1 /* Full and partial redundancy elimination and code hoisting on SSA GIMPLE.
2 Copyright (C) 2001-2022 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-iterator.h"
38 #include "gimple-fold.h"
42 #include "tree-into-ssa.h"
46 #include "tree-ssa-sccvn.h"
47 #include "tree-scalar-evolution.h"
50 #include "tree-ssa-propagate.h"
51 #include "tree-ssa-dce.h"
52 #include "tree-cfgcleanup.h"
54 #include "gimple-range.h"
56 /* Even though this file is called tree-ssa-pre.cc, 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.cc.
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
;
264 /* hash_table support. */
265 static inline hashval_t
hash (const pre_expr_d
*);
266 static inline int equal (const pre_expr_d
*, const pre_expr_d
*);
269 #define PRE_EXPR_NAME(e) (e)->u.name
270 #define PRE_EXPR_NARY(e) (e)->u.nary
271 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
272 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
274 /* Compare E1 and E1 for equality. */
277 pre_expr_d::equal (const pre_expr_d
*e1
, const pre_expr_d
*e2
)
279 if (e1
->kind
!= e2
->kind
)
285 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1
),
286 PRE_EXPR_CONSTANT (e2
));
288 return PRE_EXPR_NAME (e1
) == PRE_EXPR_NAME (e2
);
290 return vn_nary_op_eq (PRE_EXPR_NARY (e1
), PRE_EXPR_NARY (e2
));
292 return vn_reference_eq (PRE_EXPR_REFERENCE (e1
),
293 PRE_EXPR_REFERENCE (e2
));
302 pre_expr_d::hash (const pre_expr_d
*e
)
307 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e
));
309 return SSA_NAME_VERSION (PRE_EXPR_NAME (e
));
311 return PRE_EXPR_NARY (e
)->hashcode
;
313 return PRE_EXPR_REFERENCE (e
)->hashcode
;
319 /* Next global expression id number. */
320 static unsigned int next_expression_id
;
322 /* Mapping from expression to id number we can use in bitmap sets. */
323 static vec
<pre_expr
> expressions
;
324 static hash_table
<pre_expr_d
> *expression_to_id
;
325 static vec
<unsigned> name_to_id
;
326 static obstack pre_expr_obstack
;
328 /* Allocate an expression id for EXPR. */
330 static inline unsigned int
331 alloc_expression_id (pre_expr expr
)
333 struct pre_expr_d
**slot
;
334 /* Make sure we won't overflow. */
335 gcc_assert (next_expression_id
+ 1 > next_expression_id
);
336 expr
->id
= next_expression_id
++;
337 expressions
.safe_push (expr
);
338 if (expr
->kind
== NAME
)
340 unsigned version
= SSA_NAME_VERSION (PRE_EXPR_NAME (expr
));
341 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
342 re-allocations by using vec::reserve upfront. */
343 unsigned old_len
= name_to_id
.length ();
344 name_to_id
.reserve (num_ssa_names
- old_len
);
345 name_to_id
.quick_grow_cleared (num_ssa_names
);
346 gcc_assert (name_to_id
[version
] == 0);
347 name_to_id
[version
] = expr
->id
;
351 slot
= expression_to_id
->find_slot (expr
, INSERT
);
355 return next_expression_id
- 1;
358 /* Return the expression id for tree EXPR. */
360 static inline unsigned int
361 get_expression_id (const pre_expr expr
)
366 static inline unsigned int
367 lookup_expression_id (const pre_expr expr
)
369 struct pre_expr_d
**slot
;
371 if (expr
->kind
== NAME
)
373 unsigned version
= SSA_NAME_VERSION (PRE_EXPR_NAME (expr
));
374 if (name_to_id
.length () <= version
)
376 return name_to_id
[version
];
380 slot
= expression_to_id
->find_slot (expr
, NO_INSERT
);
383 return ((pre_expr
)*slot
)->id
;
387 /* Return the expression that has expression id ID */
389 static inline pre_expr
390 expression_for_id (unsigned int id
)
392 return expressions
[id
];
395 static object_allocator
<pre_expr_d
> pre_expr_pool ("pre_expr nodes");
397 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
400 get_or_alloc_expr_for_name (tree name
)
402 struct pre_expr_d expr
;
404 unsigned int result_id
;
408 PRE_EXPR_NAME (&expr
) = name
;
409 result_id
= lookup_expression_id (&expr
);
411 return expression_for_id (result_id
);
413 result
= pre_expr_pool
.allocate ();
415 result
->loc
= UNKNOWN_LOCATION
;
416 result
->value_id
= VN_INFO (name
)->value_id
;
417 PRE_EXPR_NAME (result
) = name
;
418 alloc_expression_id (result
);
422 /* Given an NARY, get or create a pre_expr to represent it. Assign
423 VALUE_ID to it or allocate a new value-id if it is zero. Record
424 LOC as the original location of the expression. */
427 get_or_alloc_expr_for_nary (vn_nary_op_t nary
, unsigned value_id
,
428 location_t loc
= UNKNOWN_LOCATION
)
430 struct pre_expr_d expr
;
432 unsigned int result_id
;
434 gcc_assert (value_id
== 0 || !value_id_constant_p (value_id
));
438 nary
->hashcode
= vn_nary_op_compute_hash (nary
);
439 PRE_EXPR_NARY (&expr
) = nary
;
440 result_id
= lookup_expression_id (&expr
);
442 return expression_for_id (result_id
);
444 result
= pre_expr_pool
.allocate ();
447 result
->value_id
= value_id
? value_id
: get_next_value_id ();
448 PRE_EXPR_NARY (result
)
449 = alloc_vn_nary_op_noinit (nary
->length
, &pre_expr_obstack
);
450 memcpy (PRE_EXPR_NARY (result
), nary
, sizeof_vn_nary_op (nary
->length
));
451 alloc_expression_id (result
);
455 /* Given an REFERENCE, get or create a pre_expr to represent it. */
458 get_or_alloc_expr_for_reference (vn_reference_t reference
,
459 location_t loc
= UNKNOWN_LOCATION
)
461 struct pre_expr_d expr
;
463 unsigned int result_id
;
465 expr
.kind
= REFERENCE
;
467 PRE_EXPR_REFERENCE (&expr
) = reference
;
468 result_id
= lookup_expression_id (&expr
);
470 return expression_for_id (result_id
);
472 result
= pre_expr_pool
.allocate ();
473 result
->kind
= REFERENCE
;
475 result
->value_id
= reference
->value_id
;
476 PRE_EXPR_REFERENCE (result
) = reference
;
477 alloc_expression_id (result
);
482 /* An unordered bitmap set. One bitmap tracks values, the other,
484 typedef class bitmap_set
487 bitmap_head expressions
;
491 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
492 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
494 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
495 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
497 /* Mapping from value id to expressions with that value_id. */
498 static vec
<bitmap
> value_expressions
;
499 /* We just record a single expression for each constant value,
500 one of kind CONSTANT. */
501 static vec
<pre_expr
> constant_value_expressions
;
504 /* This structure is used to keep track of statistics on what
505 optimization PRE was able to perform. */
508 /* The number of new expressions/temporaries generated by PRE. */
511 /* The number of inserts found due to partial anticipation */
514 /* The number of inserts made for code hoisting. */
517 /* The number of new PHI nodes added by PRE. */
521 static bool do_partial_partial
;
522 static pre_expr
bitmap_find_leader (bitmap_set_t
, unsigned int);
523 static void bitmap_value_insert_into_set (bitmap_set_t
, pre_expr
);
524 static bool bitmap_value_replace_in_set (bitmap_set_t
, pre_expr
);
525 static void bitmap_set_copy (bitmap_set_t
, bitmap_set_t
);
526 static bool bitmap_set_contains_value (bitmap_set_t
, unsigned int);
527 static void bitmap_insert_into_set (bitmap_set_t
, pre_expr
);
528 static bitmap_set_t
bitmap_set_new (void);
529 static tree
create_expression_by_pieces (basic_block
, pre_expr
, gimple_seq
*,
531 static tree
find_or_generate_expression (basic_block
, tree
, gimple_seq
*);
532 static unsigned int get_expr_value_id (pre_expr
);
534 /* We can add and remove elements and entries to and from sets
535 and hash tables, so we use alloc pools for them. */
537 static object_allocator
<bitmap_set
> bitmap_set_pool ("Bitmap sets");
538 static bitmap_obstack grand_bitmap_obstack
;
540 /* A three tuple {e, pred, v} used to cache phi translations in the
541 phi_translate_table. */
543 typedef struct expr_pred_trans_d
: public typed_noop_remove
<expr_pred_trans_d
>
545 typedef expr_pred_trans_d value_type
;
546 typedef expr_pred_trans_d compare_type
;
548 /* The expression ID. */
551 /* The value expression ID that resulted from the translation. */
554 /* hash_table support. */
555 static inline void mark_empty (expr_pred_trans_d
&);
556 static inline bool is_empty (const expr_pred_trans_d
&);
557 static inline void mark_deleted (expr_pred_trans_d
&);
558 static inline bool is_deleted (const expr_pred_trans_d
&);
559 static const bool empty_zero_p
= true;
560 static inline hashval_t
hash (const expr_pred_trans_d
&);
561 static inline int equal (const expr_pred_trans_d
&, const expr_pred_trans_d
&);
562 } *expr_pred_trans_t
;
563 typedef const struct expr_pred_trans_d
*const_expr_pred_trans_t
;
566 expr_pred_trans_d::is_empty (const expr_pred_trans_d
&e
)
572 expr_pred_trans_d::is_deleted (const expr_pred_trans_d
&e
)
578 expr_pred_trans_d::mark_empty (expr_pred_trans_d
&e
)
584 expr_pred_trans_d::mark_deleted (expr_pred_trans_d
&e
)
590 expr_pred_trans_d::hash (const expr_pred_trans_d
&e
)
596 expr_pred_trans_d::equal (const expr_pred_trans_d
&ve1
,
597 const expr_pred_trans_d
&ve2
)
599 return ve1
.e
== ve2
.e
;
602 /* Sets that we need to keep track of. */
603 typedef struct bb_bitmap_sets
605 /* The EXP_GEN set, which represents expressions/values generated in
607 bitmap_set_t exp_gen
;
609 /* The PHI_GEN set, which represents PHI results generated in a
611 bitmap_set_t phi_gen
;
613 /* The TMP_GEN set, which represents results/temporaries generated
614 in a basic block. IE the LHS of an expression. */
615 bitmap_set_t tmp_gen
;
617 /* The AVAIL_OUT set, which represents which values are available in
618 a given basic block. */
619 bitmap_set_t avail_out
;
621 /* The ANTIC_IN set, which represents which values are anticipatable
622 in a given basic block. */
623 bitmap_set_t antic_in
;
625 /* The PA_IN set, which represents which values are
626 partially anticipatable in a given basic block. */
629 /* The NEW_SETS set, which is used during insertion to augment the
630 AVAIL_OUT set of blocks with the new insertions performed during
631 the current iteration. */
632 bitmap_set_t new_sets
;
634 /* A cache for value_dies_in_block_x. */
637 /* The live virtual operand on successor edges. */
640 /* PHI translate cache for the single successor edge. */
641 hash_table
<expr_pred_trans_d
> *phi_translate_table
;
643 /* True if we have visited this block during ANTIC calculation. */
644 unsigned int visited
: 1;
646 /* True when the block contains a call that might not return. */
647 unsigned int contains_may_not_return_call
: 1;
650 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
651 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
652 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
653 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
654 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
655 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
656 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
657 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
658 #define PHI_TRANS_TABLE(BB) ((bb_value_sets_t) ((BB)->aux))->phi_translate_table
659 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
660 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
661 #define BB_LIVE_VOP_ON_EXIT(BB) ((bb_value_sets_t) ((BB)->aux))->vop_on_exit
664 /* Add the tuple mapping from {expression E, basic block PRED} to
665 the phi translation table and return whether it pre-existed. */
668 phi_trans_add (expr_pred_trans_t
*entry
, pre_expr e
, basic_block pred
)
670 if (!PHI_TRANS_TABLE (pred
))
671 PHI_TRANS_TABLE (pred
) = new hash_table
<expr_pred_trans_d
> (11);
673 expr_pred_trans_t slot
;
674 expr_pred_trans_d tem
;
675 unsigned id
= get_expression_id (e
);
677 slot
= PHI_TRANS_TABLE (pred
)->find_slot_with_hash (tem
, id
, INSERT
);
690 /* Add expression E to the expression set of value id V. */
693 add_to_value (unsigned int v
, pre_expr e
)
695 gcc_checking_assert (get_expr_value_id (e
) == v
);
697 if (value_id_constant_p (v
))
699 if (e
->kind
!= CONSTANT
)
702 if (-v
>= constant_value_expressions
.length ())
703 constant_value_expressions
.safe_grow_cleared (-v
+ 1);
705 pre_expr leader
= constant_value_expressions
[-v
];
707 constant_value_expressions
[-v
] = e
;
711 if (v
>= value_expressions
.length ())
712 value_expressions
.safe_grow_cleared (v
+ 1);
714 bitmap set
= value_expressions
[v
];
717 set
= BITMAP_ALLOC (&grand_bitmap_obstack
);
718 value_expressions
[v
] = set
;
720 bitmap_set_bit (set
, get_expression_id (e
));
724 /* Create a new bitmap set and return it. */
727 bitmap_set_new (void)
729 bitmap_set_t ret
= bitmap_set_pool
.allocate ();
730 bitmap_initialize (&ret
->expressions
, &grand_bitmap_obstack
);
731 bitmap_initialize (&ret
->values
, &grand_bitmap_obstack
);
735 /* Return the value id for a PRE expression EXPR. */
738 get_expr_value_id (pre_expr expr
)
740 /* ??? We cannot assert that expr has a value-id (it can be 0), because
741 we assign value-ids only to expressions that have a result
742 in set_hashtable_value_ids. */
743 return expr
->value_id
;
746 /* Return a VN valnum (SSA name or constant) for the PRE value-id VAL. */
749 vn_valnum_from_value_id (unsigned int val
)
751 if (value_id_constant_p (val
))
753 pre_expr vexpr
= constant_value_expressions
[-val
];
755 return PRE_EXPR_CONSTANT (vexpr
);
759 bitmap exprset
= value_expressions
[val
];
762 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
764 pre_expr vexpr
= expression_for_id (i
);
765 if (vexpr
->kind
== NAME
)
766 return VN_INFO (PRE_EXPR_NAME (vexpr
))->valnum
;
771 /* Insert an expression EXPR into a bitmapped set. */
774 bitmap_insert_into_set (bitmap_set_t set
, pre_expr expr
)
776 unsigned int val
= get_expr_value_id (expr
);
777 if (! value_id_constant_p (val
))
779 /* Note this is the only function causing multiple expressions
780 for the same value to appear in a set. This is needed for
781 TMP_GEN, PHI_GEN and NEW_SETs. */
782 bitmap_set_bit (&set
->values
, val
);
783 bitmap_set_bit (&set
->expressions
, get_expression_id (expr
));
787 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
790 bitmap_set_copy (bitmap_set_t dest
, bitmap_set_t orig
)
792 bitmap_copy (&dest
->expressions
, &orig
->expressions
);
793 bitmap_copy (&dest
->values
, &orig
->values
);
797 /* Free memory used up by SET. */
799 bitmap_set_free (bitmap_set_t set
)
801 bitmap_clear (&set
->expressions
);
802 bitmap_clear (&set
->values
);
806 pre_expr_DFS (pre_expr expr
, bitmap_set_t set
, bitmap val_visited
,
807 vec
<pre_expr
> &post
);
809 /* DFS walk leaders of VAL to their operands with leaders in SET, collecting
810 expressions in SET in postorder into POST. */
813 pre_expr_DFS (unsigned val
, bitmap_set_t set
, bitmap val_visited
,
819 /* Iterate over all leaders and DFS recurse. Borrowed from
820 bitmap_find_leader. */
821 bitmap exprset
= value_expressions
[val
];
822 if (!exprset
->first
->next
)
824 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
825 if (bitmap_bit_p (&set
->expressions
, i
))
826 pre_expr_DFS (expression_for_id (i
), set
, val_visited
, post
);
830 EXECUTE_IF_AND_IN_BITMAP (exprset
, &set
->expressions
, 0, i
, bi
)
831 pre_expr_DFS (expression_for_id (i
), set
, val_visited
, post
);
834 /* DFS walk EXPR to its operands with leaders in SET, collecting
835 expressions in SET in postorder into POST. */
838 pre_expr_DFS (pre_expr expr
, bitmap_set_t set
, bitmap val_visited
,
845 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
846 for (unsigned i
= 0; i
< nary
->length
; i
++)
848 if (TREE_CODE (nary
->op
[i
]) != SSA_NAME
)
850 unsigned int op_val_id
= VN_INFO (nary
->op
[i
])->value_id
;
851 /* If we already found a leader for the value we've
852 recursed already. Avoid the costly bitmap_find_leader. */
853 if (bitmap_bit_p (&set
->values
, op_val_id
)
854 && bitmap_set_bit (val_visited
, op_val_id
))
855 pre_expr_DFS (op_val_id
, set
, val_visited
, post
);
861 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
862 vec
<vn_reference_op_s
> operands
= ref
->operands
;
863 vn_reference_op_t operand
;
864 for (unsigned i
= 0; operands
.iterate (i
, &operand
); i
++)
867 op
[0] = operand
->op0
;
868 op
[1] = operand
->op1
;
869 op
[2] = operand
->op2
;
870 for (unsigned n
= 0; n
< 3; ++n
)
872 if (!op
[n
] || TREE_CODE (op
[n
]) != SSA_NAME
)
874 unsigned op_val_id
= VN_INFO (op
[n
])->value_id
;
875 if (bitmap_bit_p (&set
->values
, op_val_id
)
876 && bitmap_set_bit (val_visited
, op_val_id
))
877 pre_expr_DFS (op_val_id
, set
, val_visited
, post
);
884 post
.quick_push (expr
);
887 /* Generate an topological-ordered array of bitmap set SET. */
890 sorted_array_from_bitmap_set (bitmap_set_t set
)
894 vec
<pre_expr
> result
;
896 /* Pre-allocate enough space for the array. */
897 result
.create (bitmap_count_bits (&set
->expressions
));
899 auto_bitmap
val_visited (&grand_bitmap_obstack
);
900 bitmap_tree_view (val_visited
);
901 FOR_EACH_VALUE_ID_IN_SET (set
, i
, bi
)
902 if (bitmap_set_bit (val_visited
, i
))
903 pre_expr_DFS (i
, set
, val_visited
, result
);
908 /* Subtract all expressions contained in ORIG from DEST. */
911 bitmap_set_subtract_expressions (bitmap_set_t dest
, bitmap_set_t orig
)
913 bitmap_set_t result
= bitmap_set_new ();
917 bitmap_and_compl (&result
->expressions
, &dest
->expressions
,
920 FOR_EACH_EXPR_ID_IN_SET (result
, i
, bi
)
922 pre_expr expr
= expression_for_id (i
);
923 unsigned int value_id
= get_expr_value_id (expr
);
924 bitmap_set_bit (&result
->values
, value_id
);
930 /* Subtract all values in bitmap set B from bitmap set A. */
933 bitmap_set_subtract_values (bitmap_set_t a
, bitmap_set_t b
)
937 unsigned to_remove
= -1U;
938 bitmap_and_compl_into (&a
->values
, &b
->values
);
939 FOR_EACH_EXPR_ID_IN_SET (a
, i
, bi
)
941 if (to_remove
!= -1U)
943 bitmap_clear_bit (&a
->expressions
, to_remove
);
946 pre_expr expr
= expression_for_id (i
);
947 if (! bitmap_bit_p (&a
->values
, get_expr_value_id (expr
)))
950 if (to_remove
!= -1U)
951 bitmap_clear_bit (&a
->expressions
, to_remove
);
955 /* Return true if bitmapped set SET contains the value VALUE_ID. */
958 bitmap_set_contains_value (bitmap_set_t set
, unsigned int value_id
)
960 if (value_id_constant_p (value_id
))
963 return bitmap_bit_p (&set
->values
, value_id
);
966 /* Return true if two bitmap sets are equal. */
969 bitmap_set_equal (bitmap_set_t a
, bitmap_set_t b
)
971 return bitmap_equal_p (&a
->values
, &b
->values
);
974 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
975 and add it otherwise. Return true if any changes were made. */
978 bitmap_value_replace_in_set (bitmap_set_t set
, pre_expr expr
)
980 unsigned int val
= get_expr_value_id (expr
);
981 if (value_id_constant_p (val
))
984 if (bitmap_set_contains_value (set
, val
))
986 /* The number of expressions having a given value is usually
987 significantly less than the total number of expressions in SET.
