re PR c++/19476 (Missed null checking elimination with new)
[official-gcc.git] / gcc / tree-ssa-pre.c
blob50027ebffe8c0a053edd80bed115b2f2e060b6be
1 /* SSA-PRE for trees.
2 Copyright (C) 2001-2013 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
4 <stevenb@suse.de>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
11 any later version.
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "tree.h"
27 #include "basic-block.h"
28 #include "gimple-pretty-print.h"
29 #include "tree-inline.h"
30 #include "tree-ssa.h"
31 #include "gimple.h"
32 #include "hash-table.h"
33 #include "tree-iterator.h"
34 #include "alloc-pool.h"
35 #include "obstack.h"
36 #include "tree-pass.h"
37 #include "flags.h"
38 #include "bitmap.h"
39 #include "langhooks.h"
40 #include "cfgloop.h"
41 #include "tree-ssa-sccvn.h"
42 #include "tree-scalar-evolution.h"
43 #include "params.h"
44 #include "dbgcnt.h"
45 #include "domwalk.h"
46 #include "ipa-prop.h"
47 #include "tree-ssa-propagate.h"
49 /* TODO:
51 1. Avail sets can be shared by making an avail_find_leader that
52 walks up the dominator tree and looks in those avail sets.
53 This might affect code optimality, it's unclear right now.
54 2. Strength reduction can be performed by anticipating expressions
55 we can repair later on.
56 3. We can do back-substitution or smarter value numbering to catch
57 commutative expressions split up over multiple statements.
60 /* For ease of terminology, "expression node" in the below refers to
61 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
62 represent the actual statement containing the expressions we care about,
63 and we cache the value number by putting it in the expression. */
65 /* Basic algorithm
67 First we walk the statements to generate the AVAIL sets, the
68 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
69 generation of values/expressions by a given block. We use them
70 when computing the ANTIC sets. The AVAIL sets consist of
71 SSA_NAME's that represent values, so we know what values are
72 available in what blocks. AVAIL is a forward dataflow problem. In
73 SSA, values are never killed, so we don't need a kill set, or a
74 fixpoint iteration, in order to calculate the AVAIL sets. In
75 traditional parlance, AVAIL sets tell us the downsafety of the
76 expressions/values.
78 Next, we generate the ANTIC sets. These sets represent the
79 anticipatable expressions. ANTIC is a backwards dataflow
80 problem. An expression is anticipatable in a given block if it could
81 be generated in that block. This means that if we had to perform
82 an insertion in that block, of the value of that expression, we
83 could. Calculating the ANTIC sets requires phi translation of
84 expressions, because the flow goes backwards through phis. We must
85 iterate to a fixpoint of the ANTIC sets, because we have a kill
86 set. Even in SSA form, values are not live over the entire
87 function, only from their definition point onwards. So we have to
88 remove values from the ANTIC set once we go past the definition
89 point of the leaders that make them up.
90 compute_antic/compute_antic_aux performs this computation.
92 Third, we perform insertions to make partially redundant
93 expressions fully redundant.
95 An expression is partially redundant (excluding partial
96 anticipation) if:
98 1. It is AVAIL in some, but not all, of the predecessors of a
99 given block.
100 2. It is ANTIC in all the predecessors.
102 In order to make it fully redundant, we insert the expression into
103 the predecessors where it is not available, but is ANTIC.
105 For the partial anticipation case, we only perform insertion if it
106 is partially anticipated in some block, and fully available in all
107 of the predecessors.
109 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
110 performs these steps.
112 Fourth, we eliminate fully redundant expressions.
113 This is a simple statement walk that replaces redundant
114 calculations with the now available values. */
116 /* Representations of value numbers:
118 Value numbers are represented by a representative SSA_NAME. We
119 will create fake SSA_NAME's in situations where we need a
120 representative but do not have one (because it is a complex
121 expression). In order to facilitate storing the value numbers in
122 bitmaps, and keep the number of wasted SSA_NAME's down, we also
123 associate a value_id with each value number, and create full blown
124 ssa_name's only where we actually need them (IE in operands of
125 existing expressions).
127 Theoretically you could replace all the value_id's with
128 SSA_NAME_VERSION, but this would allocate a large number of
129 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
130 It would also require an additional indirection at each point we
131 use the value id. */
133 /* Representation of expressions on value numbers:
135 Expressions consisting of value numbers are represented the same
136 way as our VN internally represents them, with an additional
137 "pre_expr" wrapping around them in order to facilitate storing all
138 of the expressions in the same sets. */
140 /* Representation of sets:
142 The dataflow sets do not need to be sorted in any particular order
143 for the majority of their lifetime, are simply represented as two
144 bitmaps, one that keeps track of values present in the set, and one
145 that keeps track of expressions present in the set.
147 When we need them in topological order, we produce it on demand by
148 transforming the bitmap into an array and sorting it into topo
149 order. */
151 /* Type of expression, used to know which member of the PRE_EXPR union
152 is valid. */
154 enum pre_expr_kind
156 NAME,
157 NARY,
158 REFERENCE,
159 CONSTANT
162 typedef union pre_expr_union_d
164 tree name;
165 tree constant;
166 vn_nary_op_t nary;
167 vn_reference_t reference;
168 } pre_expr_union;
170 typedef struct pre_expr_d : typed_noop_remove <pre_expr_d>
172 enum pre_expr_kind kind;
173 unsigned int id;
174 pre_expr_union u;
176 /* hash_table support. */
177 typedef pre_expr_d value_type;
178 typedef pre_expr_d compare_type;
179 static inline hashval_t hash (const pre_expr_d *);
180 static inline int equal (const pre_expr_d *, const pre_expr_d *);
181 } *pre_expr;
183 #define PRE_EXPR_NAME(e) (e)->u.name
184 #define PRE_EXPR_NARY(e) (e)->u.nary
185 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
186 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
188 /* Compare E1 and E1 for equality. */
190 inline int
191 pre_expr_d::equal (const value_type *e1, const compare_type *e2)
193 if (e1->kind != e2->kind)
194 return false;
196 switch (e1->kind)
198 case CONSTANT:
199 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
200 PRE_EXPR_CONSTANT (e2));
201 case NAME:
202 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
203 case NARY:
204 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
205 case REFERENCE:
206 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
207 PRE_EXPR_REFERENCE (e2));
208 default:
209 gcc_unreachable ();
213 /* Hash E. */
215 inline hashval_t
216 pre_expr_d::hash (const value_type *e)
218 switch (e->kind)
220 case CONSTANT:
221 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
222 case NAME:
223 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
224 case NARY:
225 return PRE_EXPR_NARY (e)->hashcode;
226 case REFERENCE:
227 return PRE_EXPR_REFERENCE (e)->hashcode;
228 default:
229 gcc_unreachable ();
233 /* Next global expression id number. */
234 static unsigned int next_expression_id;
236 /* Mapping from expression to id number we can use in bitmap sets. */
237 static vec<pre_expr> expressions;
238 static hash_table <pre_expr_d> expression_to_id;
239 static vec<unsigned> name_to_id;
241 /* Allocate an expression id for EXPR. */
243 static inline unsigned int
244 alloc_expression_id (pre_expr expr)
246 struct pre_expr_d **slot;
247 /* Make sure we won't overflow. */
248 gcc_assert (next_expression_id + 1 > next_expression_id);
249 expr->id = next_expression_id++;
250 expressions.safe_push (expr);
251 if (expr->kind == NAME)
253 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
254 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
255 re-allocations by using vec::reserve upfront. There is no
256 vec::quick_grow_cleared unfortunately. */
257 unsigned old_len = name_to_id.length ();
258 name_to_id.reserve (num_ssa_names - old_len);
259 name_to_id.safe_grow_cleared (num_ssa_names);
260 gcc_assert (name_to_id[version] == 0);
261 name_to_id[version] = expr->id;
263 else
265 slot = expression_to_id.find_slot (expr, INSERT);
266 gcc_assert (!*slot);
267 *slot = expr;
269 return next_expression_id - 1;
272 /* Return the expression id for tree EXPR. */
274 static inline unsigned int
275 get_expression_id (const pre_expr expr)
277 return expr->id;
280 static inline unsigned int
281 lookup_expression_id (const pre_expr expr)
283 struct pre_expr_d **slot;
285 if (expr->kind == NAME)
287 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
288 if (name_to_id.length () <= version)
289 return 0;
290 return name_to_id[version];
292 else
294 slot = expression_to_id.find_slot (expr, NO_INSERT);
295 if (!slot)
296 return 0;
297 return ((pre_expr)*slot)->id;
301 /* Return the existing expression id for EXPR, or create one if one
302 does not exist yet. */
304 static inline unsigned int
305 get_or_alloc_expression_id (pre_expr expr)
307 unsigned int id = lookup_expression_id (expr);
308 if (id == 0)
309 return alloc_expression_id (expr);
310 return expr->id = id;
313 /* Return the expression that has expression id ID */
315 static inline pre_expr
316 expression_for_id (unsigned int id)
318 return expressions[id];
321 /* Free the expression id field in all of our expressions,
322 and then destroy the expressions array. */
324 static void
325 clear_expression_ids (void)
327 expressions.release ();
330 static alloc_pool pre_expr_pool;
332 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
334 static pre_expr
335 get_or_alloc_expr_for_name (tree name)
337 struct pre_expr_d expr;
338 pre_expr result;
339 unsigned int result_id;
341 expr.kind = NAME;
342 expr.id = 0;
343 PRE_EXPR_NAME (&expr) = name;
344 result_id = lookup_expression_id (&expr);
345 if (result_id != 0)
346 return expression_for_id (result_id);
348 result = (pre_expr) pool_alloc (pre_expr_pool);
349 result->kind = NAME;
350 PRE_EXPR_NAME (result) = name;
351 alloc_expression_id (result);
352 return result;
355 /* An unordered bitmap set. One bitmap tracks values, the other,
356 expressions. */
357 typedef struct bitmap_set
359 bitmap_head expressions;
360 bitmap_head values;
361 } *bitmap_set_t;
363 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
364 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
366 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
367 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
369 /* Mapping from value id to expressions with that value_id. */
370 static vec<bitmap> value_expressions;
372 /* Sets that we need to keep track of. */
373 typedef struct bb_bitmap_sets
375 /* The EXP_GEN set, which represents expressions/values generated in
376 a basic block. */
377 bitmap_set_t exp_gen;
379 /* The PHI_GEN set, which represents PHI results generated in a
380 basic block. */
381 bitmap_set_t phi_gen;
383 /* The TMP_GEN set, which represents results/temporaries generated
384 in a basic block. IE the LHS of an expression. */
385 bitmap_set_t tmp_gen;
387 /* The AVAIL_OUT set, which represents which values are available in
388 a given basic block. */
389 bitmap_set_t avail_out;
391 /* The ANTIC_IN set, which represents which values are anticipatable
392 in a given basic block. */
393 bitmap_set_t antic_in;
395 /* The PA_IN set, which represents which values are
396 partially anticipatable in a given basic block. */
397 bitmap_set_t pa_in;
399 /* The NEW_SETS set, which is used during insertion to augment the
400 AVAIL_OUT set of blocks with the new insertions performed during
401 the current iteration. */
402 bitmap_set_t new_sets;
404 /* A cache for value_dies_in_block_x. */
405 bitmap expr_dies;
407 /* True if we have visited this block during ANTIC calculation. */
408 unsigned int visited : 1;
410 /* True we have deferred processing this block during ANTIC
411 calculation until its successor is processed. */
412 unsigned int deferred : 1;
414 /* True when the block contains a call that might not return. */
415 unsigned int contains_may_not_return_call : 1;
416 } *bb_value_sets_t;
418 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
419 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
420 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
421 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
422 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
423 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
424 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
425 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
426 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
427 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
428 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
431 /* Basic block list in postorder. */
432 static int *postorder;
433 static int postorder_num;
435 /* This structure is used to keep track of statistics on what
436 optimization PRE was able to perform. */
437 static struct
439 /* The number of RHS computations eliminated by PRE. */
440 int eliminations;
442 /* The number of new expressions/temporaries generated by PRE. */
443 int insertions;
445 /* The number of inserts found due to partial anticipation */
446 int pa_insert;
448 /* The number of new PHI nodes added by PRE. */
449 int phis;
450 } pre_stats;
452 static bool do_partial_partial;
453 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int);
454 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
455 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
456 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
457 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
458 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
459 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
460 unsigned int, bool);
461 static bitmap_set_t bitmap_set_new (void);
462 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
463 tree);
464 static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
465 static unsigned int get_expr_value_id (pre_expr);
467 /* We can add and remove elements and entries to and from sets
468 and hash tables, so we use alloc pools for them. */
470 static alloc_pool bitmap_set_pool;
471 static bitmap_obstack grand_bitmap_obstack;
473 /* Set of blocks with statements that have had their EH properties changed. */
474 static bitmap need_eh_cleanup;
476 /* Set of blocks with statements that have had their AB properties changed. */
477 static bitmap need_ab_cleanup;
479 /* A three tuple {e, pred, v} used to cache phi translations in the
480 phi_translate_table. */
482 typedef struct expr_pred_trans_d : typed_free_remove<expr_pred_trans_d>
484 /* The expression. */
485 pre_expr e;
487 /* The predecessor block along which we translated the expression. */
488 basic_block pred;
490 /* The value that resulted from the translation. */
491 pre_expr v;
493 /* The hashcode for the expression, pred pair. This is cached for
494 speed reasons. */
495 hashval_t hashcode;
497 /* hash_table support. */
498 typedef expr_pred_trans_d value_type;
499 typedef expr_pred_trans_d compare_type;
500 static inline hashval_t hash (const value_type *);
501 static inline int equal (const value_type *, const compare_type *);
502 } *expr_pred_trans_t;
503 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
505 inline hashval_t
506 expr_pred_trans_d::hash (const expr_pred_trans_d *e)
508 return e->hashcode;
511 inline int
512 expr_pred_trans_d::equal (const value_type *ve1,
513 const compare_type *ve2)
515 basic_block b1 = ve1->pred;
516 basic_block b2 = ve2->pred;
518 /* If they are not translations for the same basic block, they can't
519 be equal. */
520 if (b1 != b2)
521 return false;
522 return pre_expr_d::equal (ve1->e, ve2->e);
525 /* The phi_translate_table caches phi translations for a given
526 expression and predecessor. */
527 static hash_table <expr_pred_trans_d> phi_translate_table;
529 /* Add the tuple mapping from {expression E, basic block PRED} to
530 the phi translation table and return whether it pre-existed. */
532 static inline bool
533 phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred)
535 expr_pred_trans_t *slot;
536 expr_pred_trans_d tem;
537 hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e),
538 pred->index);
539 tem.e = e;
540 tem.pred = pred;
541 tem.hashcode = hash;
542 slot = phi_translate_table.find_slot_with_hash (&tem, hash, INSERT);
543 if (*slot)
545 *entry = *slot;
546 return true;
549 *entry = *slot = XNEW (struct expr_pred_trans_d);
550 (*entry)->e = e;
551 (*entry)->pred = pred;
552 (*entry)->hashcode = hash;
553 return false;
557 /* Add expression E to the expression set of value id V. */
559 static void
560 add_to_value (unsigned int v, pre_expr e)
562 bitmap set;
564 gcc_checking_assert (get_expr_value_id (e) == v);
566 if (v >= value_expressions.length ())
568 value_expressions.safe_grow_cleared (v + 1);
571 set = value_expressions[v];
572 if (!set)
574 set = BITMAP_ALLOC (&grand_bitmap_obstack);
575 value_expressions[v] = set;
578 bitmap_set_bit (set, get_or_alloc_expression_id (e));
581 /* Create a new bitmap set and return it. */
583 static bitmap_set_t
584 bitmap_set_new (void)
586 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
587 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
588 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
589 return ret;
592 /* Return the value id for a PRE expression EXPR. */
594 static unsigned int
595 get_expr_value_id (pre_expr expr)
597 unsigned int id;
598 switch (expr->kind)
600 case CONSTANT:
601 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
602 break;
603 case NAME:
604 id = VN_INFO (PRE_EXPR_NAME (expr))->value_id;
605 break;
606 case NARY:
607 id = PRE_EXPR_NARY (expr)->value_id;
608 break;
609 case REFERENCE:
610 id = PRE_EXPR_REFERENCE (expr)->value_id;
611 break;
612 default:
613 gcc_unreachable ();
615 /* ??? We cannot assert that expr has a value-id (it can be 0), because
616 we assign value-ids only to expressions that have a result
617 in set_hashtable_value_ids. */
618 return id;
621 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
623 static tree
624 sccvn_valnum_from_value_id (unsigned int val)
626 bitmap_iterator bi;
627 unsigned int i;
628 bitmap exprset = value_expressions[val];
629 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
631 pre_expr vexpr = expression_for_id (i);
632 if (vexpr->kind == NAME)
633 return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
634 else if (vexpr->kind == CONSTANT)
635 return PRE_EXPR_CONSTANT (vexpr);
637 return NULL_TREE;
640 /* Remove an expression EXPR from a bitmapped set. */
642 static void
643 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
645 unsigned int val = get_expr_value_id (expr);
646 if (!value_id_constant_p (val))
648 bitmap_clear_bit (&set->values, val);
649 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
653 static void
654 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
655 unsigned int val, bool allow_constants)
657 if (allow_constants || !value_id_constant_p (val))
659 /* We specifically expect this and only this function to be able to
660 insert constants into a set. */
661 bitmap_set_bit (&set->values, val);
662 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
666 /* Insert an expression EXPR into a bitmapped set. */
668 static void
669 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
671 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
674 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
676 static void
677 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
679 bitmap_copy (&dest->expressions, &orig->expressions);
680 bitmap_copy (&dest->values, &orig->values);
684 /* Free memory used up by SET. */
685 static void
686 bitmap_set_free (bitmap_set_t set)
688 bitmap_clear (&set->expressions);
689 bitmap_clear (&set->values);
693 /* Generate an topological-ordered array of bitmap set SET. */
695 static vec<pre_expr>
696 sorted_array_from_bitmap_set (bitmap_set_t set)
698 unsigned int i, j;
699 bitmap_iterator bi, bj;
700 vec<pre_expr> result;
702 /* Pre-allocate roughly enough space for the array. */
703 result.create (bitmap_count_bits (&set->values));
705 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
707 /* The number of expressions having a given value is usually
708 relatively small. Thus, rather than making a vector of all
709 the expressions and sorting it by value-id, we walk the values
710 and check in the reverse mapping that tells us what expressions
711 have a given value, to filter those in our set. As a result,
712 the expressions are inserted in value-id order, which means
713 topological order.
