3900 print_bitmap_set (dump_file, TMP_GEN (block),
3901 "tmp_gen", block->index);
3902 print_bitmap_set (dump_file, AVAIL_OUT (block),
3903 "avail_out", block->index);
3906 /* Put the dominator children of BLOCK on the worklist of blocks
3907 to compute available sets for. */
3908 for (son = first_dom_son (CDI_DOMINATORS, block);
3909 son;
3910 son = next_dom_son (CDI_DOMINATORS, son))
3911 worklist[sp++] = son;
3914 free (worklist);
3918 /* Local state for the eliminate domwalk. */
3919 static vec<gimple> el_to_remove;
3920 static unsigned int el_todo;
3921 static vec<tree> el_avail;
3922 static vec<tree> el_avail_stack;
3924 /* Return a leader for OP that is available at the current point of the
3925 eliminate domwalk. */
3927 static tree
3928 eliminate_avail (tree op)
3930 tree valnum = VN_INFO (op)->valnum;
3931 if (TREE_CODE (valnum) == SSA_NAME)
3933 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
3934 return valnum;
3935 if (el_avail.length () > SSA_NAME_VERSION (valnum))
3936 return el_avail[SSA_NAME_VERSION (valnum)];
3938 else if (is_gimple_min_invariant (valnum))
3939 return valnum;
3940 return NULL_TREE;
3943 /* At the current point of the eliminate domwalk make OP available. */
3945 static void
3946 eliminate_push_avail (tree op)
3948 tree valnum = VN_INFO (op)->valnum;
3949 if (TREE_CODE (valnum) == SSA_NAME)
3951 if (el_avail.length () <= SSA_NAME_VERSION (valnum))
3952 el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1);
3953 el_avail[SSA_NAME_VERSION (valnum)] = op;
3954 el_avail_stack.safe_push (op);
3958 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
3959 the leader for the expression if insertion was successful. */
3961 static tree
3962 eliminate_insert (gimple_stmt_iterator *gsi, tree val)
3964 tree expr = vn_get_expr_for (val);
3965 if (!CONVERT_EXPR_P (expr)
3966 && TREE_CODE (expr) != VIEW_CONVERT_EXPR)
3967 return NULL_TREE;
3969 tree op = TREE_OPERAND (expr, 0);
3970 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op;
3971 if (!leader)
3972 return NULL_TREE;
3974 tree res = make_temp_ssa_name (TREE_TYPE (val), NULL, "pretmp");
3975 gimple tem = gimple_build_assign (res,
3976 fold_build1 (TREE_CODE (expr),
3977 TREE_TYPE (expr), leader));
3978 gsi_insert_before (gsi, tem, GSI_SAME_STMT);
3979 VN_INFO_GET (res)->valnum = val;
3981 if (TREE_CODE (leader) == SSA_NAME)
3982 gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true);
3984 pre_stats.insertions++;
3985 if (dump_file && (dump_flags & TDF_DETAILS))
3987 fprintf (dump_file, "Inserted ");
3988 print_gimple_stmt (dump_file, tem, 0, 0);
3991 return res;
3994 class eliminate_dom_walker : public dom_walker
3996 public:
3997 eliminate_dom_walker (cdi_direction direction, bool do_pre_)
3998 : dom_walker (direction), do_pre (do_pre_) {}
4000 virtual void before_dom_children (basic_block);
4001 virtual void after_dom_children (basic_block);
4003 bool do_pre;
4006 /* Perform elimination for the basic-block B during the domwalk. */
4008 void
4009 eliminate_dom_walker::before_dom_children (basic_block b)
4011 gimple_stmt_iterator gsi;
4012 gimple stmt;
4014 /* Mark new bb. */
4015 el_avail_stack.safe_push (NULL_TREE);
4017 /* ??? If we do nothing for unreachable blocks then this will confuse
4018 tailmerging. Eventually we can reduce its reliance on SCCVN now
4019 that we fully copy/constant-propagate (most) things. */
4021 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4023 gimple phi = gsi_stmt (gsi);
4024 tree res = PHI_RESULT (phi);
4026 if (virtual_operand_p (res))
4028 gsi_next (&gsi);
4029 continue;
4032 tree sprime = eliminate_avail (res);
4033 if (sprime
4034 && sprime != res)
4036 if (dump_file && (dump_flags & TDF_DETAILS))
4038 fprintf (dump_file, "Replaced redundant PHI node defining ");