988 Thus, rather than check, for each expression in SET, whether it
989 has the value LOOKFOR, we walk the reverse mapping that tells us
990 what expressions have a given value, and see if any of those
991 expressions are in our set. For large testcases, this is about
992 5-10x faster than walking the bitmap. If this is somehow a
993 significant lose for some cases, we can choose which set to walk
994 based on the set size. */
997 bitmap exprset
= value_expressions
[val
];
998 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
1000 if (bitmap_clear_bit (&set
->expressions
, i
))
1002 bitmap_set_bit (&set
->expressions
, get_expression_id (expr
));
1003 return i
!= get_expression_id (expr
);
1009 bitmap_insert_into_set (set
, expr
);
1013 /* Insert EXPR into SET if EXPR's value is not already present in
1017 bitmap_value_insert_into_set (bitmap_set_t set
, pre_expr expr
)
1019 unsigned int val
= get_expr_value_id (expr
);
1021 gcc_checking_assert (expr
->id
== get_expression_id (expr
));
1023 /* Constant values are always considered to be part of the set. */
1024 if (value_id_constant_p (val
))
1027 /* If the value membership changed, add the expression. */
1028 if (bitmap_set_bit (&set
->values
, val
))
1029 bitmap_set_bit (&set
->expressions
, expr
->id
);
1032 /* Print out EXPR to outfile. */
1035 print_pre_expr (FILE *outfile
, const pre_expr expr
)
1039 fprintf (outfile
, "NULL");
1045 print_generic_expr (outfile
, PRE_EXPR_CONSTANT (expr
));
1048 print_generic_expr (outfile
, PRE_EXPR_NAME (expr
));
1053 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
1054 fprintf (outfile
, "{%s,", get_tree_code_name (nary
->opcode
));
1055 for (i
= 0; i
< nary
->length
; i
++)
1057 print_generic_expr (outfile
, nary
->op
[i
]);
1058 if (i
!= (unsigned) nary
->length
- 1)
1059 fprintf (outfile
, ",");
1061 fprintf (outfile
, "}");
1067 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
1068 print_vn_reference_ops (outfile
, ref
->operands
);
1071 fprintf (outfile
, "@");
1072 print_generic_expr (outfile
, ref
->vuse
);
1078 void debug_pre_expr (pre_expr
);
1080 /* Like print_pre_expr but always prints to stderr. */
1082 debug_pre_expr (pre_expr e
)
1084 print_pre_expr (stderr
, e
);
1085 fprintf (stderr
, "\n");
1088 /* Print out SET to OUTFILE. */
1091 print_bitmap_set (FILE *outfile
, bitmap_set_t set
,
1092 const char *setname
, int blockindex
)
1094 fprintf (outfile
, "%s[%d] := { ", setname
, blockindex
);
1101 FOR_EACH_EXPR_ID_IN_SET (set
, i
, bi
)
1103 const pre_expr expr
= expression_for_id (i
);
1106 fprintf (outfile
, ", ");
1108 print_pre_expr (outfile
, expr
);
1110 fprintf (outfile
, " (%04d)", get_expr_value_id (expr
));
1113 fprintf (outfile
, " }\n");
1116 void debug_bitmap_set (bitmap_set_t
);
1119 debug_bitmap_set (bitmap_set_t set
)
1121 print_bitmap_set (stderr
, set
, "debug", 0);
1124 void debug_bitmap_sets_for (basic_block
);
1127 debug_bitmap_sets_for (basic_block bb
)
1129 print_bitmap_set (stderr
, AVAIL_OUT (bb
), "avail_out", bb
->index
);
1130 print_bitmap_set (stderr
, EXP_GEN (bb
), "exp_gen", bb
->index
);
1131 print_bitmap_set (stderr
, PHI_GEN (bb
), "phi_gen", bb
->index
);
1132 print_bitmap_set (stderr
, TMP_GEN (bb
), "tmp_gen", bb
->index
);
1133 print_bitmap_set (stderr
, ANTIC_IN (bb
), "antic_in", bb
->index
);
1134 if (do_partial_partial
)
1135 print_bitmap_set (stderr
, PA_IN (bb
), "pa_in", bb
->index
);
1136 print_bitmap_set (stderr
, NEW_SETS (bb
), "new_sets", bb
->index
);
1139 /* Print out the expressions that have VAL to OUTFILE. */
1142 print_value_expressions (FILE *outfile
, unsigned int val
)
1144 bitmap set
= value_expressions
[val
];
1149 sprintf (s
, "%04d", val
);
1150 x
.expressions
= *set
;
1151 print_bitmap_set (outfile
, &x
, s
, 0);
1157 debug_value_expressions (unsigned int val
)
1159 print_value_expressions (stderr
, val
);
1162 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1166 get_or_alloc_expr_for_constant (tree constant
)
1168 unsigned int result_id
;
1169 struct pre_expr_d expr
;
1172 expr
.kind
= CONSTANT
;
1173 PRE_EXPR_CONSTANT (&expr
) = constant
;
1174 result_id
= lookup_expression_id (&expr
);
1176 return expression_for_id (result_id
);
1178 newexpr
= pre_expr_pool
.allocate ();
1179 newexpr
->kind
= CONSTANT
;
1180 newexpr
->loc
= UNKNOWN_LOCATION
;
1181 PRE_EXPR_CONSTANT (newexpr
) = constant
;
1182 alloc_expression_id (newexpr
);
1183 newexpr
->value_id
= get_or_alloc_constant_value_id (constant
);
1184 add_to_value (newexpr
->value_id
, newexpr
);
1188 /* Return the folded version of T if T, when folded, is a gimple
1189 min_invariant or an SSA name. Otherwise, return T. */
1192 fully_constant_expression (pre_expr e
)
1200 vn_nary_op_t nary
= PRE_EXPR_NARY (e
);
1201 tree res
= vn_nary_simplify (nary
);
1204 if (is_gimple_min_invariant (res
))
1205 return get_or_alloc_expr_for_constant (res
);
1206 if (TREE_CODE (res
) == SSA_NAME
)
1207 return get_or_alloc_expr_for_name (res
);
1212 vn_reference_t ref
= PRE_EXPR_REFERENCE (e
);
1214 if ((folded
= fully_constant_vn_reference_p (ref
)))
1215 return get_or_alloc_expr_for_constant (folded
);
1223 /* Translate the VUSE backwards through phi nodes in E->dest, so that
1224 it has the value it would have in E->src. Set *SAME_VALID to true
1225 in case the new vuse doesn't change the value id of the OPERANDS. */
1228 translate_vuse_through_block (vec
<vn_reference_op_s
> operands
,
1229 alias_set_type set
, alias_set_type base_set
,
1230 tree type
, tree vuse
, edge e
, bool *same_valid
)
1232 basic_block phiblock
= e
->dest
;
1233 gimple
*phi
= SSA_NAME_DEF_STMT (vuse
);
1239 /* If value-numbering provided a memory state for this
1240 that dominates PHIBLOCK we can just use that. */
1241 if (gimple_nop_p (phi
)
1242 || (gimple_bb (phi
) != phiblock
1243 && dominated_by_p (CDI_DOMINATORS
, phiblock
, gimple_bb (phi
))))
1246 /* We have pruned expressions that are killed in PHIBLOCK via
1247 prune_clobbered_mems but we have not rewritten the VUSE to the one
1248 live at the start of the block. If there is no virtual PHI to translate
1249 through return the VUSE live at entry. Otherwise the VUSE to translate
1250 is the def of the virtual PHI node. */
1251 phi
= get_virtual_phi (phiblock
);
1253 return BB_LIVE_VOP_ON_EXIT
1254 (get_immediate_dominator (CDI_DOMINATORS
, phiblock
));
1257 && ao_ref_init_from_vn_reference (&ref
, set
, base_set
, type
, operands
))
1259 bitmap visited
= NULL
;
1260 /* Try to find a vuse that dominates this phi node by skipping
1261 non-clobbering statements. */
1262 unsigned int cnt
= param_sccvn_max_alias_queries_per_access
;
1263 vuse
= get_continuation_for_phi (phi
, &ref
, true,
1264 cnt
, &visited
, false, NULL
, NULL
);
1266 BITMAP_FREE (visited
);
1270 /* If we didn't find any, the value ID can't stay the same. */
1271 if (!vuse
&& same_valid
)
1272 *same_valid
= false;
1274 /* ??? We would like to return vuse here as this is the canonical
1275 upmost vdef that this reference is associated with. But during
1276 insertion of the references into the hash tables we only ever
1277 directly insert with their direct gimple_vuse, hence returning
1278 something else would make us not find the other expression. */
1279 return PHI_ARG_DEF (phi
, e
->dest_idx
);
1282 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1283 SET2 *or* SET3. This is used to avoid making a set consisting of the union
1284 of PA_IN and ANTIC_IN during insert and phi-translation. */
1286 static inline pre_expr
1287 find_leader_in_sets (unsigned int val
, bitmap_set_t set1
, bitmap_set_t set2
,
1288 bitmap_set_t set3
= NULL
)
1290 pre_expr result
= NULL
;
1293 result
= bitmap_find_leader (set1
, val
);
1294 if (!result
&& set2
)
1295 result
= bitmap_find_leader (set2
, val
);
1296 if (!result
&& set3
)
1297 result
= bitmap_find_leader (set3
, val
);
1301 /* Get the tree type for our PRE expression e. */
1304 get_expr_type (const pre_expr e
)
1309 return TREE_TYPE (PRE_EXPR_NAME (e
));
1311 return TREE_TYPE (PRE_EXPR_CONSTANT (e
));
1313 return PRE_EXPR_REFERENCE (e
)->type
;
1315 return PRE_EXPR_NARY (e
)->type
;
1320 /* Get a representative SSA_NAME for a given expression that is available in B.
1321 Since all of our sub-expressions are treated as values, we require
1322 them to be SSA_NAME's for simplicity.
1323 Prior versions of GVNPRE used to use "value handles" here, so that
1324 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1325 either case, the operands are really values (IE we do not expect
1326 them to be usable without finding leaders). */
1329 get_representative_for (const pre_expr e
, basic_block b
= NULL
)
1331 tree name
, valnum
= NULL_TREE
;
1332 unsigned int value_id
= get_expr_value_id (e
);
1337 return PRE_EXPR_NAME (e
);
1339 return PRE_EXPR_CONSTANT (e
);
1343 /* Go through all of the expressions representing this value
1344 and pick out an SSA_NAME. */
1347 bitmap exprs
= value_expressions
[value_id
];
1348 EXECUTE_IF_SET_IN_BITMAP (exprs
, 0, i
, bi
)
1350 pre_expr rep
= expression_for_id (i
);
1351 if (rep
->kind
== NAME
)
1353 tree name
= PRE_EXPR_NAME (rep
);
1354 valnum
= VN_INFO (name
)->valnum
;
1355 gimple
*def
= SSA_NAME_DEF_STMT (name
);
1356 /* We have to return either a new representative or one
1357 that can be used for expression simplification and thus
1358 is available in B. */
1360 || gimple_nop_p (def
)
1361 || dominated_by_p (CDI_DOMINATORS
, b
, gimple_bb (def
)))
1364 else if (rep
->kind
== CONSTANT
)
1365 return PRE_EXPR_CONSTANT (rep
);
1371 /* If we reached here we couldn't find an SSA_NAME. This can
1372 happen when we've discovered a value that has never appeared in
1373 the program as set to an SSA_NAME, as the result of phi translation.
