1 /* Full and partial redundancy elimination and code hoisting on SSA GIMPLE.
2 Copyright (C) 2001-2017 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
24 #include "coretypes.h"
30 #include "alloc-pool.h"
31 #include "tree-pass.h"
34 #include "gimple-pretty-print.h"
35 #include "fold-const.h"
37 #include "gimple-fold.h"
40 #include "gimple-iterator.h"
42 #include "tree-ssa-loop.h"
43 #include "tree-into-ssa.h"
47 #include "tree-ssa-sccvn.h"
48 #include "tree-scalar-evolution.h"
52 #include "tree-ssa-propagate.h"
53 #include "ipa-utils.h"
54 #include "tree-cfgcleanup.h"
55 #include "langhooks.h"
58 /* Even though this file is called tree-ssa-pre.c, we actually
59 implement a bit more than just PRE here. All of them piggy-back
60 on GVN which is implemented in tree-ssa-sccvn.c.
62 1. Full Redundancy Elimination (FRE)
63 This is the elimination phase of GVN.
65 2. Partial Redundancy Elimination (PRE)
66 This is adds computation of AVAIL_OUT and ANTIC_IN and
67 doing expression insertion to form GVN-PRE.
70 This optimization uses the ANTIC_IN sets computed for PRE
71 to move expressions further up than PRE would do, to make
72 multiple computations of the same value fully redundant.
73 This pass is explained below (after the explanation of the
74 basic algorithm for PRE).
79 1. Avail sets can be shared by making an avail_find_leader that
80 walks up the dominator tree and looks in those avail sets.
81 This might affect code optimality, it's unclear right now.
82 Currently the AVAIL_OUT sets are the remaining quadraticness in
84 2. Strength reduction can be performed by anticipating expressions
85 we can repair later on.
86 3. We can do back-substitution or smarter value numbering to catch
87 commutative expressions split up over multiple statements.
90 /* For ease of terminology, "expression node" in the below refers to
91 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
92 represent the actual statement containing the expressions we care about,
93 and we cache the value number by putting it in the expression. */
95 /* Basic algorithm for Partial Redundancy Elimination:
97 First we walk the statements to generate the AVAIL sets, the
98 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
99 generation of values/expressions by a given block. We use them
100 when computing the ANTIC sets. The AVAIL sets consist of
101 SSA_NAME's that represent values, so we know what values are
102 available in what blocks. AVAIL is a forward dataflow problem. In
103 SSA, values are never killed, so we don't need a kill set, or a
104 fixpoint iteration, in order to calculate the AVAIL sets. In
105 traditional parlance, AVAIL sets tell us the downsafety of the
108 Next, we generate the ANTIC sets. These sets represent the
109 anticipatable expressions. ANTIC is a backwards dataflow
110 problem. An expression is anticipatable in a given block if it could
111 be generated in that block. This means that if we had to perform
112 an insertion in that block, of the value of that expression, we
113 could. Calculating the ANTIC sets requires phi translation of
114 expressions, because the flow goes backwards through phis. We must
115 iterate to a fixpoint of the ANTIC sets, because we have a kill
116 set. Even in SSA form, values are not live over the entire
117 function, only from their definition point onwards. So we have to
118 remove values from the ANTIC set once we go past the definition
119 point of the leaders that make them up.
120 compute_antic/compute_antic_aux performs this computation.
122 Third, we perform insertions to make partially redundant
123 expressions fully redundant.
125 An expression is partially redundant (excluding partial
128 1. It is AVAIL in some, but not all, of the predecessors of a
130 2. It is ANTIC in all the predecessors.
132 In order to make it fully redundant, we insert the expression into
133 the predecessors where it is not available, but is ANTIC.
135 When optimizing for size, we only eliminate the partial redundancy
136 if we need to insert in only one predecessor. This avoids almost
137 completely the code size increase that PRE usually causes.
139 For the partial anticipation case, we only perform insertion if it
140 is partially anticipated in some block, and fully available in all
143 do_pre_regular_insertion/do_pre_partial_partial_insertion
144 performs these steps, driven by insert/insert_aux.
146 Fourth, we eliminate fully redundant expressions.
147 This is a simple statement walk that replaces redundant
148 calculations with the now available values. */
150 /* Basic algorithm for Code Hoisting:
152 Code hoisting is: Moving value computations up in the control flow
153 graph to make multiple copies redundant. Typically this is a size
154 optimization, but there are cases where it also is helpful for speed.
