1 /* Full and partial redundancy elimination and code hoisting on SSA GIMPLE.
2 Copyright (C) 2001-2018 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
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)
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/>. */
24 #include "coretypes.h"
30 #include "alloc-pool.h"
31 #include "tree-pass.h"
34 #include "gimple-pretty-print.h"
35 #include "fold-const.h"
37 #include "gimple-fold.h"
40 #include "gimple-iterator.h"
42 #include "tree-into-ssa.h"
46 #include "tree-ssa-sccvn.h"
47 #include "tree-scalar-evolution.h"
51 #include "tree-ssa-propagate.h"
52 #include "tree-ssa-dce.h"
53 #include "tree-cfgcleanup.h"
56 /* Even though this file is called tree-ssa-pre.c, we actually
57 implement a bit more than just PRE here. All of them piggy-back
58 on GVN which is implemented in tree-ssa-sccvn.c.
60 1. Full Redundancy Elimination (FRE)
61 This is the elimination phase of GVN.
63 2. Partial Redundancy Elimination (PRE)
64 This is adds computation of AVAIL_OUT and ANTIC_IN and
65 doing expression insertion to form GVN-PRE.
68 This optimization uses the ANTIC_IN sets computed for PRE
69 to move expressions further up than PRE would do, to make
70 multiple computations of the same value fully redundant.
71 This pass is explained below (after the explanation of the
72 basic algorithm for PRE).
77 1. Avail sets can be shared by making an avail_find_leader that
78 walks up the dominator tree and looks in those avail sets.
79 This might affect code optimality, it's unclear right now.
80 Currently the AVAIL_OUT sets are the remaining quadraticness in
82 2. Strength reduction can be performed by anticipating expressions
83 we can repair later on.
84 3. We can do back-substitution or smarter value numbering to catch
85 commutative expressions split up over multiple statements.
88 /* For ease of terminology, "expression node" in the below refers to
89 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
90 represent the actual statement containing the expressions we care about,
91 and we cache the value number by putting it in the expression. */
93 /* Basic algorithm for Partial Redundancy Elimination:
95 First we walk the statements to generate the AVAIL sets, the
96 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
97 generation of values/expressions by a given block. We use them
98 when computing the ANTIC sets. The AVAIL sets consist of
99 SSA_NAME's that represent values, so we know what values are
100 available in what blocks. AVAIL is a forward dataflow problem. In
101 SSA, values are never killed, so we don't need a kill set, or a
102 fixpoint iteration, in order to calculate the AVAIL sets. In
103 traditional parlance, AVAIL sets tell us the downsafety of the
106 Next, we generate the ANTIC sets. These sets represent the
107 anticipatable expressions. ANTIC is a backwards dataflow
108 problem. An expression is anticipatable in a given block if it could
109 be generated in that block. This means that if we had to perform
110 an insertion in that block, of the value of that expression, we
111 could. Calculating the ANTIC sets requires phi translation of
112 expressions, because the flow goes backwards through phis. We must
113 iterate to a fixpoint of the ANTIC sets, because we have a kill
114 set. Even in SSA form, values are not live over the entire
115 function, only from their definition point onwards. So we have to
116 remove values from the ANTIC set once we go past the definition
117 point of the leaders that make them up.
118 compute_antic/compute_antic_aux performs this computation.
120 Third, we perform insertions to make partially redundant
121 expressions fully redundant.
123 An expression is partially redundant (excluding partial
126 1. It is AVAIL in some, but not all, of the predecessors of a
128 2. It is ANTIC in all the predecessors.
130 In order to make it fully redundant, we insert the expression into
131 the predecessors where it is not available, but is ANTIC.
133 When optimizing for size, we only eliminate the partial redundancy
134 if we need to insert in only one predecessor. This avoids almost
135 completely the code size increase that PRE usually causes.
137 For the partial anticipation case, we only perform insertion if it
138 is partially anticipated in some block, and fully available in all
141 do_pre_regular_insertion/do_pre_partial_partial_insertion
142 performs these steps, driven by insert/insert_aux.
144 Fourth, we eliminate fully redundant expressions.
145 This is a simple statement walk that replaces redundant
146 calculations with the now available values. */
148 /* Basic algorithm for Code Hoisting:
150 Code hoisting is: Moving value computations up in the control flow
151 graph to make multiple copies redundant. Typically this is a size
152 optimization, but there are cases where it also is helpful for speed.
154 A simple code hoisting algorithm is implemented that piggy-backs on
155 the PRE infrastructure. For code hoisting, we have to know ANTIC_OUT
156 which is effectively ANTIC_IN - AVAIL_OUT. The latter two have to be
157 computed for PRE, and we can use them to perform a limited version of
160 For the purpose of this implementation, a value is hoistable to a basic
161 block B if the following properties are met:
163 1. The value is in ANTIC_IN(B) -- the value will be computed on all
164 paths from B to function exit and it can be computed in B);
166 2. The value is not in AVAIL_OUT(B) -- there would be no need to
167 compute the value again and make it available twice;
169 3. All successors of B are dominated by B -- makes sure that inserting
170 a computation of the value in B will make the remaining
171 computations fully redundant;
173 4. At least one successor has the value in AVAIL_OUT -- to avoid
174 hoisting values up too far;
176 5. There are at least two successors of B -- hoisting in straight
177 line code is pointless.
179 The third condition is not strictly necessary, but it would complicate
180 the hoisting pass a lot. In fact, I don't know of any code hoisting
181 algorithm that does not have this requirement. Fortunately, experiments
182 have show that most candidate hoistable values are in regions that meet
183 this condition (e.g. diamond-shape regions).
185 The forth condition is necessary to avoid hoisting things up too far
186 away from the uses of the value. Nothing else limits the algorithm
187 from hoisting everything up as far as ANTIC_IN allows. Experiments
188 with SPEC and CSiBE have shown that hoisting up too far results in more
189 spilling, less benefits for code size, and worse benchmark scores.
190 Fortunately, in practice most of the interesting hoisting opportunities
191 are caught despite this limitation.
193 For hoistable values that meet all conditions, expressions are inserted
194 to make the calculation of the hoistable value fully redundant. We
195 perform code hoisting insertions after each round of PRE insertions,
196 because code hoisting never exposes new PRE opportunities, but PRE can
197 create new code hoisting opportunities.
199 The code hoisting algorithm is implemented in do_hoist_insert, driven
200 by insert/insert_aux. */
202 /* Representations of value numbers:
204 Value numbers are represented by a representative SSA_NAME. We
205 will create fake SSA_NAME's in situations where we need a
206 representative but do not have one (because it is a complex
207 expression). In order to facilitate storing the value numbers in
208 bitmaps, and keep the number of wasted SSA_NAME's down, we also
209 associate a value_id with each value number, and create full blown
210 ssa_name's only where we actually need them (IE in operands of
211 existing expressions).
213 Theoretically you could replace all the value_id's with
214 SSA_NAME_VERSION, but this would allocate a large number of
215 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
216 It would also require an additional indirection at each point we
219 /* Representation of expressions on value numbers:
221 Expressions consisting of value numbers are represented the same
222 way as our VN internally represents them, with an additional
223 "pre_expr" wrapping around them in order to facilitate storing all
224 of the expressions in the same sets. */
226 /* Representation of sets:
228 The dataflow sets do not need to be sorted in any particular order
229 for the majority of their lifetime, are simply represented as two
230 bitmaps, one that keeps track of values present in the set, and one
231 that keeps track of expressions present in the set.
233 When we need them in topological order, we produce it on demand by
234 transforming the bitmap into an array and sorting it into topo
237 /* Type of expression, used to know which member of the PRE_EXPR union
253 vn_reference_t reference
;
256 typedef struct pre_expr_d
: nofree_ptr_hash
<pre_expr_d
>
258 enum pre_expr_kind kind
;
262 /* hash_table support. */
263 static inline hashval_t
hash (const pre_expr_d
*);
264 static inline int equal (const pre_expr_d
*, const pre_expr_d
*);
267 #define PRE_EXPR_NAME(e) (e)->u.name
268 #define PRE_EXPR_NARY(e) (e)->u.nary
269 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
270 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
272 /* Compare E1 and E1 for equality. */
275 pre_expr_d::equal (const pre_expr_d
*e1
, const pre_expr_d
*e2
)
277 if (e1
->kind
!= e2
->kind
)
283 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1
),
284 PRE_EXPR_CONSTANT (e2
));
286 return PRE_EXPR_NAME (e1
) == PRE_EXPR_NAME (e2
);
288 return vn_nary_op_eq (PRE_EXPR_NARY (e1
), PRE_EXPR_NARY (e2
));
290 return vn_reference_eq (PRE_EXPR_REFERENCE (e1
),
291 PRE_EXPR_REFERENCE (e2
));
300 pre_expr_d::hash (const pre_expr_d
*e
)
305 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e
));
307 return SSA_NAME_VERSION (PRE_EXPR_NAME (e
));
309 return PRE_EXPR_NARY (e
)->hashcode
;
311 return PRE_EXPR_REFERENCE (e
)->hashcode
;
317 /* Next global expression id number. */
318 static unsigned int next_expression_id
;
320 /* Mapping from expression to id number we can use in bitmap sets. */
321 static vec
<pre_expr
> expressions
;
322 static hash_table
<pre_expr_d
> *expression_to_id
;
323 static vec
<unsigned> name_to_id
;
325 /* Allocate an expression id for EXPR. */
327 static inline unsigned int
328 alloc_expression_id (pre_expr expr
)
330 struct pre_expr_d
**slot
;
331 /* Make sure we won't overflow. */
332 gcc_assert (next_expression_id
+ 1 > next_expression_id
);
333 expr
->id
= next_expression_id
++;
334 expressions
.safe_push (expr
);
335 if (expr
->kind
== NAME
)
337 unsigned version
= SSA_NAME_VERSION (PRE_EXPR_NAME (expr
));
338 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
339 re-allocations by using vec::reserve upfront. */
340 unsigned old_len
= name_to_id
.length ();
341 name_to_id
.reserve (num_ssa_names
- old_len
);
342 name_to_id
.quick_grow_cleared (num_ssa_names
);
343 gcc_assert (name_to_id
[version
] == 0);
344 name_to_id
[version
] = expr
->id
;
348 slot
= expression_to_id
->find_slot (expr
, INSERT
);
352 return next_expression_id
- 1;
355 /* Return the expression id for tree EXPR. */
357 static inline unsigned int
358 get_expression_id (const pre_expr expr
)
363 static inline unsigned int
364 lookup_expression_id (const pre_expr expr
)
366 struct pre_expr_d
**slot
;
368 if (expr
->kind
== NAME
)
370 unsigned version
= SSA_NAME_VERSION (PRE_EXPR_NAME (expr
));
371 if (name_to_id
.length () <= version
)
373 return name_to_id
[version
];
377 slot
= expression_to_id
->find_slot (expr
, NO_INSERT
);
380 return ((pre_expr
)*slot
)->id
;
384 /* Return the existing expression id for EXPR, or create one if one
385 does not exist yet. */
387 static inline unsigned int
388 get_or_alloc_expression_id (pre_expr expr
)
390 unsigned int id
= lookup_expression_id (expr
);
392 return alloc_expression_id (expr
);
393 return expr
->id
= id
;
396 /* Return the expression that has expression id ID */
398 static inline pre_expr
399 expression_for_id (unsigned int id
)
401 return expressions
[id
];
404 static object_allocator
<pre_expr_d
> pre_expr_pool ("pre_expr nodes");
406 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
409 get_or_alloc_expr_for_name (tree name
)
411 struct pre_expr_d expr
;
413 unsigned int result_id
;
417 PRE_EXPR_NAME (&expr
) = name
;
418 result_id
= lookup_expression_id (&expr
);
420 return expression_for_id (result_id
);
422 result
= pre_expr_pool
.allocate ();
424 PRE_EXPR_NAME (result
) = name
;
425 alloc_expression_id (result
);
429 /* An unordered bitmap set. One bitmap tracks values, the other,
431 typedef struct bitmap_set
433 bitmap_head expressions
;
437 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
438 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
440 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
441 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
443 /* Mapping from value id to expressions with that value_id. */
444 static vec
<bitmap
> value_expressions
;
446 /* Sets that we need to keep track of. */
447 typedef struct bb_bitmap_sets
449 /* The EXP_GEN set, which represents expressions/values generated in
451 bitmap_set_t exp_gen
;
453 /* The PHI_GEN set, which represents PHI results generated in a
455 bitmap_set_t phi_gen
;
457 /* The TMP_GEN set, which represents results/temporaries generated
458 in a basic block. IE the LHS of an expression. */
459 bitmap_set_t tmp_gen
;
461 /* The AVAIL_OUT set, which represents which values are available in
462 a given basic block. */
463 bitmap_set_t avail_out
;
465 /* The ANTIC_IN set, which represents which values are anticipatable
466 in a given basic block. */
467 bitmap_set_t antic_in
;
469 /* The PA_IN set, which represents which values are
470 partially anticipatable in a given basic block. */
473 /* The NEW_SETS set, which is used during insertion to augment the
474 AVAIL_OUT set of blocks with the new insertions performed during
475 the current iteration. */
476 bitmap_set_t new_sets
;
478 /* A cache for value_dies_in_block_x. */
481 /* The live virtual operand on successor edges. */
484 /* True if we have visited this block during ANTIC calculation. */
485 unsigned int visited
: 1;
487 /* True if we have visited this block after all successors have been
489 unsigned int visited_with_visited_succs
: 1;
491 /* True when the block contains a call that might not return. */
492 unsigned int contains_may_not_return_call
: 1;
495 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
496 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
497 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
498 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
499 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
500 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
501 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
502 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
503 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
504 #define BB_VISITED_WITH_VISITED_SUCCS(BB) \
505 ((bb_value_sets_t) ((BB)->aux))->visited_with_visited_succs
506 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
507 #define BB_LIVE_VOP_ON_EXIT(BB) ((bb_value_sets_t) ((BB)->aux))->vop_on_exit
510 /* This structure is used to keep track of statistics on what
511 optimization PRE was able to perform. */
514 /* The number of new expressions/temporaries generated by PRE. */
517 /* The number of inserts found due to partial anticipation */
520 /* The number of inserts made for code hoisting. */
523 /* The number of new PHI nodes added by PRE. */
527 static bool do_partial_partial
;
528 static pre_expr
bitmap_find_leader (bitmap_set_t
, unsigned int);
529 static void bitmap_value_insert_into_set (bitmap_set_t
, pre_expr
);
530 static void bitmap_value_replace_in_set (bitmap_set_t
, pre_expr
);
531 static void bitmap_set_copy (bitmap_set_t
, bitmap_set_t
);
532 static bool bitmap_set_contains_value (bitmap_set_t
, unsigned int);
533 static void bitmap_insert_into_set (bitmap_set_t
, pre_expr
);
534 static bitmap_set_t
bitmap_set_new (void);
535 static tree
create_expression_by_pieces (basic_block
, pre_expr
, gimple_seq
*,
537 static tree
find_or_generate_expression (basic_block
, tree
, gimple_seq
*);