1375 ??? We should be able to re-use this when we insert the statement
1377 name
= make_temp_ssa_name (get_expr_type (e
), gimple_build_nop (), "pretmp");
1378 vn_ssa_aux_t vn_info
= VN_INFO (name
);
1379 vn_info
->value_id
= value_id
;
1380 vn_info
->valnum
= valnum
? valnum
: name
;
1381 vn_info
->visited
= true;
1382 /* ??? For now mark this SSA name for release by VN. */
1383 vn_info
->needs_insertion
= true;
1384 add_to_value (value_id
, get_or_alloc_expr_for_name (name
));
1385 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1387 fprintf (dump_file
, "Created SSA_NAME representative ");
1388 print_generic_expr (dump_file
, name
);
1389 fprintf (dump_file
, " for expression:");
1390 print_pre_expr (dump_file
, e
);
1391 fprintf (dump_file
, " (%04d)\n", value_id
);
1399 phi_translate (bitmap_set_t
, pre_expr
, bitmap_set_t
, bitmap_set_t
, edge
);
1401 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1402 the phis in PRED. Return NULL if we can't find a leader for each part
1403 of the translated expression. */
1406 phi_translate_1 (bitmap_set_t dest
,
1407 pre_expr expr
, bitmap_set_t set1
, bitmap_set_t set2
, edge e
)
1409 basic_block pred
= e
->src
;
1410 basic_block phiblock
= e
->dest
;
1411 location_t expr_loc
= expr
->loc
;
1417 bool changed
= false;
1418 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
1419 vn_nary_op_t newnary
= XALLOCAVAR (struct vn_nary_op_s
,
1420 sizeof_vn_nary_op (nary
->length
));
1421 memcpy (newnary
, nary
, sizeof_vn_nary_op (nary
->length
));
1423 for (i
= 0; i
< newnary
->length
; i
++)
1425 if (TREE_CODE (newnary
->op
[i
]) != SSA_NAME
)
1429 pre_expr leader
, result
;
1430 unsigned int op_val_id
= VN_INFO (newnary
->op
[i
])->value_id
;
1431 leader
= find_leader_in_sets (op_val_id
, set1
, set2
);
1432 result
= phi_translate (dest
, leader
, set1
, set2
, e
);
1433 if (result
&& result
!= leader
)
1434 /* If op has a leader in the sets we translate make
1435 sure to use the value of the translated expression.
1436 We might need a new representative for that. */
1437 newnary
->op
[i
] = get_representative_for (result
, pred
);
1441 changed
|= newnary
->op
[i
] != nary
->op
[i
];
1447 unsigned int new_val_id
;
1449 PRE_EXPR_NARY (expr
) = newnary
;
1450 constant
= fully_constant_expression (expr
);
1451 PRE_EXPR_NARY (expr
) = nary
;
1452 if (constant
!= expr
)
1454 /* For non-CONSTANTs we have to make sure we can eventually
1455 insert the expression. Which means we need to have a
1457 if (constant
->kind
!= CONSTANT
)
1459 /* Do not allow simplifications to non-constants over
1460 backedges as this will likely result in a loop PHI node
1461 to be inserted and increased register pressure.
1462 See PR77498 - this avoids doing predcoms work in
1463 a less efficient way. */
1464 if (e
->flags
& EDGE_DFS_BACK
)
1468 unsigned value_id
= get_expr_value_id (constant
);
1469 /* We want a leader in ANTIC_OUT or AVAIL_OUT here.
1470 dest has what we computed into ANTIC_OUT sofar
1471 so pick from that - since topological sorting
1472 by sorted_array_from_bitmap_set isn't perfect
1473 we may lose some cases here. */
1474 constant
= find_leader_in_sets (value_id
, dest
,
1478 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1480 fprintf (dump_file
, "simplifying ");
1481 print_pre_expr (dump_file
, expr
);
1482 fprintf (dump_file
, " translated %d -> %d to ",
1483 phiblock
->index
, pred
->index
);
1484 PRE_EXPR_NARY (expr
) = newnary
;
1485 print_pre_expr (dump_file
, expr
);
1486 PRE_EXPR_NARY (expr
) = nary
;
1487 fprintf (dump_file
, " to ");
1488 print_pre_expr (dump_file
, constant
);
1489 fprintf (dump_file
, "\n");
1499 tree result
= vn_nary_op_lookup_pieces (newnary
->length
,
1504 if (result
&& is_gimple_min_invariant (result
))
1505 return get_or_alloc_expr_for_constant (result
);
1507 if (!nary
|| nary
->predicated_values
)
1510 new_val_id
= nary
->value_id
;
1511 expr
= get_or_alloc_expr_for_nary (newnary
, new_val_id
, expr_loc
);
1512 add_to_value (get_expr_value_id (expr
), expr
);
1520 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
1521 vec
<vn_reference_op_s
> operands
= ref
->operands
;
1522 tree vuse
= ref
->vuse
;
1523 tree newvuse
= vuse
;
1524 vec
<vn_reference_op_s
> newoperands
= vNULL
;
1525 bool changed
= false, same_valid
= true;
1527 vn_reference_op_t operand
;
1528 vn_reference_t newref
;
1530 for (i
= 0; operands
.iterate (i
, &operand
); i
++)
1535 tree type
= operand
->type
;
1536 vn_reference_op_s newop
= *operand
;
1537 op
[0] = operand
->op0
;
1538 op
[1] = operand
->op1
;
1539 op
[2] = operand
->op2
;
1540 for (n
= 0; n
< 3; ++n
)
1542 unsigned int op_val_id
;
1545 if (TREE_CODE (op
[n
]) != SSA_NAME
)
1547 /* We can't possibly insert these. */
1549 && !is_gimple_min_invariant (op
[n
]))
1553 op_val_id
= VN_INFO (op
[n
])->value_id
;
1554 leader
= find_leader_in_sets (op_val_id
, set1
, set2
);
1555 opresult
= phi_translate (dest
, leader
, set1
, set2
, e
);
1556 if (opresult
&& opresult
!= leader
)
1558 tree name
= get_representative_for (opresult
);
1559 changed
|= name
!= op
[n
];
1567 newoperands
.release ();
1570 /* When we translate a MEM_REF across a backedge and we have
1571 restrict info that's not from our functions parameters
1572 we have to remap it since we now may deal with a different
1573 instance where the dependence info is no longer valid.
1574 See PR102970. Note instead of keeping a remapping table
1575 per backedge we simply throw away restrict info. */
1576 if ((newop
.opcode
== MEM_REF
1577 || newop
.opcode
== TARGET_MEM_REF
)
1579 && (e
->flags
& EDGE_DFS_BACK
))
1587 if (!newoperands
.exists ())
1588 newoperands
= operands
.copy ();
1589 /* We may have changed from an SSA_NAME to a constant */
1590 if (newop
.opcode
== SSA_NAME
&& TREE_CODE (op
[0]) != SSA_NAME
)
1591 newop
.opcode
= TREE_CODE (op
[0]);
1596 newoperands
[i
] = newop
;
1598 gcc_checking_assert (i
== operands
.length ());
1602 newvuse
= translate_vuse_through_block (newoperands
.exists ()
1603 ? newoperands
: operands
,
1604 ref
->set
, ref
->base_set
,
1607 ? NULL
: &same_valid
);
1608 if (newvuse
== NULL_TREE
)
1610 newoperands
.release ();
1615 if (changed
|| newvuse
!= vuse
)
1617 unsigned int new_val_id
;
1619 tree result
= vn_reference_lookup_pieces (newvuse
, ref
->set
,
1622 newoperands
.exists ()
1623 ? newoperands
: operands
,
1626 newoperands
.release ();
1628 /* We can always insert constants, so if we have a partial
1629 redundant constant load of another type try to translate it
1630 to a constant of appropriate type. */
1631 if (result
&& is_gimple_min_invariant (result
))
1634 if (!useless_type_conversion_p (ref
->type
, TREE_TYPE (result
)))
1636 tem
= fold_unary (VIEW_CONVERT_EXPR
, ref
->type
, result
);
1637 if (tem
&& !is_gimple_min_invariant (tem
))
1641 return get_or_alloc_expr_for_constant (tem
);
1644 /* If we'd have to convert things we would need to validate
1645 if we can insert the translated expression. So fail
1646 here for now - we cannot insert an alias with a different
1647 type in the VN tables either, as that would assert. */
1649 && !useless_type_conversion_p (ref
->type
, TREE_TYPE (result
)))
1651 else if (!result
&& newref
1652 && !useless_type_conversion_p (ref
->type
, newref
->type
))
1654 newoperands
.release ();
1659 new_val_id
= newref
->value_id
;
1662 if (changed
|| !same_valid
)
1663 new_val_id
= get_next_value_id ();
1665 new_val_id
= ref
->value_id
;
1666 if (!newoperands
.exists ())
1667 newoperands
= operands
.copy ();
1668 newref
= vn_reference_insert_pieces (newvuse
, ref
->set
,
1669 ref
->base_set
, ref
->type
,
1671 result
, new_val_id
);
1672 newoperands
= vNULL
;
1674 expr
= get_or_alloc_expr_for_reference (newref
, expr_loc
);
1675 add_to_value (new_val_id
, expr
);
1677 newoperands
.release ();
1684 tree name
= PRE_EXPR_NAME (expr
);
1685 gimple
*def_stmt
= SSA_NAME_DEF_STMT (name
);
1686 /* If the SSA name is defined by a PHI node in this block,
1688 if (gimple_code (def_stmt
) == GIMPLE_PHI
1689 && gimple_bb (def_stmt
) == phiblock
)
1691 tree def
= PHI_ARG_DEF (def_stmt
, e
->dest_idx
);
1693 /* Handle constant. */
1694 if (is_gimple_min_invariant (def
))
1695 return get_or_alloc_expr_for_constant (def
);
1697 return get_or_alloc_expr_for_name (def
);
1699 /* Otherwise return it unchanged - it will get removed if its
1700 value is not available in PREDs AVAIL_OUT set of expressions
1701 by the subtraction of TMP_GEN. */
1710 /* Wrapper around phi_translate_1 providing caching functionality. */
1713 phi_translate (bitmap_set_t dest
, pre_expr expr
,
1714 bitmap_set_t set1
, bitmap_set_t set2
, edge e
)
1716 expr_pred_trans_t slot
= NULL
;
1722 /* Constants contain no values that need translation. */
1723 if (expr
->kind
== CONSTANT
)
1726 if (value_id_constant_p (get_expr_value_id (expr
)))
1729 /* Don't add translations of NAMEs as those are cheap to translate. */
1730 if (expr
->kind
!= NAME
)
1732 if (phi_trans_add (&slot
, expr
, e
->src
))
1733 return slot
->v
== 0 ? NULL
: expression_for_id (slot
->v
);
1734 /* Store NULL for the value we want to return in the case of
1740 basic_block saved_valueize_bb
= vn_context_bb
;
1741 vn_context_bb
= e
->src
;
1742 phitrans
= phi_translate_1 (dest
, expr
, set1
, set2
, e
);
1743 vn_context_bb
= saved_valueize_bb
;
1747 /* We may have reallocated. */
1748 phi_trans_add (&slot
, expr
, e
->src
);
1750 slot
->v
= get_expression_id (phitrans
);
1752 /* Remove failed translations again, they cause insert
1753 iteration to not pick up new opportunities reliably. */
1754 PHI_TRANS_TABLE (e
->src
)->clear_slot (slot
);
1761 /* For each expression in SET, translate the values through phi nodes
1762 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1763 expressions in DEST. */
1766 phi_translate_set (bitmap_set_t dest
, bitmap_set_t set
, edge e
)
1771 if (gimple_seq_empty_p (phi_nodes (e
->dest
)))
1773 bitmap_set_copy (dest
, set
);
1777 /* Allocate the phi-translation cache where we have an idea about
1778 its size. hash-table implementation internals tell us that
1779 allocating the table to fit twice the number of elements will
1780 make sure we do not usually re-allocate. */
1781 if (!PHI_TRANS_TABLE (e
->src
))
1782 PHI_TRANS_TABLE (e
->src
) = new hash_table
<expr_pred_trans_d
>
1783 (2 * bitmap_count_bits (&set
->expressions
));
1784 FOR_EACH_EXPR_ID_IN_SET (set
, i
, bi
)
1786 pre_expr expr
= expression_for_id (i
);
1787 pre_expr translated
= phi_translate (dest
, expr
, set
, NULL
, e
);
1791 bitmap_insert_into_set (dest
, translated
);
1795 /* Find the leader for a value (i.e., the name representing that
1796 value) in a given set, and return it. Return NULL if no leader
1800 bitmap_find_leader (bitmap_set_t set
, unsigned int val
)
1802 if (value_id_constant_p (val
))
1803 return constant_value_expressions
[-val
];
1805 if (bitmap_set_contains_value (set
, val
))
1807 /* Rather than walk the entire bitmap of expressions, and see
1808 whether any of them has the value we are looking for, we look
1809 at the reverse mapping, which tells us the set of expressions
1810 that have a given value (IE value->expressions with that
1811 value) and see if any of those expressions are in our set.
1812 The number of expressions per value is usually significantly
1813 less than the number of expressions in the set. In fact, for
1814 large testcases, doing it this way is roughly 5-10x faster
1815 than walking the bitmap.
1816 If this is somehow a significant lose for some cases, we can
1817 choose which set to walk based on which set is smaller. */
1820 bitmap exprset
= value_expressions
[val
];
1822 if (!exprset
->first
->next
)
1823 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
1824 if (bitmap_bit_p (&set
->expressions
, i
))
1825 return expression_for_id (i
);
1827 EXECUTE_IF_AND_IN_BITMAP (exprset
, &set
->expressions
, 0, i
, bi
)
1828 return expression_for_id (i
);
1833 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1834 BLOCK by seeing if it is not killed in the block. Note that we are
1835 only determining whether there is a store that kills it. Because
1836 of the order in which clean iterates over values, we are guaranteed
1837 that altered operands will have caused us to be eliminated from the
1838 ANTIC_IN set already. */
1841 value_dies_in_block_x (pre_expr expr
, basic_block block
)
1843 tree vuse
= PRE_EXPR_REFERENCE (expr
)->vuse
;
1844 vn_reference_t refx
= PRE_EXPR_REFERENCE (expr
);
1846 gimple_stmt_iterator gsi
;
1847 unsigned id
= get_expression_id (expr
);
1854 /* Lookup a previously calculated result. */
1855 if (EXPR_DIES (block
)
1856 && bitmap_bit_p (EXPR_DIES (block
), id
* 2))
1857 return bitmap_bit_p (EXPR_DIES (block
), id
* 2 + 1);
1859 /* A memory expression {e, VUSE} dies in the block if there is a
1860 statement that may clobber e. If, starting statement walk from the
1861 top of the basic block, a statement uses VUSE there can be no kill
1862 inbetween that use and the original statement that loaded {e, VUSE},
1863 so we can stop walking. */
1864 ref
.base
= NULL_TREE
;
1865 for (gsi
= gsi_start_bb (block
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1867 tree def_vuse
, def_vdef
;
1868 def
= gsi_stmt (gsi
);
1869 def_vuse
= gimple_vuse (def
);
1870 def_vdef
= gimple_vdef (def
);
1872 /* Not a memory statement. */
1876 /* Not a may-def. */
1879 /* A load with the same VUSE, we're done. */
1880 if (def_vuse
== vuse
)
1886 /* Init ref only if we really need it. */
1887 if (ref
.base
== NULL_TREE
1888 && !ao_ref_init_from_vn_reference (&ref
, refx
->set
, refx
->base_set
,
1889 refx
->type
, refx
->operands
))
1894 /* If the statement may clobber expr, it dies. */
1895 if (stmt_may_clobber_ref_p_1 (def
, &ref
))
1902 /* Remember the result. */
1903 if (!EXPR_DIES (block
))
1904 EXPR_DIES (block
) = BITMAP_ALLOC (&grand_bitmap_obstack
);
1905 bitmap_set_bit (EXPR_DIES (block
), id
* 2);
1907 bitmap_set_bit (EXPR_DIES (block
), id
* 2 + 1);
1913 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1914 contains its value-id. */
1917 op_valid_in_sets (bitmap_set_t set1
, bitmap_set_t set2
, tree op
)
1919 if (op
&& TREE_CODE (op
) == SSA_NAME
)
1921 unsigned int value_id
= VN_INFO (op
)->value_id
;
1922 if (!(bitmap_set_contains_value (set1
, value_id
)
1923 || (set2
&& bitmap_set_contains_value (set2
, value_id
))))
1929 /* Determine if the expression EXPR is valid in SET1 U SET2.
1930 ONLY SET2 CAN BE NULL.
1931 This means that we have a leader for each part of the expression
1932 (if it consists of values), or the expression is an SSA_NAME.