4039 print_generic_expr (dump_file, res, 0);
4040 fprintf (dump_file, " with ");
4041 print_generic_expr (dump_file, sprime, 0);
4042 fprintf (dump_file, "\n");
4045 /* If we inserted this PHI node ourself, it's not an elimination. */
4046 if (inserted_exprs
4047 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4048 pre_stats.phis--;
4049 else
4050 pre_stats.eliminations++;
4052 /* If we will propagate into all uses don't bother to do
4053 anything. */
4054 if (may_propagate_copy (res, sprime))
4056 /* Mark the PHI for removal. */
4057 el_to_remove.safe_push (phi);
4058 gsi_next (&gsi);
4059 continue;
4062 remove_phi_node (&gsi, false);
4064 if (inserted_exprs
4065 && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4066 && TREE_CODE (sprime) == SSA_NAME)
4067 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4069 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4070 sprime = fold_convert (TREE_TYPE (res), sprime);
4071 gimple stmt = gimple_build_assign (res, sprime);
4072 /* ??? It cannot yet be necessary (DOM walk). */
4073 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4075 gimple_stmt_iterator gsi2 = gsi_after_labels (b);
4076 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4077 continue;
4080 eliminate_push_avail (res);
4081 gsi_next (&gsi);
4084 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4086 tree sprime = NULL_TREE;
4087 stmt = gsi_stmt (gsi);
4088 tree lhs = gimple_get_lhs (stmt);
4089 if (lhs && TREE_CODE (lhs) == SSA_NAME
4090 && !gimple_has_volatile_ops (stmt)
4091 /* See PR43491. Do not replace a global register variable when
4092 it is a the RHS of an assignment. Do replace local register
4093 variables since gcc does not guarantee a local variable will
4094 be allocated in register.
4095 ??? The fix isn't effective here. This should instead
4096 be ensured by not value-numbering them the same but treating
4097 them like volatiles? */
4098 && !(gimple_assign_single_p (stmt)
4099 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL
4100 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt))
4101 && is_global_var (gimple_assign_rhs1 (stmt)))))
4103 sprime = eliminate_avail (lhs);
4104 if (!sprime)
4106 /* If there is no existing usable leader but SCCVN thinks
4107 it has an expression it wants to use as replacement,
4108 insert that. */
4109 tree val = VN_INFO (lhs)->valnum;
4110 if (val != VN_TOP
4111 && TREE_CODE (val) == SSA_NAME
4112 && VN_INFO (val)->needs_insertion
4113 && VN_INFO (val)->expr != NULL_TREE
4114 && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE)
4115 eliminate_push_avail (sprime);
4118 /* If this now constitutes a copy duplicate points-to
4119 and range info appropriately. This is especially
4120 important for inserted code. See tree-ssa-copy.c
4121 for similar code. */
4122 if (sprime
4123 && TREE_CODE (sprime) == SSA_NAME)
4125 basic_block sprime_b = gimple_bb (SSA_NAME_DEF_STMT (sprime));
4126 if (POINTER_TYPE_P (TREE_TYPE (lhs))
4127 && SSA_NAME_PTR_INFO (lhs)
4128 && !SSA_NAME_PTR_INFO (sprime))
4130 duplicate_ssa_name_ptr_info (sprime,
4131 SSA_NAME_PTR_INFO (lhs));
4132 if (b != sprime_b)
4133 mark_ptr_info_alignment_unknown
4134 (SSA_NAME_PTR_INFO (sprime));
4136 else if (!POINTER_TYPE_P (TREE_TYPE (lhs))
4137 && SSA_NAME_RANGE_INFO (lhs)
4138 && !SSA_NAME_RANGE_INFO (sprime)
4139 && b == sprime_b)
4140 duplicate_ssa_name_range_info (sprime,
4141 SSA_NAME_RANGE_TYPE (lhs),
4142 SSA_NAME_RANGE_INFO (lhs));
4145 /* Inhibit the use of an inserted PHI on a loop header when
4146 the address of the memory reference is a simple induction
4147 variable. In other cases the vectorizer won't do anything
4148 anyway (either it's loop invariant or a complicated
4149 expression). */
4150 if (sprime
4151 && TREE_CODE (sprime) == SSA_NAME
4152 && do_pre
4153 && flag_tree_loop_vectorize