538 static unsigned int get_expr_value_id (pre_expr
);
540 /* We can add and remove elements and entries to and from sets
541 and hash tables, so we use alloc pools for them. */
543 static object_allocator
<bitmap_set
> bitmap_set_pool ("Bitmap sets");
544 static bitmap_obstack grand_bitmap_obstack
;
546 /* A three tuple {e, pred, v} used to cache phi translations in the
547 phi_translate_table. */
549 typedef struct expr_pred_trans_d
: free_ptr_hash
<expr_pred_trans_d
>
551 /* The expression. */
554 /* The predecessor block along which we translated the expression. */
557 /* The value that resulted from the translation. */
560 /* The hashcode for the expression, pred pair. This is cached for
564 /* hash_table support. */
565 static inline hashval_t
hash (const expr_pred_trans_d
*);
566 static inline int equal (const expr_pred_trans_d
*, const expr_pred_trans_d
*);
567 } *expr_pred_trans_t
;
568 typedef const struct expr_pred_trans_d
*const_expr_pred_trans_t
;
571 expr_pred_trans_d::hash (const expr_pred_trans_d
*e
)
577 expr_pred_trans_d::equal (const expr_pred_trans_d
*ve1
,
578 const expr_pred_trans_d
*ve2
)
580 basic_block b1
= ve1
->pred
;
581 basic_block b2
= ve2
->pred
;
583 /* If they are not translations for the same basic block, they can't
587 return pre_expr_d::equal (ve1
->e
, ve2
->e
);
590 /* The phi_translate_table caches phi translations for a given
591 expression and predecessor. */
592 static hash_table
<expr_pred_trans_d
> *phi_translate_table
;
594 /* Add the tuple mapping from {expression E, basic block PRED} to
595 the phi translation table and return whether it pre-existed. */
598 phi_trans_add (expr_pred_trans_t
*entry
, pre_expr e
, basic_block pred
)
600 expr_pred_trans_t
*slot
;
601 expr_pred_trans_d tem
;
602 hashval_t hash
= iterative_hash_hashval_t (pre_expr_d::hash (e
),
607 slot
= phi_translate_table
->find_slot_with_hash (&tem
, hash
, INSERT
);
614 *entry
= *slot
= XNEW (struct expr_pred_trans_d
);
616 (*entry
)->pred
= pred
;
617 (*entry
)->hashcode
= hash
;
622 /* Add expression E to the expression set of value id V. */
625 add_to_value (unsigned int v
, pre_expr e
)
629 gcc_checking_assert (get_expr_value_id (e
) == v
);
631 if (v
>= value_expressions
.length ())
633 value_expressions
.safe_grow_cleared (v
+ 1);
636 set
= value_expressions
[v
];
639 set
= BITMAP_ALLOC (&grand_bitmap_obstack
);
640 value_expressions
[v
] = set
;
643 bitmap_set_bit (set
, get_or_alloc_expression_id (e
));
646 /* Create a new bitmap set and return it. */
649 bitmap_set_new (void)
651 bitmap_set_t ret
= bitmap_set_pool
.allocate ();
652 bitmap_initialize (&ret
->expressions
, &grand_bitmap_obstack
);
653 bitmap_initialize (&ret
->values
, &grand_bitmap_obstack
);
657 /* Return the value id for a PRE expression EXPR. */
660 get_expr_value_id (pre_expr expr
)
666 id
= get_constant_value_id (PRE_EXPR_CONSTANT (expr
));
669 id
= VN_INFO (PRE_EXPR_NAME (expr
))->value_id
;
672 id
= PRE_EXPR_NARY (expr
)->value_id
;
675 id
= PRE_EXPR_REFERENCE (expr
)->value_id
;
680 /* ??? We cannot assert that expr has a value-id (it can be 0), because
681 we assign value-ids only to expressions that have a result
682 in set_hashtable_value_ids. */
686 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
689 sccvn_valnum_from_value_id (unsigned int val
)
693 bitmap exprset
= value_expressions
[val
];
694 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
696 pre_expr vexpr
= expression_for_id (i
);
697 if (vexpr
->kind
== NAME
)
698 return VN_INFO (PRE_EXPR_NAME (vexpr
))->valnum
;
699 else if (vexpr
->kind
== CONSTANT
)
700 return PRE_EXPR_CONSTANT (vexpr
);
705 /* Insert an expression EXPR into a bitmapped set. */
708 bitmap_insert_into_set (bitmap_set_t set
, pre_expr expr
)
710 unsigned int val
= get_expr_value_id (expr
);
711 if (! value_id_constant_p (val
))
713 /* Note this is the only function causing multiple expressions
714 for the same value to appear in a set. This is needed for
715 TMP_GEN, PHI_GEN and NEW_SETs. */
716 bitmap_set_bit (&set
->values
, val
);
717 bitmap_set_bit (&set
->expressions
, get_or_alloc_expression_id (expr
));
721 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
724 bitmap_set_copy (bitmap_set_t dest
, bitmap_set_t orig
)
726 bitmap_copy (&dest
->expressions
, &orig
->expressions
);
727 bitmap_copy (&dest
->values
, &orig
->values
);
731 /* Free memory used up by SET. */
733 bitmap_set_free (bitmap_set_t set
)
735 bitmap_clear (&set
->expressions
);
736 bitmap_clear (&set
->values
);
740 /* Generate an topological-ordered array of bitmap set SET. */
743 sorted_array_from_bitmap_set (bitmap_set_t set
)
746 bitmap_iterator bi
, bj
;
747 vec
<pre_expr
> result
;
749 /* Pre-allocate enough space for the array. */
750 result
.create (bitmap_count_bits (&set
->expressions
));
752 FOR_EACH_VALUE_ID_IN_SET (set
, i
, bi
)
754 /* The number of expressions having a given value is usually
755 relatively small. Thus, rather than making a vector of all
756 the expressions and sorting it by value-id, we walk the values
757 and check in the reverse mapping that tells us what expressions
758 have a given value, to filter those in our set. As a result,
759 the expressions are inserted in value-id order, which means
762 If this is somehow a significant lose for some cases, we can
763 choose which set to walk based on the set size. */
764 bitmap exprset
= value_expressions
[i
];
765 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, j
, bj
)
767 if (bitmap_bit_p (&set
->expressions
, j
))
768 result
.quick_push (expression_for_id (j
));
775 /* Subtract all expressions contained in ORIG from DEST. */
778 bitmap_set_subtract_expressions (bitmap_set_t dest
, bitmap_set_t orig
)
780 bitmap_set_t result
= bitmap_set_new ();
784 bitmap_and_compl (&result
->expressions
, &dest
->expressions
,
787 FOR_EACH_EXPR_ID_IN_SET (result
, i
, bi
)
789 pre_expr expr
= expression_for_id (i
);
790 unsigned int value_id
= get_expr_value_id (expr
);
791 bitmap_set_bit (&result
->values
, value_id
);
797 /* Subtract all values in bitmap set B from bitmap set A. */
800 bitmap_set_subtract_values (bitmap_set_t a
, bitmap_set_t b
)
804 unsigned to_remove
= -1U;
805 bitmap_and_compl_into (&a
->values
, &b
->values
);
806 FOR_EACH_EXPR_ID_IN_SET (a
, i
, bi
)
808 if (to_remove
!= -1U)
810 bitmap_clear_bit (&a
->expressions
, to_remove
);
813 pre_expr expr
= expression_for_id (i
);
814 if (! bitmap_bit_p (&a
->values
, get_expr_value_id (expr
)))
817 if (to_remove
!= -1U)
818 bitmap_clear_bit (&a
->expressions
, to_remove
);
822 /* Return true if bitmapped set SET contains the value VALUE_ID. */
825 bitmap_set_contains_value (bitmap_set_t set
, unsigned int value_id
)
827 if (value_id_constant_p (value_id
))
830 return bitmap_bit_p (&set
->values
, value_id
);
834 bitmap_set_contains_expr (bitmap_set_t set
, const pre_expr expr
)
836 return bitmap_bit_p (&set
->expressions
, get_expression_id (expr
));
839 /* Return true if two bitmap sets are equal. */
842 bitmap_set_equal (bitmap_set_t a
, bitmap_set_t b
)
844 return bitmap_equal_p (&a
->values
, &b
->values
);
847 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
848 and add it otherwise. */
851 bitmap_value_replace_in_set (bitmap_set_t set
, pre_expr expr
)
853 unsigned int val
= get_expr_value_id (expr
);
854 if (value_id_constant_p (val
))
857 if (bitmap_set_contains_value (set
, val
))
859 /* The number of expressions having a given value is usually
860 significantly less than the total number of expressions in SET.
861 Thus, rather than check, for each expression in SET, whether it
862 has the value LOOKFOR, we walk the reverse mapping that tells us
863 what expressions have a given value, and see if any of those
864 expressions are in our set. For large testcases, this is about
865 5-10x faster than walking the bitmap. If this is somehow a
866 significant lose for some cases, we can choose which set to walk
867 based on the set size. */
870 bitmap exprset
= value_expressions
[val
];
871 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
873 if (bitmap_clear_bit (&set
->expressions
, i
))
875 bitmap_set_bit (&set
->expressions
, get_expression_id (expr
));
882 bitmap_insert_into_set (set
, expr
);
885 /* Insert EXPR into SET if EXPR's value is not already present in
889 bitmap_value_insert_into_set (bitmap_set_t set
, pre_expr expr
)
891 unsigned int val
= get_expr_value_id (expr
);
893 gcc_checking_assert (expr
->id
== get_or_alloc_expression_id (expr
));
895 /* Constant values are always considered to be part of the set. */
896 if (value_id_constant_p (val
))
899 /* If the value membership changed, add the expression. */
900 if (bitmap_set_bit (&set
->values
, val
))
901 bitmap_set_bit (&set
->expressions
, expr
->id
);
904 /* Print out EXPR to outfile. */
907 print_pre_expr (FILE *outfile
, const pre_expr expr
)
911 fprintf (outfile
, "NULL");
917 print_generic_expr (outfile
, PRE_EXPR_CONSTANT (expr
));
920 print_generic_expr (outfile
, PRE_EXPR_NAME (expr
));
925 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
926 fprintf (outfile
, "{%s,", get_tree_code_name (nary
->opcode
));
927 for (i
= 0; i
< nary
->length
; i
++)
929 print_generic_expr (outfile
, nary
->op
[i
]);
930 if (i
!= (unsigned) nary
->length
- 1)
931 fprintf (outfile
, ",");
933 fprintf (outfile
, "}");
939 vn_reference_op_t vro
;
941 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
942 fprintf (outfile
, "{");
944 ref
->operands
.iterate (i
, &vro
);
947 bool closebrace
= false;
948 if (vro
->opcode
!= SSA_NAME
949 && TREE_CODE_CLASS (vro
->opcode
) != tcc_declaration
)
951 fprintf (outfile
, "%s", get_tree_code_name (vro
->opcode
));
954 fprintf (outfile
, "<");
960 print_generic_expr (outfile
, vro
->op0
);
963 fprintf (outfile
, ",");
964 print_generic_expr (outfile
, vro
->op1
);
968 fprintf (outfile
, ",");
969 print_generic_expr (outfile
, vro
->op2
);
973 fprintf (outfile
, ">");
974 if (i
!= ref
->operands
.length () - 1)
975 fprintf (outfile
, ",");
977 fprintf (outfile
, "}");
980 fprintf (outfile
, "@");
981 print_generic_expr (outfile
, ref
->vuse
);
987 void debug_pre_expr (pre_expr
);
989 /* Like print_pre_expr but always prints to stderr. */
991 debug_pre_expr (pre_expr e
)
993 print_pre_expr (stderr
, e
);
994 fprintf (stderr
, "\n");
997 /* Print out SET to OUTFILE. */
1000 print_bitmap_set (FILE *outfile
, bitmap_set_t set
,
1001 const char *setname
, int blockindex
)
1003 fprintf (outfile
, "%s[%d] := { ", setname
, blockindex
);
1010 FOR_EACH_EXPR_ID_IN_SET (set
, i
, bi
)
1012 const pre_expr expr
= expression_for_id (i
);
1015 fprintf (outfile
, ", ");
1017 print_pre_expr (outfile
, expr
);
1019 fprintf (outfile
, " (%04d)", get_expr_value_id (expr
));
1022 fprintf (outfile
, " }\n");
1025 void debug_bitmap_set (bitmap_set_t
);
1028 debug_bitmap_set (bitmap_set_t set
)
1030 print_bitmap_set (stderr
, set
, "debug", 0);
1033 void debug_bitmap_sets_for (basic_block
);
1036 debug_bitmap_sets_for (basic_block bb
)
1038 print_bitmap_set (stderr
, AVAIL_OUT (bb
), "avail_out", bb
->index
);
1039 print_bitmap_set (stderr
, EXP_GEN (bb
), "exp_gen", bb
->index
);
1040 print_bitmap_set (stderr
, PHI_GEN (bb
), "phi_gen", bb
->index
);
1041 print_bitmap_set (stderr
, TMP_GEN (bb
), "tmp_gen", bb
->index
);
1042 print_bitmap_set (stderr
, ANTIC_IN (bb
), "antic_in", bb
->index
);
1043 if (do_partial_partial
)
1044 print_bitmap_set (stderr
, PA_IN (bb
), "pa_in", bb
->index
);
1045 print_bitmap_set (stderr
, NEW_SETS (bb
), "new_sets", bb
->index
);
1048 /* Print out the expressions that have VAL to OUTFILE. */
1051 print_value_expressions (FILE *outfile
, unsigned int val
)
1053 bitmap set
= value_expressions
[val
];
1058 sprintf (s
, "%04d", val
);
1059 x
.expressions
= *set
;
1060 print_bitmap_set (outfile
, &x
, s
, 0);
1066 debug_value_expressions (unsigned int val
)
1068 print_value_expressions (stderr
, val
);
1071 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1075 get_or_alloc_expr_for_constant (tree constant
)
1077 unsigned int result_id
;
1078 unsigned int value_id
;
1079 struct pre_expr_d expr
;
1082 expr
.kind
= CONSTANT
;
1083 PRE_EXPR_CONSTANT (&expr
) = constant
;
1084 result_id
= lookup_expression_id (&expr
);
1086 return expression_for_id (result_id
);
1088 newexpr
= pre_expr_pool
.allocate ();
1089 newexpr
->kind
= CONSTANT
;
1090 PRE_EXPR_CONSTANT (newexpr
) = constant
;
1091 alloc_expression_id (newexpr
);
1092 value_id
= get_or_alloc_constant_value_id (constant
);
1093 add_to_value (value_id
, newexpr
);
1097 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1098 Currently only supports constants and SSA_NAMES. */
1100 get_or_alloc_expr_for (tree t
)
1102 if (TREE_CODE (t
) == SSA_NAME
)
1103 return get_or_alloc_expr_for_name (t
);
1104 else if (is_gimple_min_invariant (t
))
1105 return get_or_alloc_expr_for_constant (t
);
1109 /* Return the folded version of T if T, when folded, is a gimple
1110 min_invariant or an SSA name. Otherwise, return T. */
1113 fully_constant_expression (pre_expr e
)
1121 vn_nary_op_t nary
= PRE_EXPR_NARY (e
);
1122 tree res
= vn_nary_simplify (nary
);
1125 if (is_gimple_min_invariant (res
))
1126 return get_or_alloc_expr_for_constant (res
);
1127 if (TREE_CODE (res
) == SSA_NAME
)
1128 return get_or_alloc_expr_for_name (res
);
1133 vn_reference_t ref
= PRE_EXPR_REFERENCE (e
);
1135 if ((folded
= fully_constant_vn_reference_p (ref
)))
1136 return get_or_alloc_expr_for_constant (folded
);
1145 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1146 it has the value it would have in BLOCK. Set *SAME_VALID to true
1147 in case the new vuse doesn't change the value id of the OPERANDS. */
1150 translate_vuse_through_block (vec
<vn_reference_op_s
> operands
,
1151 alias_set_type set
, tree type
, tree vuse
,
1152 basic_block phiblock
,
1153 basic_block block
, bool *same_valid
)
1155 gimple
*phi
= SSA_NAME_DEF_STMT (vuse
);
1162 if (gimple_bb (phi
) != phiblock
)
1165 use_oracle
= ao_ref_init_from_vn_reference (&ref
, set
, type
, operands
);
1167 /* Use the alias-oracle to find either the PHI node in this block,
1168 the first VUSE used in this block that is equivalent to vuse or
1169 the first VUSE which definition in this block kills the value. */
1170 if (gimple_code (phi
) == GIMPLE_PHI
)
1171 e
= find_edge (block
, phiblock
);
1172 else if (use_oracle
)
1173 while (!stmt_may_clobber_ref_p_1 (phi
, &ref
))
1175 vuse
= gimple_vuse (phi
);
1176 phi
= SSA_NAME_DEF_STMT (vuse
);
1177 if (gimple_bb (phi
) != phiblock
)
1179 if (gimple_code (phi
) == GIMPLE_PHI
)
1181 e
= find_edge (block
, phiblock
);
1192 bitmap visited
= NULL
;
1194 /* Try to find a vuse that dominates this phi node by skipping
1195 non-clobbering statements. */
1196 vuse
= get_continuation_for_phi (phi
, &ref
, &cnt
, &visited
, false,
1199 BITMAP_FREE (visited
);
1205 /* If we didn't find any, the value ID can't stay the same,
1206 but return the translated vuse. */
1207 *same_valid
= false;
1208 vuse
= PHI_ARG_DEF (phi
, e
->dest_idx
);
1210 /* ??? We would like to return vuse here as this is the canonical
1211 upmost vdef that this reference is associated with. But during
1212 insertion of the references into the hash tables we only ever
1213 directly insert with their direct gimple_vuse, hence returning
1214 something else would make us not find the other expression. */
1215 return PHI_ARG_DEF (phi
, e
->dest_idx
);
1221 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1222 SET2 *or* SET3. This is used to avoid making a set consisting of the union
1223 of PA_IN and ANTIC_IN during insert and phi-translation. */
1225 static inline pre_expr
1226 find_leader_in_sets (unsigned int val
, bitmap_set_t set1
, bitmap_set_t set2
,
1227 bitmap_set_t set3
= NULL
)
1231 result
= bitmap_find_leader (set1
, val
);
1232 if (!result
&& set2
)
1233 result
= bitmap_find_leader (set2
, val
);
1234 if (!result
&& set3
)
1235 result
= bitmap_find_leader (set3
, val
);
1239 /* Get the tree type for our PRE expression e. */
1242 get_expr_type (const pre_expr e
)
1247 return TREE_TYPE (PRE_EXPR_NAME (e
));
1249 return TREE_TYPE (PRE_EXPR_CONSTANT (e
));
1251 return PRE_EXPR_REFERENCE (e
)->type
;
1253 return PRE_EXPR_NARY (e
)->type
;
1258 /* Get a representative SSA_NAME for a given expression that is available in B.