1933 For loads/calls, we also see if the vuse is killed in this block. */
1936 valid_in_sets (bitmap_set_t set1
, bitmap_set_t set2
, pre_expr expr
)
1941 /* By construction all NAMEs are available. Non-available
1942 NAMEs are removed by subtracting TMP_GEN from the sets. */
1947 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
1948 for (i
= 0; i
< nary
->length
; i
++)
1949 if (!op_valid_in_sets (set1
, set2
, nary
->op
[i
]))
1956 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
1957 vn_reference_op_t vro
;
1960 FOR_EACH_VEC_ELT (ref
->operands
, i
, vro
)
1962 if (!op_valid_in_sets (set1
, set2
, vro
->op0
)
1963 || !op_valid_in_sets (set1
, set2
, vro
->op1
)
1964 || !op_valid_in_sets (set1
, set2
, vro
->op2
))
1974 /* Clean the set of expressions SET1 that are no longer valid in SET1 or SET2.
1975 This means expressions that are made up of values we have no leaders for
1979 clean (bitmap_set_t set1
, bitmap_set_t set2
= NULL
)
1981 vec
<pre_expr
> exprs
= sorted_array_from_bitmap_set (set1
);
1985 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
1987 if (!valid_in_sets (set1
, set2
, expr
))
1989 unsigned int val
= get_expr_value_id (expr
);
1990 bitmap_clear_bit (&set1
->expressions
, get_expression_id (expr
));
1991 /* We are entered with possibly multiple expressions for a value
1992 so before removing a value from the set see if there's an
1993 expression for it left. */
1994 if (! bitmap_find_leader (set1
, val
))
1995 bitmap_clear_bit (&set1
->values
, val
);
2004 FOR_EACH_EXPR_ID_IN_SET (set1
, j
, bi
)
2005 gcc_assert (valid_in_sets (set1
, set2
, expression_for_id (j
)));
2009 /* Clean the set of expressions that are no longer valid in SET because
2010 they are clobbered in BLOCK or because they trap and may not be executed. */
2013 prune_clobbered_mems (bitmap_set_t set
, basic_block block
)
2017 unsigned to_remove
= -1U;
2018 bool any_removed
= false;
2020 FOR_EACH_EXPR_ID_IN_SET (set
, i
, bi
)
2022 /* Remove queued expr. */
2023 if (to_remove
!= -1U)
2025 bitmap_clear_bit (&set
->expressions
, to_remove
);
2030 pre_expr expr
= expression_for_id (i
);
2031 if (expr
->kind
== REFERENCE
)
2033 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
2036 gimple
*def_stmt
= SSA_NAME_DEF_STMT (ref
->vuse
);
2037 if (!gimple_nop_p (def_stmt
)
2038 /* If value-numbering provided a memory state for this
2039 that dominates BLOCK we're done, otherwise we have
2040 to check if the value dies in BLOCK. */
2041 && !(gimple_bb (def_stmt
) != block
2042 && dominated_by_p (CDI_DOMINATORS
,
2043 block
, gimple_bb (def_stmt
)))
2044 && value_dies_in_block_x (expr
, block
))
2047 /* If the REFERENCE may trap make sure the block does not contain
2048 a possible exit point.
2049 ??? This is overly conservative if we translate AVAIL_OUT
2050 as the available expression might be after the exit point. */
2051 if (BB_MAY_NOTRETURN (block
)
2052 && vn_reference_may_trap (ref
))
2055 else if (expr
->kind
== NARY
)
2057 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
2058 /* If the NARY may trap make sure the block does not contain
2059 a possible exit point.
2060 ??? This is overly conservative if we translate AVAIL_OUT
2061 as the available expression might be after the exit point. */
2062 if (BB_MAY_NOTRETURN (block
)
2063 && vn_nary_may_trap (nary
))
2068 /* Remove queued expr. */
2069 if (to_remove
!= -1U)
2071 bitmap_clear_bit (&set
->expressions
, to_remove
);
2075 /* Above we only removed expressions, now clean the set of values
2076 which no longer have any corresponding expression. We cannot
2077 clear the value at the time we remove an expression since there
2078 may be multiple expressions per value.
2079 If we'd queue possibly to be removed values we could use
2080 the bitmap_find_leader way to see if there's still an expression
2081 for it. For some ratio of to be removed values and number of
2082 values/expressions in the set this might be faster than rebuilding
2086 bitmap_clear (&set
->values
);
2087 FOR_EACH_EXPR_ID_IN_SET (set
, i
, bi
)
2089 pre_expr expr
= expression_for_id (i
);
2090 unsigned int value_id
= get_expr_value_id (expr
);
2091 bitmap_set_bit (&set
->values
, value_id
);
2096 /* Compute the ANTIC set for BLOCK.
2098 If succs(BLOCK) > 1 then
2099 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2100 else if succs(BLOCK) == 1 then
2101 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2103 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2105 Note that clean() is deferred until after the iteration. */
2108 compute_antic_aux (basic_block block
, bool block_has_abnormal_pred_edge
)
2110 bitmap_set_t S
, old
, ANTIC_OUT
;
2114 bool was_visited
= BB_VISITED (block
);
2115 bool changed
= ! BB_VISITED (block
);
2116 BB_VISITED (block
) = 1;
2117 old
= ANTIC_OUT
= S
= NULL
;
2119 /* If any edges from predecessors are abnormal, antic_in is empty,
2121 if (block_has_abnormal_pred_edge
)
2122 goto maybe_dump_sets
;
2124 old
= ANTIC_IN (block
);
2125 ANTIC_OUT
= bitmap_set_new ();
2127 /* If the block has no successors, ANTIC_OUT is empty. */
2128 if (EDGE_COUNT (block
->succs
) == 0)
2130 /* If we have one successor, we could have some phi nodes to
2131 translate through. */
2132 else if (single_succ_p (block
))
2134 e
= single_succ_edge (block
);
2135 gcc_assert (BB_VISITED (e
->dest
));
2136 phi_translate_set (ANTIC_OUT
, ANTIC_IN (e
->dest
), e
);
2138 /* If we have multiple successors, we take the intersection of all of
2139 them. Note that in the case of loop exit phi nodes, we may have
2140 phis to translate through. */
2146 auto_vec
<edge
> worklist (EDGE_COUNT (block
->succs
));
2147 FOR_EACH_EDGE (e
, ei
, block
->succs
)
2150 && BB_VISITED (e
->dest
))
2152 else if (BB_VISITED (e
->dest
))
2153 worklist
.quick_push (e
);
2156 /* Unvisited successors get their ANTIC_IN replaced by the
2157 maximal set to arrive at a maximum ANTIC_IN solution.
2158 We can ignore them in the intersection operation and thus
2159 need not explicitely represent that maximum solution. */
2160 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2161 fprintf (dump_file
, "ANTIC_IN is MAX on %d->%d\n",
2162 e
->src
->index
, e
->dest
->index
);
2166 /* Of multiple successors we have to have visited one already
2167 which is guaranteed by iteration order. */
2168 gcc_assert (first
!= NULL
);
2170 phi_translate_set (ANTIC_OUT
, ANTIC_IN (first
->dest
), first
);
2172 /* If we have multiple successors we need to intersect the ANTIC_OUT
2173 sets. For values that's a simple intersection but for
2174 expressions it is a union. Given we want to have a single
2175 expression per value in our sets we have to canonicalize.
2176 Avoid randomness and running into cycles like for PR82129 and
2177 canonicalize the expression we choose to the one with the
2178 lowest id. This requires we actually compute the union first. */
2179 FOR_EACH_VEC_ELT (worklist
, i
, e
)
2181 if (!gimple_seq_empty_p (phi_nodes (e
->dest
)))
2183 bitmap_set_t tmp
= bitmap_set_new ();
2184 phi_translate_set (tmp
, ANTIC_IN (e
->dest
), e
);
2185 bitmap_and_into (&ANTIC_OUT
->values
, &tmp
->values
);
2186 bitmap_ior_into (&ANTIC_OUT
->expressions
, &tmp
->expressions
);
2187 bitmap_set_free (tmp
);
2191 bitmap_and_into (&ANTIC_OUT
->values
, &ANTIC_IN (e
->dest
)->values
);
2192 bitmap_ior_into (&ANTIC_OUT
->expressions
,
2193 &ANTIC_IN (e
->dest
)->expressions
);
2196 if (! worklist
.is_empty ())
2198 /* Prune expressions not in the value set. */
2201 unsigned int to_clear
= -1U;
2202 FOR_EACH_EXPR_ID_IN_SET (ANTIC_OUT
, i
, bi
)
2204 if (to_clear
!= -1U)
2206 bitmap_clear_bit (&ANTIC_OUT
->expressions
, to_clear
);
2209 pre_expr expr
= expression_for_id (i
);
2210 unsigned int value_id
= get_expr_value_id (expr
);
2211 if (!bitmap_bit_p (&ANTIC_OUT
->values
, value_id
))
2214 if (to_clear
!= -1U)
2215 bitmap_clear_bit (&ANTIC_OUT
->expressions
, to_clear
);
2219 /* Prune expressions that are clobbered in block and thus become
2220 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2221 prune_clobbered_mems (ANTIC_OUT
, block
);
2223 /* Generate ANTIC_OUT - TMP_GEN. */
2224 S
= bitmap_set_subtract_expressions (ANTIC_OUT
, TMP_GEN (block
));
2226 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2227 ANTIC_IN (block
) = bitmap_set_subtract_expressions (EXP_GEN (block
),
2230 /* Then union in the ANTIC_OUT - TMP_GEN values,
2231 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2232 bitmap_ior_into (&ANTIC_IN (block
)->values
, &S
->values
);
2233 bitmap_ior_into (&ANTIC_IN (block
)->expressions
, &S
->expressions
);
2235 /* clean (ANTIC_IN (block)) is defered to after the iteration converged
2236 because it can cause non-convergence, see for example PR81181. */
2238 /* Intersect ANTIC_IN with the old ANTIC_IN. This is required until
2239 we properly represent the maximum expression set, thus not prune
2240 values without expressions during the iteration. */
2242 && bitmap_and_into (&ANTIC_IN (block
)->values
, &old
->values
))
2244 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2245 fprintf (dump_file
, "warning: intersecting with old ANTIC_IN "
2246 "shrinks the set\n");
2247 /* Prune expressions not in the value set. */
2250 unsigned int to_clear
= -1U;
2251 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (block
), i
, bi
)
2253 if (to_clear
!= -1U)
2255 bitmap_clear_bit (&ANTIC_IN (block
)->expressions
, to_clear
);
2258 pre_expr expr
= expression_for_id (i
);
2259 unsigned int value_id
= get_expr_value_id (expr
);
2260 if (!bitmap_bit_p (&ANTIC_IN (block
)->values
, value_id
))
2263 if (to_clear
!= -1U)
2264 bitmap_clear_bit (&ANTIC_IN (block
)->expressions
, to_clear
);
2267 if (!bitmap_set_equal (old
, ANTIC_IN (block
)))
2271 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2274 print_bitmap_set (dump_file
, ANTIC_OUT
, "ANTIC_OUT", block
->index
);
2277 fprintf (dump_file
, "[changed] ");
2278 print_bitmap_set (dump_file
, ANTIC_IN (block
), "ANTIC_IN",
2282 print_bitmap_set (dump_file
, S
, "S", block
->index
);
2285 bitmap_set_free (old
);
2287 bitmap_set_free (S
);
2289 bitmap_set_free (ANTIC_OUT
);
2293 /* Compute PARTIAL_ANTIC for BLOCK.
2295 If succs(BLOCK) > 1 then
2296 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2297 in ANTIC_OUT for all succ(BLOCK)
2298 else if succs(BLOCK) == 1 then
2299 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2301 PA_IN[BLOCK] = clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK] - ANTIC_IN[BLOCK])
2305 compute_partial_antic_aux (basic_block block
,
2306 bool block_has_abnormal_pred_edge
)
2308 bitmap_set_t old_PA_IN
;
2309 bitmap_set_t PA_OUT
;
2312 unsigned long max_pa
= param_max_partial_antic_length
;
2314 old_PA_IN
= PA_OUT
= NULL
;
2316 /* If any edges from predecessors are abnormal, antic_in is empty,
2318 if (block_has_abnormal_pred_edge
)
2319 goto maybe_dump_sets
;
2321 /* If there are too many partially anticipatable values in the
2322 block, phi_translate_set can take an exponential time: stop
2323 before the translation starts. */
2325 && single_succ_p (block
)
2326 && bitmap_count_bits (&PA_IN (single_succ (block
))->values
) > max_pa
)
2327 goto maybe_dump_sets
;
2329 old_PA_IN
= PA_IN (block
);
2330 PA_OUT
= bitmap_set_new ();
2332 /* If the block has no successors, ANTIC_OUT is empty. */
2333 if (EDGE_COUNT (block
->succs
) == 0)
2335 /* If we have one successor, we could have some phi nodes to
2336 translate through. Note that we can't phi translate across DFS
2337 back edges in partial antic, because it uses a union operation on
2338 the successors. For recurrences like IV's, we will end up
2339 generating a new value in the set on each go around (i + 3 (VH.1)
2340 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2341 else if (single_succ_p (block
))
2343 e
= single_succ_edge (block
);
2344 if (!(e
->flags
& EDGE_DFS_BACK
))
2345 phi_translate_set (PA_OUT
, PA_IN (e
->dest
), e
);
2347 /* If we have multiple successors, we take the union of all of
2353 auto_vec
<edge
> worklist (EDGE_COUNT (block
->succs
));
2354 FOR_EACH_EDGE (e
, ei
, block
->succs
)
2356 if (e
->flags
& EDGE_DFS_BACK
)
2358 worklist
.quick_push (e
);
2360 if (worklist
.length () > 0)
2362 FOR_EACH_VEC_ELT (worklist
, i
, e
)
2367 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (e
->dest
), i
, bi
)
2368 bitmap_value_insert_into_set (PA_OUT
,
2369 expression_for_id (i
));
2370 if (!gimple_seq_empty_p (phi_nodes (e
->dest
)))
2372 bitmap_set_t pa_in
= bitmap_set_new ();
2373 phi_translate_set (pa_in
, PA_IN (e
->dest
), e
);
2374 FOR_EACH_EXPR_ID_IN_SET (pa_in
, i
, bi
)
2375 bitmap_value_insert_into_set (PA_OUT
,
2376 expression_for_id (i
));
2377 bitmap_set_free (pa_in
);
2380 FOR_EACH_EXPR_ID_IN_SET (PA_IN (e
->dest
), i
, bi
)
2381 bitmap_value_insert_into_set (PA_OUT
,
2382 expression_for_id (i
));
2387 /* Prune expressions that are clobbered in block and thus become
2388 invalid if translated from PA_OUT to PA_IN. */
2389 prune_clobbered_mems (PA_OUT
, block
);
2391 /* PA_IN starts with PA_OUT - TMP_GEN.
2392 Then we subtract things from ANTIC_IN. */
2393 PA_IN (block
) = bitmap_set_subtract_expressions (PA_OUT
, TMP_GEN (block
));
2395 /* For partial antic, we want to put back in the phi results, since
2396 we will properly avoid making them partially antic over backedges. */
2397 bitmap_ior_into (&PA_IN (block
)->values
, &PHI_GEN (block
)->values
);
2398 bitmap_ior_into (&PA_IN (block
)->expressions
, &PHI_GEN (block
)->expressions
);
2400 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2401 bitmap_set_subtract_values (PA_IN (block
), ANTIC_IN (block
));
2403 clean (PA_IN (block
), ANTIC_IN (block
));
2406 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2409 print_bitmap_set (dump_file
, PA_OUT
, "PA_OUT", block
->index
);
2411 print_bitmap_set (dump_file
, PA_IN (block
), "PA_IN", block
->index
);
2414 bitmap_set_free (old_PA_IN
);
2416 bitmap_set_free (PA_OUT
);
2419 /* Compute ANTIC and partial ANTIC sets. */
2422 compute_antic (void)
2424 bool changed
= true;
2425 int num_iterations
= 0;
2431 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2432 We pre-build the map of blocks with incoming abnormal edges here. */
2433 auto_sbitmap
has_abnormal_preds (last_basic_block_for_fn (cfun
));
2434 bitmap_clear (has_abnormal_preds
);
2436 FOR_ALL_BB_FN (block
, cfun
)
2438 BB_VISITED (block
) = 0;
2440 FOR_EACH_EDGE (e
, ei
, block
->preds
)
2441 if (e
->flags
& EDGE_ABNORMAL
)
2443 bitmap_set_bit (has_abnormal_preds
, block
->index
);
2447 /* While we are here, give empty ANTIC_IN sets to each block. */
2448 ANTIC_IN (block
) = bitmap_set_new ();
2449 if (do_partial_partial
)
2450 PA_IN (block
) = bitmap_set_new ();
2453 /* At the exit block we anticipate nothing. */
2454 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun
)) = 1;
2456 /* For ANTIC computation we need a postorder that also guarantees that
2457 a block with a single successor is visited after its successor.