156 A simple code hoisting algorithm is implemented that piggy-backs on
157 the PRE infrastructure. For code hoisting, we have to know ANTIC_OUT
158 which is effectively ANTIC_IN - AVAIL_OUT. The latter two have to be
159 computed for PRE, and we can use them to perform a limited version of
162 For the purpose of this implementation, a value is hoistable to a basic
163 block B if the following properties are met:
165 1. The value is in ANTIC_IN(B) -- the value will be computed on all
166 paths from B to function exit and it can be computed in B);
168 2. The value is not in AVAIL_OUT(B) -- there would be no need to
169 compute the value again and make it available twice;
171 3. All successors of B are dominated by B -- makes sure that inserting
172 a computation of the value in B will make the remaining
173 computations fully redundant;
175 4. At least one successor has the value in AVAIL_OUT -- to avoid
176 hoisting values up too far;
178 5. There are at least two successors of B -- hoisting in straight
179 line code is pointless.
181 The third condition is not strictly necessary, but it would complicate
182 the hoisting pass a lot. In fact, I don't know of any code hoisting
183 algorithm that does not have this requirement. Fortunately, experiments
184 have show that most candidate hoistable values are in regions that meet
185 this condition (e.g. diamond-shape regions).
187 The forth condition is necessary to avoid hoisting things up too far
188 away from the uses of the value. Nothing else limits the algorithm
189 from hoisting everything up as far as ANTIC_IN allows. Experiments
190 with SPEC and CSiBE have shown that hoisting up too far results in more
191 spilling, less benefits for code size, and worse benchmark scores.
192 Fortunately, in practice most of the interesting hoisting opportunities
193 are caught despite this limitation.
195 For hoistable values that meet all conditions, expressions are inserted
196 to make the calculation of the hoistable value fully redundant. We
197 perform code hoisting insertions after each round of PRE insertions,
198 because code hoisting never exposes new PRE opportunities, but PRE can
199 create new code hoisting opportunities.
201 The code hoisting algorithm is implemented in do_hoist_insert, driven
202 by insert/insert_aux. */
204 /* Representations of value numbers:
206 Value numbers are represented by a representative SSA_NAME. We
207 will create fake SSA_NAME's in situations where we need a
208 representative but do not have one (because it is a complex
209 expression). In order to facilitate storing the value numbers in
210 bitmaps, and keep the number of wasted SSA_NAME's down, we also
211 associate a value_id with each value number, and create full blown
212 ssa_name's only where we actually need them (IE in operands of
213 existing expressions).
215 Theoretically you could replace all the value_id's with
216 SSA_NAME_VERSION, but this would allocate a large number of
217 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
218 It would also require an additional indirection at each point we
221 /* Representation of expressions on value numbers:
223 Expressions consisting of value numbers are represented the same
224 way as our VN internally represents them, with an additional
225 "pre_expr" wrapping around them in order to facilitate storing all
226 of the expressions in the same sets. */
228 /* Representation of sets:
230 The dataflow sets do not need to be sorted in any particular order
231 for the majority of their lifetime, are simply represented as two
232 bitmaps, one that keeps track of values present in the set, and one
233 that keeps track of expressions present in the set.