4154 && loop_outer (b->loop_father)
4155 && has_zero_uses (sprime)
4156 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))
4157 && gimple_assign_load_p (stmt))
4159 gimple def_stmt = SSA_NAME_DEF_STMT (sprime);
4160 basic_block def_bb = gimple_bb (def_stmt);
4161 if (gimple_code (def_stmt) == GIMPLE_PHI
4162 && b->loop_father->header == def_bb)
4164 ssa_op_iter iter;
4165 tree op;
4166 bool found = false;
4167 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4169 affine_iv iv;
4170 def_bb = gimple_bb (SSA_NAME_DEF_STMT (op));
4171 if (def_bb
4172 && flow_bb_inside_loop_p (b->loop_father, def_bb)
4173 && simple_iv (b->loop_father,
4174 b->loop_father, op, &iv, true))
4176 found = true;
4177 break;
4180 if (found)
4182 if (dump_file && (dump_flags & TDF_DETAILS))
4184 fprintf (dump_file, "Not replacing ");
4185 print_gimple_expr (dump_file, stmt, 0, 0);
4186 fprintf (dump_file, " with ");
4187 print_generic_expr (dump_file, sprime, 0);
4188 fprintf (dump_file, " which would add a loop"
4189 " carried dependence to loop %d\n",
4190 b->loop_father->num);
4192 /* Don't keep sprime available. */
4193 sprime = NULL_TREE;
4198 if (sprime)
4200 /* If we can propagate the value computed for LHS into
4201 all uses don't bother doing anything with this stmt. */
4202 if (may_propagate_copy (lhs, sprime))
4204 /* Mark it for removal. */
4205 el_to_remove.safe_push (stmt);
4207 /* ??? Don't count copy/constant propagations. */
4208 if (gimple_assign_single_p (stmt)
4209 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4210 || gimple_assign_rhs1 (stmt) == sprime))
4211 continue;
4213 if (dump_file && (dump_flags & TDF_DETAILS))
4215 fprintf (dump_file, "Replaced ");
4216 print_gimple_expr (dump_file, stmt, 0, 0);
4217 fprintf (dump_file, " with ");
4218 print_generic_expr (dump_file, sprime, 0);
4219 fprintf (dump_file, " in all uses of ");
4220 print_gimple_stmt (dump_file, stmt, 0, 0);
4223 pre_stats.eliminations++;
4224 continue;
4227 /* If this is an assignment from our leader (which
4228 happens in the case the value-number is a constant)
4229 then there is nothing to do. */
4230 if (gimple_assign_single_p (stmt)
4231 && sprime == gimple_assign_rhs1 (stmt))
4232 continue;
4234 /* Else replace its RHS. */
4235 bool can_make_abnormal_goto
4236 = is_gimple_call (stmt)
4237 && stmt_can_make_abnormal_goto (stmt);
4239 if (dump_file && (dump_flags & TDF_DETAILS))
4241 fprintf (dump_file, "Replaced ");
4242 print_gimple_expr (dump_file, stmt, 0, 0);
4243 fprintf (dump_file, " with ");
4244 print_generic_expr (dump_file, sprime, 0);
4245 fprintf (dump_file, " in ");
4246 print_gimple_stmt (dump_file, stmt, 0, 0);
4249 if (TREE_CODE (sprime) == SSA_NAME)
4250 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4251 NECESSARY, true);
4253 pre_stats.eliminations++;
4254 gimple orig_stmt = stmt;
4255 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4256 TREE_TYPE (sprime)))
4257 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4258 tree vdef = gimple_vdef (stmt);
4259 tree vuse = gimple_vuse (stmt);
4260 propagate_tree_value_into_stmt (&gsi, sprime);
4261 stmt = gsi_stmt (gsi);
4262 update_stmt (stmt);
4263 if (vdef != gimple_vdef (stmt))
4264 VN_INFO (vdef)->valnum = vuse;
4266 /* If we removed EH side-effects from the statement, clean
4267 its EH information. */
4268 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4270 bitmap_set_bit (need_eh_cleanup,
4271 gimple_bb (stmt)->index);
4272 if (dump_file && (dump_flags & TDF_DETAILS))
4273 fprintf (dump_file, " Removed EH side-effects.\n");
4276 /* Likewise for AB side-effects. */
4277 if (can_make_abnormal_goto
4278 && !stmt_can_make_abnormal_goto (stmt))
4280 bitmap_set_bit (need_ab_cleanup,
4281 gimple_bb (stmt)->index);
4282 if (dump_file && (dump_flags & TDF_DETAILS))