715 If this is somehow a significant lose for some cases, we can
716 choose which set to walk based on the set size. */
717 bitmap exprset = value_expressions[i];
718 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj)
720 if (bitmap_bit_p (&set->expressions, j))
721 result.safe_push (expression_for_id (j));
725 return result;
728 /* Perform bitmapped set operation DEST &= ORIG. */
730 static void
731 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
733 bitmap_iterator bi;
734 unsigned int i;
736 if (dest != orig)
738 bitmap_head temp;
739 bitmap_initialize (&temp, &grand_bitmap_obstack);
741 bitmap_and_into (&dest->values, &orig->values);
742 bitmap_copy (&temp, &dest->expressions);
743 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
745 pre_expr expr = expression_for_id (i);
746 unsigned int value_id = get_expr_value_id (expr);
747 if (!bitmap_bit_p (&dest->values, value_id))
748 bitmap_clear_bit (&dest->expressions, i);
750 bitmap_clear (&temp);
754 /* Subtract all values and expressions contained in ORIG from DEST. */
756 static bitmap_set_t
757 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
759 bitmap_set_t result = bitmap_set_new ();
760 bitmap_iterator bi;
761 unsigned int i;
763 bitmap_and_compl (&result->expressions, &dest->expressions,
764 &orig->expressions);
766 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
768 pre_expr expr = expression_for_id (i);
769 unsigned int value_id = get_expr_value_id (expr);
770 bitmap_set_bit (&result->values, value_id);
773 return result;
776 /* Subtract all the values in bitmap set B from bitmap set A. */
778 static void
779 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
781 unsigned int i;
782 bitmap_iterator bi;
783 bitmap_head temp;
785 bitmap_initialize (&temp, &grand_bitmap_obstack);
787 bitmap_copy (&temp, &a->expressions);
788 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
790 pre_expr expr = expression_for_id (i);
791 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
792 bitmap_remove_from_set (a, expr);
794 bitmap_clear (&temp);
798 /* Return true if bitmapped set SET contains the value VALUE_ID. */
800 static bool
801 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
803 if (value_id_constant_p (value_id))
804 return true;
806 if (!set || bitmap_empty_p (&set->expressions))
807 return false;
809 return bitmap_bit_p (&set->values, value_id);
812 static inline bool
813 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
815 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
818 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
820 static void
821 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
822 const pre_expr expr)
824 bitmap exprset;
825 unsigned int i;
826 bitmap_iterator bi;
828 if (value_id_constant_p (lookfor))
829 return;
831 if (!bitmap_set_contains_value (set, lookfor))
832 return;
834 /* The number of expressions having a given value is usually
835 significantly less than the total number of expressions in SET.
836 Thus, rather than check, for each expression in SET, whether it
837 has the value LOOKFOR, we walk the reverse mapping that tells us
838 what expressions have a given value, and see if any of those
839 expressions are in our set. For large testcases, this is about
840 5-10x faster than walking the bitmap. If this is somehow a
841 significant lose for some cases, we can choose which set to walk
842 based on the set size. */
843 exprset = value_expressions[lookfor];
844 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
846 if (bitmap_clear_bit (&set->expressions, i))
848 bitmap_set_bit (&set->expressions, get_expression_id (expr));
849 return;
853 gcc_unreachable ();
856 /* Return true if two bitmap sets are equal. */
858 static bool
859 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
861 return bitmap_equal_p (&a->values, &b->values);
864 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
865 and add it otherwise. */
867 static void
868 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
870 unsigned int val = get_expr_value_id (expr);
872 if (bitmap_set_contains_value (set, val))
873 bitmap_set_replace_value (set, val, expr);
874 else
875 bitmap_insert_into_set (set, expr);
878 /* Insert EXPR into SET if EXPR's value is not already present in
879 SET. */
881 static void
882 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
884 unsigned int val = get_expr_value_id (expr);
886 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
888 /* Constant values are always considered to be part of the set. */
889 if (value_id_constant_p (val))
890 return;
892 /* If the value membership changed, add the expression. */
893 if (bitmap_set_bit (&set->values, val))
894 bitmap_set_bit (&set->expressions, expr->id);
897 /* Print out EXPR to outfile. */
899 static void
900 print_pre_expr (FILE *outfile, const pre_expr expr)
902 switch (expr->kind)
904 case CONSTANT:
905 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
906 break;
907 case NAME:
908 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
909 break;
910 case NARY:
912 unsigned int i;
913 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
914 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
915 for (i = 0; i < nary->length; i++)
917 print_generic_expr (outfile, nary->op[i], 0);
918 if (i != (unsigned) nary->length - 1)
919 fprintf (outfile, ",");
921 fprintf (outfile, "}");
923 break;
925 case REFERENCE:
927 vn_reference_op_t vro;
928 unsigned int i;
929 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
930 fprintf (outfile, "{");
931 for (i = 0;
932 ref->operands.iterate (i, &vro);
933 i++)
935 bool closebrace = false;
936 if (vro->opcode != SSA_NAME
937 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
939 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
940 if (vro->op0)
942 fprintf (outfile, "<");
943 closebrace = true;
946 if (vro->op0)
948 print_generic_expr (outfile, vro->op0, 0);
949 if (vro->op1)
951 fprintf (outfile, ",");
952 print_generic_expr (outfile, vro->op1, 0);
954 if (vro->op2)
956 fprintf (outfile, ",");
957 print_generic_expr (outfile, vro->op2, 0);
960 if (closebrace)
961 fprintf (outfile, ">");
962 if (i != ref->operands.length () - 1)
963 fprintf (outfile, ",");
965 fprintf (outfile, "}");
966 if (ref->vuse)
968 fprintf (outfile, "@");
969 print_generic_expr (outfile, ref->vuse, 0);
972 break;
975 void debug_pre_expr (pre_expr);
977 /* Like print_pre_expr but always prints to stderr. */
978 DEBUG_FUNCTION void
979 debug_pre_expr (pre_expr e)
981 print_pre_expr (stderr, e);
982 fprintf (stderr, "\n");
985 /* Print out SET to OUTFILE. */
987 static void
988 print_bitmap_set (FILE *outfile, bitmap_set_t set,
989 const char *setname, int blockindex)
991 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
992 if (set)
994 bool first = true;
995 unsigned i;
996 bitmap_iterator bi;
998 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1000 const pre_expr expr = expression_for_id (i);
1002 if (!first)
1003 fprintf (outfile, ", ");
1004 first = false;
1005 print_pre_expr (outfile, expr);
1007 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1010 fprintf (outfile, " }\n");
1013 void debug_bitmap_set (bitmap_set_t);
1015 DEBUG_FUNCTION void
1016 debug_bitmap_set (bitmap_set_t set)
1018 print_bitmap_set (stderr, set, "debug", 0);
1021 void debug_bitmap_sets_for (basic_block);
1023 DEBUG_FUNCTION void
1024 debug_bitmap_sets_for (basic_block bb)
1026 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1027 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1028 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1029 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1030 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1031 if (do_partial_partial)
1032 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1033 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1036 /* Print out the expressions that have VAL to OUTFILE. */
1038 static void
1039 print_value_expressions (FILE *outfile, unsigned int val)
1041 bitmap set = value_expressions[val];
1042 if (set)
1044 bitmap_set x;
1045 char s[10];
1046 sprintf (s, "%04d", val);
1047 x.expressions = *set;
1048 print_bitmap_set (outfile, &x, s, 0);
1053 DEBUG_FUNCTION void
1054 debug_value_expressions (unsigned int val)
1056 print_value_expressions (stderr, val);
1059 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1060 represent it. */
1062 static pre_expr
1063 get_or_alloc_expr_for_constant (tree constant)
1065 unsigned int result_id;
1066 unsigned int value_id;
1067 struct pre_expr_d expr;
1068 pre_expr newexpr;
1070 expr.kind = CONSTANT;
1071 PRE_EXPR_CONSTANT (&expr) = constant;
1072 result_id = lookup_expression_id (&expr);
1073 if (result_id != 0)
1074 return expression_for_id (result_id);
1076 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1077 newexpr->kind = CONSTANT;
1078 PRE_EXPR_CONSTANT (newexpr) = constant;
1079 alloc_expression_id (newexpr);
1080 value_id = get_or_alloc_constant_value_id (constant);
1081 add_to_value (value_id, newexpr);
1082 return newexpr;
1085 /* Given a value id V, find the actual tree representing the constant
1086 value if there is one, and return it. Return NULL if we can't find
1087 a constant. */
1089 static tree
1090 get_constant_for_value_id (unsigned int v)
1092 if (value_id_constant_p (v))
1094 unsigned int i;
1095 bitmap_iterator bi;
1096 bitmap exprset = value_expressions[v];
1098 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1100 pre_expr expr = expression_for_id (i);
1101 if (expr->kind == CONSTANT)
1102 return PRE_EXPR_CONSTANT (expr);
1105 return NULL;
1108 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1109 Currently only supports constants and SSA_NAMES. */
1110 static pre_expr
1111 get_or_alloc_expr_for (tree t)
1113 if (TREE_CODE (t) == SSA_NAME)
1114 return get_or_alloc_expr_for_name (t);
1115 else if (is_gimple_min_invariant (t))
1116 return get_or_alloc_expr_for_constant (t);
1117 else
1119 /* More complex expressions can result from SCCVN expression
1120 simplification that inserts values for them. As they all
1121 do not have VOPs the get handled by the nary ops struct. */
1122 vn_nary_op_t result;
1123 unsigned int result_id;
1124 vn_nary_op_lookup (t, &result);
1125 if (result != NULL)
1127 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1128 e->kind = NARY;
1129 PRE_EXPR_NARY (e) = result;
1130 result_id = lookup_expression_id (e);
1131 if (result_id != 0)
1133 pool_free (pre_expr_pool, e);
1134 e = expression_for_id (result_id);
1135 return e;
1137 alloc_expression_id (e);
1138 return e;
1141 return NULL;
1144 /* Return the folded version of T if T, when folded, is a gimple
1145 min_invariant. Otherwise, return T. */
1147 static pre_expr
1148 fully_constant_expression (pre_expr e)
1150 switch (e->kind)
1152 case CONSTANT:
1153 return e;
1154 case NARY:
1156 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1157 switch (TREE_CODE_CLASS (nary->opcode))
1159 case tcc_binary:
1160 case tcc_comparison:
1162 /* We have to go from trees to pre exprs to value ids to
1163 constants. */
1164 tree naryop0 = nary->op[0];
1165 tree naryop1 = nary->op[1];
1166 tree result;
1167 if (!is_gimple_min_invariant (naryop0))
1169 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1170 unsigned int vrep0 = get_expr_value_id (rep0);
1171 tree const0 = get_constant_for_value_id (vrep0);
1172 if (const0)
1173 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1175 if (!is_gimple_min_invariant (naryop1))
1177 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1178 unsigned int vrep1 = get_expr_value_id (rep1);
1179 tree const1 = get_constant_for_value_id (vrep1);
1180 if (const1)
1181 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1183 result = fold_binary (nary->opcode, nary->type,
1184 naryop0, naryop1);
1185 if (result && is_gimple_min_invariant (result))
1186 return get_or_alloc_expr_for_constant (result);
1187 /* We might have simplified the expression to a
1188 SSA_NAME for example from x_1 * 1. But we cannot
1189 insert a PHI for x_1 unconditionally as x_1 might
1190 not be available readily. */
1191 return e;
1193 case tcc_reference:
1194 if (nary->opcode != REALPART_EXPR
1195 && nary->opcode != IMAGPART_EXPR
1196 && nary->opcode != VIEW_CONVERT_EXPR)
1197 return e;
1198 /* Fallthrough. */
1199 case tcc_unary:
1201 /* We have to go from trees to pre exprs to value ids to
1202 constants. */
1203 tree naryop0 = nary->op[0];
1204 tree const0, result;
1205 if (is_gimple_min_invariant (naryop0))
1206 const0 = naryop0;
1207 else
1209 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1210 unsigned int vrep0 = get_expr_value_id (rep0);
1211 const0 = get_constant_for_value_id (vrep0);
1213 result = NULL;
1214 if (const0)
1216 tree type1 = TREE_TYPE (nary->op[0]);
1217 const0 = fold_convert (type1, const0);
1218 result = fold_unary (nary->opcode, nary->type, const0);
1220 if (result && is_gimple_min_invariant (result))
1221 return get_or_alloc_expr_for_constant (result);
1222 return e;
1224 default:
1225 return e;
1228 case REFERENCE:
1230 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1231 tree folded;
1232 if ((folded = fully_constant_vn_reference_p (ref)))
1233 return get_or_alloc_expr_for_constant (folded);
1234 return e;
1236 default:
1237 return e;
1239 return e;
1242 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1243 it has the value it would have in BLOCK. Set *SAME_VALID to true
1244 in case the new vuse doesn't change the value id of the OPERANDS. */
1246 static tree
1247 translate_vuse_through_block (vec<vn_reference_op_s> operands,
1248 alias_set_type set, tree type, tree vuse,
1249 basic_block phiblock,
1250 basic_block block, bool *same_valid)
1252 gimple phi = SSA_NAME_DEF_STMT (vuse);
1253 ao_ref ref;
1254 edge e = NULL;
1255 bool use_oracle;
1257 *same_valid = true;
1259 if (gimple_bb (phi) != phiblock)
1260 return vuse;
1262 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1264 /* Use the alias-oracle to find either the PHI node in this block,
1265 the first VUSE used in this block that is equivalent to vuse or
1266 the first VUSE which definition in this block kills the value. */
1267 if (gimple_code (phi) == GIMPLE_PHI)
1268 e = find_edge (block, phiblock);
1269 else if (use_oracle)
1270 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1272 vuse = gimple_vuse (phi);
1273 phi = SSA_NAME_DEF_STMT (vuse);
1274 if (gimple_bb (phi) != phiblock)
1275 return vuse;
1276 if (gimple_code (phi) == GIMPLE_PHI)
1278 e = find_edge (block, phiblock);
1279 break;
1282 else
1283 return NULL_TREE;
1285 if (e)
1287 if (use_oracle)
1289 bitmap visited = NULL;
1290 unsigned int cnt;
1291 /* Try to find a vuse that dominates this phi node by skipping
1292 non-clobbering statements. */
1293 vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false);
1294 if (visited)
1295 BITMAP_FREE (visited);
1297 else
1298 vuse = NULL_TREE;
1299 if (!vuse)
1301 /* If we didn't find any, the value ID can't stay the same,
1302 but return the translated vuse. */
1303 *same_valid = false;
1304 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1306 /* ??? We would like to return vuse here as this is the canonical
1307 upmost vdef that this reference is associated with. But during
1308 insertion of the references into the hash tables we only ever
1309 directly insert with their direct gimple_vuse, hence returning
1310 something else would make us not find the other expression. */
1311 return PHI_ARG_DEF (phi, e->dest_idx);
1314 return NULL_TREE;
1317 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1318 SET2. This is used to avoid making a set consisting of the union
1319 of PA_IN and ANTIC_IN during insert. */
1321 static inline pre_expr
1322 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1324 pre_expr result;
1326 result = bitmap_find_leader (set1, val);
1327 if (!result && set2)
1328 result = bitmap_find_leader (set2, val);
1329 return result;
1332 /* Get the tree type for our PRE expression e. */
1334 static tree
1335 get_expr_type (const pre_expr e)
1337 switch (e->kind)
1339 case NAME:
1340 return TREE_TYPE (PRE_EXPR_NAME (e));
1341 case CONSTANT:
1342 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1343 case REFERENCE:
1344 return PRE_EXPR_REFERENCE (e)->type;
1345 case NARY:
1346 return PRE_EXPR_NARY (e)->type;
1348 gcc_unreachable ();
1351 /* Get a representative SSA_NAME for a given expression.