235 When we need them in topological order, we produce it on demand by
236 transforming the bitmap into an array and sorting it into topo
239 /* Type of expression, used to know which member of the PRE_EXPR union
255 vn_reference_t reference
;
258 typedef struct pre_expr_d
: nofree_ptr_hash
<pre_expr_d
>
260 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
;
327 /* Allocate an expression id for EXPR. */
329 static inline unsigned int
330 alloc_expression_id (pre_expr expr
)
332 struct pre_expr_d
**slot
;
333 /* Make sure we won't overflow. */
334 gcc_assert (next_expression_id
+ 1 > next_expression_id
);
335 expr
->id
= next_expression_id
++;
336 expressions
.safe_push (expr
);
337 if (expr
->kind
== NAME
)
339 unsigned version
= SSA_NAME_VERSION (PRE_EXPR_NAME (expr
));
340 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
341 re-allocations by using vec::reserve upfront. */
342 unsigned old_len
= name_to_id
.length ();
343 name_to_id
.reserve (num_ssa_names
- old_len
);
344 name_to_id
.quick_grow_cleared (num_ssa_names
);
345 gcc_assert (name_to_id
[version
] == 0);
346 name_to_id
[version
] = expr
->id
;
350 slot
= expression_to_id
->find_slot (expr
, INSERT
);
354 return next_expression_id
- 1;
357 /* Return the expression id for tree EXPR. */
359 static inline unsigned int
360 get_expression_id (const pre_expr expr
)
365 static inline unsigned int
366 lookup_expression_id (const pre_expr expr
)
368 struct pre_expr_d
**slot
;
370 if (expr
->kind
== NAME
)
372 unsigned version
= SSA_NAME_VERSION (PRE_EXPR_NAME (expr
));
373 if (name_to_id
.length () <= version
)
375 return name_to_id
[version
];
379 slot
= expression_to_id
->find_slot (expr
, NO_INSERT
);
382 return ((pre_expr
)*slot
)->id
;
386 /* Return the existing expression id for EXPR, or create one if one
387 does not exist yet. */
389 static inline unsigned int
390 get_or_alloc_expression_id (pre_expr expr
)
392 unsigned int id
= lookup_expression_id (expr
);
394 return alloc_expression_id (expr
);
395 return expr
->id
= id
;
398 /* Return the expression that has expression id ID */
400 static inline pre_expr
401 expression_for_id (unsigned int id
)
403 return expressions
[id
];
406 /* Free the expression id field in all of our expressions,
407 and then destroy the expressions array. */
410 clear_expression_ids (void)
412 expressions
.release ();
415 static object_allocator
<pre_expr_d
> pre_expr_pool ("pre_expr nodes");
417 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
420 get_or_alloc_expr_for_name (tree name
)
422 struct pre_expr_d expr
;
424 unsigned int result_id
;
428 PRE_EXPR_NAME (&expr
) = name
;
429 result_id
= lookup_expression_id (&expr
);
431 return expression_for_id (result_id
);
433 result
= pre_expr_pool
.allocate ();
435 PRE_EXPR_NAME (result
) = name
;
436 alloc_expression_id (result
);
440 /* An unordered bitmap set. One bitmap tracks values, the other,
442 typedef struct bitmap_set
444 bitmap_head expressions
;
448 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
449 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
451 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
452 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
454 /* Mapping from value id to expressions with that value_id. */
455 static vec
<bitmap
> value_expressions
;
457 /* Sets that we need to keep track of. */
458 typedef struct bb_bitmap_sets
460 /* The EXP_GEN set, which represents expressions/values generated in
462 bitmap_set_t exp_gen
;
464 /* The PHI_GEN set, which represents PHI results generated in a
466 bitmap_set_t phi_gen
;
468 /* The TMP_GEN set, which represents results/temporaries generated
469 in a basic block. IE the LHS of an expression. */
470 bitmap_set_t tmp_gen
;
472 /* The AVAIL_OUT set, which represents which values are available in
473 a given basic block. */
474 bitmap_set_t avail_out
;
476 /* The ANTIC_IN set, which represents which values are anticipatable
477 in a given basic block. */
478 bitmap_set_t antic_in
;
480 /* The PA_IN set, which represents which values are
481 partially anticipatable in a given basic block. */
484 /* The NEW_SETS set, which is used during insertion to augment the
485 AVAIL_OUT set of blocks with the new insertions performed during
486 the current iteration. */
487 bitmap_set_t new_sets
;
489 /* A cache for value_dies_in_block_x. */
492 /* The live virtual operand on successor edges. */
495 /* True if we have visited this block during ANTIC calculation. */
496 unsigned int visited
: 1;
498 /* True when the block contains a call that might not return. */
499 unsigned int contains_may_not_return_call
: 1;
502 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
503 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