2458 RPO on the inverted CFG has this property. */
2459 auto_vec
<int, 20> postorder
;
2460 inverted_post_order_compute (&postorder
);
2462 auto_sbitmap
worklist (last_basic_block_for_fn (cfun
) + 1);
2463 bitmap_clear (worklist
);
2464 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR_FOR_FN (cfun
)->preds
)
2465 bitmap_set_bit (worklist
, e
->src
->index
);
2468 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2469 fprintf (dump_file
, "Starting iteration %d\n", num_iterations
);
2470 /* ??? We need to clear our PHI translation cache here as the
2471 ANTIC sets shrink and we restrict valid translations to
2472 those having operands with leaders in ANTIC. Same below
2473 for PA ANTIC computation. */
2476 for (i
= postorder
.length () - 1; i
>= 0; i
--)
2478 if (bitmap_bit_p (worklist
, postorder
[i
]))
2480 basic_block block
= BASIC_BLOCK_FOR_FN (cfun
, postorder
[i
]);
2481 bitmap_clear_bit (worklist
, block
->index
);
2482 if (compute_antic_aux (block
,
2483 bitmap_bit_p (has_abnormal_preds
,
2486 FOR_EACH_EDGE (e
, ei
, block
->preds
)
2487 bitmap_set_bit (worklist
, e
->src
->index
);
2492 /* Theoretically possible, but *highly* unlikely. */
2493 gcc_checking_assert (num_iterations
< 500);
2496 /* We have to clean after the dataflow problem converged as cleaning
2497 can cause non-convergence because it is based on expressions
2498 rather than values. */
2499 FOR_EACH_BB_FN (block
, cfun
)
2500 clean (ANTIC_IN (block
));
2502 statistics_histogram_event (cfun
, "compute_antic iterations",
2505 if (do_partial_partial
)
2507 /* For partial antic we ignore backedges and thus we do not need
2508 to perform any iteration when we process blocks in postorder. */
2509 for (i
= postorder
.length () - 1; i
>= 0; i
--)
2511 basic_block block
= BASIC_BLOCK_FOR_FN (cfun
, postorder
[i
]);
2512 compute_partial_antic_aux (block
,
2513 bitmap_bit_p (has_abnormal_preds
,
2520 /* Inserted expressions are placed onto this worklist, which is used
2521 for performing quick dead code elimination of insertions we made
2522 that didn't turn out to be necessary. */
2523 static bitmap inserted_exprs
;
2525 /* The actual worker for create_component_ref_by_pieces. */
2528 create_component_ref_by_pieces_1 (basic_block block
, vn_reference_t ref
,
2529 unsigned int *operand
, gimple_seq
*stmts
)
2531 vn_reference_op_t currop
= &ref
->operands
[*operand
];
2534 switch (currop
->opcode
)
2541 tree baseop
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2545 tree offset
= currop
->op0
;
2546 if (TREE_CODE (baseop
) == ADDR_EXPR
2547 && handled_component_p (TREE_OPERAND (baseop
, 0)))
2551 base
= get_addr_base_and_unit_offset (TREE_OPERAND (baseop
, 0),
2554 offset
= int_const_binop (PLUS_EXPR
, offset
,
2555 build_int_cst (TREE_TYPE (offset
),
2557 baseop
= build_fold_addr_expr (base
);
2559 genop
= build2 (MEM_REF
, currop
->type
, baseop
, offset
);
2560 MR_DEPENDENCE_CLIQUE (genop
) = currop
->clique
;
2561 MR_DEPENDENCE_BASE (genop
) = currop
->base
;
2562 REF_REVERSE_STORAGE_ORDER (genop
) = currop
->reverse
;
2566 case TARGET_MEM_REF
:
2568 tree genop0
= NULL_TREE
, genop1
= NULL_TREE
;
2569 vn_reference_op_t nextop
= &ref
->operands
[(*operand
)++];
2570 tree baseop
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2576 genop0
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2582 genop1
= find_or_generate_expression (block
, nextop
->op0
, stmts
);
2586 genop
= build5 (TARGET_MEM_REF
, currop
->type
,
2587 baseop
, currop
->op2
, genop0
, currop
->op1
, genop1
);
2589 MR_DEPENDENCE_CLIQUE (genop
) = currop
->clique
;
2590 MR_DEPENDENCE_BASE (genop
) = currop
->base
;
2597 gcc_assert (is_gimple_min_invariant (currop
->op0
));
2603 case VIEW_CONVERT_EXPR
:
2605 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2609 return fold_build1 (currop
->opcode
, currop
->type
, genop0
);
2612 case WITH_SIZE_EXPR
:
2614 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2618 tree genop1
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2621 return fold_build2 (currop
->opcode
, currop
->type
, genop0
, genop1
);
2626 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2630 tree op1
= currop
->op0
;
2631 tree op2
= currop
->op1
;
2632 tree t
= build3 (BIT_FIELD_REF
, currop
->type
, genop0
, op1
, op2
);
2633 REF_REVERSE_STORAGE_ORDER (t
) = currop
->reverse
;
2637 /* For array ref vn_reference_op's, operand 1 of the array ref
2638 is op0 of the reference op and operand 3 of the array ref is
2640 case ARRAY_RANGE_REF
:
2644 tree genop1
= currop
->op0
;
2645 tree genop2
= currop
->op1
;
2646 tree genop3
= currop
->op2
;
2647 genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2651 genop1
= find_or_generate_expression (block
, genop1
, stmts
);
2656 tree domain_type
= TYPE_DOMAIN (TREE_TYPE (genop0
));
2657 /* Drop zero minimum index if redundant. */
2658 if (integer_zerop (genop2
)
2660 || integer_zerop (TYPE_MIN_VALUE (domain_type
))))
2664 genop2
= find_or_generate_expression (block
, genop2
, stmts
);
2671 tree elmt_type
= TREE_TYPE (TREE_TYPE (genop0
));
2672 /* We can't always put a size in units of the element alignment
2673 here as the element alignment may be not visible. See
2674 PR43783. Simply drop the element size for constant
2676 if (TREE_CODE (genop3
) == INTEGER_CST
2677 && TREE_CODE (TYPE_SIZE_UNIT (elmt_type
)) == INTEGER_CST
2678 && wi::eq_p (wi::to_offset (TYPE_SIZE_UNIT (elmt_type
)),
2679 (wi::to_offset (genop3
)
2680 * vn_ref_op_align_unit (currop
))))
2684 genop3
= find_or_generate_expression (block
, genop3
, stmts
);
2689 return build4 (currop
->opcode
, currop
->type
, genop0
, genop1
,
2696 tree genop2
= currop
->op1
;
2697 op0
= create_component_ref_by_pieces_1 (block
, ref
, operand
, stmts
);
2700 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2704 genop2
= find_or_generate_expression (block
, genop2
, stmts
);
2708 return fold_build3 (COMPONENT_REF
, TREE_TYPE (op1
), op0
, op1
, genop2
);
2713 genop
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2735 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2736 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2737 trying to rename aggregates into ssa form directly, which is a no no.
2739 Thus, this routine doesn't create temporaries, it just builds a
2740 single access expression for the array, calling
2741 find_or_generate_expression to build the innermost pieces.
2743 This function is a subroutine of create_expression_by_pieces, and
2744 should not be called on it's own unless you really know what you
2748 create_component_ref_by_pieces (basic_block block
, vn_reference_t ref
,
2751 unsigned int op
= 0;
2752 return create_component_ref_by_pieces_1 (block
, ref
, &op
, stmts
);
2755 /* Find a simple leader for an expression, or generate one using
2756 create_expression_by_pieces from a NARY expression for the value.
2757 BLOCK is the basic_block we are looking for leaders in.
2758 OP is the tree expression to find a leader for or generate.
2759 Returns the leader or NULL_TREE on failure. */
2762 find_or_generate_expression (basic_block block
, tree op
, gimple_seq
*stmts
)
2764 /* Constants are always leaders. */
2765 if (is_gimple_min_invariant (op
))
2768 gcc_assert (TREE_CODE (op
) == SSA_NAME
);
2769 vn_ssa_aux_t info
= VN_INFO (op
);
2770 unsigned int lookfor
= info
->value_id
;
2771 if (value_id_constant_p (lookfor
))
2772 return info
->valnum
;
2774 pre_expr leader
= bitmap_find_leader (AVAIL_OUT (block
), lookfor
);
2777 if (leader
->kind
== NAME
)
2778 return PRE_EXPR_NAME (leader
);
2779 else if (leader
->kind
== CONSTANT
)
2780 return PRE_EXPR_CONSTANT (leader
);
2785 gcc_assert (!value_id_constant_p (lookfor
));
2787 /* It must be a complex expression, so generate it recursively. Note
2788 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2789 where the insert algorithm fails to insert a required expression. */
2790 bitmap exprset
= value_expressions
[lookfor
];
2793 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
2795 pre_expr temp
= expression_for_id (i
);
2796 /* We cannot insert random REFERENCE expressions at arbitrary
2797 places. We can insert NARYs which eventually re-materializes
2798 its operand values. */
2799 if (temp
->kind
== NARY
)
2800 return create_expression_by_pieces (block
, temp
, stmts
,
2808 /* Create an expression in pieces, so that we can handle very complex
2809 expressions that may be ANTIC, but not necessary GIMPLE.
2810 BLOCK is the basic block the expression will be inserted into,
2811 EXPR is the expression to insert (in value form)
2812 STMTS is a statement list to append the necessary insertions into.
2814 This function will die if we hit some value that shouldn't be
2815 ANTIC but is (IE there is no leader for it, or its components).
2816 The function returns NULL_TREE in case a different antic expression
2817 has to be inserted first.
2818 This function may also generate expressions that are themselves
2819 partially or fully redundant. Those that are will be either made
2820 fully redundant during the next iteration of insert (for partially
2821 redundant ones), or eliminated by eliminate (for fully redundant
2825 create_expression_by_pieces (basic_block block
, pre_expr expr
,
2826 gimple_seq
*stmts
, tree type
)
2830 gimple_seq forced_stmts
= NULL
;
2831 unsigned int value_id
;
2832 gimple_stmt_iterator gsi
;
2833 tree exprtype
= type
? type
: get_expr_type (expr
);
2839 /* We may hit the NAME/CONSTANT case if we have to convert types
2840 that value numbering saw through. */
2842 folded
= PRE_EXPR_NAME (expr
);
2843 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (folded
))
2845 if (useless_type_conversion_p (exprtype
, TREE_TYPE (folded
)))
2850 folded
= PRE_EXPR_CONSTANT (expr
);
2851 tree tem
= fold_convert (exprtype
, folded
);
2852 if (is_gimple_min_invariant (tem
))
2857 if (PRE_EXPR_REFERENCE (expr
)->operands
[0].opcode
== CALL_EXPR
)
2859 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
2860 unsigned int operand
= 1;
2861 vn_reference_op_t currop
= &ref
->operands
[0];
2862 tree sc
= NULL_TREE
;
2863 tree fn
= NULL_TREE
;
2866 fn
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2872 sc
= find_or_generate_expression (block
, currop
->op1
, stmts
);
2876 auto_vec
<tree
> args (ref
->operands
.length () - 1);
2877 while (operand
< ref
->operands
.length ())
2879 tree arg
= create_component_ref_by_pieces_1 (block
, ref
,
2883 args
.quick_push (arg
);
2888 call
= gimple_build_call_vec (fn
, args
);
2889 gimple_call_set_fntype (call
, currop
->type
);
2892 call
= gimple_build_call_internal_vec ((internal_fn
)currop
->clique
,
2894 gimple_set_location (call
, expr
->loc
);
2896 gimple_call_set_chain (call
, sc
);
2897 tree forcedname
= make_ssa_name (ref
->type
);
2898 gimple_call_set_lhs (call
, forcedname
);
2899 /* There's no CCP pass after PRE which would re-compute alignment
2900 information so make sure we re-materialize this here. */
2901 if (gimple_call_builtin_p (call
, BUILT_IN_ASSUME_ALIGNED
)
2902 && args
.length () - 2 <= 1
2903 && tree_fits_uhwi_p (args
[1])
2904 && (args
.length () != 3 || tree_fits_uhwi_p (args
[2])))
2906 unsigned HOST_WIDE_INT halign
= tree_to_uhwi (args
[1]);
2907 unsigned HOST_WIDE_INT hmisalign
2908 = args
.length () == 3 ? tree_to_uhwi (args
[2]) : 0;
2909 if ((halign
& (halign
- 1)) == 0
2910 && (hmisalign
& ~(halign
- 1)) == 0
2911 && (unsigned int)halign
!= 0)
2912 set_ptr_info_alignment (get_ptr_info (forcedname
),
2915 gimple_set_vuse (call
, BB_LIVE_VOP_ON_EXIT (block
));
2916 gimple_seq_add_stmt_without_update (&forced_stmts
, call
);
2917 folded
= forcedname
;
2921 folded
= create_component_ref_by_pieces (block
,
2922 PRE_EXPR_REFERENCE (expr
),
2926 name
= make_temp_ssa_name (exprtype
, NULL
, "pretmp");
2927 newstmt
= gimple_build_assign (name
, folded
);
2928 gimple_set_location (newstmt
, expr
->loc
);
2929 gimple_seq_add_stmt_without_update (&forced_stmts
, newstmt
);
2930 gimple_set_vuse (newstmt
, BB_LIVE_VOP_ON_EXIT (block
));
2936 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
2937 tree
*genop
= XALLOCAVEC (tree
, nary
->length
);
2939 for (i
= 0; i
< nary
->length
; ++i
)
2941 genop
[i
] = find_or_generate_expression (block
, nary
->op
[i
], stmts
);
2944 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2945 may have conversions stripped. */
2946 if (nary
->opcode
== POINTER_PLUS_EXPR
)
2949 genop
[i
] = gimple_convert (&forced_stmts
,
2950 nary
->type
, genop
[i
]);
2952 genop
[i
] = gimple_convert (&forced_stmts
,
2953 sizetype
, genop
[i
]);
2956 genop
[i
] = gimple_convert (&forced_stmts
,
2957 TREE_TYPE (nary
->op
[i
]), genop
[i
]);
2959 if (nary
->opcode
== CONSTRUCTOR
)
2961 vec
<constructor_elt
, va_gc
> *elts
= NULL
;
2962 for (i
= 0; i
< nary
->length
; ++i
)
2963 CONSTRUCTOR_APPEND_ELT (elts
, NULL_TREE
, genop
[i
]);
2964 folded
= build_constructor (nary
->type
, elts
);
2965 name
= make_temp_ssa_name (exprtype
, NULL
, "pretmp");
2966 newstmt
= gimple_build_assign (name
, folded
);
2967 gimple_set_location (newstmt
, expr
->loc
);
2968 gimple_seq_add_stmt_without_update (&forced_stmts
, newstmt
);
2973 switch (nary
->length
)
2976 folded
= gimple_build (&forced_stmts
, expr
->loc
,
2977 nary
->opcode
, nary
->type
, genop
[0]);
2980 folded
= gimple_build (&forced_stmts
, expr
->loc
, nary
->opcode
,
2981 nary
->type
, genop
[0], genop
[1]);
2984 folded
= gimple_build (&forced_stmts
, expr
->loc
, nary
->opcode
,
2985 nary
->type
, genop
[0], genop
[1],
2998 folded
= gimple_convert (&forced_stmts
, exprtype
, folded
);
3000 /* If there is nothing to insert, return the simplified result. */
3001 if (gimple_seq_empty_p (forced_stmts
))
3003 /* If we simplified to a constant return it and discard eventually
3005 if (is_gimple_min_invariant (folded
))
3007 gimple_seq_discard (forced_stmts
);
3010 /* Likewise if we simplified to sth not queued for insertion. */
3012 gsi
= gsi_last (forced_stmts
);
3013 for (; !gsi_end_p (gsi
); gsi_prev (&gsi
))
3015 gimple
*stmt
= gsi_stmt (gsi
);
3016 tree forcedname
= gimple_get_lhs (stmt
);
3017 if (forcedname
== folded
)
3025 gimple_seq_discard (forced_stmts
);
3028 gcc_assert (TREE_CODE (folded
) == SSA_NAME
);
3030 /* If we have any intermediate expressions to the value sets, add them
3031 to the value sets and chain them in the instruction stream. */
3034 gsi
= gsi_start (forced_stmts
);
3035 for (; !gsi_end_p (gsi
); gsi_next (&gsi
))
3037 gimple
*stmt
= gsi_stmt (gsi
);
3038 tree forcedname
= gimple_get_lhs (stmt
);
3041 if (forcedname
!= folded
)
3043 vn_ssa_aux_t vn_info
= VN_INFO (forcedname
);
3044 vn_info
->valnum
= forcedname
;
3045 vn_info
->value_id
= get_next_value_id ();
3046 nameexpr
= get_or_alloc_expr_for_name (forcedname
);
3047 add_to_value (vn_info
->value_id
, nameexpr
);
3048 if (NEW_SETS (block
))
3049 bitmap_value_replace_in_set (NEW_SETS (block
), nameexpr
);
3050 bitmap_value_replace_in_set (AVAIL_OUT (block
), nameexpr
);
3053 bitmap_set_bit (inserted_exprs
, SSA_NAME_VERSION (forcedname
));
3055 gimple_seq_add_seq (stmts
, forced_stmts
);
3060 /* Fold the last statement. */
3061 gsi
= gsi_last (*stmts
);
3062 if (fold_stmt_inplace (&gsi
))
3063 update_stmt (gsi_stmt (gsi
));
3065 /* Add a value number to the temporary.