1352 Since all of our sub-expressions are treated as values, we require
1353 them to be SSA_NAME's for simplicity.
1354 Prior versions of GVNPRE used to use "value handles" here, so that
1355 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1356 either case, the operands are really values (IE we do not expect
1357 them to be usable without finding leaders). */
1359 static tree
1360 get_representative_for (const pre_expr e)
1362 tree name;
1363 unsigned int value_id = get_expr_value_id (e);
1365 switch (e->kind)
1367 case NAME:
1368 return PRE_EXPR_NAME (e);
1369 case CONSTANT:
1370 return PRE_EXPR_CONSTANT (e);
1371 case NARY:
1372 case REFERENCE:
1374 /* Go through all of the expressions representing this value
1375 and pick out an SSA_NAME. */
1376 unsigned int i;
1377 bitmap_iterator bi;
1378 bitmap exprs = value_expressions[value_id];
1379 EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1381 pre_expr rep = expression_for_id (i);
1382 if (rep->kind == NAME)
1383 return PRE_EXPR_NAME (rep);
1384 else if (rep->kind == CONSTANT)
1385 return PRE_EXPR_CONSTANT (rep);
1388 break;
1391 /* If we reached here we couldn't find an SSA_NAME. This can
1392 happen when we've discovered a value that has never appeared in
1393 the program as set to an SSA_NAME, as the result of phi translation.
1394 Create one here.
1395 ??? We should be able to re-use this when we insert the statement
1396 to compute it. */
1397 name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1398 VN_INFO_GET (name)->value_id = value_id;
1399 VN_INFO (name)->valnum = name;
1400 /* ??? For now mark this SSA name for release by SCCVN. */
1401 VN_INFO (name)->needs_insertion = true;
1402 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1403 if (dump_file && (dump_flags & TDF_DETAILS))
1405 fprintf (dump_file, "Created SSA_NAME representative ");
1406 print_generic_expr (dump_file, name, 0);
1407 fprintf (dump_file, " for expression:");
1408 print_pre_expr (dump_file, e);
1409 fprintf (dump_file, " (%04d)\n", value_id);
1412 return name;
1417 static pre_expr
1418 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1419 basic_block pred, basic_block phiblock);
1421 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1422 the phis in PRED. Return NULL if we can't find a leader for each part
1423 of the translated expression. */
1425 static pre_expr
1426 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1427 basic_block pred, basic_block phiblock)
1429 switch (expr->kind)
1431 case NARY:
1433 unsigned int i;
1434 bool changed = false;
1435 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1436 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1437 sizeof_vn_nary_op (nary->length));
1438 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1440 for (i = 0; i < newnary->length; i++)
1442 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1443 continue;
1444 else
1446 pre_expr leader, result;
1447 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1448 leader = find_leader_in_sets (op_val_id, set1, set2);
1449 result = phi_translate (leader, set1, set2, pred, phiblock);
1450 if (result && result != leader)
1452 tree name = get_representative_for (result);
1453 if (!name)
1454 return NULL;
1455 newnary->op[i] = name;
1457 else if (!result)
1458 return NULL;
1460 changed |= newnary->op[i] != nary->op[i];
1463 if (changed)
1465 pre_expr constant;
1466 unsigned int new_val_id;
1468 tree result = vn_nary_op_lookup_pieces (newnary->length,
1469 newnary->opcode,
1470 newnary->type,
1471 &newnary->op[0],
1472 &nary);
1473 if (result && is_gimple_min_invariant (result))
1474 return get_or_alloc_expr_for_constant (result);
1476 expr = (pre_expr) pool_alloc (pre_expr_pool);
1477 expr->kind = NARY;
1478 expr->id = 0;
1479 if (nary)
1481 PRE_EXPR_NARY (expr) = nary;
1482 constant = fully_constant_expression (expr);
1483 if (constant != expr)
1484 return constant;
1486 new_val_id = nary->value_id;
1487 get_or_alloc_expression_id (expr);
1489 else
1491 new_val_id = get_next_value_id ();
1492 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
1493 nary = vn_nary_op_insert_pieces (newnary->length,
1494 newnary->opcode,
1495 newnary->type,
1496 &newnary->op[0],
1497 result, new_val_id);
1498 PRE_EXPR_NARY (expr) = nary;
1499 constant = fully_constant_expression (expr);
1500 if (constant != expr)
1501 return constant;
1502 get_or_alloc_expression_id (expr);
1504 add_to_value (new_val_id, expr);
1506 return expr;
1508 break;
1510 case REFERENCE:
1512 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1513 vec<vn_reference_op_s> operands = ref->operands;
1514 tree vuse = ref->vuse;
1515 tree newvuse = vuse;
1516 vec<vn_reference_op_s> newoperands = vNULL;
1517 bool changed = false, same_valid = true;
1518 unsigned int i, j, n;
1519 vn_reference_op_t operand;
1520 vn_reference_t newref;
1522 for (i = 0, j = 0;
1523 operands.iterate (i, &operand); i++, j++)
1525 pre_expr opresult;
1526 pre_expr leader;
1527 tree op[3];
1528 tree type = operand->type;
1529 vn_reference_op_s newop = *operand;
1530 op[0] = operand->op0;
1531 op[1] = operand->op1;
1532 op[2] = operand->op2;
1533 for (n = 0; n < 3; ++n)
1535 unsigned int op_val_id;
1536 if (!op[n])
1537 continue;
1538 if (TREE_CODE (op[n]) != SSA_NAME)
1540 /* We can't possibly insert these. */
1541 if (n != 0
1542 && !is_gimple_min_invariant (op[n]))
1543 break;
1544 continue;
1546 op_val_id = VN_INFO (op[n])->value_id;
1547 leader = find_leader_in_sets (op_val_id, set1, set2);
1548 if (!leader)
1549 break;
1550 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1551 if (!opresult)
1552 break;
1553 if (opresult != leader)
1555 tree name = get_representative_for (opresult);
1556 if (!name)
1557 break;
1558 changed |= name != op[n];
1559 op[n] = name;
1562 if (n != 3)
1564 newoperands.release ();
1565 return NULL;
1567 if (!newoperands.exists ())
1568 newoperands = operands.copy ();
1569 /* We may have changed from an SSA_NAME to a constant */
1570 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1571 newop.opcode = TREE_CODE (op[0]);
1572 newop.type = type;
1573 newop.op0 = op[0];
1574 newop.op1 = op[1];
1575 newop.op2 = op[2];
1576 /* If it transforms a non-constant ARRAY_REF into a constant
1577 one, adjust the constant offset. */
1578 if (newop.opcode == ARRAY_REF
1579 && newop.off == -1
1580 && TREE_CODE (op[0]) == INTEGER_CST
1581 && TREE_CODE (op[1]) == INTEGER_CST
1582 && TREE_CODE (op[2]) == INTEGER_CST)
1584 double_int off = tree_to_double_int (op[0]);
1585 off += -tree_to_double_int (op[1]);
1586 off *= tree_to_double_int (op[2]);
1587 if (off.fits_shwi ())
1588 newop.off = off.low;
1590 newoperands[j] = newop;
1591 /* If it transforms from an SSA_NAME to an address, fold with
1592 a preceding indirect reference. */
1593 if (j > 0 && op[0] && TREE_CODE (op[0]) == ADDR_EXPR
1594 && newoperands[j - 1].opcode == MEM_REF)
1595 vn_reference_fold_indirect (&newoperands, &j);
1597 if (i != operands.length ())
1599 newoperands.release ();
1600 return NULL;
1603 if (vuse)
1605 newvuse = translate_vuse_through_block (newoperands,
1606 ref->set, ref->type,
1607 vuse, phiblock, pred,
1608 &same_valid);
1609 if (newvuse == NULL_TREE)
1611 newoperands.release ();
1612 return NULL;
1616 if (changed || newvuse != vuse)
1618 unsigned int new_val_id;
1619 pre_expr constant;
1621 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1622 ref->type,
1623 newoperands,
1624 &newref, VN_WALK);
1625 if (result)
1626 newoperands.release ();
1628 /* We can always insert constants, so if we have a partial
1629 redundant constant load of another type try to translate it
1630 to a constant of appropriate type. */
1631 if (result && is_gimple_min_invariant (result))
1633 tree tem = result;
1634 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1636 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1637 if (tem && !is_gimple_min_invariant (tem))
1638 tem = NULL_TREE;
1640 if (tem)
1641 return get_or_alloc_expr_for_constant (tem);
1644 /* If we'd have to convert things we would need to validate
1645 if we can insert the translated expression. So fail
1646 here for now - we cannot insert an alias with a different
1647 type in the VN tables either, as that would assert. */
1648 if (result
1649 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1650 return NULL;
1651 else if (!result && newref
1652 && !useless_type_conversion_p (ref->type, newref->type))
1654 newoperands.release ();
1655 return NULL;
1658 expr = (pre_expr) pool_alloc (pre_expr_pool);
1659 expr->kind = REFERENCE;
1660 expr->id = 0;
1662 if (newref)
1664 PRE_EXPR_REFERENCE (expr) = newref;
1665 constant = fully_constant_expression (expr);
1666 if (constant != expr)
1667 return constant;
1669 new_val_id = newref->value_id;
1670 get_or_alloc_expression_id (expr);
1672 else
1674 if (changed || !same_valid)
1676 new_val_id = get_next_value_id ();
1677 value_expressions.safe_grow_cleared
1678 (get_max_value_id () + 1);
1680 else
1681 new_val_id = ref->value_id;
1682 newref = vn_reference_insert_pieces (newvuse, ref->set,
1683 ref->type,
1684 newoperands,
1685 result, new_val_id);
1686 newoperands.create (0);
1687 PRE_EXPR_REFERENCE (expr) = newref;
1688 constant = fully_constant_expression (expr);
1689 if (constant != expr)
1690 return constant;
1691 get_or_alloc_expression_id (expr);
1693 add_to_value (new_val_id, expr);
1695 newoperands.release ();
1696 return expr;
1698 break;
1700 case NAME:
1702 tree name = PRE_EXPR_NAME (expr);
1703 gimple def_stmt = SSA_NAME_DEF_STMT (name);
1704 /* If the SSA name is defined by a PHI node in this block,
1705 translate it. */
1706 if (gimple_code (def_stmt) == GIMPLE_PHI
1707 && gimple_bb (def_stmt) == phiblock)
1709 edge e = find_edge (pred, gimple_bb (def_stmt));
1710 tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1712 /* Handle constant. */
1713 if (is_gimple_min_invariant (def))
1714 return get_or_alloc_expr_for_constant (def);
1716 return get_or_alloc_expr_for_name (def);
1718 /* Otherwise return it unchanged - it will get cleaned if its
1719 value is not available in PREDs AVAIL_OUT set of expressions. */
1720 return expr;
1723 default:
1724 gcc_unreachable ();
1728 /* Wrapper around phi_translate_1 providing caching functionality. */
1730 static pre_expr
1731 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1732 basic_block pred, basic_block phiblock)
1734 expr_pred_trans_t slot = NULL;
1735 pre_expr phitrans;
1737 if (!expr)
1738 return NULL;
1740 /* Constants contain no values that need translation. */
1741 if (expr->kind == CONSTANT)
1742 return expr;
1744 if (value_id_constant_p (get_expr_value_id (expr)))
1745 return expr;
1747 /* Don't add translations of NAMEs as those are cheap to translate. */
1748 if (expr->kind != NAME)
1750 if (phi_trans_add (&slot, expr, pred))
1751 return slot->v;
1752 /* Store NULL for the value we want to return in the case of
1753 recursing. */
1754 slot->v = NULL;
1757 /* Translate. */
1758 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1760 if (slot)
1761 slot->v = phitrans;
1763 return phitrans;
1767 /* For each expression in SET, translate the values through phi nodes
1768 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1769 expressions in DEST. */
1771 static void
1772 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1773 basic_block phiblock)
1775 vec<pre_expr> exprs;
1776 pre_expr expr;
1777 int i;
1779 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1781 bitmap_set_copy (dest, set);
1782 return;
1785 exprs = sorted_array_from_bitmap_set (set);
1786 FOR_EACH_VEC_ELT (exprs, i, expr)
1788 pre_expr translated;
1789 translated = phi_translate (expr, set, NULL, pred, phiblock);
1790 if (!translated)
1791 continue;
1793 /* We might end up with multiple expressions from SET being
1794 translated to the same value. In this case we do not want
1795 to retain the NARY or REFERENCE expression but prefer a NAME
1796 which would be the leader. */
1797 if (translated->kind == NAME)
1798 bitmap_value_replace_in_set (dest, translated);
1799 else
1800 bitmap_value_insert_into_set (dest, translated);
1802 exprs.release ();
1805 /* Find the leader for a value (i.e., the name representing that
1806 value) in a given set, and return it. Return NULL if no leader
1807 is found. */
1809 static pre_expr
1810 bitmap_find_leader (bitmap_set_t set, unsigned int val)
1812 if (value_id_constant_p (val))
1814 unsigned int i;
1815 bitmap_iterator bi;
1816 bitmap exprset = value_expressions[val];
1818 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1820 pre_expr expr = expression_for_id (i);
1821 if (expr->kind == CONSTANT)
1822 return expr;
1825 if (bitmap_set_contains_value (set, val))
1827 /* Rather than walk the entire bitmap of expressions, and see
1828 whether any of them has the value we are looking for, we look
1829 at the reverse mapping, which tells us the set of expressions
1830 that have a given value (IE value->expressions with that
1831 value) and see if any of those expressions are in our set.
1832 The number of expressions per value is usually significantly
1833 less than the number of expressions in the set. In fact, for
1834 large testcases, doing it this way is roughly 5-10x faster
1835 than walking the bitmap.
1836 If this is somehow a significant lose for some cases, we can
1837 choose which set to walk based on which set is smaller. */
1838 unsigned int i;
1839 bitmap_iterator bi;
1840 bitmap exprset = value_expressions[val];
1842 EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1843 return expression_for_id (i);
1845 return NULL;
1848 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1849 BLOCK by seeing if it is not killed in the block. Note that we are
1850 only determining whether there is a store that kills it. Because
1851 of the order in which clean iterates over values, we are guaranteed
1852 that altered operands will have caused us to be eliminated from the
1853 ANTIC_IN set already. */
1855 static bool
1856 value_dies_in_block_x (pre_expr expr, basic_block block)
1858 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1859 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1860 gimple def;
1861 gimple_stmt_iterator gsi;
1862 unsigned id = get_expression_id (expr);
1863 bool res = false;
1864 ao_ref ref;
1866 if (!vuse)
1867 return false;
1869 /* Lookup a previously calculated result. */
1870 if (EXPR_DIES (block)
1871 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1872 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1874 /* A memory expression {e, VUSE} dies in the block if there is a
1875 statement that may clobber e. If, starting statement walk from the
1876 top of the basic block, a statement uses VUSE there can be no kill
1877 inbetween that use and the original statement that loaded {e, VUSE},
1878 so we can stop walking. */
1879 ref.base = NULL_TREE;
1880 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1882 tree def_vuse, def_vdef;
1883 def = gsi_stmt (gsi);
1884 def_vuse = gimple_vuse (def);
1885 def_vdef = gimple_vdef (def);
1887 /* Not a memory statement. */
1888 if (!def_vuse)
1889 continue;
1891 /* Not a may-def. */
1892 if (!def_vdef)
1894 /* A load with the same VUSE, we're done. */
1895 if (def_vuse == vuse)
1896 break;
1898 continue;
1901 /* Init ref only if we really need it. */
1902 if (ref.base == NULL_TREE
1903 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1904 refx->operands))
1906 res = true;
1907 break;
1909 /* If the statement may clobber expr, it dies. */
1910 if (stmt_may_clobber_ref_p_1 (def, &ref))
1912 res = true;
1913 break;
1917 /* Remember the result. */
1918 if (!EXPR_DIES (block))
1919 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1920 bitmap_set_bit (EXPR_DIES (block), id * 2);
1921 if (res)
1922 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1924 return res;
1928 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1929 contains its value-id. */
1931 static bool
1932 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1934 if (op && TREE_CODE (op) == SSA_NAME)
1936 unsigned int value_id = VN_INFO (op)->value_id;
1937 if (!(bitmap_set_contains_value (set1, value_id)
1938 || (set2 && bitmap_set_contains_value (set2, value_id))))
1939 return false;
1941 return true;
1944 /* Determine if the expression EXPR is valid in SET1 U SET2.
1945 ONLY SET2 CAN BE NULL.
1946 This means that we have a leader for each part of the expression
1947 (if it consists of values), or the expression is an SSA_NAME.