1259 Since all of our sub-expressions are treated as values, we require
1260 them to be SSA_NAME's for simplicity.
1261 Prior versions of GVNPRE used to use "value handles" here, so that
1262 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1263 either case, the operands are really values (IE we do not expect
1264 them to be usable without finding leaders). */
1267 get_representative_for (const pre_expr e
, basic_block b
= NULL
)
1269 tree name
, valnum
= NULL_TREE
;
1270 unsigned int value_id
= get_expr_value_id (e
);
1275 return VN_INFO (PRE_EXPR_NAME (e
))->valnum
;
1277 return PRE_EXPR_CONSTANT (e
);
1281 /* Go through all of the expressions representing this value
1282 and pick out an SSA_NAME. */
1285 bitmap exprs
= value_expressions
[value_id
];
1286 EXECUTE_IF_SET_IN_BITMAP (exprs
, 0, i
, bi
)
1288 pre_expr rep
= expression_for_id (i
);
1289 if (rep
->kind
== NAME
)
1291 tree name
= PRE_EXPR_NAME (rep
);
1292 valnum
= VN_INFO (name
)->valnum
;
1293 gimple
*def
= SSA_NAME_DEF_STMT (name
);
1294 /* We have to return either a new representative or one
1295 that can be used for expression simplification and thus
1296 is available in B. */
1298 || gimple_nop_p (def
)
1299 || dominated_by_p (CDI_DOMINATORS
, b
, gimple_bb (def
)))
1302 else if (rep
->kind
== CONSTANT
)
1303 return PRE_EXPR_CONSTANT (rep
);
1309 /* If we reached here we couldn't find an SSA_NAME. This can
1310 happen when we've discovered a value that has never appeared in
1311 the program as set to an SSA_NAME, as the result of phi translation.
1313 ??? We should be able to re-use this when we insert the statement
1315 name
= make_temp_ssa_name (get_expr_type (e
), gimple_build_nop (), "pretmp");
1316 VN_INFO_GET (name
)->value_id
= value_id
;
1317 VN_INFO (name
)->valnum
= valnum
? valnum
: name
;
1318 /* ??? For now mark this SSA name for release by SCCVN. */
1319 VN_INFO (name
)->needs_insertion
= true;
1320 add_to_value (value_id
, get_or_alloc_expr_for_name (name
));
1321 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1323 fprintf (dump_file
, "Created SSA_NAME representative ");
1324 print_generic_expr (dump_file
, name
);
1325 fprintf (dump_file
, " for expression:");
1326 print_pre_expr (dump_file
, e
);
1327 fprintf (dump_file
, " (%04d)\n", value_id
);
1335 phi_translate (pre_expr expr
, bitmap_set_t set1
, bitmap_set_t set2
, edge e
);
1337 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1338 the phis in PRED. Return NULL if we can't find a leader for each part
1339 of the translated expression. */
1342 phi_translate_1 (pre_expr expr
, bitmap_set_t set1
, bitmap_set_t set2
, edge e
)
1344 basic_block pred
= e
->src
;
1345 basic_block phiblock
= e
->dest
;
1351 bool changed
= false;
1352 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
1353 vn_nary_op_t newnary
= XALLOCAVAR (struct vn_nary_op_s
,
1354 sizeof_vn_nary_op (nary
->length
));
1355 memcpy (newnary
, nary
, sizeof_vn_nary_op (nary
->length
));
1357 for (i
= 0; i
< newnary
->length
; i
++)
1359 if (TREE_CODE (newnary
->op
[i
]) != SSA_NAME
)
1363 pre_expr leader
, result
;
1364 unsigned int op_val_id
= VN_INFO (newnary
->op
[i
])->value_id
;
1365 leader
= find_leader_in_sets (op_val_id
, set1
, set2
);
1366 result
= phi_translate (leader
, set1
, set2
, e
);
1367 if (result
&& result
!= leader
)
1368 /* Force a leader as well as we are simplifying this
1370 newnary
->op
[i
] = get_representative_for (result
, pred
);
1374 changed
|= newnary
->op
[i
] != nary
->op
[i
];
1380 unsigned int new_val_id
;
1382 PRE_EXPR_NARY (expr
) = newnary
;
1383 constant
= fully_constant_expression (expr
);
1384 PRE_EXPR_NARY (expr
) = nary
;
1385 if (constant
!= expr
)
1387 /* For non-CONSTANTs we have to make sure we can eventually
1388 insert the expression. Which means we need to have a
1390 if (constant
->kind
!= CONSTANT
)
1392 /* Do not allow simplifications to non-constants over
1393 backedges as this will likely result in a loop PHI node
1394 to be inserted and increased register pressure.
1395 See PR77498 - this avoids doing predcoms work in
1396 a less efficient way. */
1397 if (e
->flags
& EDGE_DFS_BACK
)
1401 unsigned value_id
= get_expr_value_id (constant
);
1402 constant
= find_leader_in_sets (value_id
, set1
, set2
,
1412 /* vn_nary_* do not valueize operands. */
1413 for (i
= 0; i
< newnary
->length
; ++i
)
1414 if (TREE_CODE (newnary
->op
[i
]) == SSA_NAME
)
1415 newnary
->op
[i
] = VN_INFO (newnary
->op
[i
])->valnum
;
1416 tree result
= vn_nary_op_lookup_pieces (newnary
->length
,
1421 if (result
&& is_gimple_min_invariant (result
))
1422 return get_or_alloc_expr_for_constant (result
);
1424 expr
= pre_expr_pool
.allocate ();
1429 PRE_EXPR_NARY (expr
) = nary
;
1430 new_val_id
= nary
->value_id
;
1431 get_or_alloc_expression_id (expr
);
1435 new_val_id
= get_next_value_id ();
1436 value_expressions
.safe_grow_cleared (get_max_value_id () + 1);
1437 nary
= vn_nary_op_insert_pieces (newnary
->length
,
1441 result
, new_val_id
);
1442 PRE_EXPR_NARY (expr
) = nary
;
1443 get_or_alloc_expression_id (expr
);
1445 add_to_value (new_val_id
, expr
);
1453 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
1454 vec
<vn_reference_op_s
> operands
= ref
->operands
;
1455 tree vuse
= ref
->vuse
;
1456 tree newvuse
= vuse
;
1457 vec
<vn_reference_op_s
> newoperands
= vNULL
;
1458 bool changed
= false, same_valid
= true;
1460 vn_reference_op_t operand
;
1461 vn_reference_t newref
;
1463 for (i
= 0; operands
.iterate (i
, &operand
); i
++)
1468 tree type
= operand
->type
;
1469 vn_reference_op_s newop
= *operand
;
1470 op
[0] = operand
->op0
;
1471 op
[1] = operand
->op1
;
1472 op
[2] = operand
->op2
;
1473 for (n
= 0; n
< 3; ++n
)
1475 unsigned int op_val_id
;
1478 if (TREE_CODE (op
[n
]) != SSA_NAME
)
1480 /* We can't possibly insert these. */
1482 && !is_gimple_min_invariant (op
[n
]))
1486 op_val_id
= VN_INFO (op
[n
])->value_id
;
1487 leader
= find_leader_in_sets (op_val_id
, set1
, set2
);
1488 opresult
= phi_translate (leader
, set1
, set2
, e
);
1489 if (opresult
&& opresult
!= leader
)
1491 tree name
= get_representative_for (opresult
);
1492 changed
|= name
!= op
[n
];
1500 newoperands
.release ();
1505 if (!newoperands
.exists ())
1506 newoperands
= operands
.copy ();
1507 /* We may have changed from an SSA_NAME to a constant */
1508 if (newop
.opcode
== SSA_NAME
&& TREE_CODE (op
[0]) != SSA_NAME
)
1509 newop
.opcode
= TREE_CODE (op
[0]);
1514 newoperands
[i
] = newop
;
1516 gcc_checking_assert (i
== operands
.length ());
1520 newvuse
= translate_vuse_through_block (newoperands
.exists ()
1521 ? newoperands
: operands
,
1522 ref
->set
, ref
->type
,
1523 vuse
, phiblock
, pred
,
1525 if (newvuse
== NULL_TREE
)
1527 newoperands
.release ();
1532 if (changed
|| newvuse
!= vuse
)
1534 unsigned int new_val_id
;
1536 tree result
= vn_reference_lookup_pieces (newvuse
, ref
->set
,
1538 newoperands
.exists ()
1539 ? newoperands
: operands
,
1542 newoperands
.release ();
1544 /* We can always insert constants, so if we have a partial
1545 redundant constant load of another type try to translate it
1546 to a constant of appropriate type. */
1547 if (result
&& is_gimple_min_invariant (result
))
1550 if (!useless_type_conversion_p (ref
->type
, TREE_TYPE (result
)))
1552 tem
= fold_unary (VIEW_CONVERT_EXPR
, ref
->type
, result
);
1553 if (tem
&& !is_gimple_min_invariant (tem
))
1557 return get_or_alloc_expr_for_constant (tem
);
1560 /* If we'd have to convert things we would need to validate
1561 if we can insert the translated expression. So fail
1562 here for now - we cannot insert an alias with a different
1563 type in the VN tables either, as that would assert. */
1565 && !useless_type_conversion_p (ref
->type
, TREE_TYPE (result
)))
1567 else if (!result
&& newref
1568 && !useless_type_conversion_p (ref
->type
, newref
->type
))
1570 newoperands
.release ();
1574 expr
= pre_expr_pool
.allocate ();
1575 expr
->kind
= REFERENCE
;
1579 new_val_id
= newref
->value_id
;
1582 if (changed
|| !same_valid
)
1584 new_val_id
= get_next_value_id ();
1585 value_expressions
.safe_grow_cleared
1586 (get_max_value_id () + 1);
1589 new_val_id
= ref
->value_id
;
1590 if (!newoperands
.exists ())
1591 newoperands
= operands
.copy ();
1592 newref
= vn_reference_insert_pieces (newvuse
, ref
->set
,
1595 result
, new_val_id
);
1596 newoperands
= vNULL
;
1598 PRE_EXPR_REFERENCE (expr
) = newref
;
1599 get_or_alloc_expression_id (expr
);
1600 add_to_value (new_val_id
, expr
);
1602 newoperands
.release ();
1609 tree name
= PRE_EXPR_NAME (expr
);
1610 gimple
*def_stmt
= SSA_NAME_DEF_STMT (name
);
1611 /* If the SSA name is defined by a PHI node in this block,
1613 if (gimple_code (def_stmt
) == GIMPLE_PHI
1614 && gimple_bb (def_stmt
) == phiblock
)
1616 tree def
= PHI_ARG_DEF (def_stmt
, e
->dest_idx
);
1618 /* Handle constant. */
1619 if (is_gimple_min_invariant (def
))
1620 return get_or_alloc_expr_for_constant (def
);
1622 return get_or_alloc_expr_for_name (def
);
1624 /* Otherwise return it unchanged - it will get removed if its
1625 value is not available in PREDs AVAIL_OUT set of expressions
1626 by the subtraction of TMP_GEN. */
1635 /* Wrapper around phi_translate_1 providing caching functionality. */
1638 phi_translate (pre_expr expr
, bitmap_set_t set1
, bitmap_set_t set2
, edge e
)
1640 expr_pred_trans_t slot
= NULL
;
1646 /* Constants contain no values that need translation. */
1647 if (expr
->kind
== CONSTANT
)
1650 if (value_id_constant_p (get_expr_value_id (expr
)))
1653 /* Don't add translations of NAMEs as those are cheap to translate. */
1654 if (expr
->kind
!= NAME
)
1656 if (phi_trans_add (&slot
, expr
, e
->src
))
1658 /* Store NULL for the value we want to return in the case of
1664 phitrans
= phi_translate_1 (expr
, set1
, set2
, e
);
1671 /* Remove failed translations again, they cause insert
1672 iteration to not pick up new opportunities reliably. */
1673 phi_translate_table
->remove_elt_with_hash (slot
, slot
->hashcode
);
1680 /* For each expression in SET, translate the values through phi nodes
1681 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1682 expressions in DEST. */
1685 phi_translate_set (bitmap_set_t dest
, bitmap_set_t set
, edge e
)
1687 vec
<pre_expr
> exprs
;
1691 if (gimple_seq_empty_p (phi_nodes (e
->dest
)))
1693 bitmap_set_copy (dest
, set
);
1697 exprs
= sorted_array_from_bitmap_set (set
);
1698 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
1700 pre_expr translated
;
1701 translated
= phi_translate (expr
, set
, NULL
, e
);
1705 bitmap_insert_into_set (dest
, translated
);
1710 /* Find the leader for a value (i.e., the name representing that
1711 value) in a given set, and return it. Return NULL if no leader
1715 bitmap_find_leader (bitmap_set_t set
, unsigned int val
)
1717 if (value_id_constant_p (val
))
1721 bitmap exprset
= value_expressions
[val
];
1723 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
1725 pre_expr expr
= expression_for_id (i
);
1726 if (expr
->kind
== CONSTANT
)
1730 if (bitmap_set_contains_value (set
, val
))
1732 /* Rather than walk the entire bitmap of expressions, and see
1733 whether any of them has the value we are looking for, we look
1734 at the reverse mapping, which tells us the set of expressions
1735 that have a given value (IE value->expressions with that
1736 value) and see if any of those expressions are in our set.