4283 fprintf (dump_file, " Removed AB side-effects.\n");
4286 continue;
4290 /* If the statement is a scalar store, see if the expression
4291 has the same value number as its rhs. If so, the store is
4292 dead. */
4293 if (gimple_assign_single_p (stmt)
4294 && !gimple_has_volatile_ops (stmt)
4295 && !is_gimple_reg (gimple_assign_lhs (stmt))
4296 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4297 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4299 tree val;
4300 tree rhs = gimple_assign_rhs1 (stmt);
4301 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4302 gimple_vuse (stmt), VN_WALK, NULL);
4303 if (TREE_CODE (rhs) == SSA_NAME)
4304 rhs = VN_INFO (rhs)->valnum;
4305 if (val
4306 && operand_equal_p (val, rhs, 0))
4308 if (dump_file && (dump_flags & TDF_DETAILS))
4310 fprintf (dump_file, "Deleted redundant store ");
4311 print_gimple_stmt (dump_file, stmt, 0, 0);
4314 /* Queue stmt for removal. */
4315 el_to_remove.safe_push (stmt);
4316 continue;
4320 bool can_make_abnormal_goto = stmt_can_make_abnormal_goto (stmt);
4321 bool was_noreturn = (is_gimple_call (stmt)
4322 && gimple_call_noreturn_p (stmt));
4323 tree vdef = gimple_vdef (stmt);
4324 tree vuse = gimple_vuse (stmt);
4326 /* If we didn't replace the whole stmt (or propagate the result
4327 into all uses), replace all uses on this stmt with their
4328 leaders. */
4329 use_operand_p use_p;
4330 ssa_op_iter iter;
4331 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
4333 tree use = USE_FROM_PTR (use_p);
4334 /* ??? The call code above leaves stmt operands un-updated. */
4335 if (TREE_CODE (use) != SSA_NAME)
4336 continue;
4337 tree sprime = eliminate_avail (use);
4338 if (sprime && sprime != use
4339 && may_propagate_copy (use, sprime)
4340 /* We substitute into debug stmts to avoid excessive
4341 debug temporaries created by removed stmts, but we need
4342 to avoid doing so for inserted sprimes as we never want
4343 to create debug temporaries for them. */
4344 && (!inserted_exprs
4345 || TREE_CODE (sprime) != SSA_NAME
4346 || !is_gimple_debug (stmt)
4347 || !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))))
4349 propagate_value (use_p, sprime);
4350 gimple_set_modified (stmt, true);
4351 if (TREE_CODE (sprime) == SSA_NAME
4352 && !is_gimple_debug (stmt))
4353 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4354 NECESSARY, true);
4358 /* Visit indirect calls and turn them into direct calls if
4359 possible using the devirtualization machinery. */
4360 if (is_gimple_call (stmt))
4362 tree fn = gimple_call_fn (stmt);
4363 if (fn
4364 && flag_devirtualize
4365 && virtual_method_call_p (fn))
4367 tree otr_type;
4368 HOST_WIDE_INT otr_token;
4369 ipa_polymorphic_call_context context;
4370 tree instance;
4371 bool final;
4373 instance = get_polymorphic_call_info (current_function_decl,
4375 &otr_type, &otr_token, &context, stmt);
4377 context.get_dynamic_type (instance, OBJ_TYPE_REF_OBJECT (fn), otr_type, stmt);
4379 vec <cgraph_node *>targets
4380 = possible_polymorphic_call_targets (obj_type_ref_class (fn),
4381 tree_to_uhwi
4382 (OBJ_TYPE_REF_TOKEN (fn)),
4383 context,
4384 &final);
4385 if (dump_enabled_p ())
4386 dump_possible_polymorphic_call_targets (dump_file,
4387 obj_type_ref_class (fn),
4388 tree_to_uhwi
4389 (OBJ_TYPE_REF_TOKEN (fn)),
4390 context);
4391 if (final && targets.length () <= 1 && dbg_cnt (devirt))
4393 tree fn;
4394 if (targets.length () == 1)
4395 fn = targets[0]->decl;
4396 else
4397 fn = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
4398 if (dump_enabled_p ())
4400 location_t loc = gimple_location_safe (stmt);
4401 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc,
4402 "converting indirect call to "
4403 "function %s\n",
4404 cgraph_node::get (fn)->name ());
4406 gimple_call_set_fndecl (stmt, fn);