1948 For loads/calls, we also see if the vuse is killed in this block. */
1950 static bool
1951 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
1952 basic_block block)
1954 switch (expr->kind)
1956 case NAME:
1957 return bitmap_find_leader (AVAIL_OUT (block),
1958 get_expr_value_id (expr)) != NULL;
1959 case NARY:
1961 unsigned int i;
1962 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1963 for (i = 0; i < nary->length; i++)
1964 if (!op_valid_in_sets (set1, set2, nary->op[i]))
1965 return false;
1966 return true;
1968 break;
1969 case REFERENCE:
1971 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1972 vn_reference_op_t vro;
1973 unsigned int i;
1975 FOR_EACH_VEC_ELT (ref->operands, i, vro)
1977 if (!op_valid_in_sets (set1, set2, vro->op0)
1978 || !op_valid_in_sets (set1, set2, vro->op1)
1979 || !op_valid_in_sets (set1, set2, vro->op2))
1980 return false;
1982 return true;
1984 default:
1985 gcc_unreachable ();
1989 /* Clean the set of expressions that are no longer valid in SET1 or
1990 SET2. This means expressions that are made up of values we have no
1991 leaders for in SET1 or SET2. This version is used for partial
1992 anticipation, which means it is not valid in either ANTIC_IN or
1993 PA_IN. */
1995 static void
1996 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
1998 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
1999 pre_expr expr;
2000 int i;
2002 FOR_EACH_VEC_ELT (exprs, i, expr)
2004 if (!valid_in_sets (set1, set2, expr, block))
2005 bitmap_remove_from_set (set1, expr);
2007 exprs.release ();
2010 /* Clean the set of expressions that are no longer valid in SET. This
2011 means expressions that are made up of values we have no leaders for
2012 in SET. */
2014 static void
2015 clean (bitmap_set_t set, basic_block block)
2017 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set);
2018 pre_expr expr;
2019 int i;
2021 FOR_EACH_VEC_ELT (exprs, i, expr)
2023 if (!valid_in_sets (set, NULL, expr, block))
2024 bitmap_remove_from_set (set, expr);
2026 exprs.release ();
2029 /* Clean the set of expressions that are no longer valid in SET because
2030 they are clobbered in BLOCK or because they trap and may not be executed. */
2032 static void
2033 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2035 bitmap_iterator bi;
2036 unsigned i;
2038 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2040 pre_expr expr = expression_for_id (i);
2041 if (expr->kind == REFERENCE)
2043 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2044 if (ref->vuse)
2046 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2047 if (!gimple_nop_p (def_stmt)
2048 && ((gimple_bb (def_stmt) != block
2049 && !dominated_by_p (CDI_DOMINATORS,
2050 block, gimple_bb (def_stmt)))
2051 || (gimple_bb (def_stmt) == block
2052 && value_dies_in_block_x (expr, block))))
2053 bitmap_remove_from_set (set, expr);
2056 else if (expr->kind == NARY)
2058 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2059 /* If the NARY may trap make sure the block does not contain
2060 a possible exit point.
2061 ??? This is overly conservative if we translate AVAIL_OUT
2062 as the available expression might be after the exit point. */
2063 if (BB_MAY_NOTRETURN (block)
2064 && vn_nary_may_trap (nary))
2065 bitmap_remove_from_set (set, expr);
2070 static sbitmap has_abnormal_preds;
2072 /* List of blocks that may have changed during ANTIC computation and
2073 thus need to be iterated over. */
2075 static sbitmap changed_blocks;
2077 /* Decide whether to defer a block for a later iteration, or PHI
2078 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2079 should defer the block, and true if we processed it. */
2081 static bool
2082 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2083 basic_block block, basic_block phiblock)
2085 if (!BB_VISITED (phiblock))
2087 bitmap_set_bit (changed_blocks, block->index);
2088 BB_VISITED (block) = 0;
2089 BB_DEFERRED (block) = 1;
2090 return false;
2092 else
2093 phi_translate_set (dest, source, block, phiblock);
2094 return true;
2097 /* Compute the ANTIC set for BLOCK.
2099 If succs(BLOCK) > 1 then
2100 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2101 else if succs(BLOCK) == 1 then
2102 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2104 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2107 static bool
2108 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2110 bool changed = false;
2111 bitmap_set_t S, old, ANTIC_OUT;
2112 bitmap_iterator bi;
2113 unsigned int bii;
2114 edge e;
2115 edge_iterator ei;
2117 old = ANTIC_OUT = S = NULL;
2118 BB_VISITED (block) = 1;
2120 /* If any edges from predecessors are abnormal, antic_in is empty,
2121 so do nothing. */
2122 if (block_has_abnormal_pred_edge)
2123 goto maybe_dump_sets;
2125 old = ANTIC_IN (block);
2126 ANTIC_OUT = bitmap_set_new ();
2128 /* If the block has no successors, ANTIC_OUT is empty. */
2129 if (EDGE_COUNT (block->succs) == 0)
2131 /* If we have one successor, we could have some phi nodes to
2132 translate through. */
2133 else if (single_succ_p (block))
2135 basic_block succ_bb = single_succ (block);
2137 /* We trade iterations of the dataflow equations for having to
2138 phi translate the maximal set, which is incredibly slow
2139 (since the maximal set often has 300+ members, even when you
2140 have a small number of blocks).
2141 Basically, we defer the computation of ANTIC for this block
2142 until we have processed it's successor, which will inevitably
2143 have a *much* smaller set of values to phi translate once
2144 clean has been run on it.
2145 The cost of doing this is that we technically perform more
2146 iterations, however, they are lower cost iterations.
2148 Timings for PRE on tramp3d-v4:
2149 without maximal set fix: 11 seconds
2150 with maximal set fix/without deferring: 26 seconds
2151 with maximal set fix/with deferring: 11 seconds
2154 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2155 block, succ_bb))
2157 changed = true;
2158 goto maybe_dump_sets;
2161 /* If we have multiple successors, we take the intersection of all of
2162 them. Note that in the case of loop exit phi nodes, we may have
2163 phis to translate through. */
2164 else
2166 vec<basic_block> worklist;
2167 size_t i;
2168 basic_block bprime, first = NULL;
2170 worklist.create (EDGE_COUNT (block->succs));
2171 FOR_EACH_EDGE (e, ei, block->succs)
2173 if (!first
2174 && BB_VISITED (e->dest))
2175 first = e->dest;
2176 else if (BB_VISITED (e->dest))
2177 worklist.quick_push (e->dest);
2180 /* Of multiple successors we have to have visited one already. */
2181 if (!first)
2183 bitmap_set_bit (changed_blocks, block->index);
2184 BB_VISITED (block) = 0;
2185 BB_DEFERRED (block) = 1;
2186 changed = true;
2187 worklist.release ();
2188 goto maybe_dump_sets;
2191 if (!gimple_seq_empty_p (phi_nodes (first)))
2192 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2193 else
2194 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2196 FOR_EACH_VEC_ELT (worklist, i, bprime)
2198 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2200 bitmap_set_t tmp = bitmap_set_new ();
2201 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2202 bitmap_set_and (ANTIC_OUT, tmp);
2203 bitmap_set_free (tmp);
2205 else
2206 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2208 worklist.release ();
2211 /* Prune expressions that are clobbered in block and thus become
2212 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2213 prune_clobbered_mems (ANTIC_OUT, block);
2215 /* Generate ANTIC_OUT - TMP_GEN. */
2216 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2218 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2219 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2220 TMP_GEN (block));
2222 /* Then union in the ANTIC_OUT - TMP_GEN values,
2223 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2224 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2225 bitmap_value_insert_into_set (ANTIC_IN (block),
2226 expression_for_id (bii));
2228 clean (ANTIC_IN (block), block);
2230 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2232 changed = true;
2233 bitmap_set_bit (changed_blocks, block->index);
2234 FOR_EACH_EDGE (e, ei, block->preds)
2235 bitmap_set_bit (changed_blocks, e->src->index);
2237 else
2238 bitmap_clear_bit (changed_blocks, block->index);
2240 maybe_dump_sets:
2241 if (dump_file && (dump_flags & TDF_DETAILS))
2243 if (!BB_DEFERRED (block) || BB_VISITED (block))
2245 if (ANTIC_OUT)
2246 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2248 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2249 block->index);
2251 if (S)
2252 print_bitmap_set (dump_file, S, "S", block->index);
2254 else
2256 fprintf (dump_file,
2257 "Block %d was deferred for a future iteration.\n",
2258 block->index);
2261 if (old)
2262 bitmap_set_free (old);
2263 if (S)
2264 bitmap_set_free (S);
2265 if (ANTIC_OUT)
2266 bitmap_set_free (ANTIC_OUT);
2267 return changed;
2270 /* Compute PARTIAL_ANTIC for BLOCK.
2272 If succs(BLOCK) > 1 then
2273 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2274 in ANTIC_OUT for all succ(BLOCK)
2275 else if succs(BLOCK) == 1 then
2276 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2278 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2279 - ANTIC_IN[BLOCK])
2282 static bool
2283 compute_partial_antic_aux (basic_block block,
2284 bool block_has_abnormal_pred_edge)
2286 bool changed = false;
2287 bitmap_set_t old_PA_IN;
2288 bitmap_set_t PA_OUT;
2289 edge e;
2290 edge_iterator ei;
2291 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2293 old_PA_IN = PA_OUT = NULL;
2295 /* If any edges from predecessors are abnormal, antic_in is empty,
2296 so do nothing. */
2297 if (block_has_abnormal_pred_edge)
2298 goto maybe_dump_sets;
2300 /* If there are too many partially anticipatable values in the
2301 block, phi_translate_set can take an exponential time: stop
2302 before the translation starts. */
2303 if (max_pa
2304 && single_succ_p (block)
2305 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2306 goto maybe_dump_sets;
2308 old_PA_IN = PA_IN (block);
2309 PA_OUT = bitmap_set_new ();
2311 /* If the block has no successors, ANTIC_OUT is empty. */
2312 if (EDGE_COUNT (block->succs) == 0)
2314 /* If we have one successor, we could have some phi nodes to
2315 translate through. Note that we can't phi translate across DFS
2316 back edges in partial antic, because it uses a union operation on
2317 the successors. For recurrences like IV's, we will end up
2318 generating a new value in the set on each go around (i + 3 (VH.1)
2319 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2320 else if (single_succ_p (block))
2322 basic_block succ = single_succ (block);
2323 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2324 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2326 /* If we have multiple successors, we take the union of all of
2327 them. */
2328 else
2330 vec<basic_block> worklist;
2331 size_t i;
2332 basic_block bprime;
2334 worklist.create (EDGE_COUNT (block->succs));
2335 FOR_EACH_EDGE (e, ei, block->succs)
2337 if (e->flags & EDGE_DFS_BACK)
2338 continue;
2339 worklist.quick_push (e->dest);
2341 if (worklist.length () > 0)
2343 FOR_EACH_VEC_ELT (worklist, i, bprime)
2345 unsigned int i;
2346 bitmap_iterator bi;
2348 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2349 bitmap_value_insert_into_set (PA_OUT,
2350 expression_for_id (i));
2351 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2353 bitmap_set_t pa_in = bitmap_set_new ();
2354 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2355 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2356 bitmap_value_insert_into_set (PA_OUT,
2357 expression_for_id (i));
2358 bitmap_set_free (pa_in);
2360 else
2361 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2362 bitmap_value_insert_into_set (PA_OUT,
2363 expression_for_id (i));
2366 worklist.release ();
2369 /* Prune expressions that are clobbered in block and thus become
2370 invalid if translated from PA_OUT to PA_IN. */
2371 prune_clobbered_mems (PA_OUT, block);
2373 /* PA_IN starts with PA_OUT - TMP_GEN.
2374 Then we subtract things from ANTIC_IN. */
2375 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2377 /* For partial antic, we want to put back in the phi results, since
2378 we will properly avoid making them partially antic over backedges. */
2379 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2380 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2382 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2383 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2385 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2387 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2389 changed = true;
2390 bitmap_set_bit (changed_blocks, block->index);
2391 FOR_EACH_EDGE (e, ei, block->preds)
2392 bitmap_set_bit (changed_blocks, e->src->index);
2394 else
2395 bitmap_clear_bit (changed_blocks, block->index);
2397 maybe_dump_sets:
2398 if (dump_file && (dump_flags & TDF_DETAILS))
2400 if (PA_OUT)
2401 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2403 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2405 if (old_PA_IN)
2406 bitmap_set_free (old_PA_IN);
2407 if (PA_OUT)
2408 bitmap_set_free (PA_OUT);
2409 return changed;
2412 /* Compute ANTIC and partial ANTIC sets. */
2414 static void
2415 compute_antic (void)
2417 bool changed = true;
2418 int num_iterations = 0;
2419 basic_block block;
2420 int i;
2422 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2423 We pre-build the map of blocks with incoming abnormal edges here. */
2424 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2425 bitmap_clear (has_abnormal_preds);
2427 FOR_ALL_BB (block)
2429 edge_iterator ei;
2430 edge e;
2432 FOR_EACH_EDGE (e, ei, block->preds)
2434 e->flags &= ~EDGE_DFS_BACK;
2435 if (e->flags & EDGE_ABNORMAL)
2437 bitmap_set_bit (has_abnormal_preds, block->index);
2438 break;
2442 BB_VISITED (block) = 0;
2443 BB_DEFERRED (block) = 0;
2445 /* While we are here, give empty ANTIC_IN sets to each block. */
2446 ANTIC_IN (block) = bitmap_set_new ();
2447 PA_IN (block) = bitmap_set_new ();
2450 /* At the exit block we anticipate nothing. */
2451 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2453 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2454 bitmap_ones (changed_blocks);
2455 while (changed)
2457 if (dump_file && (dump_flags & TDF_DETAILS))
2458 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2459 /* ??? We need to clear our PHI translation cache here as the
2460 ANTIC sets shrink and we restrict valid translations to
2461 those having operands with leaders in ANTIC. Same below
2462 for PA ANTIC computation. */
2463 num_iterations++;
2464 changed = false;
2465 for (i = postorder_num - 1; i >= 0; i--)
2467 if (bitmap_bit_p (changed_blocks, postorder[i]))
2469 basic_block block = BASIC_BLOCK (postorder[i]);
2470 changed |= compute_antic_aux (block,
2471 bitmap_bit_p (has_abnormal_preds,
2472 block->index));
2475 /* Theoretically possible, but *highly* unlikely. */
2476 gcc_checking_assert (num_iterations < 500);
2479 statistics_histogram_event (cfun, "compute_antic iterations",
2480 num_iterations);
2482 if (do_partial_partial)
2484 bitmap_ones (changed_blocks);
2485 mark_dfs_back_edges ();
2486 num_iterations = 0;
2487 changed = true;
2488 while (changed)
2490 if (dump_file && (dump_flags & TDF_DETAILS))
2491 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2492 num_iterations++;
2493 changed = false;
2494 for (i = postorder_num - 1 ; i >= 0; i--)
2496 if (bitmap_bit_p (changed_blocks, postorder[i]))
2498 basic_block block = BASIC_BLOCK (postorder[i]);
2499 changed
2500 |= compute_partial_antic_aux (block,
2501 bitmap_bit_p (has_abnormal_preds,
2502 block->index));
2505 /* Theoretically possible, but *highly* unlikely. */
2506 gcc_checking_assert (num_iterations < 500);
2508 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2509 num_iterations);
2511 sbitmap_free (has_abnormal_preds);
2512 sbitmap_free (changed_blocks);
2516 /* Inserted expressions are placed onto this worklist, which is used
2517 for performing quick dead code elimination of insertions we made
2518 that didn't turn out to be necessary. */
2519 static bitmap inserted_exprs;
2521 /* The actual worker for create_component_ref_by_pieces. */
2523 static tree
2524 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2525 unsigned int *operand, gimple_seq *stmts)
2527 vn_reference_op_t currop = &ref->operands[*operand];
2528 tree genop;
2529 ++*operand;
2530 switch (currop->opcode)
2532 case CALL_EXPR:
2534 tree folded, sc = NULL_TREE;
2535 unsigned int nargs = 0;
2536 tree fn, *args;
2537 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2538 fn = currop->op0;
2539 else
2540 fn = find_or_generate_expression (block, currop->op0, stmts);
2541 if (!fn)
2542 return NULL_TREE;
2543 if (currop->op1)
2545 sc = find_or_generate_expression (block, currop->op1, stmts);
2546 if (!sc)
2547 return NULL_TREE;
2549 args = XNEWVEC (tree, ref->operands.length () - 1);
2550 while (*operand < ref->operands.length ())
2552 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2553 operand, stmts);
2554 if (!args[nargs])
2555 return NULL_TREE;
2556 nargs++;
2558 folded = build_call_array (currop->type,
2559 (TREE_CODE (fn) == FUNCTION_DECL
2560 ? build_fold_addr_expr (fn) : fn),
2561 nargs, args);
2562 free (args);
2563 if (sc)
2564 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2565 return folded;
2568 case MEM_REF:
2570 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2571 stmts);
2572 if (!baseop)
2573 return NULL_TREE;
2574 tree offset = currop->op0;
2575 if (TREE_CODE (baseop) == ADDR_EXPR
2576 && handled_component_p (TREE_OPERAND (baseop, 0)))
2578 HOST_WIDE_INT off;
2579 tree base;
2580 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2581 &off);
2582 gcc_assert (base);
2583 offset = int_const_binop (PLUS_EXPR, offset,
2584 build_int_cst (TREE_TYPE (offset),
2585 off));
2586 baseop = build_fold_addr_expr (base);
2588 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2591 case TARGET_MEM_REF:
2593 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2594 vn_reference_op_t nextop = &ref->operands[++*operand];
2595 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2596 stmts);
2597 if (!baseop)
2598 return NULL_TREE;
2599 if (currop->op0)
2601 genop0 = find_or_generate_expression (block, currop->op0, stmts);
2602 if (!genop0)
2603 return NULL_TREE;
2605 if (nextop->op0)
2607 genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2608 if (!genop1)
2609 return NULL_TREE;
2611 return build5 (TARGET_MEM_REF, currop->type,
2612 baseop, currop->op2, genop0, currop->op1, genop1);
2615 case ADDR_EXPR:
2616 if (currop->op0)
2618 gcc_assert (is_gimple_min_invariant (currop->op0));
2619 return currop->op0;
2621 /* Fallthrough. */
2622 case REALPART_EXPR:
2623 case IMAGPART_EXPR:
2624 case VIEW_CONVERT_EXPR:
2626 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2627 stmts);
2628 if (!genop0)
2629 return NULL_TREE;
2630 return fold_build1 (currop->opcode, currop->type, genop0);
2633 case WITH_SIZE_EXPR:
2635 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2636 stmts);
2637 if (!genop0)
2638 return NULL_TREE;
2639 tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2640 if (!genop1)
2641 return NULL_TREE;
2642 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2645 case BIT_FIELD_REF:
2647 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2648 stmts);
2649 if (!genop0)
2650 return NULL_TREE;
2651 tree op1 = currop->op0;
2652 tree op2 = currop->op1;
2653 return fold_build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2656 /* For array ref vn_reference_op's, operand 1 of the array ref
2657 is op0 of the reference op and operand 3 of the array ref is
2658 op1. */
2659 case ARRAY_RANGE_REF:
2660 case ARRAY_REF:
2662 tree genop0;
2663 tree genop1 = currop->op0;
2664 tree genop2 = currop->op1;
2665 tree genop3 = currop->op2;
2666 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2667 stmts);
2668 if (!genop0)
2669 return NULL_TREE;
2670 genop1 = find_or_generate_expression (block, genop1, stmts);
2671 if (!genop1)
2672 return NULL_TREE;
2673 if (genop2)
2675 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2676 /* Drop zero minimum index if redundant. */
2677 if (integer_zerop (genop2)
2678 && (!domain_type
2679 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2680 genop2 = NULL_TREE;
2681 else
2683 genop2 = find_or_generate_expression (block, genop2, stmts);
2684 if (!genop2)
2685 return NULL_TREE;
2688 if (genop3)
2690 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2691 /* We can't always put a size in units of the element alignment
2692 here as the element alignment may be not visible. See
2693 PR43783. Simply drop the element size for constant
2694 sizes. */
2695 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2696 genop3 = NULL_TREE;
2697 else
2699 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2700 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2701 genop3 = find_or_generate_expression (block, genop3, stmts);
2702 if (!genop3)
2703 return NULL_TREE;
2706 return build4 (currop->opcode, currop->type, genop0, genop1,
2707 genop2, genop3);
2709 case COMPONENT_REF:
2711 tree op0;
2712 tree op1;
2713 tree genop2 = currop->op1;
2714 op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2715 if (!op0)
2716 return NULL_TREE;
2717 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2718 op1 = currop->op0;
2719 if (genop2)
2721 genop2 = find_or_generate_expression (block, genop2, stmts);
2722 if (!genop2)
2723 return NULL_TREE;
2725 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2728 case SSA_NAME:
2730 genop = find_or_generate_expression (block, currop->op0, stmts);
2731 return genop;
2733 case STRING_CST:
2734 case INTEGER_CST:
2735 case COMPLEX_CST:
2736 case VECTOR_CST:
2737 case REAL_CST:
2738 case CONSTRUCTOR:
2739 case VAR_DECL:
2740 case PARM_DECL:
2741 case CONST_DECL:
2742 case RESULT_DECL:
2743 case FUNCTION_DECL:
2744 return currop->op0;
2746 default:
2747 gcc_unreachable ();
2751 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2752 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2753 trying to rename aggregates into ssa form directly, which is a no no.