1737 The number of expressions per value is usually significantly
1738 less than the number of expressions in the set. In fact, for
1739 large testcases, doing it this way is roughly 5-10x faster
1740 than walking the bitmap.
1741 If this is somehow a significant lose for some cases, we can
1742 choose which set to walk based on which set is smaller. */
1745 bitmap exprset
= value_expressions
[val
];
1747 EXECUTE_IF_AND_IN_BITMAP (exprset
, &set
->expressions
, 0, i
, bi
)
1748 return expression_for_id (i
);
1753 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1754 BLOCK by seeing if it is not killed in the block. Note that we are
1755 only determining whether there is a store that kills it. Because
1756 of the order in which clean iterates over values, we are guaranteed
1757 that altered operands will have caused us to be eliminated from the
1758 ANTIC_IN set already. */
1761 value_dies_in_block_x (pre_expr expr
, basic_block block
)
1763 tree vuse
= PRE_EXPR_REFERENCE (expr
)->vuse
;
1764 vn_reference_t refx
= PRE_EXPR_REFERENCE (expr
);
1766 gimple_stmt_iterator gsi
;
1767 unsigned id
= get_expression_id (expr
);
1774 /* Lookup a previously calculated result. */
1775 if (EXPR_DIES (block
)
1776 && bitmap_bit_p (EXPR_DIES (block
), id
* 2))
1777 return bitmap_bit_p (EXPR_DIES (block
), id
* 2 + 1);
1779 /* A memory expression {e, VUSE} dies in the block if there is a
1780 statement that may clobber e. If, starting statement walk from the
1781 top of the basic block, a statement uses VUSE there can be no kill
1782 inbetween that use and the original statement that loaded {e, VUSE},
1783 so we can stop walking. */
1784 ref
.base
= NULL_TREE
;
1785 for (gsi
= gsi_start_bb (block
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1787 tree def_vuse
, def_vdef
;
1788 def
= gsi_stmt (gsi
);
1789 def_vuse
= gimple_vuse (def
);
1790 def_vdef
= gimple_vdef (def
);
1792 /* Not a memory statement. */
1796 /* Not a may-def. */
1799 /* A load with the same VUSE, we're done. */
1800 if (def_vuse
== vuse
)
1806 /* Init ref only if we really need it. */
1807 if (ref
.base
== NULL_TREE
1808 && !ao_ref_init_from_vn_reference (&ref
, refx
->set
, refx
->type
,
1814 /* If the statement may clobber expr, it dies. */
1815 if (stmt_may_clobber_ref_p_1 (def
, &ref
))
1822 /* Remember the result. */
1823 if (!EXPR_DIES (block
))
1824 EXPR_DIES (block
) = BITMAP_ALLOC (&grand_bitmap_obstack
);
1825 bitmap_set_bit (EXPR_DIES (block
), id
* 2);
1827 bitmap_set_bit (EXPR_DIES (block
), id
* 2 + 1);
1833 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1834 contains its value-id. */
1837 op_valid_in_sets (bitmap_set_t set1
, bitmap_set_t set2
, tree op
)
1839 if (op
&& TREE_CODE (op
) == SSA_NAME
)
1841 unsigned int value_id
= VN_INFO (op
)->value_id
;
1842 if (!(bitmap_set_contains_value (set1
, value_id
)
1843 || (set2
&& bitmap_set_contains_value (set2
, value_id
))))
1849 /* Determine if the expression EXPR is valid in SET1 U SET2.
1850 ONLY SET2 CAN BE NULL.
1851 This means that we have a leader for each part of the expression
1852 (if it consists of values), or the expression is an SSA_NAME.
1853 For loads/calls, we also see if the vuse is killed in this block. */
1856 valid_in_sets (bitmap_set_t set1
, bitmap_set_t set2
, pre_expr expr
)
1861 /* By construction all NAMEs are available. Non-available
1862 NAMEs are removed by subtracting TMP_GEN from the sets. */
1867 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
1868 for (i
= 0; i
< nary
->length
; i
++)
1869 if (!op_valid_in_sets (set1
, set2
, nary
->op
[i
]))
1876 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
1877 vn_reference_op_t vro
;
1880 FOR_EACH_VEC_ELT (ref
->operands
, i
, vro
)
1882 if (!op_valid_in_sets (set1
, set2
, vro
->op0
)
1883 || !op_valid_in_sets (set1
, set2
, vro
->op1
)
1884 || !op_valid_in_sets (set1
, set2
, vro
->op2
))
1894 /* Clean the set of expressions SET1 that are no longer valid in SET1 or SET2.
1895 This means expressions that are made up of values we have no leaders for
1899 clean (bitmap_set_t set1
, bitmap_set_t set2
= NULL
)
1901 vec
<pre_expr
> exprs
= sorted_array_from_bitmap_set (set1
);
1905 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
1907 if (!valid_in_sets (set1
, set2
, expr
))
1909 unsigned int val
= get_expr_value_id (expr
);
1910 bitmap_clear_bit (&set1
->expressions
, get_expression_id (expr
));
1911 /* We are entered with possibly multiple expressions for a value
1912 so before removing a value from the set see if there's an
1913 expression for it left. */
1914 if (! bitmap_find_leader (set1
, val
))
1915 bitmap_clear_bit (&set1
->values
, val
);
1921 /* Clean the set of expressions that are no longer valid in SET because
1922 they are clobbered in BLOCK or because they trap and may not be executed. */
1925 prune_clobbered_mems (bitmap_set_t set
, basic_block block
)
1929 unsigned to_remove
= -1U;
1930 bool any_removed
= false;
1932 FOR_EACH_EXPR_ID_IN_SET (set
, i
, bi
)
1934 /* Remove queued expr. */
1935 if (to_remove
!= -1U)
1937 bitmap_clear_bit (&set
->expressions
, to_remove
);
1942 pre_expr expr
= expression_for_id (i
);
1943 if (expr
->kind
== REFERENCE
)
1945 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
1948 gimple
*def_stmt
= SSA_NAME_DEF_STMT (ref
->vuse
);
1949 if (!gimple_nop_p (def_stmt
)
1950 && ((gimple_bb (def_stmt
) != block
1951 && !dominated_by_p (CDI_DOMINATORS
,
1952 block
, gimple_bb (def_stmt
)))
1953 || (gimple_bb (def_stmt
) == block
1954 && value_dies_in_block_x (expr
, block
))))
1958 else if (expr
->kind
== NARY
)
1960 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
1961 /* If the NARY may trap make sure the block does not contain
1962 a possible exit point.
1963 ??? This is overly conservative if we translate AVAIL_OUT
1964 as the available expression might be after the exit point. */
1965 if (BB_MAY_NOTRETURN (block
)
1966 && vn_nary_may_trap (nary
))
1971 /* Remove queued expr. */
1972 if (to_remove
!= -1U)
1974 bitmap_clear_bit (&set
->expressions
, to_remove
);
1978 /* Above we only removed expressions, now clean the set of values
1979 which no longer have any corresponding expression. We cannot
1980 clear the value at the time we remove an expression since there
1981 may be multiple expressions per value.
1982 If we'd queue possibly to be removed values we could use
1983 the bitmap_find_leader way to see if there's still an expression
1984 for it. For some ratio of to be removed values and number of
1985 values/expressions in the set this might be faster than rebuilding
1989 bitmap_clear (&set
->values
);
1990 FOR_EACH_EXPR_ID_IN_SET (set
, i
, bi
)
1992 pre_expr expr
= expression_for_id (i
);
1993 unsigned int value_id
= get_expr_value_id (expr
);
1994 bitmap_set_bit (&set
->values
, value_id
);
1999 static sbitmap has_abnormal_preds
;
2001 /* Compute the ANTIC set for BLOCK.
2003 If succs(BLOCK) > 1 then
2004 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2005 else if succs(BLOCK) == 1 then
2006 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2008 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2010 Note that clean() is deferred until after the iteration. */
2013 compute_antic_aux (basic_block block
, bool block_has_abnormal_pred_edge
)
2015 bitmap_set_t S
, old
, ANTIC_OUT
;
2019 bool was_visited
= BB_VISITED (block
);
2020 bool changed
= ! BB_VISITED (block
);
2021 BB_VISITED (block
) = 1;
2022 old
= ANTIC_OUT
= S
= NULL
;
2024 /* If any edges from predecessors are abnormal, antic_in is empty,
2026 if (block_has_abnormal_pred_edge
)
2027 goto maybe_dump_sets
;
2029 old
= ANTIC_IN (block
);
2030 ANTIC_OUT
= bitmap_set_new ();
2032 /* If the block has no successors, ANTIC_OUT is empty. */
2033 if (EDGE_COUNT (block
->succs
) == 0)
2035 /* If we have one successor, we could have some phi nodes to
2036 translate through. */
2037 else if (single_succ_p (block
))
2039 e
= single_succ_edge (block
);
2040 gcc_assert (BB_VISITED (e
->dest
));
2041 BB_VISITED_WITH_VISITED_SUCCS (block
)
2042 = BB_VISITED_WITH_VISITED_SUCCS (e
->dest
);
2043 phi_translate_set (ANTIC_OUT
, ANTIC_IN (e
->dest
), e
);
2045 /* If we have multiple successors, we take the intersection of all of
2046 them. Note that in the case of loop exit phi nodes, we may have
2047 phis to translate through. */
2053 BB_VISITED_WITH_VISITED_SUCCS (block
) = true;
2054 auto_vec
<edge
> worklist (EDGE_COUNT (block
->succs
));
2055 FOR_EACH_EDGE (e
, ei
, block
->succs
)
2058 && BB_VISITED (e
->dest
))
2060 else if (BB_VISITED (e
->dest
))
2061 worklist
.quick_push (e
);
2064 /* Unvisited successors get their ANTIC_IN replaced by the
2065 maximal set to arrive at a maximum ANTIC_IN solution.
2066 We can ignore them in the intersection operation and thus
2067 need not explicitely represent that maximum solution. */
2068 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2069 fprintf (dump_file
, "ANTIC_IN is MAX on %d->%d\n",
2070 e
->src
->index
, e
->dest
->index
);
2072 BB_VISITED_WITH_VISITED_SUCCS (block
)
2073 &= BB_VISITED_WITH_VISITED_SUCCS (e
->dest
);
2076 /* Of multiple successors we have to have visited one already
2077 which is guaranteed by iteration order. */
2078 gcc_assert (first
!= NULL
);
2080 phi_translate_set (ANTIC_OUT
, ANTIC_IN (first
->dest
), first
);
2082 /* If we have multiple successors we need to intersect the ANTIC_OUT
2083 sets. For values that's a simple intersection but for
2084 expressions it is a union. Given we want to have a single
2085 expression per value in our sets we have to canonicalize.
2086 Avoid randomness and running into cycles like for PR82129 and
2087 canonicalize the expression we choose to the one with the
2088 lowest id. This requires we actually compute the union first. */
2089 FOR_EACH_VEC_ELT (worklist
, i
, e
)
2091 if (!gimple_seq_empty_p (phi_nodes (e
->dest
)))
2093 bitmap_set_t tmp
= bitmap_set_new ();
2094 phi_translate_set (tmp
, ANTIC_IN (e
->dest
), e
);
2095 bitmap_and_into (&ANTIC_OUT
->values
, &tmp
->values
);
2096 bitmap_ior_into (&ANTIC_OUT
->expressions
, &tmp
->expressions
);
2097 bitmap_set_free (tmp
);
2101 bitmap_and_into (&ANTIC_OUT
->values
, &ANTIC_IN (e
->dest
)->values
);
2102 bitmap_ior_into (&ANTIC_OUT
->expressions
,
2103 &ANTIC_IN (e
->dest
)->expressions
);
2106 if (! worklist
.is_empty ())
2108 /* Prune expressions not in the value set. */
2111 unsigned int to_clear
= -1U;
2112 FOR_EACH_EXPR_ID_IN_SET (ANTIC_OUT
, i
, bi
)
2114 if (to_clear
!= -1U)
2116 bitmap_clear_bit (&ANTIC_OUT
->expressions
, to_clear
);
2119 pre_expr expr
= expression_for_id (i
);
2120 unsigned int value_id
= get_expr_value_id (expr
);
2121 if (!bitmap_bit_p (&ANTIC_OUT
->values
, value_id
))
2124 if (to_clear
!= -1U)
2125 bitmap_clear_bit (&ANTIC_OUT
->expressions
, to_clear
);
2129 /* Prune expressions that are clobbered in block and thus become
2130 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2131 prune_clobbered_mems (ANTIC_OUT
, block
);
2133 /* Generate ANTIC_OUT - TMP_GEN. */
2134 S
= bitmap_set_subtract_expressions (ANTIC_OUT
, TMP_GEN (block
));
2136 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2137 ANTIC_IN (block
) = bitmap_set_subtract_expressions (EXP_GEN (block
),
2140 /* Then union in the ANTIC_OUT - TMP_GEN values,
2141 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2142 bitmap_ior_into (&ANTIC_IN (block
)->values
, &S
->values
);
2143 bitmap_ior_into (&ANTIC_IN (block
)->expressions
, &S
->expressions
);
2145 /* clean (ANTIC_IN (block)) is defered to after the iteration converged
2146 because it can cause non-convergence, see for example PR81181. */
2148 if (!bitmap_set_equal (old
, ANTIC_IN (block
)))
2151 /* After the initial value set computation the value set may
2152 only shrink during the iteration. */
2153 if (was_visited
&& BB_VISITED_WITH_VISITED_SUCCS (block
) && flag_checking
)
2157 EXECUTE_IF_AND_COMPL_IN_BITMAP (&ANTIC_IN (block
)->values
,
2158 &old
->values
, 0, i
, bi
)
2164 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2167 print_bitmap_set (dump_file
, ANTIC_OUT
, "ANTIC_OUT", block
->index
);
2170 fprintf (dump_file
, "[changed] ");
2171 print_bitmap_set (dump_file
, ANTIC_IN (block
), "ANTIC_IN",
2175 print_bitmap_set (dump_file
, S
, "S", block
->index
);
2178 bitmap_set_free (old
);
2180 bitmap_set_free (S
);
2182 bitmap_set_free (ANTIC_OUT
);
2186 /* Compute PARTIAL_ANTIC for BLOCK.