4407 gimple_set_modified (stmt, true);
4409 else
4410 gcc_assert (!ipa_intraprocedural_devirtualization (stmt));
4414 if (gimple_modified_p (stmt))
4416 /* If a formerly non-invariant ADDR_EXPR is turned into an
4417 invariant one it was on a separate stmt. */
4418 if (gimple_assign_single_p (stmt)
4419 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR)
4420 recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
4421 gimple old_stmt = stmt;
4422 if (is_gimple_call (stmt))
4424 /* ??? Only fold calls inplace for now, this may create new
4425 SSA names which in turn will confuse free_scc_vn SSA name
4426 release code. */
4427 fold_stmt_inplace (&gsi);
4428 /* When changing a call into a noreturn call, cfg cleanup
4429 is needed to fix up the noreturn call. */
4430 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4431 el_todo |= TODO_cleanup_cfg;
4433 else
4435 fold_stmt (&gsi);
4436 stmt = gsi_stmt (gsi);
4437 if ((gimple_code (stmt) == GIMPLE_COND
4438 && (gimple_cond_true_p (stmt)
4439 || gimple_cond_false_p (stmt)))
4440 || (gimple_code (stmt) == GIMPLE_SWITCH
4441 && TREE_CODE (gimple_switch_index (stmt)) == INTEGER_CST))
4442 el_todo |= TODO_cleanup_cfg;
4444 /* If we removed EH side-effects from the statement, clean
4445 its EH information. */
4446 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
4448 bitmap_set_bit (need_eh_cleanup,
4449 gimple_bb (stmt)->index);
4450 if (dump_file && (dump_flags & TDF_DETAILS))
4451 fprintf (dump_file, " Removed EH side-effects.\n");
4453 /* Likewise for AB side-effects. */
4454 if (can_make_abnormal_goto
4455 && !stmt_can_make_abnormal_goto (stmt))
4457 bitmap_set_bit (need_ab_cleanup,
4458 gimple_bb (stmt)->index);
4459 if (dump_file && (dump_flags & TDF_DETAILS))
4460 fprintf (dump_file, " Removed AB side-effects.\n");
4462 update_stmt (stmt);
4463 if (vdef != gimple_vdef (stmt))
4464 VN_INFO (vdef)->valnum = vuse;
4467 /* Make new values available - for fully redundant LHS we
4468 continue with the next stmt above and skip this. */
4469 def_operand_p defp;
4470 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_DEF)
4471 eliminate_push_avail (DEF_FROM_PTR (defp));
4474 /* Replace destination PHI arguments. */
4475 edge_iterator ei;
4476 edge e;
4477 FOR_EACH_EDGE (e, ei, b->succs)
4479 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
4481 gimple phi = gsi_stmt (gsi);
4482 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
4483 tree arg = USE_FROM_PTR (use_p);
4484 if (TREE_CODE (arg) != SSA_NAME
4485 || virtual_operand_p (arg))
4486 continue;
4487 tree sprime = eliminate_avail (arg);
4488 if (sprime && may_propagate_copy (arg, sprime))
4490 propagate_value (use_p, sprime);
4491 if (TREE_CODE (sprime) == SSA_NAME)
4492 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4498 /* Make no longer available leaders no longer available. */
4500 void
4501 eliminate_dom_walker::after_dom_children (basic_block)
4503 tree entry;
4504 while ((entry = el_avail_stack.pop ()) != NULL_TREE)
4505 el_avail[SSA_NAME_VERSION (VN_INFO (entry)->valnum)] = NULL_TREE;
4508 /* Eliminate fully redundant computations. */
4510 static unsigned int
4511 eliminate (bool do_pre)
4513 gimple_stmt_iterator gsi;
4514 gimple stmt;
4516 need_eh_cleanup = BITMAP_ALLOC (NULL);
4517 need_ab_cleanup = BITMAP_ALLOC (NULL);
4519 el_to_remove.create (0);
4520 el_todo = 0;
4521 el_avail.create (0);
4522 el_avail_stack.create (0);
4524 eliminate_dom_walker (CDI_DOMINATORS,
4525 do_pre).walk (cfun->cfg->x_entry_block_ptr);
4527 el_avail.release ();
4528 el_avail_stack.release ();
4530 /* We cannot remove stmts during BB walk, especially not release SSA
4531 names there as this confuses the VN machinery. The stmts ending
4532 up in el_to_remove are either stores or simple copies.