2755 Thus, this routine doesn't create temporaries, it just builds a
2756 single access expression for the array, calling
2757 find_or_generate_expression to build the innermost pieces.
2759 This function is a subroutine of create_expression_by_pieces, and
2760 should not be called on it's own unless you really know what you
2761 are doing. */
2763 static tree
2764 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2765 gimple_seq *stmts)
2767 unsigned int op = 0;
2768 return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2771 /* Find a simple leader for an expression, or generate one using
2772 create_expression_by_pieces from a NARY expression for the value.
2773 BLOCK is the basic_block we are looking for leaders in.
2774 OP is the tree expression to find a leader for or generate.
2775 Returns the leader or NULL_TREE on failure. */
2777 static tree
2778 find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2780 pre_expr expr = get_or_alloc_expr_for (op);
2781 unsigned int lookfor = get_expr_value_id (expr);
2782 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2783 if (leader)
2785 if (leader->kind == NAME)
2786 return PRE_EXPR_NAME (leader);
2787 else if (leader->kind == CONSTANT)
2788 return PRE_EXPR_CONSTANT (leader);
2790 /* Defer. */
2791 return NULL_TREE;
2794 /* It must be a complex expression, so generate it recursively. Note
2795 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2796 where the insert algorithm fails to insert a required expression. */
2797 bitmap exprset = value_expressions[lookfor];
2798 bitmap_iterator bi;
2799 unsigned int i;
2800 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2802 pre_expr temp = expression_for_id (i);
2803 /* We cannot insert random REFERENCE expressions at arbitrary
2804 places. We can insert NARYs which eventually re-materializes
2805 its operand values. */
2806 if (temp->kind == NARY)
2807 return create_expression_by_pieces (block, temp, stmts,
2808 get_expr_type (expr));
2811 /* Defer. */
2812 return NULL_TREE;
2815 #define NECESSARY GF_PLF_1
2817 /* Create an expression in pieces, so that we can handle very complex
2818 expressions that may be ANTIC, but not necessary GIMPLE.
2819 BLOCK is the basic block the expression will be inserted into,
2820 EXPR is the expression to insert (in value form)
2821 STMTS is a statement list to append the necessary insertions into.
2823 This function will die if we hit some value that shouldn't be
2824 ANTIC but is (IE there is no leader for it, or its components).
2825 The function returns NULL_TREE in case a different antic expression
2826 has to be inserted first.
2827 This function may also generate expressions that are themselves
2828 partially or fully redundant. Those that are will be either made
2829 fully redundant during the next iteration of insert (for partially
2830 redundant ones), or eliminated by eliminate (for fully redundant
2831 ones). */
2833 static tree
2834 create_expression_by_pieces (basic_block block, pre_expr expr,
2835 gimple_seq *stmts, tree type)
2837 tree name;
2838 tree folded;
2839 gimple_seq forced_stmts = NULL;
2840 unsigned int value_id;
2841 gimple_stmt_iterator gsi;
2842 tree exprtype = type ? type : get_expr_type (expr);
2843 pre_expr nameexpr;
2844 gimple newstmt;
2846 switch (expr->kind)
2848 /* We may hit the NAME/CONSTANT case if we have to convert types
2849 that value numbering saw through. */
2850 case NAME:
2851 folded = PRE_EXPR_NAME (expr);
2852 break;
2853 case CONSTANT:
2854 folded = PRE_EXPR_CONSTANT (expr);
2855 break;
2856 case REFERENCE:
2858 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2859 folded = create_component_ref_by_pieces (block, ref, stmts);
2860 if (!folded)
2861 return NULL_TREE;
2863 break;
2864 case NARY:
2866 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2867 tree *genop = XALLOCAVEC (tree, nary->length);
2868 unsigned i;
2869 for (i = 0; i < nary->length; ++i)
2871 genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2872 if (!genop[i])
2873 return NULL_TREE;
2874 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2875 may have conversions stripped. */
2876 if (nary->opcode == POINTER_PLUS_EXPR)
2878 if (i == 0)
2879 genop[i] = fold_convert (nary->type, genop[i]);
2880 else if (i == 1)
2881 genop[i] = convert_to_ptrofftype (genop[i]);
2883 else
2884 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]);
2886 if (nary->opcode == CONSTRUCTOR)
2888 vec<constructor_elt, va_gc> *elts = NULL;
2889 for (i = 0; i < nary->length; ++i)
2890 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2891 folded = build_constructor (nary->type, elts);
2893 else
2895 switch (nary->length)
2897 case 1:
2898 folded = fold_build1 (nary->opcode, nary->type,
2899 genop[0]);
2900 break;
2901 case 2:
2902 folded = fold_build2 (nary->opcode, nary->type,
2903 genop[0], genop[1]);
2904 break;
2905 case 3:
2906 folded = fold_build3 (nary->opcode, nary->type,
2907 genop[0], genop[1], genop[2]);
2908 break;
2909 default:
2910 gcc_unreachable ();
2914 break;
2915 default:
2916 gcc_unreachable ();
2919 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2920 folded = fold_convert (exprtype, folded);
2922 /* Force the generated expression to be a sequence of GIMPLE
2923 statements.
2924 We have to call unshare_expr because force_gimple_operand may
2925 modify the tree we pass to it. */
2926 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
2927 false, NULL);
2929 /* If we have any intermediate expressions to the value sets, add them
2930 to the value sets and chain them in the instruction stream. */
2931 if (forced_stmts)
2933 gsi = gsi_start (forced_stmts);
2934 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2936 gimple stmt = gsi_stmt (gsi);
2937 tree forcedname = gimple_get_lhs (stmt);
2938 pre_expr nameexpr;
2940 if (TREE_CODE (forcedname) == SSA_NAME)
2942 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
2943 VN_INFO_GET (forcedname)->valnum = forcedname;
2944 VN_INFO (forcedname)->value_id = get_next_value_id ();
2945 nameexpr = get_or_alloc_expr_for_name (forcedname);
2946 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2947 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2948 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2951 gimple_seq_add_seq (stmts, forced_stmts);
2954 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2955 newstmt = gimple_build_assign (name, folded);
2956 gimple_set_plf (newstmt, NECESSARY, false);
2958 gimple_seq_add_stmt (stmts, newstmt);
2959 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
2961 /* Fold the last statement. */
2962 gsi = gsi_last (*stmts);
2963 if (fold_stmt_inplace (&gsi))
2964 update_stmt (gsi_stmt (gsi));
2966 /* Add a value number to the temporary.
2967 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2968 we are creating the expression by pieces, and this particular piece of
2969 the expression may have been represented. There is no harm in replacing
2970 here. */
2971 value_id = get_expr_value_id (expr);
2972 VN_INFO_GET (name)->value_id = value_id;
2973 VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id);
2974 if (VN_INFO (name)->valnum == NULL_TREE)
2975 VN_INFO (name)->valnum = name;
2976 gcc_assert (VN_INFO (name)->valnum != NULL_TREE);
2977 nameexpr = get_or_alloc_expr_for_name (name);
2978 add_to_value (value_id, nameexpr);
2979 if (NEW_SETS (block))
2980 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2981 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2983 pre_stats.insertions++;
2984 if (dump_file && (dump_flags & TDF_DETAILS))
2986 fprintf (dump_file, "Inserted ");
2987 print_gimple_stmt (dump_file, newstmt, 0, 0);
2988 fprintf (dump_file, " in predecessor %d (%04d)\n",
2989 block->index, value_id);
2992 return name;
2996 /* Returns true if we want to inhibit the insertions of PHI nodes
2997 for the given EXPR for basic block BB (a member of a loop).
2998 We want to do this, when we fear that the induction variable we
2999 create might inhibit vectorization. */
3001 static bool
3002 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3004 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3005 vec<vn_reference_op_s> ops = vr->operands;
3006 vn_reference_op_t op;
3007 unsigned i;
3009 /* If we aren't going to vectorize we don't inhibit anything. */
3010 if (!flag_tree_loop_vectorize)
3011 return false;
3013 /* Otherwise we inhibit the insertion when the address of the
3014 memory reference is a simple induction variable. In other
3015 cases the vectorizer won't do anything anyway (either it's
3016 loop invariant or a complicated expression). */
3017 FOR_EACH_VEC_ELT (ops, i, op)
3019 switch (op->opcode)
3021 case CALL_EXPR:
3022 /* Calls are not a problem. */
3023 return false;
3025 case ARRAY_REF:
3026 case ARRAY_RANGE_REF:
3027 if (TREE_CODE (op->op0) != SSA_NAME)
3028 break;
3029 /* Fallthru. */
3030 case SSA_NAME:
3032 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3033 affine_iv iv;
3034 /* Default defs are loop invariant. */
3035 if (!defbb)
3036 break;
3037 /* Defined outside this loop, also loop invariant. */
3038 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3039 break;
3040 /* If it's a simple induction variable inhibit insertion,
3041 the vectorizer might be interested in this one. */
3042 if (simple_iv (bb->loop_father, bb->loop_father,
3043 op->op0, &iv, true))
3044 return true;
3045 /* No simple IV, vectorizer can't do anything, hence no
3046 reason to inhibit the transformation for this operand. */
3047 break;
3049 default:
3050 break;
3053 return false;
3056 /* Insert the to-be-made-available values of expression EXPRNUM for each
3057 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3058 merge the result with a phi node, given the same value number as
3059 NODE. Return true if we have inserted new stuff. */
3061 static bool
3062 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3063 vec<pre_expr> avail)
3065 pre_expr expr = expression_for_id (exprnum);
3066 pre_expr newphi;
3067 unsigned int val = get_expr_value_id (expr);
3068 edge pred;
3069 bool insertions = false;
3070 bool nophi = false;
3071 basic_block bprime;
3072 pre_expr eprime;
3073 edge_iterator ei;
3074 tree type = get_expr_type (expr);
3075 tree temp;
3076 gimple phi;
3078 /* Make sure we aren't creating an induction variable. */
3079 if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
3081 bool firstinsideloop = false;
3082 bool secondinsideloop = false;
3083 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3084 EDGE_PRED (block, 0)->src);
3085 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3086 EDGE_PRED (block, 1)->src);
3087 /* Induction variables only have one edge inside the loop. */
3088 if ((firstinsideloop ^ secondinsideloop)
3089 && (expr->kind != REFERENCE
3090 || inhibit_phi_insertion (block, expr)))
3092 if (dump_file && (dump_flags & TDF_DETAILS))
3093 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3094 nophi = true;
3098 /* Make the necessary insertions. */
3099 FOR_EACH_EDGE (pred, ei, block->preds)
3101 gimple_seq stmts = NULL;
3102 tree builtexpr;
3103 bprime = pred->src;
3104 eprime = avail[pred->dest_idx];
3106 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3108 builtexpr = create_expression_by_pieces (bprime, eprime,
3109 &stmts, type);
3110 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3111 gsi_insert_seq_on_edge (pred, stmts);
3112 if (!builtexpr)
3114 /* We cannot insert a PHI node if we failed to insert
3115 on one edge. */
3116 nophi = true;
3117 continue;
3119 avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3120 insertions = true;
3122 else if (eprime->kind == CONSTANT)
3124 /* Constants may not have the right type, fold_convert
3125 should give us back a constant with the right type. */
3126 tree constant = PRE_EXPR_CONSTANT (eprime);
3127 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3129 tree builtexpr = fold_convert (type, constant);
3130 if (!is_gimple_min_invariant (builtexpr))
3132 tree forcedexpr = force_gimple_operand (builtexpr,
3133 &stmts, true,
3134 NULL);
3135 if (!is_gimple_min_invariant (forcedexpr))
3137 if (forcedexpr != builtexpr)
3139 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3140 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3142 if (stmts)
3144 gimple_stmt_iterator gsi;
3145 gsi = gsi_start (stmts);
3146 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3148 gimple stmt = gsi_stmt (gsi);
3149 tree lhs = gimple_get_lhs (stmt);
3150 if (TREE_CODE (lhs) == SSA_NAME)
3151 bitmap_set_bit (inserted_exprs,
3152 SSA_NAME_VERSION (lhs));
3153 gimple_set_plf (stmt, NECESSARY, false);
3155 gsi_insert_seq_on_edge (pred, stmts);
3157 avail[pred->dest_idx]
3158 = get_or_alloc_expr_for_name (forcedexpr);
3161 else
3162 avail[pred->dest_idx]
3163 = get_or_alloc_expr_for_constant (builtexpr);
3166 else if (eprime->kind == NAME)
3168 /* We may have to do a conversion because our value
3169 numbering can look through types in certain cases, but
3170 our IL requires all operands of a phi node have the same
3171 type. */
3172 tree name = PRE_EXPR_NAME (eprime);
3173 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3175 tree builtexpr;
3176 tree forcedexpr;
3177 builtexpr = fold_convert (type, name);
3178 forcedexpr = force_gimple_operand (builtexpr,
3179 &stmts, true,
3180 NULL);
3182 if (forcedexpr != name)
3184 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3185 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3188 if (stmts)
3190 gimple_stmt_iterator gsi;
3191 gsi = gsi_start (stmts);
3192 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3194 gimple stmt = gsi_stmt (gsi);
3195 tree lhs = gimple_get_lhs (stmt);
3196 if (TREE_CODE (lhs) == SSA_NAME)
3197 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3198 gimple_set_plf (stmt, NECESSARY, false);
3200 gsi_insert_seq_on_edge (pred, stmts);
3202 avail[pred->dest_idx] = get_or_alloc_expr_for_name (forcedexpr);
3206 /* If we didn't want a phi node, and we made insertions, we still have
3207 inserted new stuff, and thus return true. If we didn't want a phi node,
3208 and didn't make insertions, we haven't added anything new, so return
3209 false. */
3210 if (nophi && insertions)
3211 return true;
3212 else if (nophi && !insertions)
3213 return false;
3215 /* Now build a phi for the new variable. */
3216 temp = make_temp_ssa_name (type, NULL, "prephitmp");
3217 phi = create_phi_node (temp, block);
3219 gimple_set_plf (phi, NECESSARY, false);
3220 VN_INFO_GET (temp)->value_id = val;
3221 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3222 if (VN_INFO (temp)->valnum == NULL_TREE)
3223 VN_INFO (temp)->valnum = temp;
3224 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3225 FOR_EACH_EDGE (pred, ei, block->preds)
3227 pre_expr ae = avail[pred->dest_idx];
3228 gcc_assert (get_expr_type (ae) == type
3229 || useless_type_conversion_p (type, get_expr_type (ae)));
3230 if (ae->kind == CONSTANT)
3231 add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3232 pred, UNKNOWN_LOCATION);
3233 else
3234 add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3237 newphi = get_or_alloc_expr_for_name (temp);
3238 add_to_value (val, newphi);
3240 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3241 this insertion, since we test for the existence of this value in PHI_GEN
3242 before proceeding with the partial redundancy checks in insert_aux.