2188 If succs(BLOCK) > 1 then
2189 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2190 in ANTIC_OUT for all succ(BLOCK)
2191 else if succs(BLOCK) == 1 then
2192 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2194 PA_IN[BLOCK] = clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK] - ANTIC_IN[BLOCK])
2198 compute_partial_antic_aux (basic_block block
,
2199 bool block_has_abnormal_pred_edge
)
2201 bitmap_set_t old_PA_IN
;
2202 bitmap_set_t PA_OUT
;
2205 unsigned long max_pa
= PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH
);
2207 old_PA_IN
= PA_OUT
= NULL
;
2209 /* If any edges from predecessors are abnormal, antic_in is empty,
2211 if (block_has_abnormal_pred_edge
)
2212 goto maybe_dump_sets
;
2214 /* If there are too many partially anticipatable values in the
2215 block, phi_translate_set can take an exponential time: stop
2216 before the translation starts. */
2218 && single_succ_p (block
)
2219 && bitmap_count_bits (&PA_IN (single_succ (block
))->values
) > max_pa
)
2220 goto maybe_dump_sets
;
2222 old_PA_IN
= PA_IN (block
);
2223 PA_OUT
= bitmap_set_new ();
2225 /* If the block has no successors, ANTIC_OUT is empty. */
2226 if (EDGE_COUNT (block
->succs
) == 0)
2228 /* If we have one successor, we could have some phi nodes to
2229 translate through. Note that we can't phi translate across DFS
2230 back edges in partial antic, because it uses a union operation on
2231 the successors. For recurrences like IV's, we will end up
2232 generating a new value in the set on each go around (i + 3 (VH.1)
2233 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2234 else if (single_succ_p (block
))
2236 e
= single_succ_edge (block
);
2237 if (!(e
->flags
& EDGE_DFS_BACK
))
2238 phi_translate_set (PA_OUT
, PA_IN (e
->dest
), e
);
2240 /* If we have multiple successors, we take the union of all of
2246 auto_vec
<edge
> worklist (EDGE_COUNT (block
->succs
));
2247 FOR_EACH_EDGE (e
, ei
, block
->succs
)
2249 if (e
->flags
& EDGE_DFS_BACK
)
2251 worklist
.quick_push (e
);
2253 if (worklist
.length () > 0)
2255 FOR_EACH_VEC_ELT (worklist
, i
, e
)
2260 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (e
->dest
), i
, bi
)
2261 bitmap_value_insert_into_set (PA_OUT
,
2262 expression_for_id (i
));
2263 if (!gimple_seq_empty_p (phi_nodes (e
->dest
)))
2265 bitmap_set_t pa_in
= bitmap_set_new ();
2266 phi_translate_set (pa_in
, PA_IN (e
->dest
), e
);
2267 FOR_EACH_EXPR_ID_IN_SET (pa_in
, i
, bi
)
2268 bitmap_value_insert_into_set (PA_OUT
,
2269 expression_for_id (i
));
2270 bitmap_set_free (pa_in
);
2273 FOR_EACH_EXPR_ID_IN_SET (PA_IN (e
->dest
), i
, bi
)
2274 bitmap_value_insert_into_set (PA_OUT
,
2275 expression_for_id (i
));
2280 /* Prune expressions that are clobbered in block and thus become
2281 invalid if translated from PA_OUT to PA_IN. */
2282 prune_clobbered_mems (PA_OUT
, block
);
2284 /* PA_IN starts with PA_OUT - TMP_GEN.
2285 Then we subtract things from ANTIC_IN. */
2286 PA_IN (block
) = bitmap_set_subtract_expressions (PA_OUT
, TMP_GEN (block
));
2288 /* For partial antic, we want to put back in the phi results, since
2289 we will properly avoid making them partially antic over backedges. */
2290 bitmap_ior_into (&PA_IN (block
)->values
, &PHI_GEN (block
)->values
);
2291 bitmap_ior_into (&PA_IN (block
)->expressions
, &PHI_GEN (block
)->expressions
);
2293 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2294 bitmap_set_subtract_values (PA_IN (block
), ANTIC_IN (block
));
2296 clean (PA_IN (block
), ANTIC_IN (block
));
2299 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2302 print_bitmap_set (dump_file
, PA_OUT
, "PA_OUT", block
->index
);
2304 print_bitmap_set (dump_file
, PA_IN (block
), "PA_IN", block
->index
);
2307 bitmap_set_free (old_PA_IN
);
2309 bitmap_set_free (PA_OUT
);
2312 /* Compute ANTIC and partial ANTIC sets. */
2315 compute_antic (void)
2317 bool changed
= true;
2318 int num_iterations
= 0;
2324 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2325 We pre-build the map of blocks with incoming abnormal edges here. */
2326 has_abnormal_preds
= sbitmap_alloc (last_basic_block_for_fn (cfun
));
2327 bitmap_clear (has_abnormal_preds
);
2329 FOR_ALL_BB_FN (block
, cfun
)
2331 BB_VISITED (block
) = 0;
2332 BB_VISITED_WITH_VISITED_SUCCS (block
) = 0;
2334 FOR_EACH_EDGE (e
, ei
, block
->preds
)
2335 if (e
->flags
& EDGE_ABNORMAL
)
2337 bitmap_set_bit (has_abnormal_preds
, block
->index
);
2341 /* While we are here, give empty ANTIC_IN sets to each block. */
2342 ANTIC_IN (block
) = bitmap_set_new ();
2343 if (do_partial_partial
)
2344 PA_IN (block
) = bitmap_set_new ();
2347 /* At the exit block we anticipate nothing. */
2348 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun
)) = 1;
2349 BB_VISITED_WITH_VISITED_SUCCS (EXIT_BLOCK_PTR_FOR_FN (cfun
)) = 1;
2351 /* For ANTIC computation we need a postorder that also guarantees that
2352 a block with a single successor is visited after its successor.
2353 RPO on the inverted CFG has this property. */
2354 auto_vec
<int, 20> postorder
;
2355 inverted_post_order_compute (&postorder
);
2357 auto_sbitmap
worklist (last_basic_block_for_fn (cfun
) + 1);
2358 bitmap_clear (worklist
);
2359 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR_FOR_FN (cfun
)->preds
)
2360 bitmap_set_bit (worklist
, e
->src
->index
);
2363 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2364 fprintf (dump_file
, "Starting iteration %d\n", num_iterations
);
2365 /* ??? We need to clear our PHI translation cache here as the
2366 ANTIC sets shrink and we restrict valid translations to
2367 those having operands with leaders in ANTIC. Same below
2368 for PA ANTIC computation. */
2371 for (i
= postorder
.length () - 1; i
>= 0; i
--)
2373 if (bitmap_bit_p (worklist
, postorder
[i
]))
2375 basic_block block
= BASIC_BLOCK_FOR_FN (cfun
, postorder
[i
]);
2376 bitmap_clear_bit (worklist
, block
->index
);
2377 if (compute_antic_aux (block
,
2378 bitmap_bit_p (has_abnormal_preds
,
2381 FOR_EACH_EDGE (e
, ei
, block
->preds
)
2382 bitmap_set_bit (worklist
, e
->src
->index
);
2387 /* Theoretically possible, but *highly* unlikely. */
2388 gcc_checking_assert (num_iterations
< 500);
2391 /* We have to clean after the dataflow problem converged as cleaning
2392 can cause non-convergence because it is based on expressions
2393 rather than values. */
2394 FOR_EACH_BB_FN (block
, cfun
)
2395 clean (ANTIC_IN (block
));
2397 statistics_histogram_event (cfun
, "compute_antic iterations",
2400 if (do_partial_partial
)
2402 /* For partial antic we ignore backedges and thus we do not need
2403 to perform any iteration when we process blocks in postorder. */
2405 = pre_and_rev_post_order_compute (NULL
, postorder
.address (), false);
2406 for (i
= postorder_num
- 1 ; i
>= 0; i
--)
2408 basic_block block
= BASIC_BLOCK_FOR_FN (cfun
, postorder
[i
]);
2409 compute_partial_antic_aux (block
,
2410 bitmap_bit_p (has_abnormal_preds
,
2415 sbitmap_free (has_abnormal_preds
);
2419 /* Inserted expressions are placed onto this worklist, which is used
2420 for performing quick dead code elimination of insertions we made
2421 that didn't turn out to be necessary. */
2422 static bitmap inserted_exprs
;
2424 /* The actual worker for create_component_ref_by_pieces. */
2427 create_component_ref_by_pieces_1 (basic_block block
, vn_reference_t ref
,
2428 unsigned int *operand
, gimple_seq
*stmts
)
2430 vn_reference_op_t currop
= &ref
->operands
[*operand
];
2433 switch (currop
->opcode
)
2440 tree baseop
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2444 tree offset
= currop
->op0
;
2445 if (TREE_CODE (baseop
) == ADDR_EXPR
2446 && handled_component_p (TREE_OPERAND (baseop
, 0)))
2450 base
= get_addr_base_and_unit_offset (TREE_OPERAND (baseop
, 0),
2453 offset
= int_const_binop (PLUS_EXPR
, offset
,
2454 build_int_cst (TREE_TYPE (offset
),
2456 baseop
= build_fold_addr_expr (base
);
2458 genop
= build2 (MEM_REF
, currop
->type
, baseop
, offset
);
2459 MR_DEPENDENCE_CLIQUE (genop
) = currop
->clique
;
2460 MR_DEPENDENCE_BASE (genop
) = currop
->base
;
2461 REF_REVERSE_STORAGE_ORDER (genop
) = currop
->reverse
;
2465 case TARGET_MEM_REF
:
2467 tree genop0
= NULL_TREE
, genop1
= NULL_TREE
;
2468 vn_reference_op_t nextop
= &ref
->operands
[++*operand
];
2469 tree baseop
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2475 genop0
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2481 genop1
= find_or_generate_expression (block
, nextop
->op0
, stmts
);
2485 genop
= build5 (TARGET_MEM_REF
, currop
->type
,
2486 baseop
, currop
->op2
, genop0
, currop
->op1
, genop1
);
2488 MR_DEPENDENCE_CLIQUE (genop
) = currop
->clique
;
2489 MR_DEPENDENCE_BASE (genop
) = currop
->base
;
2496 gcc_assert (is_gimple_min_invariant (currop
->op0
));
2502 case VIEW_CONVERT_EXPR
:
2504 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2508 return fold_build1 (currop
->opcode
, currop
->type
, genop0
);
2511 case WITH_SIZE_EXPR
:
2513 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2517 tree genop1
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2520 return fold_build2 (currop
->opcode
, currop
->type
, genop0
, genop1
);
2525 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2529 tree op1
= currop
->op0
;
2530 tree op2
= currop
->op1
;
2531 tree t
= build3 (BIT_FIELD_REF
, currop
->type
, genop0
, op1
, op2
);
2532 REF_REVERSE_STORAGE_ORDER (t
) = currop
->reverse
;
2536 /* For array ref vn_reference_op's, operand 1 of the array ref
2537 is op0 of the reference op and operand 3 of the array ref is
2539 case ARRAY_RANGE_REF
:
2543 tree genop1
= currop
->op0
;
2544 tree genop2
= currop
->op1
;
2545 tree genop3
= currop
->op2
;
2546 genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2550 genop1
= find_or_generate_expression (block
, genop1
, stmts
);
2555 tree domain_type
= TYPE_DOMAIN (TREE_TYPE (genop0
));
2556 /* Drop zero minimum index if redundant. */
2557 if (integer_zerop (genop2
)
2559 || integer_zerop (TYPE_MIN_VALUE (domain_type
))))
2563 genop2
= find_or_generate_expression (block
, genop2
, stmts
);
2570 tree elmt_type
= TREE_TYPE (TREE_TYPE (genop0
));
2571 /* We can't always put a size in units of the element alignment
2572 here as the element alignment may be not visible. See
2573 PR43783. Simply drop the element size for constant
2575 if (TREE_CODE (genop3
) == INTEGER_CST
2576 && TREE_CODE (TYPE_SIZE_UNIT (elmt_type
)) == INTEGER_CST
2577 && wi::eq_p (wi::to_offset (TYPE_SIZE_UNIT (elmt_type
)),
2578 (wi::to_offset (genop3
)
2579 * vn_ref_op_align_unit (currop
))))
2583 genop3
= find_or_generate_expression (block
, genop3
, stmts
);
2588 return build4 (currop
->opcode
, currop
->type
, genop0
, genop1
,
2595 tree genop2
= currop
->op1
;
2596 op0
= create_component_ref_by_pieces_1 (block
, ref
, operand
, stmts
);
2599 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2603 genop2
= find_or_generate_expression (block
, genop2
, stmts
);
2607 return fold_build3 (COMPONENT_REF
, TREE_TYPE (op1
), op0
, op1
, genop2
);
2612 genop
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2633 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2634 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2635 trying to rename aggregates into ssa form directly, which is a no no.
2637 Thus, this routine doesn't create temporaries, it just builds a
2638 single access expression for the array, calling
2639 find_or_generate_expression to build the innermost pieces.
2641 This function is a subroutine of create_expression_by_pieces, and
2642 should not be called on it's own unless you really know what you
2646 create_component_ref_by_pieces (basic_block block
, vn_reference_t ref
,
2649 unsigned int op
= 0;
2650 return create_component_ref_by_pieces_1 (block
, ref
, &op
, stmts
);
2653 /* Find a simple leader for an expression, or generate one using
2654 create_expression_by_pieces from a NARY expression for the value.
2655 BLOCK is the basic_block we are looking for leaders in.
2656 OP is the tree expression to find a leader for or generate.
2657 Returns the leader or NULL_TREE on failure. */
2660 find_or_generate_expression (basic_block block
, tree op
, gimple_seq
*stmts
)
2662 pre_expr expr
= get_or_alloc_expr_for (op
);
2663 unsigned int lookfor
= get_expr_value_id (expr
);
2664 pre_expr leader
= bitmap_find_leader (AVAIL_OUT (block
), lookfor
);
2667 if (leader
->kind
== NAME
)
2668 return PRE_EXPR_NAME (leader
);
2669 else if (leader
->kind
== CONSTANT
)
2670 return PRE_EXPR_CONSTANT (leader
);
2676 /* It must be a complex expression, so generate it recursively. Note
2677 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2678 where the insert algorithm fails to insert a required expression. */
2679 bitmap exprset
= value_expressions
[lookfor
];
2682 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
2684 pre_expr temp
= expression_for_id (i
);
2685 /* We cannot insert random REFERENCE expressions at arbitrary
2686 places. We can insert NARYs which eventually re-materializes
2687 its operand values. */
2688 if (temp
->kind
== NARY
)
2689 return create_expression_by_pieces (block
, temp
, stmts
,
2690 get_expr_type (expr
));
2697 /* Create an expression in pieces, so that we can handle very complex
2698 expressions that may be ANTIC, but not necessary GIMPLE.
2699 BLOCK is the basic block the expression will be inserted into,
2700 EXPR is the expression to insert (in value form)
2701 STMTS is a statement list to append the necessary insertions into.
2703 This function will die if we hit some value that shouldn't be
2704 ANTIC but is (IE there is no leader for it, or its components).
2705 The function returns NULL_TREE in case a different antic expression
2706 has to be inserted first.