3066 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3067 we are creating the expression by pieces, and this particular piece of
3068 the expression may have been represented. There is no harm in replacing
3070 value_id
= get_expr_value_id (expr
);
3071 vn_ssa_aux_t vn_info
= VN_INFO (name
);
3072 vn_info
->value_id
= value_id
;
3073 vn_info
->valnum
= vn_valnum_from_value_id (value_id
);
3074 if (vn_info
->valnum
== NULL_TREE
)
3075 vn_info
->valnum
= name
;
3076 gcc_assert (vn_info
->valnum
!= NULL_TREE
);
3077 nameexpr
= get_or_alloc_expr_for_name (name
);
3078 add_to_value (value_id
, nameexpr
);
3079 if (NEW_SETS (block
))
3080 bitmap_value_replace_in_set (NEW_SETS (block
), nameexpr
);
3081 bitmap_value_replace_in_set (AVAIL_OUT (block
), nameexpr
);
3083 pre_stats
.insertions
++;
3084 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3086 fprintf (dump_file
, "Inserted ");
3087 print_gimple_stmt (dump_file
, gsi_stmt (gsi_last (*stmts
)), 0);
3088 fprintf (dump_file
, " in predecessor %d (%04d)\n",
3089 block
->index
, value_id
);
3096 /* Insert the to-be-made-available values of expression EXPRNUM for each
3097 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3098 merge the result with a phi node, given the same value number as
3099 NODE. Return true if we have inserted new stuff. */
3102 insert_into_preds_of_block (basic_block block
, unsigned int exprnum
,
3103 vec
<pre_expr
> &avail
)
3105 pre_expr expr
= expression_for_id (exprnum
);
3107 unsigned int val
= get_expr_value_id (expr
);
3109 bool insertions
= false;
3114 tree type
= get_expr_type (expr
);
3118 /* Make sure we aren't creating an induction variable. */
3119 if (bb_loop_depth (block
) > 0 && EDGE_COUNT (block
->preds
) == 2)
3121 bool firstinsideloop
= false;
3122 bool secondinsideloop
= false;
3123 firstinsideloop
= flow_bb_inside_loop_p (block
->loop_father
,
3124 EDGE_PRED (block
, 0)->src
);
3125 secondinsideloop
= flow_bb_inside_loop_p (block
->loop_father
,
3126 EDGE_PRED (block
, 1)->src
);
3127 /* Induction variables only have one edge inside the loop. */
3128 if ((firstinsideloop
^ secondinsideloop
)
3129 && expr
->kind
!= REFERENCE
)
3131 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3132 fprintf (dump_file
, "Skipping insertion of phi for partial "
3133 "redundancy: Looks like an induction variable\n");
3138 /* Make the necessary insertions. */
3139 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3141 /* When we are not inserting a PHI node do not bother inserting
3142 into places that do not dominate the anticipated computations. */
3143 if (nophi
&& !dominated_by_p (CDI_DOMINATORS
, block
, pred
->src
))
3145 gimple_seq stmts
= NULL
;
3148 eprime
= avail
[pred
->dest_idx
];
3149 builtexpr
= create_expression_by_pieces (bprime
, eprime
,
3151 gcc_assert (!(pred
->flags
& EDGE_ABNORMAL
));
3152 if (!gimple_seq_empty_p (stmts
))
3154 basic_block new_bb
= gsi_insert_seq_on_edge_immediate (pred
, stmts
);
3155 gcc_assert (! new_bb
);
3160 /* We cannot insert a PHI node if we failed to insert
3165 if (is_gimple_min_invariant (builtexpr
))
3166 avail
[pred
->dest_idx
] = get_or_alloc_expr_for_constant (builtexpr
);
3168 avail
[pred
->dest_idx
] = get_or_alloc_expr_for_name (builtexpr
);
3170 /* If we didn't want a phi node, and we made insertions, we still have
3171 inserted new stuff, and thus return true. If we didn't want a phi node,
3172 and didn't make insertions, we haven't added anything new, so return
3174 if (nophi
&& insertions
)
3176 else if (nophi
&& !insertions
)
3179 /* Now build a phi for the new variable. */
3180 temp
= make_temp_ssa_name (type
, NULL
, "prephitmp");
3181 phi
= create_phi_node (temp
, block
);
3183 vn_ssa_aux_t vn_info
= VN_INFO (temp
);
3184 vn_info
->value_id
= val
;
3185 vn_info
->valnum
= vn_valnum_from_value_id (val
);
3186 if (vn_info
->valnum
== NULL_TREE
)
3187 vn_info
->valnum
= temp
;
3188 bitmap_set_bit (inserted_exprs
, SSA_NAME_VERSION (temp
));
3189 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3191 pre_expr ae
= avail
[pred
->dest_idx
];
3192 gcc_assert (get_expr_type (ae
) == type
3193 || useless_type_conversion_p (type
, get_expr_type (ae
)));
3194 if (ae
->kind
== CONSTANT
)
3195 add_phi_arg (phi
, unshare_expr (PRE_EXPR_CONSTANT (ae
)),
3196 pred
, UNKNOWN_LOCATION
);
3198 add_phi_arg (phi
, PRE_EXPR_NAME (ae
), pred
, UNKNOWN_LOCATION
);
3201 newphi
= get_or_alloc_expr_for_name (temp
);
3202 add_to_value (val
, newphi
);
3204 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3205 this insertion, since we test for the existence of this value in PHI_GEN
3206 before proceeding with the partial redundancy checks in insert_aux.
3208 The value may exist in AVAIL_OUT, in particular, it could be represented
3209 by the expression we are trying to eliminate, in which case we want the
3210 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3213 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3214 this block, because if it did, it would have existed in our dominator's
3215 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3218 bitmap_insert_into_set (PHI_GEN (block
), newphi
);
3219 bitmap_value_replace_in_set (AVAIL_OUT (block
),
3221 if (NEW_SETS (block
))
3222 bitmap_insert_into_set (NEW_SETS (block
), newphi
);
3224 /* If we insert a PHI node for a conversion of another PHI node
3225 in the same basic-block try to preserve range information.
3226 This is important so that followup loop passes receive optimal
3227 number of iteration analysis results. See PR61743. */
3228 if (expr
->kind
== NARY
3229 && CONVERT_EXPR_CODE_P (expr
->u
.nary
->opcode
)
3230 && TREE_CODE (expr
->u
.nary
->op
[0]) == SSA_NAME
3231 && gimple_bb (SSA_NAME_DEF_STMT (expr
->u
.nary
->op
[0])) == block
3232 && INTEGRAL_TYPE_P (type
)
3233 && INTEGRAL_TYPE_P (TREE_TYPE (expr
->u
.nary
->op
[0]))
3234 && (TYPE_PRECISION (type
)
3235 >= TYPE_PRECISION (TREE_TYPE (expr
->u
.nary
->op
[0])))
3236 && SSA_NAME_RANGE_INFO (expr
->u
.nary
->op
[0]))
3239 if (get_range_query (cfun
)->range_of_expr (r
, expr
->u
.nary
->op
[0])
3240 && r
.kind () == VR_RANGE
3241 && !wi::neg_p (r
.lower_bound (), SIGNED
)
3242 && !wi::neg_p (r
.upper_bound (), SIGNED
))
3244 /* Just handle extension and sign-changes of all-positive ranges. */
3245 range_cast (r
, type
);
3246 set_range_info (temp
, r
);
3250 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3252 fprintf (dump_file
, "Created phi ");
3253 print_gimple_stmt (dump_file
, phi
, 0);
3254 fprintf (dump_file
, " in block %d (%04d)\n", block
->index
, val
);
3262 /* Perform insertion of partially redundant or hoistable values.
3263 For BLOCK, do the following:
3264 1. Propagate the NEW_SETS of the dominator into the current block.
3265 If the block has multiple predecessors,
3266 2a. Iterate over the ANTIC expressions for the block to see if
3267 any of them are partially redundant.
3268 2b. If so, insert them into the necessary predecessors to make
3269 the expression fully redundant.
3270 2c. Insert a new PHI merging the values of the predecessors.
3271 2d. Insert the new PHI, and the new expressions, into the
3273 If the block has multiple successors,
3274 3a. Iterate over the ANTIC values for the block to see if
3275 any of them are good candidates for hoisting.
3276 3b. If so, insert expressions computing the values in BLOCK,
3277 and add the new expressions into the NEW_SETS set.
3278 4. Recursively call ourselves on the dominator children of BLOCK.
3280 Steps 1, 2a, and 4 are done by insert_aux. 2b, 2c and 2d are done by
3281 do_pre_regular_insertion and do_partial_insertion. 3a and 3b are
3282 done in do_hoist_insertion.
3286 do_pre_regular_insertion (basic_block block
, basic_block dom
,
3287 vec
<pre_expr
> exprs
)
3289 bool new_stuff
= false;
3291 auto_vec
<pre_expr
, 2> avail
;
3294 avail
.safe_grow (EDGE_COUNT (block
->preds
), true);
3296 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
3298 if (expr
->kind
== NARY
3299 || expr
->kind
== REFERENCE
)
3302 bool by_some
= false;
3303 bool cant_insert
= false;
3304 bool all_same
= true;
3305 pre_expr first_s
= NULL
;
3308 pre_expr eprime
= NULL
;
3310 pre_expr edoubleprime
= NULL
;
3311 bool do_insertion
= false;
3313 val
= get_expr_value_id (expr
);
3314 if (bitmap_set_contains_value (PHI_GEN (block
), val
))
3316 if (bitmap_set_contains_value (AVAIL_OUT (dom
), val
))
3318 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3320 fprintf (dump_file
, "Found fully redundant value: ");
3321 print_pre_expr (dump_file
, expr
);
3322 fprintf (dump_file
, "\n");
3327 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3329 unsigned int vprime
;
3331 /* We should never run insertion for the exit block
3332 and so not come across fake pred edges. */
3333 gcc_assert (!(pred
->flags
& EDGE_FAKE
));
3335 /* We are looking at ANTIC_OUT of bprime. */
3336 eprime
= phi_translate (NULL
, expr
, ANTIC_IN (block
), NULL
, pred
);
3338 /* eprime will generally only be NULL if the
3339 value of the expression, translated
3340 through the PHI for this predecessor, is
3341 undefined. If that is the case, we can't
3342 make the expression fully redundant,
3343 because its value is undefined along a
3344 predecessor path. We can thus break out
3345 early because it doesn't matter what the
3346 rest of the results are. */
3349 avail
[pred
->dest_idx
] = NULL
;
3354 vprime
= get_expr_value_id (eprime
);
3355 edoubleprime
= bitmap_find_leader (AVAIL_OUT (bprime
),
3357 if (edoubleprime
== NULL
)
3359 avail
[pred
->dest_idx
] = eprime
;
3364 avail
[pred
->dest_idx
] = edoubleprime
;
3366 /* We want to perform insertions to remove a redundancy on
3367 a path in the CFG we want to optimize for speed. */
3368 if (optimize_edge_for_speed_p (pred
))
3369 do_insertion
= true;
3370 if (first_s
== NULL
)
3371 first_s
= edoubleprime
;
3372 else if (!pre_expr_d::equal (first_s
, edoubleprime
))
3376 /* If we can insert it, it's not the same value
3377 already existing along every predecessor, and
3378 it's defined by some predecessor, it is
3379 partially redundant. */
3380 if (!cant_insert
&& !all_same
&& by_some
)
3384 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3386 fprintf (dump_file
, "Skipping partial redundancy for "
3388 print_pre_expr (dump_file
, expr
);
3389 fprintf (dump_file
, " (%04d), no redundancy on to be "
3390 "optimized for speed edge\n", val
);
3393 else if (dbg_cnt (treepre_insert
))
3395 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3397 fprintf (dump_file
, "Found partial redundancy for "
3399 print_pre_expr (dump_file
, expr
);
3400 fprintf (dump_file
, " (%04d)\n",
3401 get_expr_value_id (expr
));
3403 if (insert_into_preds_of_block (block
,
3404 get_expression_id (expr
),
3409 /* If all edges produce the same value and that value is
3410 an invariant, then the PHI has the same value on all
3411 edges. Note this. */
3412 else if (!cant_insert
3414 && (edoubleprime
->kind
!= NAME
3415 || !SSA_NAME_OCCURS_IN_ABNORMAL_PHI
3416 (PRE_EXPR_NAME (edoubleprime
))))
3418 gcc_assert (edoubleprime
->kind
== CONSTANT
3419 || edoubleprime
->kind
== NAME
);
3421 tree temp
= make_temp_ssa_name (get_expr_type (expr
),
3424 = gimple_build_assign (temp
,
3425 edoubleprime
->kind
== CONSTANT
?
3426 PRE_EXPR_CONSTANT (edoubleprime
) :
3427 PRE_EXPR_NAME (edoubleprime
));
3428 gimple_stmt_iterator gsi
= gsi_after_labels (block
);
3429 gsi_insert_before (&gsi
, assign
, GSI_NEW_STMT
);
3431 vn_ssa_aux_t vn_info
= VN_INFO (temp
);
3432 vn_info
->value_id
= val
;
3433 vn_info
->valnum
= vn_valnum_from_value_id (val
);
3434 if (vn_info
->valnum
== NULL_TREE
)
3435 vn_info
->valnum
= temp
;
3436 bitmap_set_bit (inserted_exprs
, SSA_NAME_VERSION (temp
));
3437 pre_expr newe
= get_or_alloc_expr_for_name (temp
);
3438 add_to_value (val
, newe
);
3439 bitmap_value_replace_in_set (AVAIL_OUT (block
), newe
);
3440 bitmap_insert_into_set (NEW_SETS (block
), newe
);
3441 bitmap_insert_into_set (PHI_GEN (block
), newe
);
3450 /* Perform insertion for partially anticipatable expressions. There
3451 is only one case we will perform insertion for these. This case is
3452 if the expression is partially anticipatable, and fully available.