3244 The value may exist in AVAIL_OUT, in particular, it could be represented
3245 by the expression we are trying to eliminate, in which case we want the
3246 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3247 inserted there.
3249 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3250 this block, because if it did, it would have existed in our dominator's
3251 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3254 bitmap_insert_into_set (PHI_GEN (block), newphi);
3255 bitmap_value_replace_in_set (AVAIL_OUT (block),
3256 newphi);
3257 bitmap_insert_into_set (NEW_SETS (block),
3258 newphi);
3260 if (dump_file && (dump_flags & TDF_DETAILS))
3262 fprintf (dump_file, "Created phi ");
3263 print_gimple_stmt (dump_file, phi, 0, 0);
3264 fprintf (dump_file, " in block %d (%04d)\n", block->index, val);
3266 pre_stats.phis++;
3267 return true;
3272 /* Perform insertion of partially redundant values.
3273 For BLOCK, do the following:
3274 1. Propagate the NEW_SETS of the dominator into the current block.
3275 If the block has multiple predecessors,
3276 2a. Iterate over the ANTIC expressions for the block to see if
3277 any of them are partially redundant.
3278 2b. If so, insert them into the necessary predecessors to make
3279 the expression fully redundant.
3280 2c. Insert a new PHI merging the values of the predecessors.
3281 2d. Insert the new PHI, and the new expressions, into the
3282 NEW_SETS set.
3283 3. Recursively call ourselves on the dominator children of BLOCK.
3285 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3286 do_regular_insertion and do_partial_insertion.
3290 static bool
3291 do_regular_insertion (basic_block block, basic_block dom)
3293 bool new_stuff = false;
3294 vec<pre_expr> exprs;
3295 pre_expr expr;
3296 vec<pre_expr> avail = vNULL;
3297 int i;
3299 exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3300 avail.safe_grow (EDGE_COUNT (block->preds));
3302 FOR_EACH_VEC_ELT (exprs, i, expr)
3304 if (expr->kind == NARY
3305 || expr->kind == REFERENCE)
3307 unsigned int val;
3308 bool by_some = false;
3309 bool cant_insert = false;
3310 bool all_same = true;
3311 pre_expr first_s = NULL;
3312 edge pred;
3313 basic_block bprime;
3314 pre_expr eprime = NULL;
3315 edge_iterator ei;
3316 pre_expr edoubleprime = NULL;
3317 bool do_insertion = false;
3319 val = get_expr_value_id (expr);
3320 if (bitmap_set_contains_value (PHI_GEN (block), val))
3321 continue;
3322 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3324 if (dump_file && (dump_flags & TDF_DETAILS))
3326 fprintf (dump_file, "Found fully redundant value: ");
3327 print_pre_expr (dump_file, expr);
3328 fprintf (dump_file, "\n");
3330 continue;
3333 FOR_EACH_EDGE (pred, ei, block->preds)
3335 unsigned int vprime;
3337 /* We should never run insertion for the exit block
3338 and so not come across fake pred edges. */
3339 gcc_assert (!(pred->flags & EDGE_FAKE));
3340 bprime = pred->src;
3341 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3342 bprime, block);
3344 /* eprime will generally only be NULL if the
3345 value of the expression, translated
3346 through the PHI for this predecessor, is
3347 undefined. If that is the case, we can't
3348 make the expression fully redundant,
3349 because its value is undefined along a
3350 predecessor path. We can thus break out
3351 early because it doesn't matter what the
3352 rest of the results are. */
3353 if (eprime == NULL)
3355 avail[pred->dest_idx] = NULL;
3356 cant_insert = true;
3357 break;
3360 eprime = fully_constant_expression (eprime);
3361 vprime = get_expr_value_id (eprime);
3362 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3363 vprime);
3364 if (edoubleprime == NULL)
3366 avail[pred->dest_idx] = eprime;
3367 all_same = false;
3369 else
3371 avail[pred->dest_idx] = edoubleprime;
3372 by_some = true;
3373 /* We want to perform insertions to remove a redundancy on
3374 a path in the CFG we want to optimize for speed. */
3375 if (optimize_edge_for_speed_p (pred))
3376 do_insertion = true;
3377 if (first_s == NULL)
3378 first_s = edoubleprime;
3379 else if (!pre_expr_d::equal (first_s, edoubleprime))
3380 all_same = false;
3383 /* If we can insert it, it's not the same value
3384 already existing along every predecessor, and
3385 it's defined by some predecessor, it is
3386 partially redundant. */
3387 if (!cant_insert && !all_same && by_some)
3389 if (!do_insertion)
3391 if (dump_file && (dump_flags & TDF_DETAILS))
3393 fprintf (dump_file, "Skipping partial redundancy for "
3394 "expression ");
3395 print_pre_expr (dump_file, expr);
3396 fprintf (dump_file, " (%04d), no redundancy on to be "
3397 "optimized for speed edge\n", val);
3400 else if (dbg_cnt (treepre_insert))
3402 if (dump_file && (dump_flags & TDF_DETAILS))
3404 fprintf (dump_file, "Found partial redundancy for "
3405 "expression ");
3406 print_pre_expr (dump_file, expr);
3407 fprintf (dump_file, " (%04d)\n",
3408 get_expr_value_id (expr));
3410 if (insert_into_preds_of_block (block,
3411 get_expression_id (expr),
3412 avail))
3413 new_stuff = true;
3416 /* If all edges produce the same value and that value is
3417 an invariant, then the PHI has the same value on all
3418 edges. Note this. */
3419 else if (!cant_insert && all_same)
3421 gcc_assert (edoubleprime->kind == CONSTANT
3422 || edoubleprime->kind == NAME);
3424 tree temp = make_temp_ssa_name (get_expr_type (expr),
3425 NULL, "pretmp");
3426 gimple assign = gimple_build_assign (temp,
3427 edoubleprime->kind == CONSTANT ? PRE_EXPR_CONSTANT (edoubleprime) : PRE_EXPR_NAME (edoubleprime));
3428 gimple_stmt_iterator gsi = gsi_after_labels (block);
3429 gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3431 gimple_set_plf (assign, NECESSARY, false);
3432 VN_INFO_GET (temp)->value_id = val;
3433 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3434 if (VN_INFO (temp)->valnum == NULL_TREE)
3435 VN_INFO (temp)->valnum = temp;
3436 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3437 pre_expr newe = get_or_alloc_expr_for_name (temp);
3438 add_to_value (val, newe);
3439 bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3440 bitmap_insert_into_set (NEW_SETS (block), newe);
3445 exprs.release ();
3446 avail.release ();
3447 return new_stuff;
3451 /* Perform insertion for partially anticipatable expressions. There
3452 is only one case we will perform insertion for these. This case is
3453 if the expression is partially anticipatable, and fully available.
3454 In this case, we know that putting it earlier will enable us to
3455 remove the later computation. */
3458 static bool
3459 do_partial_partial_insertion (basic_block block, basic_block dom)
3461 bool new_stuff = false;
3462 vec<pre_expr> exprs;
3463 pre_expr expr;
3464 vec<pre_expr> avail = vNULL;
3465 int i;
3467 exprs = sorted_array_from_bitmap_set (PA_IN (block));
3468 avail.safe_grow (EDGE_COUNT (block->preds));
3470 FOR_EACH_VEC_ELT (exprs, i, expr)
3472 if (expr->kind == NARY
3473 || expr->kind == REFERENCE)
3475 unsigned int val;
3476 bool by_all = true;
3477 bool cant_insert = false;
3478 edge pred;
3479 basic_block bprime;
3480 pre_expr eprime = NULL;
3481 edge_iterator ei;
3483 val = get_expr_value_id (expr);
3484 if (bitmap_set_contains_value (PHI_GEN (block), val))
3485 continue;
3486 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3487 continue;
3489 FOR_EACH_EDGE (pred, ei, block->preds)
3491 unsigned int vprime;
3492 pre_expr edoubleprime;
3494 /* We should never run insertion for the exit block
3495 and so not come across fake pred edges. */
3496 gcc_assert (!(pred->flags & EDGE_FAKE));
3497 bprime = pred->src;
3498 eprime = phi_translate (expr, ANTIC_IN (block),
3499 PA_IN (block),
3500 bprime, block);
3502 /* eprime will generally only be NULL if the
3503 value of the expression, translated
3504 through the PHI for this predecessor, is
3505 undefined. If that is the case, we can't
3506 make the expression fully redundant,
3507 because its value is undefined along a
3508 predecessor path. We can thus break out
3509 early because it doesn't matter what the
3510 rest of the results are. */
3511 if (eprime == NULL)
3513 avail[pred->dest_idx] = NULL;
3514 cant_insert = true;
3515 break;
3518 eprime = fully_constant_expression (eprime);
3519 vprime = get_expr_value_id (eprime);
3520 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3521 avail[pred->dest_idx] = edoubleprime;
3522 if (edoubleprime == NULL)
3524 by_all = false;
3525 break;
3529 /* If we can insert it, it's not the same value
3530 already existing along every predecessor, and
3531 it's defined by some predecessor, it is
3532 partially redundant. */
3533 if (!cant_insert && by_all)
3535 edge succ;
3536 bool do_insertion = false;
3538 /* Insert only if we can remove a later expression on a path
3539 that we want to optimize for speed.
3540 The phi node that we will be inserting in BLOCK is not free,
3541 and inserting it for the sake of !optimize_for_speed successor
3542 may cause regressions on the speed path. */
3543 FOR_EACH_EDGE (succ, ei, block->succs)
3545 if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3546 || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3548 if (optimize_edge_for_speed_p (succ))
3549 do_insertion = true;
3553 if (!do_insertion)
3555 if (dump_file && (dump_flags & TDF_DETAILS))
3557 fprintf (dump_file, "Skipping partial partial redundancy "
3558 "for expression ");
3559 print_pre_expr (dump_file, expr);
3560 fprintf (dump_file, " (%04d), not (partially) anticipated "
3561 "on any to be optimized for speed edges\n", val);
3564 else if (dbg_cnt (treepre_insert))
3566 pre_stats.pa_insert++;
3567 if (dump_file && (dump_flags & TDF_DETAILS))
3569 fprintf (dump_file, "Found partial partial redundancy "
3570 "for expression ");
3571 print_pre_expr (dump_file, expr);
3572 fprintf (dump_file, " (%04d)\n",
3573 get_expr_value_id (expr));
3575 if (insert_into_preds_of_block (block,
3576 get_expression_id (expr),
3577 avail))
3578 new_stuff = true;
3584 exprs.release ();
3585 avail.release ();
3586 return new_stuff;
3589 static bool
3590 insert_aux (basic_block block)
3592 basic_block son;
3593 bool new_stuff = false;
3595 if (block)
3597 basic_block dom;
3598 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3599 if (dom)
3601 unsigned i;
3602 bitmap_iterator bi;
3603 bitmap_set_t newset = NEW_SETS (dom);
3604 if (newset)
3606 /* Note that we need to value_replace both NEW_SETS, and
3607 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3608 represented by some non-simple expression here that we want
3609 to replace it with. */
3610 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3612 pre_expr expr = expression_for_id (i);
3613 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3614 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3617 if (!single_pred_p (block))
3619 new_stuff |= do_regular_insertion (block, dom);
3620 if (do_partial_partial)
3621 new_stuff |= do_partial_partial_insertion (block, dom);
3625 for (son = first_dom_son (CDI_DOMINATORS, block);
3626 son;
3627 son = next_dom_son (CDI_DOMINATORS, son))
3629 new_stuff |= insert_aux (son);
3632 return new_stuff;
3635 /* Perform insertion of partially redundant values. */
3637 static void
3638 insert (void)
3640 bool new_stuff = true;
3641 basic_block bb;
3642 int num_iterations = 0;
3644 FOR_ALL_BB (bb)
3645 NEW_SETS (bb) = bitmap_set_new ();
3647 while (new_stuff)
3649 num_iterations++;
3650 if (dump_file && dump_flags & TDF_DETAILS)
3651 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3652 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3654 /* Clear the NEW sets before the next iteration. We have already
3655 fully propagated its contents. */
3656 if (new_stuff)
3657 FOR_ALL_BB (bb)
3658 bitmap_set_free (NEW_SETS (bb));
3660 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3664 /* Compute the AVAIL set for all basic blocks.
3666 This function performs value numbering of the statements in each basic
3667 block. The AVAIL sets are built from information we glean while doing
3668 this value numbering, since the AVAIL sets contain only one entry per
3669 value.
3671 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3672 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3674 static void
3675 compute_avail (void)
3678 basic_block block, son;
3679 basic_block *worklist;
3680 size_t sp = 0;
3681 unsigned i;
3683 /* We pretend that default definitions are defined in the entry block.
3684 This includes function arguments and the static chain decl. */
3685 for (i = 1; i < num_ssa_names; ++i)
3687 tree name = ssa_name (i);
3688 pre_expr e;
3689 if (!name
3690 || !SSA_NAME_IS_DEFAULT_DEF (name)
3691 || has_zero_uses (name)
3692 || virtual_operand_p (name))
3693 continue;
3695 e = get_or_alloc_expr_for_name (name);
3696 add_to_value (get_expr_value_id (e), e);
3697 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3698 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3701 if (dump_file && (dump_flags & TDF_DETAILS))
3703 print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR),
3704 "tmp_gen", ENTRY_BLOCK);
3705 print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR),
3706 "avail_out", ENTRY_BLOCK);
3709 /* Allocate the worklist. */
3710 worklist = XNEWVEC (basic_block, n_basic_blocks);
3712 /* Seed the algorithm by putting the dominator children of the entry
3713 block on the worklist. */
3714 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3715 son;
3716 son = next_dom_son (CDI_DOMINATORS, son))
3717 worklist[sp++] = son;
3719 /* Loop until the worklist is empty. */
3720 while (sp)
3722 gimple_stmt_iterator gsi;
3723 gimple stmt;
3724 basic_block dom;
3726 /* Pick a block from the worklist. */
3727 block = worklist[--sp];
3729 /* Initially, the set of available values in BLOCK is that of
3730 its immediate dominator. */
3731 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3732 if (dom)
3733 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3735 /* Generate values for PHI nodes. */
3736 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3738 tree result = gimple_phi_result (gsi_stmt (gsi));
3740 /* We have no need for virtual phis, as they don't represent
3741 actual computations. */
3742 if (virtual_operand_p (result))
3743 continue;
3745 pre_expr e = get_or_alloc_expr_for_name (result);
3746 add_to_value (get_expr_value_id (e), e);
3747 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3748 bitmap_insert_into_set (PHI_GEN (block), e);
3751 BB_MAY_NOTRETURN (block) = 0;
3753 /* Now compute value numbers and populate value sets with all
3754 the expressions computed in BLOCK. */
3755 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3757 ssa_op_iter iter;
3758 tree op;
3760 stmt = gsi_stmt (gsi);
3762 /* Cache whether the basic-block has any non-visible side-effect
3763 or control flow.