2707 This function may also generate expressions that are themselves
2708 partially or fully redundant. Those that are will be either made
2709 fully redundant during the next iteration of insert (for partially
2710 redundant ones), or eliminated by eliminate (for fully redundant
2714 create_expression_by_pieces (basic_block block
, pre_expr expr
,
2715 gimple_seq
*stmts
, tree type
)
2719 gimple_seq forced_stmts
= NULL
;
2720 unsigned int value_id
;
2721 gimple_stmt_iterator gsi
;
2722 tree exprtype
= type
? type
: get_expr_type (expr
);
2728 /* We may hit the NAME/CONSTANT case if we have to convert types
2729 that value numbering saw through. */
2731 folded
= PRE_EXPR_NAME (expr
);
2732 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (folded
))
2734 if (useless_type_conversion_p (exprtype
, TREE_TYPE (folded
)))
2739 folded
= PRE_EXPR_CONSTANT (expr
);
2740 tree tem
= fold_convert (exprtype
, folded
);
2741 if (is_gimple_min_invariant (tem
))
2746 if (PRE_EXPR_REFERENCE (expr
)->operands
[0].opcode
== CALL_EXPR
)
2748 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
2749 unsigned int operand
= 1;
2750 vn_reference_op_t currop
= &ref
->operands
[0];
2751 tree sc
= NULL_TREE
;
2752 tree fn
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2757 sc
= find_or_generate_expression (block
, currop
->op1
, stmts
);
2761 auto_vec
<tree
> args (ref
->operands
.length () - 1);
2762 while (operand
< ref
->operands
.length ())
2764 tree arg
= create_component_ref_by_pieces_1 (block
, ref
,
2768 args
.quick_push (arg
);
2770 gcall
*call
= gimple_build_call_vec (fn
, args
);
2771 gimple_call_set_with_bounds (call
, currop
->with_bounds
);
2773 gimple_call_set_chain (call
, sc
);
2774 tree forcedname
= make_ssa_name (currop
->type
);
2775 gimple_call_set_lhs (call
, forcedname
);
2776 /* There's no CCP pass after PRE which would re-compute alignment
2777 information so make sure we re-materialize this here. */
2778 if (gimple_call_builtin_p (call
, BUILT_IN_ASSUME_ALIGNED
)
2779 && args
.length () - 2 <= 1
2780 && tree_fits_uhwi_p (args
[1])
2781 && (args
.length () != 3 || tree_fits_uhwi_p (args
[2])))
2783 unsigned HOST_WIDE_INT halign
= tree_to_uhwi (args
[1]);
2784 unsigned HOST_WIDE_INT hmisalign
2785 = args
.length () == 3 ? tree_to_uhwi (args
[2]) : 0;
2786 if ((halign
& (halign
- 1)) == 0
2787 && (hmisalign
& ~(halign
- 1)) == 0)
2788 set_ptr_info_alignment (get_ptr_info (forcedname
),
2791 gimple_set_vuse (call
, BB_LIVE_VOP_ON_EXIT (block
));
2792 gimple_seq_add_stmt_without_update (&forced_stmts
, call
);
2793 folded
= forcedname
;
2797 folded
= create_component_ref_by_pieces (block
,
2798 PRE_EXPR_REFERENCE (expr
),
2802 name
= make_temp_ssa_name (exprtype
, NULL
, "pretmp");
2803 newstmt
= gimple_build_assign (name
, folded
);
2804 gimple_seq_add_stmt_without_update (&forced_stmts
, newstmt
);
2805 gimple_set_vuse (newstmt
, BB_LIVE_VOP_ON_EXIT (block
));
2811 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
2812 tree
*genop
= XALLOCAVEC (tree
, nary
->length
);
2814 for (i
= 0; i
< nary
->length
; ++i
)
2816 genop
[i
] = find_or_generate_expression (block
, nary
->op
[i
], stmts
);
2819 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2820 may have conversions stripped. */
2821 if (nary
->opcode
== POINTER_PLUS_EXPR
)
2824 genop
[i
] = gimple_convert (&forced_stmts
,
2825 nary
->type
, genop
[i
]);
2827 genop
[i
] = gimple_convert (&forced_stmts
,
2828 sizetype
, genop
[i
]);
2831 genop
[i
] = gimple_convert (&forced_stmts
,
2832 TREE_TYPE (nary
->op
[i
]), genop
[i
]);
2834 if (nary
->opcode
== CONSTRUCTOR
)
2836 vec
<constructor_elt
, va_gc
> *elts
= NULL
;
2837 for (i
= 0; i
< nary
->length
; ++i
)
2838 CONSTRUCTOR_APPEND_ELT (elts
, NULL_TREE
, genop
[i
]);
2839 folded
= build_constructor (nary
->type
, elts
);
2840 name
= make_temp_ssa_name (exprtype
, NULL
, "pretmp");
2841 newstmt
= gimple_build_assign (name
, folded
);
2842 gimple_seq_add_stmt_without_update (&forced_stmts
, newstmt
);
2847 switch (nary
->length
)
2850 folded
= gimple_build (&forced_stmts
, nary
->opcode
, nary
->type
,
2854 folded
= gimple_build (&forced_stmts
, nary
->opcode
, nary
->type
,
2855 genop
[0], genop
[1]);
2858 folded
= gimple_build (&forced_stmts
, nary
->opcode
, nary
->type
,
2859 genop
[0], genop
[1], genop
[2]);
2871 folded
= gimple_convert (&forced_stmts
, exprtype
, folded
);
2873 /* If there is nothing to insert, return the simplified result. */
2874 if (gimple_seq_empty_p (forced_stmts
))
2876 /* If we simplified to a constant return it and discard eventually
2878 if (is_gimple_min_invariant (folded
))
2880 gimple_seq_discard (forced_stmts
);
2883 /* Likewise if we simplified to sth not queued for insertion. */
2885 gsi
= gsi_last (forced_stmts
);
2886 for (; !gsi_end_p (gsi
); gsi_prev (&gsi
))
2888 gimple
*stmt
= gsi_stmt (gsi
);
2889 tree forcedname
= gimple_get_lhs (stmt
);
2890 if (forcedname
== folded
)
2898 gimple_seq_discard (forced_stmts
);
2901 gcc_assert (TREE_CODE (folded
) == SSA_NAME
);
2903 /* If we have any intermediate expressions to the value sets, add them
2904 to the value sets and chain them in the instruction stream. */
2907 gsi
= gsi_start (forced_stmts
);
2908 for (; !gsi_end_p (gsi
); gsi_next (&gsi
))
2910 gimple
*stmt
= gsi_stmt (gsi
);
2911 tree forcedname
= gimple_get_lhs (stmt
);
2914 if (forcedname
!= folded
)
2916 VN_INFO_GET (forcedname
)->valnum
= forcedname
;
2917 VN_INFO (forcedname
)->value_id
= get_next_value_id ();
2918 nameexpr
= get_or_alloc_expr_for_name (forcedname
);
2919 add_to_value (VN_INFO (forcedname
)->value_id
, nameexpr
);
2920 bitmap_value_replace_in_set (NEW_SETS (block
), nameexpr
);
2921 bitmap_value_replace_in_set (AVAIL_OUT (block
), nameexpr
);
2924 bitmap_set_bit (inserted_exprs
, SSA_NAME_VERSION (forcedname
));
2926 gimple_seq_add_seq (stmts
, forced_stmts
);
2931 /* Fold the last statement. */
2932 gsi
= gsi_last (*stmts
);
2933 if (fold_stmt_inplace (&gsi
))
2934 update_stmt (gsi_stmt (gsi
));
2936 /* Add a value number to the temporary.
2937 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2938 we are creating the expression by pieces, and this particular piece of
2939 the expression may have been represented. There is no harm in replacing
2941 value_id
= get_expr_value_id (expr
);
2942 VN_INFO_GET (name
)->value_id
= value_id
;
2943 VN_INFO (name
)->valnum
= sccvn_valnum_from_value_id (value_id
);
2944 if (VN_INFO (name
)->valnum
== NULL_TREE
)
2945 VN_INFO (name
)->valnum
= name
;
2946 gcc_assert (VN_INFO (name
)->valnum
!= NULL_TREE
);
2947 nameexpr
= get_or_alloc_expr_for_name (name
);
2948 add_to_value (value_id
, nameexpr
);
2949 if (NEW_SETS (block
))
2950 bitmap_value_replace_in_set (NEW_SETS (block
), nameexpr
);
2951 bitmap_value_replace_in_set (AVAIL_OUT (block
), nameexpr
);
2953 pre_stats
.insertions
++;
2954 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2956 fprintf (dump_file
, "Inserted ");
2957 print_gimple_stmt (dump_file
, gsi_stmt (gsi_last (*stmts
)), 0);
2958 fprintf (dump_file
, " in predecessor %d (%04d)\n",
2959 block
->index
, value_id
);
2966 /* Insert the to-be-made-available values of expression EXPRNUM for each
2967 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
2968 merge the result with a phi node, given the same value number as
2969 NODE. Return true if we have inserted new stuff. */
2972 insert_into_preds_of_block (basic_block block
, unsigned int exprnum
,
2973 vec
<pre_expr
> avail
)
2975 pre_expr expr
= expression_for_id (exprnum
);
2977 unsigned int val
= get_expr_value_id (expr
);
2979 bool insertions
= false;
2984 tree type
= get_expr_type (expr
);
2988 /* Make sure we aren't creating an induction variable. */
2989 if (bb_loop_depth (block
) > 0 && EDGE_COUNT (block
->preds
) == 2)
2991 bool firstinsideloop
= false;
2992 bool secondinsideloop
= false;
2993 firstinsideloop
= flow_bb_inside_loop_p (block
->loop_father
,
2994 EDGE_PRED (block
, 0)->src
);
2995 secondinsideloop
= flow_bb_inside_loop_p (block
->loop_father
,
2996 EDGE_PRED (block
, 1)->src
);
2997 /* Induction variables only have one edge inside the loop. */
2998 if ((firstinsideloop
^ secondinsideloop
)
2999 && expr
->kind
!= REFERENCE
)
3001 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3002 fprintf (dump_file
, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3007 /* Make the necessary insertions. */
3008 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3010 gimple_seq stmts
= NULL
;
3013 eprime
= avail
[pred
->dest_idx
];
3014 builtexpr
= create_expression_by_pieces (bprime
, eprime
,
3016 gcc_assert (!(pred
->flags
& EDGE_ABNORMAL
));
3017 if (!gimple_seq_empty_p (stmts
))
3019 basic_block new_bb
= gsi_insert_seq_on_edge_immediate (pred
, stmts
);
3020 gcc_assert (! new_bb
);
3025 /* We cannot insert a PHI node if we failed to insert
3030 if (is_gimple_min_invariant (builtexpr
))
3031 avail
[pred
->dest_idx
] = get_or_alloc_expr_for_constant (builtexpr
);
3033 avail
[pred
->dest_idx
] = get_or_alloc_expr_for_name (builtexpr
);
3035 /* If we didn't want a phi node, and we made insertions, we still have
3036 inserted new stuff, and thus return true. If we didn't want a phi node,
3037 and didn't make insertions, we haven't added anything new, so return
3039 if (nophi
&& insertions
)
3041 else if (nophi
&& !insertions
)
3044 /* Now build a phi for the new variable. */
3045 temp
= make_temp_ssa_name (type
, NULL
, "prephitmp");
3046 phi
= create_phi_node (temp
, block
);
3048 VN_INFO_GET (temp
)->value_id
= val
;
3049 VN_INFO (temp
)->valnum
= sccvn_valnum_from_value_id (val
);
3050 if (VN_INFO (temp
)->valnum
== NULL_TREE
)
3051 VN_INFO (temp
)->valnum
= temp
;
3052 bitmap_set_bit (inserted_exprs
, SSA_NAME_VERSION (temp
));
3053 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3055 pre_expr ae
= avail
[pred
->dest_idx
];
3056 gcc_assert (get_expr_type (ae
) == type
3057 || useless_type_conversion_p (type
, get_expr_type (ae
)));
3058 if (ae
->kind
== CONSTANT
)
3059 add_phi_arg (phi
, unshare_expr (PRE_EXPR_CONSTANT (ae
)),
3060 pred
, UNKNOWN_LOCATION
);
3062 add_phi_arg (phi
, PRE_EXPR_NAME (ae
), pred
, UNKNOWN_LOCATION
);
3065 newphi
= get_or_alloc_expr_for_name (temp
);
3066 add_to_value (val
, newphi
);
3068 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3069 this insertion, since we test for the existence of this value in PHI_GEN
3070 before proceeding with the partial redundancy checks in insert_aux.
3072 The value may exist in AVAIL_OUT, in particular, it could be represented
3073 by the expression we are trying to eliminate, in which case we want the
3074 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3077 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3078 this block, because if it did, it would have existed in our dominator's
3079 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3082 bitmap_insert_into_set (PHI_GEN (block
), newphi
);
3083 bitmap_value_replace_in_set (AVAIL_OUT (block
),
3085 bitmap_insert_into_set (NEW_SETS (block
),
3088 /* If we insert a PHI node for a conversion of another PHI node
3089 in the same basic-block try to preserve range information.
3090 This is important so that followup loop passes receive optimal
3091 number of iteration analysis results. See PR61743. */
3092 if (expr
->kind
== NARY
3093 && CONVERT_EXPR_CODE_P (expr
->u
.nary
->opcode
)
3094 && TREE_CODE (expr
->u
.nary
->op
[0]) == SSA_NAME
3095 && gimple_bb (SSA_NAME_DEF_STMT (expr
->u
.nary
->op
[0])) == block
3096 && INTEGRAL_TYPE_P (type
)
3097 && INTEGRAL_TYPE_P (TREE_TYPE (expr
->u
.nary
->op
[0]))
3098 && (TYPE_PRECISION (type
)
3099 >= TYPE_PRECISION (TREE_TYPE (expr
->u
.nary
->op
[0])))
3100 && SSA_NAME_RANGE_INFO (expr
->u
.nary
->op
[0]))
3103 if (get_range_info (expr
->u
.nary
->op
[0], &min
, &max
) == VR_RANGE
3104 && !wi::neg_p (min
, SIGNED
)
3105 && !wi::neg_p (max
, SIGNED
))
3106 /* Just handle extension and sign-changes of all-positive ranges. */
3107 set_range_info (temp
,
3108 SSA_NAME_RANGE_TYPE (expr
->u
.nary
->op
[0]),
3109 wide_int_storage::from (min
, TYPE_PRECISION (type
),
3111 wide_int_storage::from (max
, TYPE_PRECISION (type
),
3115 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3117 fprintf (dump_file
, "Created phi ");
3118 print_gimple_stmt (dump_file
, phi
, 0);
3119 fprintf (dump_file
, " in block %d (%04d)\n", block
->index
, val
);
3127 /* Perform insertion of partially redundant or hoistable values.
3128 For BLOCK, do the following:
3129 1. Propagate the NEW_SETS of the dominator into the current block.
3130 If the block has multiple predecessors,
3131 2a. Iterate over the ANTIC expressions for the block to see if
3132 any of them are partially redundant.
3133 2b. If so, insert them into the necessary predecessors to make
3134 the expression fully redundant.
3135 2c. Insert a new PHI merging the values of the predecessors.
3136 2d. Insert the new PHI, and the new expressions, into the
3138 If the block has multiple successors,
3139 3a. Iterate over the ANTIC values for the block to see if
3140 any of them are good candidates for hoisting.
3141 3b. If so, insert expressions computing the values in BLOCK,
3142 and add the new expressions into the NEW_SETS set.
3143 4. Recursively call ourselves on the dominator children of BLOCK.
3145 Steps 1, 2a, and 4 are done by insert_aux. 2b, 2c and 2d are done by
3146 do_pre_regular_insertion and do_partial_insertion. 3a and 3b are
3147 done in do_hoist_insertion.