3453 In this case, we know that putting it earlier will enable us to
3454 remove the later computation. */
3457 do_pre_partial_partial_insertion (basic_block block
, basic_block dom
,
3458 vec
<pre_expr
> exprs
)
3460 bool new_stuff
= false;
3462 auto_vec
<pre_expr
, 2> avail
;
3465 avail
.safe_grow (EDGE_COUNT (block
->preds
), true);
3467 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
3469 if (expr
->kind
== NARY
3470 || expr
->kind
== REFERENCE
)
3474 bool cant_insert
= false;
3477 pre_expr eprime
= NULL
;
3480 val
= get_expr_value_id (expr
);
3481 if (bitmap_set_contains_value (PHI_GEN (block
), val
))
3483 if (bitmap_set_contains_value (AVAIL_OUT (dom
), val
))
3486 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3488 unsigned int vprime
;
3489 pre_expr edoubleprime
;
3491 /* We should never run insertion for the exit block
3492 and so not come across fake pred edges. */
3493 gcc_assert (!(pred
->flags
& EDGE_FAKE
));
3495 eprime
= phi_translate (NULL
, expr
, ANTIC_IN (block
),
3496 PA_IN (block
), pred
);
3498 /* eprime will generally only be NULL if the
3499 value of the expression, translated
3500 through the PHI for this predecessor, is
3501 undefined. If that is the case, we can't
3502 make the expression fully redundant,
3503 because its value is undefined along a
3504 predecessor path. We can thus break out
3505 early because it doesn't matter what the
3506 rest of the results are. */
3509 avail
[pred
->dest_idx
] = NULL
;
3514 vprime
= get_expr_value_id (eprime
);
3515 edoubleprime
= bitmap_find_leader (AVAIL_OUT (bprime
), vprime
);
3516 avail
[pred
->dest_idx
] = edoubleprime
;
3517 if (edoubleprime
== NULL
)
3524 /* If we can insert it, it's not the same value
3525 already existing along every predecessor, and
3526 it's defined by some predecessor, it is
3527 partially redundant. */
3528 if (!cant_insert
&& by_all
)
3531 bool do_insertion
= false;
3533 /* Insert only if we can remove a later expression on a path
3534 that we want to optimize for speed.
3535 The phi node that we will be inserting in BLOCK is not free,
3536 and inserting it for the sake of !optimize_for_speed successor
3537 may cause regressions on the speed path. */
3538 FOR_EACH_EDGE (succ
, ei
, block
->succs
)
3540 if (bitmap_set_contains_value (PA_IN (succ
->dest
), val
)
3541 || bitmap_set_contains_value (ANTIC_IN (succ
->dest
), val
))
3543 if (optimize_edge_for_speed_p (succ
))
3544 do_insertion
= true;
3550 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3552 fprintf (dump_file
, "Skipping partial partial redundancy "
3554 print_pre_expr (dump_file
, expr
);
3555 fprintf (dump_file
, " (%04d), not (partially) anticipated "
3556 "on any to be optimized for speed edges\n", val
);
3559 else if (dbg_cnt (treepre_insert
))
3561 pre_stats
.pa_insert
++;
3562 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3564 fprintf (dump_file
, "Found partial partial redundancy "
3566 print_pre_expr (dump_file
, expr
);
3567 fprintf (dump_file
, " (%04d)\n",
3568 get_expr_value_id (expr
));
3570 if (insert_into_preds_of_block (block
,
3571 get_expression_id (expr
),
3582 /* Insert expressions in BLOCK to compute hoistable values up.
3583 Return TRUE if something was inserted, otherwise return FALSE.
3584 The caller has to make sure that BLOCK has at least two successors. */
3587 do_hoist_insertion (basic_block block
)
3591 bool new_stuff
= false;
3593 gimple_stmt_iterator last
;
3595 /* At least two successors, or else... */
3596 gcc_assert (EDGE_COUNT (block
->succs
) >= 2);
3598 /* Check that all successors of BLOCK are dominated by block.
3599 We could use dominated_by_p() for this, but actually there is a much
3600 quicker check: any successor that is dominated by BLOCK can't have
3601 more than one predecessor edge. */
3602 FOR_EACH_EDGE (e
, ei
, block
->succs
)
3603 if (! single_pred_p (e
->dest
))
3606 /* Determine the insertion point. If we cannot safely insert before
3607 the last stmt if we'd have to, bail out. */
3608 last
= gsi_last_bb (block
);
3609 if (!gsi_end_p (last
)
3610 && !is_ctrl_stmt (gsi_stmt (last
))
3611 && stmt_ends_bb_p (gsi_stmt (last
)))
3614 /* Compute the set of hoistable expressions from ANTIC_IN. First compute
3615 hoistable values. */
3616 bitmap_set hoistable_set
;
3618 /* A hoistable value must be in ANTIC_IN(block)
3619 but not in AVAIL_OUT(BLOCK). */
3620 bitmap_initialize (&hoistable_set
.values
, &grand_bitmap_obstack
);
3621 bitmap_and_compl (&hoistable_set
.values
,
3622 &ANTIC_IN (block
)->values
, &AVAIL_OUT (block
)->values
);
3624 /* Short-cut for a common case: hoistable_set is empty. */
3625 if (bitmap_empty_p (&hoistable_set
.values
))
3628 /* Compute which of the hoistable values is in AVAIL_OUT of
3629 at least one of the successors of BLOCK. */
3630 bitmap_head availout_in_some
;
3631 bitmap_initialize (&availout_in_some
, &grand_bitmap_obstack
);
3632 FOR_EACH_EDGE (e
, ei
, block
->succs
)
3633 /* Do not consider expressions solely because their availability
3634 on loop exits. They'd be ANTIC-IN throughout the whole loop
3635 and thus effectively hoisted across loops by combination of
3636 PRE and hoisting. */
3637 if (! loop_exit_edge_p (block
->loop_father
, e
))
3638 bitmap_ior_and_into (&availout_in_some
, &hoistable_set
.values
,
3639 &AVAIL_OUT (e
->dest
)->values
);
3640 bitmap_clear (&hoistable_set
.values
);
3642 /* Short-cut for a common case: availout_in_some is empty. */
3643 if (bitmap_empty_p (&availout_in_some
))
3646 /* Hack hoitable_set in-place so we can use sorted_array_from_bitmap_set. */
3647 bitmap_move (&hoistable_set
.values
, &availout_in_some
);
3648 hoistable_set
.expressions
= ANTIC_IN (block
)->expressions
;
3650 /* Now finally construct the topological-ordered expression set. */
3651 vec
<pre_expr
> exprs
= sorted_array_from_bitmap_set (&hoistable_set
);
3653 bitmap_clear (&hoistable_set
.values
);
3655 /* If there are candidate values for hoisting, insert expressions
3656 strategically to make the hoistable expressions fully redundant. */
3658 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
3660 /* While we try to sort expressions topologically above the
3661 sorting doesn't work out perfectly. Catch expressions we
3662 already inserted. */
3663 unsigned int value_id
= get_expr_value_id (expr
);
3664 if (bitmap_set_contains_value (AVAIL_OUT (block
), value_id
))
3666 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3669 "Already inserted expression for ");
3670 print_pre_expr (dump_file
, expr
);
3671 fprintf (dump_file
, " (%04d)\n", value_id
);
3676 /* If we end up with a punned expression representation and this
3677 happens to be a float typed one give up - we can't know for
3678 sure whether all paths perform the floating-point load we are
3679 about to insert and on some targets this can cause correctness
3680 issues. See PR88240. */
3681 if (expr
->kind
== REFERENCE
3682 && PRE_EXPR_REFERENCE (expr
)->punned
3683 && FLOAT_TYPE_P (get_expr_type (expr
)))
3686 /* OK, we should hoist this value. Perform the transformation. */
3687 pre_stats
.hoist_insert
++;
3688 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3691 "Inserting expression in block %d for code hoisting: ",
3693 print_pre_expr (dump_file
, expr
);
3694 fprintf (dump_file
, " (%04d)\n", value_id
);
3697 gimple_seq stmts
= NULL
;
3698 tree res
= create_expression_by_pieces (block
, expr
, &stmts
,
3699 get_expr_type (expr
));
3701 /* Do not return true if expression creation ultimately
3702 did not insert any statements. */
3703 if (gimple_seq_empty_p (stmts
))
3707 if (gsi_end_p (last
) || is_ctrl_stmt (gsi_stmt (last
)))
3708 gsi_insert_seq_before (&last
, stmts
, GSI_SAME_STMT
);
3710 gsi_insert_seq_after (&last
, stmts
, GSI_NEW_STMT
);
3713 /* Make sure to not return true if expression creation ultimately
3714 failed but also make sure to insert any stmts produced as they
3715 are tracked in inserted_exprs. */
3727 /* Perform insertion of partially redundant and hoistable values. */
3734 FOR_ALL_BB_FN (bb
, cfun
)
3735 NEW_SETS (bb
) = bitmap_set_new ();
3737 int *rpo
= XNEWVEC (int, n_basic_blocks_for_fn (cfun
));
3738 int *bb_rpo
= XNEWVEC (int, last_basic_block_for_fn (cfun
) + 1);
3739 int rpo_num
= pre_and_rev_post_order_compute (NULL
, rpo
, false);
3740 for (int i
= 0; i
< rpo_num
; ++i
)
3743 int num_iterations
= 0;
3748 if (dump_file
&& dump_flags
& TDF_DETAILS
)
3749 fprintf (dump_file
, "Starting insert iteration %d\n", num_iterations
);
3752 for (int idx
= 0; idx
< rpo_num
; ++idx
)
3754 basic_block block
= BASIC_BLOCK_FOR_FN (cfun
, rpo
[idx
]);
3755 basic_block dom
= get_immediate_dominator (CDI_DOMINATORS
, block
);
3760 bitmap_set_t newset
;
3762 /* First, update the AVAIL_OUT set with anything we may have
3763 inserted higher up in the dominator tree. */
3764 newset
= NEW_SETS (dom
);
3766 /* Note that we need to value_replace both NEW_SETS, and
3767 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3768 represented by some non-simple expression here that we want
3769 to replace it with. */
3770 bool avail_out_changed
= false;
3771 FOR_EACH_EXPR_ID_IN_SET (newset
, i
, bi
)
3773 pre_expr expr
= expression_for_id (i
);
3774 bitmap_value_replace_in_set (NEW_SETS (block
), expr
);
3776 |= bitmap_value_replace_in_set (AVAIL_OUT (block
), expr
);
3778 /* We need to iterate if AVAIL_OUT of an already processed
3779 block source changed. */
3780 if (avail_out_changed
&& !changed
)
3784 FOR_EACH_EDGE (e
, ei
, block
->succs
)
3785 if (e
->dest
->index
!= EXIT_BLOCK
3786 && bb_rpo
[e
->dest
->index
] < idx
)
3790 /* Insert expressions for partial redundancies. */
3791 if (flag_tree_pre
&& !single_pred_p (block
))
3794 = sorted_array_from_bitmap_set (ANTIC_IN (block
));
3795 /* Sorting is not perfect, iterate locally. */
3796 while (do_pre_regular_insertion (block
, dom
, exprs
))
3799 if (do_partial_partial
)
3801 exprs
= sorted_array_from_bitmap_set (PA_IN (block
));
3802 while (do_pre_partial_partial_insertion (block
, dom
,
3811 /* Clear the NEW sets before the next iteration. We have already
3812 fully propagated its contents. */
3814 FOR_ALL_BB_FN (bb
, cfun
)
3815 bitmap_set_free (NEW_SETS (bb
));
3819 statistics_histogram_event (cfun
, "insert iterations", num_iterations
);
3821 /* AVAIL_OUT is not needed after insertion so we don't have to
3822 propagate NEW_SETS from hoist insertion. */
3823 FOR_ALL_BB_FN (bb
, cfun
)
3825 bitmap_set_free (NEW_SETS (bb
));
3826 bitmap_set_pool
.remove (NEW_SETS (bb
));
3827 NEW_SETS (bb
) = NULL
;
3830 /* Insert expressions for hoisting. Do a backward walk here since
3831 inserting into BLOCK exposes new opportunities in its predecessors.
3832 Since PRE and hoist insertions can cause back-to-back iteration
3833 and we are interested in PRE insertion exposed hoisting opportunities
3834 but not in hoisting exposed PRE ones do hoist insertion only after
3835 PRE insertion iteration finished and do not iterate it. */
3836 if (flag_code_hoisting
)
3837 for (int idx
= rpo_num
- 1; idx
>= 0; --idx
)
3839 basic_block block
= BASIC_BLOCK_FOR_FN (cfun
, rpo
[idx
]);
3840 if (EDGE_COUNT (block
->succs
) >= 2)
3841 changed
|= do_hoist_insertion (block
);
3849 /* Compute the AVAIL set for all basic blocks.
3851 This function performs value numbering of the statements in each basic
3852 block. The AVAIL sets are built from information we glean while doing
3853 this value numbering, since the AVAIL sets contain only one entry per
3856 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3857 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3860 compute_avail (function
*fun
)
3863 basic_block block
, son
;
3864 basic_block
*worklist
;
3869 /* We pretend that default definitions are defined in the entry block.
3870 This includes function arguments and the static chain decl. */
3871 FOR_EACH_SSA_NAME (i
, name
, fun
)
3874 if (!SSA_NAME_IS_DEFAULT_DEF (name
)
3875 || has_zero_uses (name
)
3876 || virtual_operand_p (name
))
3879 e
= get_or_alloc_expr_for_name (name
);
3880 add_to_value (get_expr_value_id (e
), e
);
3881 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (fun
)), e
);
3882 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (fun
)),
3886 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3888 print_bitmap_set (dump_file
, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (fun
)),
3889 "tmp_gen", ENTRY_BLOCK
);
3890 print_bitmap_set (dump_file
, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (fun
)),
3891 "avail_out", ENTRY_BLOCK
);
3894 /* Allocate the worklist. */
3895 worklist
= XNEWVEC (basic_block
, n_basic_blocks_for_fn (fun
));
3897 /* Seed the algorithm by putting the dominator children of the entry
3898 block on the worklist. */
3899 for (son
= first_dom_son (CDI_DOMINATORS
, ENTRY_BLOCK_PTR_FOR_FN (fun
));
3901 son
= next_dom_son (CDI_DOMINATORS
, son
))
3902 worklist
[sp
++] = son
;
3904 BB_LIVE_VOP_ON_EXIT (ENTRY_BLOCK_PTR_FOR_FN (fun
))
3905 = ssa_default_def (fun
, gimple_vop (fun
));
3907 /* Loop until the worklist is empty. */
3913 /* Pick a block from the worklist. */
3914 block
= worklist
[--sp
];
3915 vn_context_bb
= block
;
3917 /* Initially, the set of available values in BLOCK is that of
3918 its immediate dominator. */
3919 dom
= get_immediate_dominator (CDI_DOMINATORS
, block
);
3922 bitmap_set_copy (AVAIL_OUT (block
), AVAIL_OUT (dom
));
3923 BB_LIVE_VOP_ON_EXIT (block
) = BB_LIVE_VOP_ON_EXIT (dom
);
3926 /* Generate values for PHI nodes. */
3927 for (gphi_iterator gsi
= gsi_start_phis (block
); !gsi_end_p (gsi
);
3930 tree result
= gimple_phi_result (gsi
.phi ());
3932 /* We have no need for virtual phis, as they don't represent
3933 actual computations. */
3934 if (virtual_operand_p (result
))
3936 BB_LIVE_VOP_ON_EXIT (block
) = result
;
3940 pre_expr e
= get_or_alloc_expr_for_name (result
);
3941 add_to_value (get_expr_value_id (e
), e
);
3942 bitmap_value_insert_into_set (AVAIL_OUT (block
), e
);
3943 bitmap_insert_into_set (PHI_GEN (block
), e
);
3946 BB_MAY_NOTRETURN (block
) = 0;
3948 /* Now compute value numbers and populate value sets with all
3949 the expressions computed in BLOCK. */
3950 bool set_bb_may_notreturn
= false;
3951 for (gimple_stmt_iterator gsi
= gsi_start_bb (block
); !gsi_end_p (gsi
);
3957 stmt
= gsi_stmt (gsi
);
3959 if (set_bb_may_notreturn
)
3961 BB_MAY_NOTRETURN (block
) = 1;
3962 set_bb_may_notreturn
= false;
3965 /* Cache whether the basic-block has any non-visible side-effect
3967 If this isn't a call or it is the last stmt in the
3968 basic-block then the CFG represents things correctly. */
3969 if (is_gimple_call (stmt
) && !stmt_ends_bb_p (stmt
))
3971 /* Non-looping const functions always return normally.