4533 Remove stmts in reverse order to make debug stmt creation possible. */
4534 while (!el_to_remove.is_empty ())
4536 stmt = el_to_remove.pop ();
4538 if (dump_file && (dump_flags & TDF_DETAILS))
4540 fprintf (dump_file, "Removing dead stmt ");
4541 print_gimple_stmt (dump_file, stmt, 0, 0);
4544 tree lhs;
4545 if (gimple_code (stmt) == GIMPLE_PHI)
4546 lhs = gimple_phi_result (stmt);
4547 else
4548 lhs = gimple_get_lhs (stmt);
4550 if (inserted_exprs
4551 && TREE_CODE (lhs) == SSA_NAME)
4552 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4554 gsi = gsi_for_stmt (stmt);
4555 if (gimple_code (stmt) == GIMPLE_PHI)
4556 remove_phi_node (&gsi, true);
4557 else
4559 basic_block bb = gimple_bb (stmt);
4560 unlink_stmt_vdef (stmt);
4561 if (gsi_remove (&gsi, true))
4562 bitmap_set_bit (need_eh_cleanup, bb->index);
4563 release_defs (stmt);
4566 /* Removing a stmt may expose a forwarder block. */
4567 el_todo |= TODO_cleanup_cfg;
4569 el_to_remove.release ();
4571 return el_todo;
4574 /* Perform CFG cleanups made necessary by elimination. */
4576 static unsigned
4577 fini_eliminate (void)
4579 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4580 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4582 if (do_eh_cleanup)
4583 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4585 if (do_ab_cleanup)
4586 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4588 BITMAP_FREE (need_eh_cleanup);
4589 BITMAP_FREE (need_ab_cleanup);
4591 if (do_eh_cleanup || do_ab_cleanup)
4592 return TODO_cleanup_cfg;
4593 return 0;
4596 /* Borrow a bit of tree-ssa-dce.c for the moment.
4597 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4598 this may be a bit faster, and we may want critical edges kept split. */
4600 /* If OP's defining statement has not already been determined to be necessary,
4601 mark that statement necessary. Return the stmt, if it is newly
4602 necessary. */
4604 static inline gimple
4605 mark_operand_necessary (tree op)
4607 gimple stmt;
4609 gcc_assert (op);
4611 if (TREE_CODE (op) != SSA_NAME)
4612 return NULL;
4614 stmt = SSA_NAME_DEF_STMT (op);
4615 gcc_assert (stmt);
4617 if (gimple_plf (stmt, NECESSARY)
4618 || gimple_nop_p (stmt))
4619 return NULL;
4621 gimple_set_plf (stmt, NECESSARY, true);
4622 return stmt;
4625 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4626 to insert PHI nodes sometimes, and because value numbering of casts isn't
4627 perfect, we sometimes end up inserting dead code. This simple DCE-like
4628 pass removes any insertions we made that weren't actually used. */
4630 static void
4631 remove_dead_inserted_code (void)
4633 bitmap worklist;
4634 unsigned i;
4635 bitmap_iterator bi;
4636 gimple t;
4638 worklist = BITMAP_ALLOC (NULL);
4639 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4641 t = SSA_NAME_DEF_STMT (ssa_name (i));
4642 if (gimple_plf (t, NECESSARY))
4643 bitmap_set_bit (worklist, i);
4645 while (!bitmap_empty_p (worklist))
4647 i = bitmap_first_set_bit (worklist);
4648 bitmap_clear_bit (worklist, i);
4649 t = SSA_NAME_DEF_STMT (ssa_name (i));
4651 /* PHI nodes are somewhat special in that each PHI alternative has
4652 data and control dependencies. All the statements feeding the
4653 PHI node's arguments are always necessary. */
4654 if (gimple_code (t) == GIMPLE_PHI)
4656 unsigned k;
4658 for (k = 0; k < gimple_phi_num_args (t); k++)
4660 tree arg = PHI_ARG_DEF (t, k);
4661 if (TREE_CODE (arg) == SSA_NAME)
4663 gimple n = mark_operand_necessary (arg);
4664 if (n)
4665 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4669 else
4671 /* Propagate through the operands. Examine all the USE, VUSE and
4672 VDEF operands in this statement. Mark all the statements
4673 which feed this statement's uses as necessary. */
4674 ssa_op_iter iter;
4675 tree use;
4677 /* The operands of VDEF expressions are also needed as they
4678 represent potential definitions that may reach this
4679 statement (VDEF operands allow us to follow def-def
4680 links). */
4682 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4684 gimple n = mark_operand_necessary (use);
4685 if (n)
4686 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4691 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4693 t = SSA_NAME_DEF_STMT (ssa_name (i));