3151 do_pre_regular_insertion (basic_block block
, basic_block dom
)
3153 bool new_stuff
= false;
3154 vec
<pre_expr
> exprs
;
3156 auto_vec
<pre_expr
> avail
;
3159 exprs
= sorted_array_from_bitmap_set (ANTIC_IN (block
));
3160 avail
.safe_grow (EDGE_COUNT (block
->preds
));
3162 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
3164 if (expr
->kind
== NARY
3165 || expr
->kind
== REFERENCE
)
3168 bool by_some
= false;
3169 bool cant_insert
= false;
3170 bool all_same
= true;
3171 pre_expr first_s
= NULL
;
3174 pre_expr eprime
= NULL
;
3176 pre_expr edoubleprime
= NULL
;
3177 bool do_insertion
= false;
3179 val
= get_expr_value_id (expr
);
3180 if (bitmap_set_contains_value (PHI_GEN (block
), val
))
3182 if (bitmap_set_contains_value (AVAIL_OUT (dom
), val
))
3184 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3186 fprintf (dump_file
, "Found fully redundant value: ");
3187 print_pre_expr (dump_file
, expr
);
3188 fprintf (dump_file
, "\n");
3193 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3195 unsigned int vprime
;
3197 /* We should never run insertion for the exit block
3198 and so not come across fake pred edges. */
3199 gcc_assert (!(pred
->flags
& EDGE_FAKE
));
3201 /* We are looking at ANTIC_OUT of bprime. */
3202 eprime
= phi_translate (expr
, ANTIC_IN (block
), NULL
, pred
);
3204 /* eprime will generally only be NULL if the
3205 value of the expression, translated
3206 through the PHI for this predecessor, is
3207 undefined. If that is the case, we can't
3208 make the expression fully redundant,
3209 because its value is undefined along a
3210 predecessor path. We can thus break out
3211 early because it doesn't matter what the
3212 rest of the results are. */
3215 avail
[pred
->dest_idx
] = NULL
;
3220 vprime
= get_expr_value_id (eprime
);
3221 edoubleprime
= bitmap_find_leader (AVAIL_OUT (bprime
),
3223 if (edoubleprime
== NULL
)
3225 avail
[pred
->dest_idx
] = eprime
;
3230 avail
[pred
->dest_idx
] = edoubleprime
;
3232 /* We want to perform insertions to remove a redundancy on
3233 a path in the CFG we want to optimize for speed. */
3234 if (optimize_edge_for_speed_p (pred
))
3235 do_insertion
= true;
3236 if (first_s
== NULL
)
3237 first_s
= edoubleprime
;
3238 else if (!pre_expr_d::equal (first_s
, edoubleprime
))
3242 /* If we can insert it, it's not the same value
3243 already existing along every predecessor, and
3244 it's defined by some predecessor, it is
3245 partially redundant. */
3246 if (!cant_insert
&& !all_same
&& by_some
)
3250 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3252 fprintf (dump_file
, "Skipping partial redundancy for "
3254 print_pre_expr (dump_file
, expr
);
3255 fprintf (dump_file
, " (%04d), no redundancy on to be "
3256 "optimized for speed edge\n", val
);
3259 else if (dbg_cnt (treepre_insert
))
3261 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3263 fprintf (dump_file
, "Found partial redundancy for "
3265 print_pre_expr (dump_file
, expr
);
3266 fprintf (dump_file
, " (%04d)\n",
3267 get_expr_value_id (expr
));
3269 if (insert_into_preds_of_block (block
,
3270 get_expression_id (expr
),
3275 /* If all edges produce the same value and that value is
3276 an invariant, then the PHI has the same value on all
3277 edges. Note this. */
3278 else if (!cant_insert
&& all_same
)
3280 gcc_assert (edoubleprime
->kind
== CONSTANT
3281 || edoubleprime
->kind
== NAME
);
3283 tree temp
= make_temp_ssa_name (get_expr_type (expr
),
3286 = gimple_build_assign (temp
,
3287 edoubleprime
->kind
== CONSTANT
?
3288 PRE_EXPR_CONSTANT (edoubleprime
) :
3289 PRE_EXPR_NAME (edoubleprime
));
3290 gimple_stmt_iterator gsi
= gsi_after_labels (block
);
3291 gsi_insert_before (&gsi
, assign
, GSI_NEW_STMT
);
3293 VN_INFO_GET (temp
)->value_id
= val
;
3294 VN_INFO (temp
)->valnum
= sccvn_valnum_from_value_id (val
);
3295 if (VN_INFO (temp
)->valnum
== NULL_TREE
)
3296 VN_INFO (temp
)->valnum
= temp
;
3297 bitmap_set_bit (inserted_exprs
, SSA_NAME_VERSION (temp
));
3298 pre_expr newe
= get_or_alloc_expr_for_name (temp
);
3299 add_to_value (val
, newe
);
3300 bitmap_value_replace_in_set (AVAIL_OUT (block
), newe
);
3301 bitmap_insert_into_set (NEW_SETS (block
), newe
);
3311 /* Perform insertion for partially anticipatable expressions. There
3312 is only one case we will perform insertion for these. This case is
3313 if the expression is partially anticipatable, and fully available.
3314 In this case, we know that putting it earlier will enable us to
3315 remove the later computation. */
3318 do_pre_partial_partial_insertion (basic_block block
, basic_block dom
)
3320 bool new_stuff
= false;
3321 vec
<pre_expr
> exprs
;
3323 auto_vec
<pre_expr
> avail
;
3326 exprs
= sorted_array_from_bitmap_set (PA_IN (block
));
3327 avail
.safe_grow (EDGE_COUNT (block
->preds
));
3329 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
3331 if (expr
->kind
== NARY
3332 || expr
->kind
== REFERENCE
)
3336 bool cant_insert
= false;
3339 pre_expr eprime
= NULL
;
3342 val
= get_expr_value_id (expr
);
3343 if (bitmap_set_contains_value (PHI_GEN (block
), val
))
3345 if (bitmap_set_contains_value (AVAIL_OUT (dom
), val
))
3348 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3350 unsigned int vprime
;
3351 pre_expr edoubleprime
;
3353 /* We should never run insertion for the exit block
3354 and so not come across fake pred edges. */
3355 gcc_assert (!(pred
->flags
& EDGE_FAKE
));
3357 eprime
= phi_translate (expr
, ANTIC_IN (block
),
3358 PA_IN (block
), pred
);
3360 /* eprime will generally only be NULL if the
3361 value of the expression, translated
3362 through the PHI for this predecessor, is
3363 undefined. If that is the case, we can't
3364 make the expression fully redundant,
3365 because its value is undefined along a
3366 predecessor path. We can thus break out
3367 early because it doesn't matter what the
3368 rest of the results are. */
3371 avail
[pred
->dest_idx
] = NULL
;
3376 vprime
= get_expr_value_id (eprime
);
3377 edoubleprime
= bitmap_find_leader (AVAIL_OUT (bprime
), vprime
);
3378 avail
[pred
->dest_idx
] = edoubleprime
;
3379 if (edoubleprime
== NULL
)
3386 /* If we can insert it, it's not the same value
3387 already existing along every predecessor, and
3388 it's defined by some predecessor, it is
3389 partially redundant. */
3390 if (!cant_insert
&& by_all
)
3393 bool do_insertion
= false;
3395 /* Insert only if we can remove a later expression on a path
3396 that we want to optimize for speed.
3397 The phi node that we will be inserting in BLOCK is not free,
3398 and inserting it for the sake of !optimize_for_speed successor
3399 may cause regressions on the speed path. */
3400 FOR_EACH_EDGE (succ
, ei
, block
->succs
)
3402 if (bitmap_set_contains_value (PA_IN (succ
->dest
), val
)
3403 || bitmap_set_contains_value (ANTIC_IN (succ
->dest
), val
))
3405 if (optimize_edge_for_speed_p (succ
))
3406 do_insertion
= true;
3412 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3414 fprintf (dump_file
, "Skipping partial partial redundancy "
3416 print_pre_expr (dump_file
, expr
);
3417 fprintf (dump_file
, " (%04d), not (partially) anticipated "
3418 "on any to be optimized for speed edges\n", val
);
3421 else if (dbg_cnt (treepre_insert
))
3423 pre_stats
.pa_insert
++;
3424 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3426 fprintf (dump_file
, "Found partial partial redundancy "
3428 print_pre_expr (dump_file
, expr
);
3429 fprintf (dump_file
, " (%04d)\n",
3430 get_expr_value_id (expr
));
3432 if (insert_into_preds_of_block (block
,
3433 get_expression_id (expr
),
3445 /* Insert expressions in BLOCK to compute hoistable values up.
3446 Return TRUE if something was inserted, otherwise return FALSE.
3447 The caller has to make sure that BLOCK has at least two successors. */
3450 do_hoist_insertion (basic_block block
)
3454 bool new_stuff
= false;
3456 gimple_stmt_iterator last
;
3458 /* At least two successors, or else... */
3459 gcc_assert (EDGE_COUNT (block
->succs
) >= 2);
3461 /* Check that all successors of BLOCK are dominated by block.
3462 We could use dominated_by_p() for this, but actually there is a much
3463 quicker check: any successor that is dominated by BLOCK can't have
3464 more than one predecessor edge. */
3465 FOR_EACH_EDGE (e
, ei
, block
->succs
)
3466 if (! single_pred_p (e
->dest
))
3469 /* Determine the insertion point. If we cannot safely insert before
3470 the last stmt if we'd have to, bail out. */
3471 last
= gsi_last_bb (block
);
3472 if (!gsi_end_p (last
)
3473 && !is_ctrl_stmt (gsi_stmt (last
))
3474 && stmt_ends_bb_p (gsi_stmt (last
)))
3477 /* Compute the set of hoistable expressions from ANTIC_IN. First compute
3478 hoistable values. */
3479 bitmap_set hoistable_set
;
3481 /* A hoistable value must be in ANTIC_IN(block)
3482 but not in AVAIL_OUT(BLOCK). */
3483 bitmap_initialize (&hoistable_set
.values
, &grand_bitmap_obstack
);
3484 bitmap_and_compl (&hoistable_set
.values
,
3485 &ANTIC_IN (block
)->values
, &AVAIL_OUT (block
)->values
);
3487 /* Short-cut for a common case: hoistable_set is empty. */
3488 if (bitmap_empty_p (&hoistable_set
.values
))
3491 /* Compute which of the hoistable values is in AVAIL_OUT of
3492 at least one of the successors of BLOCK. */
3493 bitmap_head availout_in_some
;
3494 bitmap_initialize (&availout_in_some
, &grand_bitmap_obstack
);
3495 FOR_EACH_EDGE (e
, ei
, block
->succs
)
3496 /* Do not consider expressions solely because their availability
3497 on loop exits. They'd be ANTIC-IN throughout the whole loop
3498 and thus effectively hoisted across loops by combination of
3499 PRE and hoisting. */
3500 if (! loop_exit_edge_p (block
->loop_father
, e
))
3501 bitmap_ior_and_into (&availout_in_some
, &hoistable_set
.values
,
3502 &AVAIL_OUT (e
->dest
)->values
);
3503 bitmap_clear (&hoistable_set
.values
);
3505 /* Short-cut for a common case: availout_in_some is empty. */
3506 if (bitmap_empty_p (&availout_in_some
))
3509 /* Hack hoitable_set in-place so we can use sorted_array_from_bitmap_set. */
3510 hoistable_set
.values
= availout_in_some
;
3511 hoistable_set
.expressions
= ANTIC_IN (block
)->expressions
;
3513 /* Now finally construct the topological-ordered expression set. */
3514 vec
<pre_expr
> exprs
= sorted_array_from_bitmap_set (&hoistable_set
);
3516 bitmap_clear (&hoistable_set
.values
);
3518 /* If there are candidate values for hoisting, insert expressions
3519 strategically to make the hoistable expressions fully redundant. */
3521 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
3523 /* While we try to sort expressions topologically above the
3524 sorting doesn't work out perfectly. Catch expressions we
3525 already inserted. */
3526 unsigned int value_id
= get_expr_value_id (expr
);
3527 if (bitmap_set_contains_value (AVAIL_OUT (block
), value_id
))
3529 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3532 "Already inserted expression for ");
3533 print_pre_expr (dump_file
, expr
);
3534 fprintf (dump_file
, " (%04d)\n", value_id
);
3539 /* OK, we should hoist this value. Perform the transformation. */
3540 pre_stats
.hoist_insert
++;
3541 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3544 "Inserting expression in block %d for code hoisting: ",
3546 print_pre_expr (dump_file
, expr
);
3547 fprintf (dump_file
, " (%04d)\n", value_id
);
3550 gimple_seq stmts
= NULL
;
3551 tree res
= create_expression_by_pieces (block
, expr
, &stmts
,
3552 get_expr_type (expr
));
3554 /* Do not return true if expression creation ultimately
3555 did not insert any statements. */
3556 if (gimple_seq_empty_p (stmts
))
3560 if (gsi_end_p (last
) || is_ctrl_stmt (gsi_stmt (last
)))
3561 gsi_insert_seq_before (&last
, stmts
, GSI_SAME_STMT
);
3563 gsi_insert_seq_after (&last
, stmts
, GSI_NEW_STMT
);
3566 /* Make sure to not return true if expression creation ultimately
3567 failed but also make sure to insert any stmts produced as they
3568 are tracked in inserted_exprs. */
3580 /* Do a dominator walk on the control flow graph, and insert computations
3581 of values as necessary for PRE and hoisting. */
3584 insert_aux (basic_block block
, bool do_pre
, bool do_hoist
)
3587 bool new_stuff
= false;
3592 dom
= get_immediate_dominator (CDI_DOMINATORS
, block
);
3597 bitmap_set_t newset
;
3599 /* First, update the AVAIL_OUT set with anything we may have
3600 inserted higher up in the dominator tree. */
3601 newset
= NEW_SETS (dom
);
3604 /* Note that we need to value_replace both NEW_SETS, and
3605 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3606 represented by some non-simple expression here that we want
3607 to replace it with. */
3608 FOR_EACH_EXPR_ID_IN_SET (newset
, i
, bi
)
3610 pre_expr expr
= expression_for_id (i
);
3611 bitmap_value_replace_in_set (NEW_SETS (block
), expr
);
3612 bitmap_value_replace_in_set (AVAIL_OUT (block
), expr
);
3616 /* Insert expressions for partial redundancies. */
3617 if (do_pre
&& !single_pred_p (block
))
3619 new_stuff
|= do_pre_regular_insertion (block
, dom
);
3620 if (do_partial_partial
)
3621 new_stuff
|= do_pre_partial_partial_insertion (block
, dom
);
3624 /* Insert expressions for hoisting. */
3625 if (do_hoist
&& EDGE_COUNT (block
->succs
) >= 2)
3626 new_stuff
|= do_hoist_insertion (block
);
3629 for (son
= first_dom_son (CDI_DOMINATORS
, block
);
3631 son
= next_dom_son (CDI_DOMINATORS
, son
))
3633 new_stuff
|= insert_aux (son
, do_pre
, do_hoist
);
3639 /* Perform insertion of partially redundant and hoistable values. */
3644 bool new_stuff
= true;
3646 int num_iterations
= 0;
3648 FOR_ALL_BB_FN (bb
, cfun
)
3649 NEW_SETS (bb
) = bitmap_set_new ();
3654 if (dump_file
&& dump_flags
& TDF_DETAILS
)
3655 fprintf (dump_file
, "Starting insert iteration %d\n", num_iterations
);
3656 new_stuff
= insert_aux (ENTRY_BLOCK_PTR_FOR_FN (cfun
), flag_tree_pre
,
3657 flag_code_hoisting
);
3659 /* Clear the NEW sets before the next iteration. We have already
3660 fully propagated its contents. */
3662 FOR_ALL_BB_FN (bb
, cfun
)
3663 bitmap_set_free (NEW_SETS (bb
));
3665 statistics_histogram_event (cfun
, "insert iterations", num_iterations
);
3669 /* Compute the AVAIL set for all basic blocks.
3671 This function performs value numbering of the statements in each basic
3672 block. The AVAIL sets are built from information we glean while doing
3673 this value numbering, since the AVAIL sets contain only one entry per
3676 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3677 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3680 compute_avail (void)
3683 basic_block block
, son
;
3684 basic_block
*worklist
;
3689 /* We pretend that default definitions are defined in the entry block.