3972 Otherwise the call might not return or have side-effects
3973 that forbids hoisting possibly trapping expressions
3975 int flags
= gimple_call_flags (stmt
);
3976 if (!(flags
& (ECF_CONST
|ECF_PURE
))
3977 || (flags
& ECF_LOOPING_CONST_OR_PURE
)
3978 || stmt_can_throw_external (fun
, stmt
))
3979 /* Defer setting of BB_MAY_NOTRETURN to avoid it
3980 influencing the processing of the call itself. */
3981 set_bb_may_notreturn
= true;
3984 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_DEF
)
3986 pre_expr e
= get_or_alloc_expr_for_name (op
);
3987 add_to_value (get_expr_value_id (e
), e
);
3988 bitmap_insert_into_set (TMP_GEN (block
), e
);
3989 bitmap_value_insert_into_set (AVAIL_OUT (block
), e
);
3992 if (gimple_vdef (stmt
))
3993 BB_LIVE_VOP_ON_EXIT (block
) = gimple_vdef (stmt
);
3995 if (gimple_has_side_effects (stmt
)
3996 || stmt_could_throw_p (fun
, stmt
)
3997 || is_gimple_debug (stmt
))
4000 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_USE
)
4002 if (ssa_undefined_value_p (op
))
4004 pre_expr e
= get_or_alloc_expr_for_name (op
);
4005 bitmap_value_insert_into_set (EXP_GEN (block
), e
);
4008 switch (gimple_code (stmt
))
4016 vn_reference_s ref1
;
4017 pre_expr result
= NULL
;
4019 vn_reference_lookup_call (as_a
<gcall
*> (stmt
), &ref
, &ref1
);
4020 /* There is no point to PRE a call without a value. */
4021 if (!ref
|| !ref
->result
)
4024 /* If the value of the call is not invalidated in
4025 this block until it is computed, add the expression
4027 if ((!gimple_vuse (stmt
)
4029 (SSA_NAME_DEF_STMT (gimple_vuse (stmt
))) == GIMPLE_PHI
4030 || gimple_bb (SSA_NAME_DEF_STMT
4031 (gimple_vuse (stmt
))) != block
)
4032 /* If the REFERENCE traps and there was a preceding
4033 point in the block that might not return avoid
4034 adding the reference to EXP_GEN. */
4035 && (!BB_MAY_NOTRETURN (block
)
4036 || !vn_reference_may_trap (ref
)))
4038 result
= get_or_alloc_expr_for_reference
4039 (ref
, gimple_location (stmt
));
4040 add_to_value (get_expr_value_id (result
), result
);
4041 bitmap_value_insert_into_set (EXP_GEN (block
), result
);
4048 pre_expr result
= NULL
;
4049 switch (vn_get_stmt_kind (stmt
))
4053 enum tree_code code
= gimple_assign_rhs_code (stmt
);
4056 /* COND_EXPR is awkward in that it contains an
4057 embedded complex expression.
4058 Don't even try to shove it through PRE. */
4059 if (code
== COND_EXPR
)
4062 vn_nary_op_lookup_stmt (stmt
, &nary
);
4063 if (!nary
|| nary
->predicated_values
)
4066 unsigned value_id
= nary
->value_id
;
4067 if (value_id_constant_p (value_id
))
4070 /* Record the un-valueized expression for EXP_GEN. */
4071 nary
= XALLOCAVAR (struct vn_nary_op_s
,
4073 (vn_nary_length_from_stmt (stmt
)));
4074 init_vn_nary_op_from_stmt (nary
, as_a
<gassign
*> (stmt
));
4076 /* If the NARY traps and there was a preceding
4077 point in the block that might not return avoid
4078 adding the nary to EXP_GEN. */
4079 if (BB_MAY_NOTRETURN (block
)
4080 && vn_nary_may_trap (nary
))
4083 result
= get_or_alloc_expr_for_nary
4084 (nary
, value_id
, gimple_location (stmt
));
4090 tree rhs1
= gimple_assign_rhs1 (stmt
);
4092 ao_ref_init (&rhs1_ref
, rhs1
);
4093 alias_set_type set
= ao_ref_alias_set (&rhs1_ref
);
4094 alias_set_type base_set
4095 = ao_ref_base_alias_set (&rhs1_ref
);
4096 vec
<vn_reference_op_s
> operands
4097 = vn_reference_operands_for_lookup (rhs1
);
4099 vn_reference_lookup_pieces (gimple_vuse (stmt
), set
,
4100 base_set
, TREE_TYPE (rhs1
),
4101 operands
, &ref
, VN_WALK
);
4104 operands
.release ();
4108 /* If the REFERENCE traps and there was a preceding
4109 point in the block that might not return avoid
4110 adding the reference to EXP_GEN. */
4111 if (BB_MAY_NOTRETURN (block
)
4112 && vn_reference_may_trap (ref
))
4114 operands
.release ();
4118 /* If the value of the reference is not invalidated in
4119 this block until it is computed, add the expression
4121 if (gimple_vuse (stmt
))
4125 def_stmt
= SSA_NAME_DEF_STMT (gimple_vuse (stmt
));
4126 while (!gimple_nop_p (def_stmt
)
4127 && gimple_code (def_stmt
) != GIMPLE_PHI
4128 && gimple_bb (def_stmt
) == block
)
4130 if (stmt_may_clobber_ref_p
4131 (def_stmt
, gimple_assign_rhs1 (stmt
)))
4137 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt
));
4141 operands
.release ();
4146 /* If the load was value-numbered to another
4147 load make sure we do not use its expression
4148 for insertion if it wouldn't be a valid
4150 /* At the momemt we have a testcase
4151 for hoist insertion of aligned vs. misaligned
4152 variants in gcc.dg/torture/pr65270-1.c thus
4153 with just alignment to be considered we can
4154 simply replace the expression in the hashtable
4155 with the most conservative one. */
4156 vn_reference_op_t ref1
= &ref
->operands
.last ();
4157 while (ref1
->opcode
!= TARGET_MEM_REF
4158 && ref1
->opcode
!= MEM_REF
4159 && ref1
!= &ref
->operands
[0])
4161 vn_reference_op_t ref2
= &operands
.last ();
4162 while (ref2
->opcode
!= TARGET_MEM_REF
4163 && ref2
->opcode
!= MEM_REF
4164 && ref2
!= &operands
[0])
4166 if ((ref1
->opcode
== TARGET_MEM_REF
4167 || ref1
->opcode
== MEM_REF
)
4168 && (TYPE_ALIGN (ref1
->type
)
4169 > TYPE_ALIGN (ref2
->type
)))
4171 = build_aligned_type (ref1
->type
,
4172 TYPE_ALIGN (ref2
->type
));
4173 /* TBAA behavior is an obvious part so make sure
4174 that the hashtable one covers this as well
4175 by adjusting the ref alias set and its base. */
4177 || alias_set_subset_of (set
, ref
->set
))
4179 else if (ref1
->opcode
!= ref2
->opcode
4180 || (ref1
->opcode
!= MEM_REF
4181 && ref1
->opcode
!= TARGET_MEM_REF
))
4183 /* With mismatching base opcodes or bases
4184 other than MEM_REF or TARGET_MEM_REF we
4185 can't do any easy TBAA adjustment. */
4186 operands
.release ();
4189 else if (alias_set_subset_of (ref
->set
, set
))
4192 if (ref1
->opcode
== MEM_REF
)
4194 = wide_int_to_tree (TREE_TYPE (ref2
->op0
),
4195 wi::to_wide (ref1
->op0
));
4198 = wide_int_to_tree (TREE_TYPE (ref2
->op2
),
4199 wi::to_wide (ref1
->op2
));
4204 if (ref1
->opcode
== MEM_REF
)
4206 = wide_int_to_tree (ptr_type_node
,
4207 wi::to_wide (ref1
->op0
));
4210 = wide_int_to_tree (ptr_type_node
,
4211 wi::to_wide (ref1
->op2
));
4213 operands
.release ();
4215 result
= get_or_alloc_expr_for_reference
4216 (ref
, gimple_location (stmt
));
4224 add_to_value (get_expr_value_id (result
), result
);
4225 bitmap_value_insert_into_set (EXP_GEN (block
), result
);
4232 if (set_bb_may_notreturn
)
4234 BB_MAY_NOTRETURN (block
) = 1;
4235 set_bb_may_notreturn
= false;
4238 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4240 print_bitmap_set (dump_file
, EXP_GEN (block
),
4241 "exp_gen", block
->index
);
4242 print_bitmap_set (dump_file
, PHI_GEN (block
),
4243 "phi_gen", block
->index
);
4244 print_bitmap_set (dump_file
, TMP_GEN (block
),
4245 "tmp_gen", block
->index
);
4246 print_bitmap_set (dump_file
, AVAIL_OUT (block
),
4247 "avail_out", block
->index
);
4250 /* Put the dominator children of BLOCK on the worklist of blocks
4251 to compute available sets for. */
4252 for (son
= first_dom_son (CDI_DOMINATORS
, block
);
4254 son
= next_dom_son (CDI_DOMINATORS
, son
))
4255 worklist
[sp
++] = son
;
4257 vn_context_bb
= NULL
;
4263 /* Initialize data structures used by PRE. */
4270 next_expression_id
= 1;
4271 expressions
.create (0);
4272 expressions
.safe_push (NULL
);
4273 value_expressions
.create (get_max_value_id () + 1);
4274 value_expressions
.quick_grow_cleared (get_max_value_id () + 1);
4275 constant_value_expressions
.create (get_max_constant_value_id () + 1);
4276 constant_value_expressions
.quick_grow_cleared (get_max_constant_value_id () + 1);
4277 name_to_id
.create (0);
4278 gcc_obstack_init (&pre_expr_obstack
);
4280 inserted_exprs
= BITMAP_ALLOC (NULL
);
4282 connect_infinite_loops_to_exit ();
4283 memset (&pre_stats
, 0, sizeof (pre_stats
));
4285 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets
));
4287 calculate_dominance_info (CDI_DOMINATORS
);
4289 bitmap_obstack_initialize (&grand_bitmap_obstack
);
4290 expression_to_id
= new hash_table
<pre_expr_d
> (num_ssa_names
* 3);
4291 FOR_ALL_BB_FN (bb
, cfun
)
4293 EXP_GEN (bb
) = bitmap_set_new ();
4294 PHI_GEN (bb
) = bitmap_set_new ();
4295 TMP_GEN (bb
) = bitmap_set_new ();
4296 AVAIL_OUT (bb
) = bitmap_set_new ();
4297 PHI_TRANS_TABLE (bb
) = NULL
;
4302 /* Deallocate data structures used by PRE. */
4307 value_expressions
.release ();
4308 constant_value_expressions
.release ();
4309 expressions
.release ();
4310 bitmap_obstack_release (&grand_bitmap_obstack
);
4311 bitmap_set_pool
.release ();
4312 pre_expr_pool
.release ();
4313 delete expression_to_id
;
4314 expression_to_id
= NULL
;
4315 name_to_id
.release ();
4316 obstack_free (&pre_expr_obstack
, NULL
);
4319 FOR_ALL_BB_FN (bb
, cfun
)
4320 if (bb
->aux
&& PHI_TRANS_TABLE (bb
))
4321 delete PHI_TRANS_TABLE (bb
);
4322 free_aux_for_blocks ();
4327 const pass_data pass_data_pre
=
4329 GIMPLE_PASS
, /* type */
4331 OPTGROUP_NONE
, /* optinfo_flags */
4332 TV_TREE_PRE
, /* tv_id */
4333 ( PROP_cfg
| PROP_ssa
), /* properties_required */
4334 0, /* properties_provided */
4335 0, /* properties_destroyed */
4336 TODO_rebuild_alias
, /* todo_flags_start */
4337 0, /* todo_flags_finish */
4340 class pass_pre
: public gimple_opt_pass
4343 pass_pre (gcc::context
*ctxt
)
4344 : gimple_opt_pass (pass_data_pre
, ctxt
)
4347 /* opt_pass methods: */
4348 bool gate (function
*) final override
4349 { return flag_tree_pre
!= 0 || flag_code_hoisting
!= 0; }
4350 unsigned int execute (function
*) final override
;
4352 }; // class pass_pre
4354 /* Valueization hook for RPO VN when we are calling back to it
4355 at ANTIC compute time. */
4358 pre_valueize (tree name
)
4360 if (TREE_CODE (name
) == SSA_NAME
)
4362 tree tem
= VN_INFO (name
)->valnum
;
4363 if (tem
!= VN_TOP
&& tem
!= name
)
4365 if (TREE_CODE (tem
) != SSA_NAME
4366 || SSA_NAME_IS_DEFAULT_DEF (tem
))
4368 /* We create temporary SSA names for representatives that
4369 do not have a definition (yet) but are not default defs either
4370 assume they are fine to use. */
4371 basic_block def_bb
= gimple_bb (SSA_NAME_DEF_STMT (tem
));
4373 || dominated_by_p (CDI_DOMINATORS
, vn_context_bb
, def_bb
))
4375 /* ??? Now we could look for a leader. Ideally we'd somehow
4376 expose RPO VN leaders and get rid of AVAIL_OUT as well... */
4383 pass_pre::execute (function
*fun
)
4385 unsigned int todo
= 0;
4387 do_partial_partial
=
4388 flag_tree_partial_pre
&& optimize_function_for_speed_p (fun
);
4390 /* This has to happen before VN runs because
4391 loop_optimizer_init may create new phis, etc. */
4392 loop_optimizer_init (LOOPS_NORMAL
);
4393 split_edges_for_insertion ();
4395 calculate_dominance_info (CDI_DOMINATORS
);
4397 run_rpo_vn (VN_WALK
);
4401 vn_valueize
= pre_valueize
;
4403 /* Insert can get quite slow on an incredibly large number of basic
4404 blocks due to some quadratic behavior. Until this behavior is
4405 fixed, don't run it when he have an incredibly large number of
4406 bb's. If we aren't going to run insert, there is no point in
4407 computing ANTIC, either, even though it's plenty fast nor do
4408 we require AVAIL. */
4409 if (n_basic_blocks_for_fn (fun
) < 4000)
4411 compute_avail (fun
);
4416 /* Make sure to remove fake edges before committing our inserts.
4417 This makes sure we don't end up with extra critical edges that
4418 we would need to split. */
4419 remove_fake_exit_edges ();
4420 gsi_commit_edge_inserts ();
4422 /* Eliminate folds statements which might (should not...) end up
4423 not keeping virtual operands up-to-date. */
4424 gcc_assert (!need_ssa_update_p (fun
));
4426 statistics_counter_event (fun
, "Insertions", pre_stats
.insertions
);
4427 statistics_counter_event (fun
, "PA inserted", pre_stats
.pa_insert
);
4428 statistics_counter_event (fun
, "HOIST inserted", pre_stats
.hoist_insert
);
4429 statistics_counter_event (fun
, "New PHIs", pre_stats
.phis
);
4431 todo
|= eliminate_with_rpo_vn (inserted_exprs
);
4438 loop_optimizer_finalize ();
4440 /* Perform a CFG cleanup before we run simple_dce_from_worklist since
4441 unreachable code regions will have not up-to-date SSA form which
4443 bool need_crit_edge_split
= false;
4444 if (todo
& TODO_cleanup_cfg
)
4446 cleanup_tree_cfg ();
4447 need_crit_edge_split
= true;
4450 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4451 to insert PHI nodes sometimes, and because value numbering of casts isn't
4452 perfect, we sometimes end up inserting dead code. This simple DCE-like
4453 pass removes any insertions we made that weren't actually used. */
4454 simple_dce_from_worklist (inserted_exprs
);
4455 BITMAP_FREE (inserted_exprs
);
4457 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4458 case we can merge the block with the remaining predecessor of the block.
4460 - call merge_blocks after each tail merge iteration
4461 - call merge_blocks after all tail merge iterations
4462 - mark TODO_cleanup_cfg when necessary. */
4463 todo
|= tail_merge_optimize (need_crit_edge_split
);
4467 /* Tail merging invalidates the virtual SSA web, together with
4468 cfg-cleanup opportunities exposed by PRE this will wreck the
4469 SSA updating machinery. So make sure to run update-ssa
4470 manually, before eventually scheduling cfg-cleanup as part of
4472 update_ssa (TODO_update_ssa_only_virtuals
);
4480 make_pass_pre (gcc::context
*ctxt
)
4482 return new pass_pre (ctxt
);