4694 if (!gimple_plf (t, NECESSARY))
4696 gimple_stmt_iterator gsi;
4698 if (dump_file && (dump_flags & TDF_DETAILS))
4700 fprintf (dump_file, "Removing unnecessary insertion:");
4701 print_gimple_stmt (dump_file, t, 0, 0);
4704 gsi = gsi_for_stmt (t);
4705 if (gimple_code (t) == GIMPLE_PHI)
4706 remove_phi_node (&gsi, true);
4707 else
4709 gsi_remove (&gsi, true);
4710 release_defs (t);
4714 BITMAP_FREE (worklist);
4718 /* Initialize data structures used by PRE. */
4720 static void
4721 init_pre (void)
4723 basic_block bb;
4725 next_expression_id = 1;
4726 expressions.create (0);
4727 expressions.safe_push (NULL);
4728 value_expressions.create (get_max_value_id () + 1);
4729 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
4730 name_to_id.create (0);
4732 inserted_exprs = BITMAP_ALLOC (NULL);
4734 connect_infinite_loops_to_exit ();
4735 memset (&pre_stats, 0, sizeof (pre_stats));
4737 postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
4738 postorder_num = inverted_post_order_compute (postorder);
4740 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4742 calculate_dominance_info (CDI_POST_DOMINATORS);
4743 calculate_dominance_info (CDI_DOMINATORS);
4745 bitmap_obstack_initialize (&grand_bitmap_obstack);
4746 phi_translate_table = new hash_table<expr_pred_trans_d> (5110);
4747 expression_to_id = new hash_table<pre_expr_d> (num_ssa_names * 3);
4748 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4749 sizeof (struct bitmap_set), 30);
4750 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4751 sizeof (struct pre_expr_d), 30);
4752 FOR_ALL_BB_FN (bb, cfun)
4754 EXP_GEN (bb) = bitmap_set_new ();
4755 PHI_GEN (bb) = bitmap_set_new ();
4756 TMP_GEN (bb) = bitmap_set_new ();
4757 AVAIL_OUT (bb) = bitmap_set_new ();
4762 /* Deallocate data structures used by PRE. */
4764 static void
4765 fini_pre ()
4767 free (postorder);
4768 value_expressions.release ();
4769 BITMAP_FREE (inserted_exprs);
4770 bitmap_obstack_release (&grand_bitmap_obstack);
4771 free_alloc_pool (bitmap_set_pool);
4772 free_alloc_pool (pre_expr_pool);
4773 delete phi_translate_table;
4774 phi_translate_table = NULL;
4775 delete expression_to_id;
4776 expression_to_id = NULL;
4777 name_to_id.release ();
4779 free_aux_for_blocks ();
4781 free_dominance_info (CDI_POST_DOMINATORS);
4784 namespace {
4786 const pass_data pass_data_pre =
4788 GIMPLE_PASS, /* type */
4789 "pre", /* name */
4790 OPTGROUP_NONE, /* optinfo_flags */
4791 TV_TREE_PRE, /* tv_id */
4792 /* PROP_no_crit_edges is ensured by placing pass_split_crit_edges before
4793 pass_pre. */
4794 ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */
4795 0, /* properties_provided */
4796 PROP_no_crit_edges, /* properties_destroyed */
4797 TODO_rebuild_alias, /* todo_flags_start */
4798 0, /* todo_flags_finish */
4801 class pass_pre : public gimple_opt_pass
4803 public:
4804 pass_pre (gcc::context *ctxt)
4805 : gimple_opt_pass (pass_data_pre, ctxt)
4808 /* opt_pass methods: */
4809 virtual bool gate (function *) { return flag_tree_pre != 0; }
4810 virtual unsigned int execute (function *);
4812 }; // class pass_pre
4814 unsigned int
4815 pass_pre::execute (function *fun)
4817 unsigned int todo = 0;
4819 do_partial_partial =
4820 flag_tree_partial_pre && optimize_function_for_speed_p (fun);
4822 /* This has to happen before SCCVN runs because
4823 loop_optimizer_init may create new phis, etc. */
4824 loop_optimizer_init (LOOPS_NORMAL);
4826 if (!run_scc_vn (VN_WALK))
4828 loop_optimizer_finalize ();
4829 return 0;
4832 init_pre ();
4833 scev_initialize ();
4835 /* Collect and value number expressions computed in each basic block. */
4836 compute_avail ();
4838 /* Insert can get quite slow on an incredibly large number of basic
4839 blocks due to some quadratic behavior. Until this behavior is
4840 fixed, don't run it when he have an incredibly large number of
4841 bb's. If we aren't going to run insert, there is no point in
4842 computing ANTIC, either, even though it's plenty fast. */
4843 if (n_basic_blocks_for_fn (fun) < 4000)
4845 compute_antic ();
4846 insert ();
4849 /* Make sure to remove fake edges before committing our inserts.