3690 This includes function arguments and the static chain decl. */
3691 FOR_EACH_SSA_NAME (i
, name
, cfun
)
3694 if (!SSA_NAME_IS_DEFAULT_DEF (name
)
3695 || has_zero_uses (name
)
3696 || virtual_operand_p (name
))
3699 e
= get_or_alloc_expr_for_name (name
);
3700 add_to_value (get_expr_value_id (e
), e
);
3701 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun
)), e
);
3702 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun
)),
3706 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3708 print_bitmap_set (dump_file
, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun
)),
3709 "tmp_gen", ENTRY_BLOCK
);
3710 print_bitmap_set (dump_file
, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun
)),
3711 "avail_out", ENTRY_BLOCK
);
3714 /* Allocate the worklist. */
3715 worklist
= XNEWVEC (basic_block
, n_basic_blocks_for_fn (cfun
));
3717 /* Seed the algorithm by putting the dominator children of the entry
3718 block on the worklist. */
3719 for (son
= first_dom_son (CDI_DOMINATORS
, ENTRY_BLOCK_PTR_FOR_FN (cfun
));
3721 son
= next_dom_son (CDI_DOMINATORS
, son
))
3722 worklist
[sp
++] = son
;
3724 BB_LIVE_VOP_ON_EXIT (ENTRY_BLOCK_PTR_FOR_FN (cfun
))
3725 = ssa_default_def (cfun
, gimple_vop (cfun
));
3727 /* Loop until the worklist is empty. */
3733 /* Pick a block from the worklist. */
3734 block
= worklist
[--sp
];
3736 /* Initially, the set of available values in BLOCK is that of
3737 its immediate dominator. */
3738 dom
= get_immediate_dominator (CDI_DOMINATORS
, block
);
3741 bitmap_set_copy (AVAIL_OUT (block
), AVAIL_OUT (dom
));
3742 BB_LIVE_VOP_ON_EXIT (block
) = BB_LIVE_VOP_ON_EXIT (dom
);
3745 /* Generate values for PHI nodes. */
3746 for (gphi_iterator gsi
= gsi_start_phis (block
); !gsi_end_p (gsi
);
3749 tree result
= gimple_phi_result (gsi
.phi ());
3751 /* We have no need for virtual phis, as they don't represent
3752 actual computations. */
3753 if (virtual_operand_p (result
))
3755 BB_LIVE_VOP_ON_EXIT (block
) = result
;
3759 pre_expr e
= get_or_alloc_expr_for_name (result
);
3760 add_to_value (get_expr_value_id (e
), e
);
3761 bitmap_value_insert_into_set (AVAIL_OUT (block
), e
);
3762 bitmap_insert_into_set (PHI_GEN (block
), e
);
3765 BB_MAY_NOTRETURN (block
) = 0;
3767 /* Now compute value numbers and populate value sets with all
3768 the expressions computed in BLOCK. */
3769 for (gimple_stmt_iterator gsi
= gsi_start_bb (block
); !gsi_end_p (gsi
);
3775 stmt
= gsi_stmt (gsi
);
3777 /* Cache whether the basic-block has any non-visible side-effect
3779 If this isn't a call or it is the last stmt in the
3780 basic-block then the CFG represents things correctly. */
3781 if (is_gimple_call (stmt
) && !stmt_ends_bb_p (stmt
))
3783 /* Non-looping const functions always return normally.
3784 Otherwise the call might not return or have side-effects
3785 that forbids hoisting possibly trapping expressions
3787 int flags
= gimple_call_flags (stmt
);
3788 if (!(flags
& ECF_CONST
)
3789 || (flags
& ECF_LOOPING_CONST_OR_PURE
))
3790 BB_MAY_NOTRETURN (block
) = 1;
3793 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_DEF
)
3795 pre_expr e
= get_or_alloc_expr_for_name (op
);
3797 add_to_value (get_expr_value_id (e
), e
);
3798 bitmap_insert_into_set (TMP_GEN (block
), e
);
3799 bitmap_value_insert_into_set (AVAIL_OUT (block
), e
);
3802 if (gimple_vdef (stmt
))
3803 BB_LIVE_VOP_ON_EXIT (block
) = gimple_vdef (stmt
);
3805 if (gimple_has_side_effects (stmt
)
3806 || stmt_could_throw_p (stmt
)
3807 || is_gimple_debug (stmt
))
3810 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_USE
)
3812 if (ssa_undefined_value_p (op
))
3814 pre_expr e
= get_or_alloc_expr_for_name (op
);
3815 bitmap_value_insert_into_set (EXP_GEN (block
), e
);
3818 switch (gimple_code (stmt
))
3826 vn_reference_s ref1
;
3827 pre_expr result
= NULL
;
3829 /* We can value number only calls to real functions. */
3830 if (gimple_call_internal_p (stmt
))
3833 vn_reference_lookup_call (as_a
<gcall
*> (stmt
), &ref
, &ref1
);
3837 /* If the value of the call is not invalidated in
3838 this block until it is computed, add the expression
3840 if (!gimple_vuse (stmt
)
3842 (SSA_NAME_DEF_STMT (gimple_vuse (stmt
))) == GIMPLE_PHI
3843 || gimple_bb (SSA_NAME_DEF_STMT
3844 (gimple_vuse (stmt
))) != block
)
3846 result
= pre_expr_pool
.allocate ();
3847 result
->kind
= REFERENCE
;
3849 PRE_EXPR_REFERENCE (result
) = ref
;
3851 get_or_alloc_expression_id (result
);
3852 add_to_value (get_expr_value_id (result
), result
);
3853 bitmap_value_insert_into_set (EXP_GEN (block
), result
);
3860 pre_expr result
= NULL
;
3861 switch (vn_get_stmt_kind (stmt
))
3865 enum tree_code code
= gimple_assign_rhs_code (stmt
);
3868 /* COND_EXPR and VEC_COND_EXPR are awkward in
3869 that they contain an embedded complex expression.
3870 Don't even try to shove those through PRE. */
3871 if (code
== COND_EXPR
3872 || code
== VEC_COND_EXPR
)
3875 vn_nary_op_lookup_stmt (stmt
, &nary
);
3879 /* If the NARY traps and there was a preceding
3880 point in the block that might not return avoid
3881 adding the nary to EXP_GEN. */
3882 if (BB_MAY_NOTRETURN (block
)
3883 && vn_nary_may_trap (nary
))
3886 result
= pre_expr_pool
.allocate ();
3887 result
->kind
= NARY
;
3889 PRE_EXPR_NARY (result
) = nary
;
3895 tree rhs1
= gimple_assign_rhs1 (stmt
);
3896 alias_set_type set
= get_alias_set (rhs1
);
3897 vec
<vn_reference_op_s
> operands
3898 = vn_reference_operands_for_lookup (rhs1
);
3900 vn_reference_lookup_pieces (gimple_vuse (stmt
), set
,
3902 operands
, &ref
, VN_WALK
);
3905 operands
.release ();
3909 /* If the value of the reference is not invalidated in
3910 this block until it is computed, add the expression
3912 if (gimple_vuse (stmt
))
3916 def_stmt
= SSA_NAME_DEF_STMT (gimple_vuse (stmt
));
3917 while (!gimple_nop_p (def_stmt
)
3918 && gimple_code (def_stmt
) != GIMPLE_PHI
3919 && gimple_bb (def_stmt
) == block
)
3921 if (stmt_may_clobber_ref_p
3922 (def_stmt
, gimple_assign_rhs1 (stmt
)))
3928 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt
));
3932 operands
.release ();
3937 /* If the load was value-numbered to another
3938 load make sure we do not use its expression
3939 for insertion if it wouldn't be a valid
3941 /* At the momemt we have a testcase
3942 for hoist insertion of aligned vs. misaligned
3943 variants in gcc.dg/torture/pr65270-1.c thus
3944 with just alignment to be considered we can
3945 simply replace the expression in the hashtable
3946 with the most conservative one. */
3947 vn_reference_op_t ref1
= &ref
->operands
.last ();
3948 while (ref1
->opcode
!= TARGET_MEM_REF
3949 && ref1
->opcode
!= MEM_REF
3950 && ref1
!= &ref
->operands
[0])
3952 vn_reference_op_t ref2
= &operands
.last ();
3953 while (ref2
->opcode
!= TARGET_MEM_REF
3954 && ref2
->opcode
!= MEM_REF
3955 && ref2
!= &operands
[0])
3957 if ((ref1
->opcode
== TARGET_MEM_REF
3958 || ref1
->opcode
== MEM_REF
)
3959 && (TYPE_ALIGN (ref1
->type
)
3960 > TYPE_ALIGN (ref2
->type
)))
3962 = build_aligned_type (ref1
->type
,
3963 TYPE_ALIGN (ref2
->type
));
3964 /* TBAA behavior is an obvious part so make sure
3965 that the hashtable one covers this as well
3966 by adjusting the ref alias set and its base. */
3968 || alias_set_subset_of (set
, ref
->set
))
3970 else if (alias_set_subset_of (ref
->set
, set
))
3973 if (ref1
->opcode
== MEM_REF
)
3975 = wide_int_to_tree (TREE_TYPE (ref2
->op0
),
3976 wi::to_wide (ref1
->op0
));
3979 = wide_int_to_tree (TREE_TYPE (ref2
->op2
),
3980 wi::to_wide (ref1
->op2
));
3985 if (ref1
->opcode
== MEM_REF
)
3987 = wide_int_to_tree (ptr_type_node
,
3988 wi::to_wide (ref1
->op0
));
3991 = wide_int_to_tree (ptr_type_node
,
3992 wi::to_wide (ref1
->op2
));
3994 operands
.release ();
3996 result
= pre_expr_pool
.allocate ();
3997 result
->kind
= REFERENCE
;
3999 PRE_EXPR_REFERENCE (result
) = ref
;
4007 get_or_alloc_expression_id (result
);
4008 add_to_value (get_expr_value_id (result
), result
);
4009 bitmap_value_insert_into_set (EXP_GEN (block
), result
);
4017 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4019 print_bitmap_set (dump_file
, EXP_GEN (block
),
4020 "exp_gen", block
->index
);
4021 print_bitmap_set (dump_file
, PHI_GEN (block
),
4022 "phi_gen", block
->index
);
4023 print_bitmap_set (dump_file
, TMP_GEN (block
),
4024 "tmp_gen", block
->index
);
4025 print_bitmap_set (dump_file
, AVAIL_OUT (block
),
4026 "avail_out", block
->index
);
4029 /* Put the dominator children of BLOCK on the worklist of blocks
4030 to compute available sets for. */
4031 for (son
= first_dom_son (CDI_DOMINATORS
, block
);
4033 son
= next_dom_son (CDI_DOMINATORS
, son
))
4034 worklist
[sp
++] = son
;
4041 /* Initialize data structures used by PRE. */
4048 next_expression_id
= 1;
4049 expressions
.create (0);
4050 expressions
.safe_push (NULL
);
4051 value_expressions
.create (get_max_value_id () + 1);
4052 value_expressions
.safe_grow_cleared (get_max_value_id () + 1);
4053 name_to_id
.create (0);
4055 inserted_exprs
= BITMAP_ALLOC (NULL
);
4057 connect_infinite_loops_to_exit ();
4058 memset (&pre_stats
, 0, sizeof (pre_stats
));
4060 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets
));
4062 calculate_dominance_info (CDI_DOMINATORS
);
4064 bitmap_obstack_initialize (&grand_bitmap_obstack
);
4065 phi_translate_table
= new hash_table
<expr_pred_trans_d
> (5110);
4066 expression_to_id
= new hash_table
<pre_expr_d
> (num_ssa_names
* 3);
4067 FOR_ALL_BB_FN (bb
, cfun
)
4069 EXP_GEN (bb
) = bitmap_set_new ();
4070 PHI_GEN (bb
) = bitmap_set_new ();
4071 TMP_GEN (bb
) = bitmap_set_new ();
4072 AVAIL_OUT (bb
) = bitmap_set_new ();
4077 /* Deallocate data structures used by PRE. */
4082 value_expressions
.release ();
4083 expressions
.release ();
4084 BITMAP_FREE (inserted_exprs
);
4085 bitmap_obstack_release (&grand_bitmap_obstack
);
4086 bitmap_set_pool
.release ();
4087 pre_expr_pool
.release ();
4088 delete phi_translate_table
;
4089 phi_translate_table
= NULL
;
4090 delete expression_to_id
;
4091 expression_to_id
= NULL
;
4092 name_to_id
.release ();
4094 free_aux_for_blocks ();
4099 const pass_data pass_data_pre
=
4101 GIMPLE_PASS
, /* type */
4103 OPTGROUP_NONE
, /* optinfo_flags */
4104 TV_TREE_PRE
, /* tv_id */
4105 ( PROP_cfg
| PROP_ssa
), /* properties_required */
4106 0, /* properties_provided */
4107 0, /* properties_destroyed */
4108 TODO_rebuild_alias
, /* todo_flags_start */
4109 0, /* todo_flags_finish */
4112 class pass_pre
: public gimple_opt_pass
4115 pass_pre (gcc::context
*ctxt
)
4116 : gimple_opt_pass (pass_data_pre
, ctxt
)
4119 /* opt_pass methods: */
4120 virtual bool gate (function
*)
4121 { return flag_tree_pre
!= 0 || flag_code_hoisting
!= 0; }
4122 virtual unsigned int execute (function
*);
4124 }; // class pass_pre
4127 pass_pre::execute (function
*fun
)
4129 unsigned int todo
= 0;
4131 do_partial_partial
=
4132 flag_tree_partial_pre
&& optimize_function_for_speed_p (fun
);
4134 /* This has to happen before SCCVN runs because
4135 loop_optimizer_init may create new phis, etc. */
4136 loop_optimizer_init (LOOPS_NORMAL
);
4137 split_critical_edges ();
4140 run_scc_vn (VN_WALK
);
4144 /* Insert can get quite slow on an incredibly large number of basic
4145 blocks due to some quadratic behavior. Until this behavior is
4146 fixed, don't run it when he have an incredibly large number of
4147 bb's. If we aren't going to run insert, there is no point in
4148 computing ANTIC, either, even though it's plenty fast nor do
4149 we require AVAIL. */
4150 if (n_basic_blocks_for_fn (fun
) < 4000)
4157 /* Make sure to remove fake edges before committing our inserts.
4158 This makes sure we don't end up with extra critical edges that
4159 we would need to split. */
4160 remove_fake_exit_edges ();
4161 gsi_commit_edge_inserts ();
4163 /* Eliminate folds statements which might (should not...) end up
4164 not keeping virtual operands up-to-date. */
4165 gcc_assert (!need_ssa_update_p (fun
));
4167 statistics_counter_event (fun
, "Insertions", pre_stats
.insertions
);
4168 statistics_counter_event (fun
, "PA inserted", pre_stats
.pa_insert
);
4169 statistics_counter_event (fun
, "HOIST inserted", pre_stats
.hoist_insert
);
4170 statistics_counter_event (fun
, "New PHIs", pre_stats
.phis
);
4172 /* Remove all the redundant expressions. */
4173 todo
|= vn_eliminate (inserted_exprs
);
4175 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4176 to insert PHI nodes sometimes, and because value numbering of casts isn't
4177 perfect, we sometimes end up inserting dead code. This simple DCE-like
4178 pass removes any insertions we made that weren't actually used. */
4179 simple_dce_from_worklist (inserted_exprs
);
4184 loop_optimizer_finalize ();
4186 /* Restore SSA info before tail-merging as that resets it as well. */
4187 scc_vn_restore_ssa_info ();
4189 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4190 case we can merge the block with the remaining predecessor of the block.
4192 - call merge_blocks after each tail merge iteration
4193 - call merge_blocks after all tail merge iterations
4194 - mark TODO_cleanup_cfg when necessary
4195 - share the cfg cleanup with fini_pre. */
4196 todo
|= tail_merge_optimize (todo
);
4200 /* Tail merging invalidates the virtual SSA web, together with
4201 cfg-cleanup opportunities exposed by PRE this will wreck the
4202 SSA updating machinery. So make sure to run update-ssa
4203 manually, before eventually scheduling cfg-cleanup as part of
4205 update_ssa (TODO_update_ssa_only_virtuals
);
4213 make_pass_pre (gcc::context
*ctxt
)
4215 return new pass_pre (ctxt
);