3764 If this isn't a call or it is the last stmt in the
3765 basic-block then the CFG represents things correctly. */
3766 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3768 /* Non-looping const functions always return normally.
3769 Otherwise the call might not return or have side-effects
3770 that forbids hoisting possibly trapping expressions
3771 before it. */
3772 int flags = gimple_call_flags (stmt);
3773 if (!(flags & ECF_CONST)
3774 || (flags & ECF_LOOPING_CONST_OR_PURE))
3775 BB_MAY_NOTRETURN (block) = 1;
3778 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3780 pre_expr e = get_or_alloc_expr_for_name (op);
3782 add_to_value (get_expr_value_id (e), e);
3783 bitmap_insert_into_set (TMP_GEN (block), e);
3784 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3787 if (gimple_has_side_effects (stmt)
3788 || stmt_could_throw_p (stmt)
3789 || is_gimple_debug (stmt))
3790 continue;
3792 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3794 if (ssa_undefined_value_p (op))
3795 continue;
3796 pre_expr e = get_or_alloc_expr_for_name (op);
3797 bitmap_value_insert_into_set (EXP_GEN (block), e);
3800 switch (gimple_code (stmt))
3802 case GIMPLE_RETURN:
3803 continue;
3805 case GIMPLE_CALL:
3807 vn_reference_t ref;
3808 pre_expr result = NULL;
3809 vec<vn_reference_op_s> ops = vNULL;
3811 /* We can value number only calls to real functions. */
3812 if (gimple_call_internal_p (stmt))
3813 continue;
3815 copy_reference_ops_from_call (stmt, &ops);
3816 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3817 gimple_expr_type (stmt),
3818 ops, &ref, VN_NOWALK);
3819 ops.release ();
3820 if (!ref)
3821 continue;
3823 /* If the value of the call is not invalidated in
3824 this block until it is computed, add the expression
3825 to EXP_GEN. */
3826 if (!gimple_vuse (stmt)
3827 || gimple_code
3828 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
3829 || gimple_bb (SSA_NAME_DEF_STMT
3830 (gimple_vuse (stmt))) != block)
3832 result = (pre_expr) pool_alloc (pre_expr_pool);
3833 result->kind = REFERENCE;
3834 result->id = 0;
3835 PRE_EXPR_REFERENCE (result) = ref;
3837 get_or_alloc_expression_id (result);
3838 add_to_value (get_expr_value_id (result), result);
3839 bitmap_value_insert_into_set (EXP_GEN (block), result);
3841 continue;
3844 case GIMPLE_ASSIGN:
3846 pre_expr result = NULL;
3847 switch (vn_get_stmt_kind (stmt))
3849 case VN_NARY:
3851 enum tree_code code = gimple_assign_rhs_code (stmt);
3852 vn_nary_op_t nary;
3854 /* COND_EXPR and VEC_COND_EXPR are awkward in
3855 that they contain an embedded complex expression.
3856 Don't even try to shove those through PRE. */
3857 if (code == COND_EXPR
3858 || code == VEC_COND_EXPR)
3859 continue;
3861 vn_nary_op_lookup_stmt (stmt, &nary);
3862 if (!nary)
3863 continue;
3865 /* If the NARY traps and there was a preceding
3866 point in the block that might not return avoid
3867 adding the nary to EXP_GEN. */
3868 if (BB_MAY_NOTRETURN (block)
3869 && vn_nary_may_trap (nary))
3870 continue;
3872 result = (pre_expr) pool_alloc (pre_expr_pool);
3873 result->kind = NARY;
3874 result->id = 0;
3875 PRE_EXPR_NARY (result) = nary;
3876 break;
3879 case VN_REFERENCE:
3881 vn_reference_t ref;
3882 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3883 gimple_vuse (stmt),
3884 VN_WALK, &ref);
3885 if (!ref)
3886 continue;
3888 /* If the value of the reference is not invalidated in
3889 this block until it is computed, add the expression
3890 to EXP_GEN. */
3891 if (gimple_vuse (stmt))
3893 gimple def_stmt;
3894 bool ok = true;
3895 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
3896 while (!gimple_nop_p (def_stmt)
3897 && gimple_code (def_stmt) != GIMPLE_PHI
3898 && gimple_bb (def_stmt) == block)
3900 if (stmt_may_clobber_ref_p
3901 (def_stmt, gimple_assign_rhs1 (stmt)))
3903 ok = false;
3904 break;
3906 def_stmt
3907 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
3909 if (!ok)
3910 continue;
3913 result = (pre_expr) pool_alloc (pre_expr_pool);
3914 result->kind = REFERENCE;
3915 result->id = 0;
3916 PRE_EXPR_REFERENCE (result) = ref;
3917 break;
3920 default:
3921 continue;
3924 get_or_alloc_expression_id (result);
3925 add_to_value (get_expr_value_id (result), result);
3926 bitmap_value_insert_into_set (EXP_GEN (block), result);
3927 continue;
3929 default:
3930 break;
3934 if (dump_file && (dump_flags & TDF_DETAILS))
3936 print_bitmap_set (dump_file, EXP_GEN (block),
3937 "exp_gen", block->index);
3938 print_bitmap_set (dump_file, PHI_GEN (block),
3939 "phi_gen", block->index);
3940 print_bitmap_set (dump_file, TMP_GEN (block),
3941 "tmp_gen", block->index);
3942 print_bitmap_set (dump_file, AVAIL_OUT (block),
3943 "avail_out", block->index);
3946 /* Put the dominator children of BLOCK on the worklist of blocks
3947 to compute available sets for. */
3948 for (son = first_dom_son (CDI_DOMINATORS, block);
3949 son;
3950 son = next_dom_son (CDI_DOMINATORS, son))
3951 worklist[sp++] = son;
3954 free (worklist);
3958 /* Local state for the eliminate domwalk. */
3959 static vec<gimple> el_to_remove;
3960 static vec<gimple> el_to_update;
3961 static unsigned int el_todo;
3962 static vec<tree> el_avail;
3963 static vec<tree> el_avail_stack;
3965 /* Return a leader for OP that is available at the current point of the
3966 eliminate domwalk. */
3968 static tree
3969 eliminate_avail (tree op)
3971 tree valnum = VN_INFO (op)->valnum;
3972 if (TREE_CODE (valnum) == SSA_NAME)
3974 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
3975 return valnum;
3976 if (el_avail.length () > SSA_NAME_VERSION (valnum))
3977 return el_avail[SSA_NAME_VERSION (valnum)];
3979 else if (is_gimple_min_invariant (valnum))
3980 return valnum;
3981 return NULL_TREE;
3984 /* At the current point of the eliminate domwalk make OP available. */
3986 static void
3987 eliminate_push_avail (tree op)
3989 tree valnum = VN_INFO (op)->valnum;
3990 if (TREE_CODE (valnum) == SSA_NAME)
3992 if (el_avail.length () <= SSA_NAME_VERSION (valnum))
3993 el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1);
3994 el_avail[SSA_NAME_VERSION (valnum)] = op;
3995 el_avail_stack.safe_push (op);
3999 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
4000 the leader for the expression if insertion was successful. */
4002 static tree
4003 eliminate_insert (gimple_stmt_iterator *gsi, tree val)
4005 tree expr = vn_get_expr_for (val);
4006 if (!CONVERT_EXPR_P (expr)
4007 && TREE_CODE (expr) != VIEW_CONVERT_EXPR)
4008 return NULL_TREE;
4010 tree op = TREE_OPERAND (expr, 0);
4011 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op;
4012 if (!leader)
4013 return NULL_TREE;
4015 tree res = make_temp_ssa_name (TREE_TYPE (val), NULL, "pretmp");
4016 gimple tem = gimple_build_assign (res,
4017 fold_build1 (TREE_CODE (expr),
4018 TREE_TYPE (expr), leader));
4019 gsi_insert_before (gsi, tem, GSI_SAME_STMT);
4020 VN_INFO_GET (res)->valnum = val;
4022 if (TREE_CODE (leader) == SSA_NAME)
4023 gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true);
4025 pre_stats.insertions++;
4026 if (dump_file && (dump_flags & TDF_DETAILS))
4028 fprintf (dump_file, "Inserted ");
4029 print_gimple_stmt (dump_file, tem, 0, 0);
4032 return res;
4035 class eliminate_dom_walker : public dom_walker
4037 public:
4038 eliminate_dom_walker (cdi_direction direction) : dom_walker (direction) {}
4040 virtual void before_dom_children (basic_block);
4041 virtual void after_dom_children (basic_block);
4044 /* Perform elimination for the basic-block B during the domwalk. */
4046 void
4047 eliminate_dom_walker::before_dom_children (basic_block b)
4049 gimple_stmt_iterator gsi;
4050 gimple stmt;
4052 /* Mark new bb. */
4053 el_avail_stack.safe_push (NULL_TREE);
4055 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4057 gimple stmt, phi = gsi_stmt (gsi);
4058 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4059 gimple_stmt_iterator gsi2;
4061 /* We want to perform redundant PHI elimination. Do so by
4062 replacing the PHI with a single copy if possible.
4063 Do not touch inserted, single-argument or virtual PHIs. */
4064 if (gimple_phi_num_args (phi) == 1
4065 || virtual_operand_p (res))
4067 gsi_next (&gsi);
4068 continue;
4071 sprime = eliminate_avail (res);
4072 if (!sprime
4073 || sprime == res)
4075 eliminate_push_avail (res);
4076 gsi_next (&gsi);
4077 continue;
4079 else if (is_gimple_min_invariant (sprime))
4081 if (!useless_type_conversion_p (TREE_TYPE (res),
4082 TREE_TYPE (sprime)))
4083 sprime = fold_convert (TREE_TYPE (res), sprime);
4086 if (dump_file && (dump_flags & TDF_DETAILS))
4088 fprintf (dump_file, "Replaced redundant PHI node defining ");
4089 print_generic_expr (dump_file, res, 0);
4090 fprintf (dump_file, " with ");
4091 print_generic_expr (dump_file, sprime, 0);
4092 fprintf (dump_file, "\n");
4095 remove_phi_node (&gsi, false);
4097 if (inserted_exprs
4098 && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4099 && TREE_CODE (sprime) == SSA_NAME)
4100 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4102 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4103 sprime = fold_convert (TREE_TYPE (res), sprime);
4104 stmt = gimple_build_assign (res, sprime);
4105 SSA_NAME_DEF_STMT (res) = stmt;
4106 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4108 gsi2 = gsi_after_labels (b);
4109 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4110 /* Queue the copy for eventual removal. */
4111 el_to_remove.safe_push (stmt);
4112 /* If we inserted this PHI node ourself, it's not an elimination. */
4113 if (inserted_exprs
4114 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4115 pre_stats.phis--;
4116 else
4117 pre_stats.eliminations++;
4120 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4122 tree lhs = NULL_TREE;
4123 tree rhs = NULL_TREE;
4125 stmt = gsi_stmt (gsi);
4127 if (gimple_has_lhs (stmt))
4128 lhs = gimple_get_lhs (stmt);
4130 if (gimple_assign_single_p (stmt))
4131 rhs = gimple_assign_rhs1 (stmt);
4133 /* Lookup the RHS of the expression, see if we have an
4134 available computation for it. If so, replace the RHS with
4135 the available computation. */
4136 if (gimple_has_lhs (stmt)
4137 && TREE_CODE (lhs) == SSA_NAME
4138 && !gimple_has_volatile_ops (stmt))
4140 tree sprime;
4141 gimple orig_stmt = stmt;
4143 sprime = eliminate_avail (lhs);
4144 /* If there is no usable leader mark lhs as leader for its value. */
4145 if (!sprime)
4146 eliminate_push_avail (lhs);
4148 /* See PR43491. Do not replace a global register variable when
4149 it is a the RHS of an assignment. Do replace local register
4150 variables since gcc does not guarantee a local variable will
4151 be allocated in register.
4152 Do not perform copy propagation or undo constant propagation. */
4153 if (gimple_assign_single_p (stmt)
4154 && (TREE_CODE (rhs) == SSA_NAME
4155 || is_gimple_min_invariant (rhs)
4156 || (TREE_CODE (rhs) == VAR_DECL
4157 && is_global_var (rhs)
4158 && DECL_HARD_REGISTER (rhs))))
4159 continue;
4161 if (!sprime)
4163 /* If there is no existing usable leader but SCCVN thinks
4164 it has an expression it wants to use as replacement,
4165 insert that. */
4166 tree val = VN_INFO (lhs)->valnum;
4167 if (val != VN_TOP
4168 && TREE_CODE (val) == SSA_NAME
4169 && VN_INFO (val)->needs_insertion
4170 && VN_INFO (val)->expr != NULL_TREE
4171 && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE)
4172 eliminate_push_avail (sprime);
4174 else if (is_gimple_min_invariant (sprime))
4176 /* If there is no existing leader but SCCVN knows this
4177 value is constant, use that constant. */
4178 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4179 TREE_TYPE (sprime)))
4180 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4182 if (dump_file && (dump_flags & TDF_DETAILS))
4184 fprintf (dump_file, "Replaced ");
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, " in ");
4189 print_gimple_stmt (dump_file, stmt, 0, 0);
4191 pre_stats.eliminations++;
4192 propagate_tree_value_into_stmt (&gsi, sprime);
4193 stmt = gsi_stmt (gsi);
4194 update_stmt (stmt);
4196 /* If we removed EH side-effects from the statement, clean
4197 its EH information. */
4198 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4200 bitmap_set_bit (need_eh_cleanup,
4201 gimple_bb (stmt)->index);
4202 if (dump_file && (dump_flags & TDF_DETAILS))
4203 fprintf (dump_file, " Removed EH side-effects.\n");
4205 continue;
4208 if (sprime
4209 && sprime != lhs
4210 && (rhs == NULL_TREE
4211 || TREE_CODE (rhs) != SSA_NAME
4212 || may_propagate_copy (rhs, sprime)))
4214 bool can_make_abnormal_goto
4215 = is_gimple_call (stmt)
4216 && stmt_can_make_abnormal_goto (stmt);
4218 gcc_assert (sprime != rhs);
4220 if (dump_file && (dump_flags & TDF_DETAILS))
4222 fprintf (dump_file, "Replaced ");
4223 print_gimple_expr (dump_file, stmt, 0, 0);
4224 fprintf (dump_file, " with ");
4225 print_generic_expr (dump_file, sprime, 0);
4226 fprintf (dump_file, " in ");
4227 print_gimple_stmt (dump_file, stmt, 0, 0);
4230 if (TREE_CODE (sprime) == SSA_NAME)
4231 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4232 NECESSARY, true);
4233 /* We need to make sure the new and old types actually match,
4234 which may require adding a simple cast, which fold_convert
4235 will do for us. */
4236 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4237 && !useless_type_conversion_p (gimple_expr_type (stmt),
4238 TREE_TYPE (sprime)))
4239 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4241 pre_stats.eliminations++;
4242 propagate_tree_value_into_stmt (&gsi, sprime);
4243 stmt = gsi_stmt (gsi);
4244 update_stmt (stmt);
4246 /* If we removed EH side-effects from the statement, clean
4247 its EH information. */
4248 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4250 bitmap_set_bit (need_eh_cleanup,
4251 gimple_bb (stmt)->index);
4252 if (dump_file && (dump_flags & TDF_DETAILS))
4253 fprintf (dump_file, " Removed EH side-effects.\n");
4256 /* Likewise for AB side-effects. */
4257 if (can_make_abnormal_goto
4258 && !stmt_can_make_abnormal_goto (stmt))
4260 bitmap_set_bit (need_ab_cleanup,
4261 gimple_bb (stmt)->index);
4262 if (dump_file && (dump_flags & TDF_DETAILS))
4263 fprintf (dump_file, " Removed AB side-effects.\n");
4267 /* If the statement is a scalar store, see if the expression
4268 has the same value number as its rhs. If so, the store is
4269 dead. */
4270 else if (gimple_assign_single_p (stmt)
4271 && !gimple_has_volatile_ops (stmt)
4272 && !is_gimple_reg (gimple_assign_lhs (stmt))
4273 && (TREE_CODE (rhs) == SSA_NAME
4274 || is_gimple_min_invariant (rhs)))
4276 tree val;
4277 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4278 gimple_vuse (stmt), VN_WALK, NULL);
4279 if (TREE_CODE (rhs) == SSA_NAME)
4280 rhs = VN_INFO (rhs)->valnum;
4281 if (val
4282 && operand_equal_p (val, rhs, 0))
4284 if (dump_file && (dump_flags & TDF_DETAILS))
4286 fprintf (dump_file, "Deleted redundant store ");
4287 print_gimple_stmt (dump_file, stmt, 0, 0);
4290 /* Queue stmt for removal. */
4291 el_to_remove.safe_push (stmt);
4294 /* Visit COND_EXPRs and fold the comparison with the
4295 available value-numbers. */
4296 else if (gimple_code (stmt) == GIMPLE_COND)
4298 tree op0 = gimple_cond_lhs (stmt);
4299 tree op1 = gimple_cond_rhs (stmt);
4300 tree result;
4302 if (TREE_CODE (op0) == SSA_NAME)
4303 op0 = VN_INFO (op0)->valnum;
4304 if (TREE_CODE (op1) == SSA_NAME)