4850 This makes sure we don't end up with extra critical edges that
4851 we would need to split. */
4852 remove_fake_exit_edges ();
4853 gsi_commit_edge_inserts ();
4855 /* Remove all the redundant expressions. */
4856 todo |= eliminate (true);
4858 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
4859 statistics_counter_event (fun, "PA inserted", pre_stats.pa_insert);
4860 statistics_counter_event (fun, "New PHIs", pre_stats.phis);
4861 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
4863 clear_expression_ids ();
4864 remove_dead_inserted_code ();
4866 scev_finalize ();
4867 fini_pre ();
4868 todo |= fini_eliminate ();
4869 loop_optimizer_finalize ();
4871 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4872 case we can merge the block with the remaining predecessor of the block.
4873 It should either:
4874 - call merge_blocks after each tail merge iteration
4875 - call merge_blocks after all tail merge iterations
4876 - mark TODO_cleanup_cfg when necessary
4877 - share the cfg cleanup with fini_pre. */
4878 todo |= tail_merge_optimize (todo);
4880 free_scc_vn ();
4882 /* Tail merging invalidates the virtual SSA web, together with
4883 cfg-cleanup opportunities exposed by PRE this will wreck the
4884 SSA updating machinery. So make sure to run update-ssa
4885 manually, before eventually scheduling cfg-cleanup as part of
4886 the todo. */
4887 update_ssa (TODO_update_ssa_only_virtuals);
4889 return todo;
4892 } // anon namespace
4894 gimple_opt_pass *
4895 make_pass_pre (gcc::context *ctxt)
4897 return new pass_pre (ctxt);
4900 namespace {
4902 const pass_data pass_data_fre =
4904 GIMPLE_PASS, /* type */
4905 "fre", /* name */
4906 OPTGROUP_NONE, /* optinfo_flags */
4907 TV_TREE_FRE, /* tv_id */
4908 ( PROP_cfg | PROP_ssa ), /* properties_required */
4909 0, /* properties_provided */
4910 0, /* properties_destroyed */
4911 0, /* todo_flags_start */
4912 0, /* todo_flags_finish */
4915 class pass_fre : public gimple_opt_pass
4917 public:
4918 pass_fre (gcc::context *ctxt)
4919 : gimple_opt_pass (pass_data_fre, ctxt)
4922 /* opt_pass methods: */
4923 opt_pass * clone () { return new pass_fre (m_ctxt); }
4924 virtual bool gate (function *) { return flag_tree_fre != 0; }
4925 virtual unsigned int execute (function *);
4927 }; // class pass_fre
4929 unsigned int
4930 pass_fre::execute (function *fun)
4932 unsigned int todo = 0;
4934 if (!run_scc_vn (VN_WALKREWRITE))
4935 return 0;
4937 memset (&pre_stats, 0, sizeof (pre_stats));
4939 /* Remove all the redundant expressions. */
4940 todo |= eliminate (false);
4942 todo |= fini_eliminate ();
4944 free_scc_vn ();
4946 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
4947 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
4949 return todo;
4952 } // anon namespace
4954 gimple_opt_pass *
4955 make_pass_fre (gcc::context *ctxt)
4957 return new pass_fre (ctxt);