4305 op1 = VN_INFO (op1)->valnum;
4306 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4307 op0, op1);
4308 if (result && TREE_CODE (result) == INTEGER_CST)
4310 if (integer_zerop (result))
4311 gimple_cond_make_false (stmt);
4312 else
4313 gimple_cond_make_true (stmt);
4314 update_stmt (stmt);
4315 el_todo = TODO_cleanup_cfg;
4318 /* Visit indirect calls and turn them into direct calls if
4319 possible. */
4320 if (is_gimple_call (stmt))
4322 tree orig_fn = gimple_call_fn (stmt);
4323 tree fn;
4324 if (!orig_fn)
4325 continue;
4326 if (TREE_CODE (orig_fn) == SSA_NAME)
4327 fn = VN_INFO (orig_fn)->valnum;
4328 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4329 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4331 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4332 if (!gimple_call_addr_fndecl (fn))
4334 fn = ipa_intraprocedural_devirtualization (stmt);
4335 if (fn)
4336 fn = build_fold_addr_expr (fn);
4339 else
4340 continue;
4341 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4342 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4343 TREE_TYPE (fn)))
4345 bool can_make_abnormal_goto
4346 = stmt_can_make_abnormal_goto (stmt);
4347 bool was_noreturn = gimple_call_noreturn_p (stmt);
4349 if (dump_file && (dump_flags & TDF_DETAILS))
4351 fprintf (dump_file, "Replacing call target with ");
4352 print_generic_expr (dump_file, fn, 0);
4353 fprintf (dump_file, " in ");
4354 print_gimple_stmt (dump_file, stmt, 0, 0);
4357 gimple_call_set_fn (stmt, fn);
4358 el_to_update.safe_push (stmt);
4360 /* When changing a call into a noreturn call, cfg cleanup
4361 is needed to fix up the noreturn call. */
4362 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4363 el_todo |= TODO_cleanup_cfg;
4365 /* If we removed EH side-effects from the statement, clean
4366 its EH information. */
4367 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4369 bitmap_set_bit (need_eh_cleanup,
4370 gimple_bb (stmt)->index);
4371 if (dump_file && (dump_flags & TDF_DETAILS))
4372 fprintf (dump_file, " Removed EH side-effects.\n");
4375 /* Likewise for AB side-effects. */
4376 if (can_make_abnormal_goto
4377 && !stmt_can_make_abnormal_goto (stmt))
4379 bitmap_set_bit (need_ab_cleanup,
4380 gimple_bb (stmt)->index);
4381 if (dump_file && (dump_flags & TDF_DETAILS))
4382 fprintf (dump_file, " Removed AB side-effects.\n");
4385 /* Changing an indirect call to a direct call may
4386 have exposed different semantics. This may
4387 require an SSA update. */
4388 el_todo |= TODO_update_ssa_only_virtuals;
4394 /* Make no longer available leaders no longer available. */
4396 void
4397 eliminate_dom_walker::after_dom_children (basic_block)
4399 tree entry;
4400 while ((entry = el_avail_stack.pop ()) != NULL_TREE)
4401 el_avail[SSA_NAME_VERSION (VN_INFO (entry)->valnum)] = NULL_TREE;
4404 /* Eliminate fully redundant computations. */
4406 static unsigned int
4407 eliminate (void)
4409 gimple_stmt_iterator gsi;
4410 gimple stmt;
4411 unsigned i;
4413 need_eh_cleanup = BITMAP_ALLOC (NULL);
4414 need_ab_cleanup = BITMAP_ALLOC (NULL);
4416 el_to_remove.create (0);
4417 el_to_update.create (0);
4418 el_todo = 0;
4419 el_avail.create (0);
4420 el_avail_stack.create (0);
4422 eliminate_dom_walker (CDI_DOMINATORS).walk (cfun->cfg->x_entry_block_ptr);
4424 el_avail.release ();
4425 el_avail_stack.release ();
4427 /* We cannot remove stmts during BB walk, especially not release SSA
4428 names there as this confuses the VN machinery. The stmts ending
4429 up in el_to_remove are either stores or simple copies. */
4430 FOR_EACH_VEC_ELT (el_to_remove, i, stmt)
4432 tree lhs = gimple_assign_lhs (stmt);
4433 tree rhs = gimple_assign_rhs1 (stmt);
4434 use_operand_p use_p;
4435 gimple use_stmt;
4437 /* If there is a single use only, propagate the equivalency
4438 instead of keeping the copy. */
4439 if (TREE_CODE (lhs) == SSA_NAME
4440 && TREE_CODE (rhs) == SSA_NAME
4441 && single_imm_use (lhs, &use_p, &use_stmt)
4442 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4444 SET_USE (use_p, rhs);
4445 update_stmt (use_stmt);
4446 if (inserted_exprs
4447 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4448 && TREE_CODE (rhs) == SSA_NAME)
4449 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4452 /* If this is a store or a now unused copy, remove it. */
4453 if (TREE_CODE (lhs) != SSA_NAME
4454 || has_zero_uses (lhs))
4456 basic_block bb = gimple_bb (stmt);
4457 gsi = gsi_for_stmt (stmt);
4458 unlink_stmt_vdef (stmt);
4459 if (gsi_remove (&gsi, true))
4460 bitmap_set_bit (need_eh_cleanup, bb->index);
4461 if (inserted_exprs
4462 && TREE_CODE (lhs) == SSA_NAME)
4463 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4464 release_defs (stmt);
4467 el_to_remove.release ();
4469 /* We cannot update call statements with virtual operands during
4470 SSA walk. This might remove them which in turn makes our
4471 VN lattice invalid. */
4472 FOR_EACH_VEC_ELT (el_to_update, i, stmt)
4473 update_stmt (stmt);
4474 el_to_update.release ();
4476 return el_todo;
4479 /* Perform CFG cleanups made necessary by elimination. */
4481 static unsigned
4482 fini_eliminate (void)
4484 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4485 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4487 if (do_eh_cleanup)
4488 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4490 if (do_ab_cleanup)
4491 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4493 BITMAP_FREE (need_eh_cleanup);
4494 BITMAP_FREE (need_ab_cleanup);
4496 if (do_eh_cleanup || do_ab_cleanup)
4497 return TODO_cleanup_cfg;
4498 return 0;
4501 /* Borrow a bit of tree-ssa-dce.c for the moment.
4502 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4503 this may be a bit faster, and we may want critical edges kept split. */
4505 /* If OP's defining statement has not already been determined to be necessary,
4506 mark that statement necessary. Return the stmt, if it is newly
4507 necessary. */
4509 static inline gimple
4510 mark_operand_necessary (tree op)
4512 gimple stmt;
4514 gcc_assert (op);
4516 if (TREE_CODE (op) != SSA_NAME)
4517 return NULL;
4519 stmt = SSA_NAME_DEF_STMT (op);
4520 gcc_assert (stmt);
4522 if (gimple_plf (stmt, NECESSARY)
4523 || gimple_nop_p (stmt))
4524 return NULL;
4526 gimple_set_plf (stmt, NECESSARY, true);
4527 return stmt;
4530 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4531 to insert PHI nodes sometimes, and because value numbering of casts isn't
4532 perfect, we sometimes end up inserting dead code. This simple DCE-like
4533 pass removes any insertions we made that weren't actually used. */
4535 static void
4536 remove_dead_inserted_code (void)
4538 bitmap worklist;
4539 unsigned i;
4540 bitmap_iterator bi;
4541 gimple t;
4543 worklist = BITMAP_ALLOC (NULL);
4544 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4546 t = SSA_NAME_DEF_STMT (ssa_name (i));
4547 if (gimple_plf (t, NECESSARY))
4548 bitmap_set_bit (worklist, i);
4550 while (!bitmap_empty_p (worklist))
4552 i = bitmap_first_set_bit (worklist);
4553 bitmap_clear_bit (worklist, i);
4554 t = SSA_NAME_DEF_STMT (ssa_name (i));
4556 /* PHI nodes are somewhat special in that each PHI alternative has
4557 data and control dependencies. All the statements feeding the
4558 PHI node's arguments are always necessary. */
4559 if (gimple_code (t) == GIMPLE_PHI)
4561 unsigned k;
4563 for (k = 0; k < gimple_phi_num_args (t); k++)
4565 tree arg = PHI_ARG_DEF (t, k);
4566 if (TREE_CODE (arg) == SSA_NAME)
4568 gimple n = mark_operand_necessary (arg);
4569 if (n)
4570 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4574 else
4576 /* Propagate through the operands. Examine all the USE, VUSE and
4577 VDEF operands in this statement. Mark all the statements
4578 which feed this statement's uses as necessary. */
4579 ssa_op_iter iter;
4580 tree use;
4582 /* The operands of VDEF expressions are also needed as they
4583 represent potential definitions that may reach this
4584 statement (VDEF operands allow us to follow def-def
4585 links). */
4587 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4589 gimple n = mark_operand_necessary (use);
4590 if (n)
4591 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4596 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4598 t = SSA_NAME_DEF_STMT (ssa_name (i));
4599 if (!gimple_plf (t, NECESSARY))
4601 gimple_stmt_iterator gsi;
4603 if (dump_file && (dump_flags & TDF_DETAILS))
4605 fprintf (dump_file, "Removing unnecessary insertion:");
4606 print_gimple_stmt (dump_file, t, 0, 0);
4609 gsi = gsi_for_stmt (t);
4610 if (gimple_code (t) == GIMPLE_PHI)
4611 remove_phi_node (&gsi, true);
4612 else
4614 gsi_remove (&gsi, true);
4615 release_defs (t);
4619 BITMAP_FREE (worklist);
4623 /* Initialize data structures used by PRE. */
4625 static void
4626 init_pre (void)
4628 basic_block bb;
4630 next_expression_id = 1;
4631 expressions.create (0);
4632 expressions.safe_push (NULL);
4633 value_expressions.create (get_max_value_id () + 1);
4634 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
4635 name_to_id.create (0);
4637 inserted_exprs = BITMAP_ALLOC (NULL);
4639 connect_infinite_loops_to_exit ();
4640 memset (&pre_stats, 0, sizeof (pre_stats));
4642 postorder = XNEWVEC (int, n_basic_blocks);
4643 postorder_num = inverted_post_order_compute (postorder);
4645 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4647 calculate_dominance_info (CDI_POST_DOMINATORS);
4648 calculate_dominance_info (CDI_DOMINATORS);
4650 bitmap_obstack_initialize (&grand_bitmap_obstack);
4651 phi_translate_table.create (5110);
4652 expression_to_id.create (num_ssa_names * 3);
4653 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4654 sizeof (struct bitmap_set), 30);
4655 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4656 sizeof (struct pre_expr_d), 30);
4657 FOR_ALL_BB (bb)
4659 EXP_GEN (bb) = bitmap_set_new ();
4660 PHI_GEN (bb) = bitmap_set_new ();
4661 TMP_GEN (bb) = bitmap_set_new ();
4662 AVAIL_OUT (bb) = bitmap_set_new ();
4667 /* Deallocate data structures used by PRE. */
4669 static void
4670 fini_pre ()
4672 free (postorder);
4673 value_expressions.release ();
4674 BITMAP_FREE (inserted_exprs);
4675 bitmap_obstack_release (&grand_bitmap_obstack);
4676 free_alloc_pool (bitmap_set_pool);
4677 free_alloc_pool (pre_expr_pool);
4678 phi_translate_table.dispose ();
4679 expression_to_id.dispose ();
4680 name_to_id.release ();
4682 free_aux_for_blocks ();
4684 free_dominance_info (CDI_POST_DOMINATORS);
4687 /* Gate and execute functions for PRE. */
4689 static unsigned int
4690 do_pre (void)
4692 unsigned int todo = 0;
4694 do_partial_partial =
4695 flag_tree_partial_pre && optimize_function_for_speed_p (cfun);
4697 /* This has to happen before SCCVN runs because
4698 loop_optimizer_init may create new phis, etc. */
4699 loop_optimizer_init (LOOPS_NORMAL);
4701 if (!run_scc_vn (VN_WALK))
4703 loop_optimizer_finalize ();
4704 return 0;
4707 init_pre ();
4708 scev_initialize ();
4710 /* Collect and value number expressions computed in each basic block. */
4711 compute_avail ();
4713 /* Insert can get quite slow on an incredibly large number of basic
4714 blocks due to some quadratic behavior. Until this behavior is
4715 fixed, don't run it when he have an incredibly large number of
4716 bb's. If we aren't going to run insert, there is no point in
4717 computing ANTIC, either, even though it's plenty fast. */
4718 if (n_basic_blocks < 4000)
4720 compute_antic ();
4721 insert ();
4724 /* Make sure to remove fake edges before committing our inserts.
4725 This makes sure we don't end up with extra critical edges that
4726 we would need to split. */
4727 remove_fake_exit_edges ();
4728 gsi_commit_edge_inserts ();
4730 /* Remove all the redundant expressions. */
4731 todo |= eliminate ();
4733 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4734 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4735 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4736 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4738 clear_expression_ids ();
4739 remove_dead_inserted_code ();
4740 todo |= TODO_verify_flow;
4742 scev_finalize ();
4743 fini_pre ();
4744 todo |= fini_eliminate ();
4745 loop_optimizer_finalize ();
4747 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4748 case we can merge the block with the remaining predecessor of the block.
4749 It should either:
4750 - call merge_blocks after each tail merge iteration
4751 - call merge_blocks after all tail merge iterations
4752 - mark TODO_cleanup_cfg when necessary
4753 - share the cfg cleanup with fini_pre. */
4754 todo |= tail_merge_optimize (todo);
4756 free_scc_vn ();
4758 /* Tail merging invalidates the virtual SSA web, together with
4759 cfg-cleanup opportunities exposed by PRE this will wreck the
4760 SSA updating machinery. So make sure to run update-ssa
4761 manually, before eventually scheduling cfg-cleanup as part of
4762 the todo. */
4763 update_ssa (TODO_update_ssa_only_virtuals);
4765 return todo;
4768 static bool
4769 gate_pre (void)
4771 return flag_tree_pre != 0;
4774 namespace {
4776 const pass_data pass_data_pre =
4778 GIMPLE_PASS, /* type */
4779 "pre", /* name */
4780 OPTGROUP_NONE, /* optinfo_flags */
4781 true, /* has_gate */
4782 true, /* has_execute */
4783 TV_TREE_PRE, /* tv_id */
4784 ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */
4785 0, /* properties_provided */
4786 0, /* properties_destroyed */
4787 TODO_rebuild_alias, /* todo_flags_start */
4788 TODO_verify_ssa, /* todo_flags_finish */
4791 class pass_pre : public gimple_opt_pass
4793 public:
4794 pass_pre (gcc::context *ctxt)
4795 : gimple_opt_pass (pass_data_pre, ctxt)
4798 /* opt_pass methods: */
4799 bool gate () { return gate_pre (); }
4800 unsigned int execute () { return do_pre (); }
4802 }; // class pass_pre
4804 } // anon namespace
4806 gimple_opt_pass *
4807 make_pass_pre (gcc::context *ctxt)
4809 return new pass_pre (ctxt);
4813 /* Gate and execute functions for FRE. */
4815 static unsigned int
4816 execute_fre (void)
4818 unsigned int todo = 0;
4820 if (!run_scc_vn (VN_WALKREWRITE))
4821 return 0;
4823 memset (&pre_stats, 0, sizeof (pre_stats));
4825 /* Remove all the redundant expressions. */
4826 todo |= eliminate ();
4828 todo |= fini_eliminate ();
4830 free_scc_vn ();
4832 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4833 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4835 return todo;
4838 static bool
4839 gate_fre (void)
4841 return flag_tree_fre != 0;
4844 namespace {
4846 const pass_data pass_data_fre =
4848 GIMPLE_PASS, /* type */
4849 "fre", /* name */
4850 OPTGROUP_NONE, /* optinfo_flags */
4851 true, /* has_gate */
4852 true, /* has_execute */
4853 TV_TREE_FRE, /* tv_id */
4854 ( PROP_cfg | PROP_ssa ), /* properties_required */
4855 0, /* properties_provided */
4856 0, /* properties_destroyed */
4857 0, /* todo_flags_start */
4858 TODO_verify_ssa, /* todo_flags_finish */
4861 class pass_fre : public gimple_opt_pass
4863 public:
4864 pass_fre (gcc::context *ctxt)
4865 : gimple_opt_pass (pass_data_fre, ctxt)
4868 /* opt_pass methods: */
4869 opt_pass * clone () { return new pass_fre (m_ctxt); }
4870 bool gate () { return gate_fre (); }
4871 unsigned int execute () { return execute_fre (); }
4873 }; // class pass_fre
4875 } // anon namespace
4877 gimple_opt_pass *
4878 make_pass_fre (gcc::context *ctxt)
4880 return new pass_fre (ctxt);