504 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
505 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
506 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
507 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
508 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
509 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
510 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
511 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
512 #define BB_LIVE_VOP_ON_EXIT(BB) ((bb_value_sets_t) ((BB)->aux))->vop_on_exit
515 /* This structure is used to keep track of statistics on what
516 optimization PRE was able to perform. */
519 /* The number of RHS computations eliminated by PRE. */
522 /* The number of new expressions/temporaries generated by PRE. */
525 /* The number of inserts found due to partial anticipation */
528 /* The number of inserts made for code hoisting. */
531 /* The number of new PHI nodes added by PRE. */
535 static bool do_partial_partial
;
536 static pre_expr
bitmap_find_leader (bitmap_set_t
, unsigned int);
537 static void bitmap_value_insert_into_set (bitmap_set_t
, pre_expr
);
538 static void bitmap_value_replace_in_set (bitmap_set_t
, pre_expr
);
539 static void bitmap_set_copy (bitmap_set_t
, bitmap_set_t
);
540 static void bitmap_set_and (bitmap_set_t
, bitmap_set_t
);
541 static bool bitmap_set_contains_value (bitmap_set_t
, unsigned int);
542 static void bitmap_insert_into_set (bitmap_set_t
, pre_expr
);
543 static bitmap_set_t
bitmap_set_new (void);
544 static tree
create_expression_by_pieces (basic_block
, pre_expr
, gimple_seq
*,
546 static tree
find_or_generate_expression (basic_block
, tree
, gimple_seq
*);
547 static unsigned int get_expr_value_id (pre_expr
);
549 /* We can add and remove elements and entries to and from sets
550 and hash tables, so we use alloc pools for them. */
552 static object_allocator
<bitmap_set
> bitmap_set_pool ("Bitmap sets");
553 static bitmap_obstack grand_bitmap_obstack
;
555 /* Set of blocks with statements that have had their EH properties changed. */
556 static bitmap need_eh_cleanup
;
558 /* Set of blocks with statements that have had their AB properties changed. */
559 static bitmap need_ab_cleanup
;
561 /* A three tuple {e, pred, v} used to cache phi translations in the
562 phi_translate_table. */
564 typedef struct expr_pred_trans_d
: free_ptr_hash
<expr_pred_trans_d
>
566 /* The expression. */
569 /* The predecessor block along which we translated the expression. */
572 /* The value that resulted from the translation. */
575 /* The hashcode for the expression, pred pair. This is cached for
579 /* hash_table support. */
580 static inline hashval_t
hash (const expr_pred_trans_d
*);
581 static inline int equal (const expr_pred_trans_d
*, const expr_pred_trans_d
*);
582 } *expr_pred_trans_t
;
583 typedef const struct expr_pred_trans_d
*const_expr_pred_trans_t
;
586 expr_pred_trans_d::hash (const expr_pred_trans_d
*e
)
592 expr_pred_trans_d::equal (const expr_pred_trans_d
*ve1
,
593 const expr_pred_trans_d
*ve2
)
595 basic_block b1
= ve1
->pred
;
596 basic_block b2
= ve2
->pred
;
598 /* If they are not translations for the same basic block, they can't
602 return pre_expr_d::equal (ve1
->e
, ve2
->e
);
605 /* The phi_translate_table caches phi translations for a given
606 expression and predecessor. */
607 static hash_table
<expr_pred_trans_d
> *phi_translate_table
;
609 /* Add the tuple mapping from {expression E, basic block PRED} to
610 the phi translation table and return whether it pre-existed. */
613 phi_trans_add (expr_pred_trans_t
*entry
, pre_expr e
, basic_block pred
)
615 expr_pred_trans_t
*slot
;
616 expr_pred_trans_d tem
;
617 hashval_t hash
= iterative_hash_hashval_t (pre_expr_d::hash (e
),
622 slot
= phi_translate_table
->find_slot_with_hash (&tem
, hash
, INSERT
);
629 *entry
= *slot
= XNEW (struct expr_pred_trans_d
);
631 (*entry
)->pred
= pred
;
632 (*entry
)->hashcode
= hash
;
637 /* Add expression E to the expression set of value id V. */
640 add_to_value (unsigned int v
, pre_expr e
)
644 gcc_checking_assert (get_expr_value_id (e
) == v
);
646 if (v
>= value_expressions
.length ())
648 value_expressions
.safe_grow_cleared (v
+ 1);
651 set
= value_expressions
[v
];
654 set
= BITMAP_ALLOC (&grand_bitmap_obstack
);
655 value_expressions
[v
] = set
;
658 bitmap_set_bit (set
, get_or_alloc_expression_id (e
));
661 /* Create a new bitmap set and return it. */
664 bitmap_set_new (void)
666 bitmap_set_t ret
= bitmap_set_pool
.allocate ();
667 bitmap_initialize (&ret
->expressions
, &grand_bitmap_obstack
);
668 bitmap_initialize (&ret
->values
, &grand_bitmap_obstack
);
672 /* Return the value id for a PRE expression EXPR. */
675 get_expr_value_id (pre_expr expr
)
681 id
= get_constant_value_id (PRE_EXPR_CONSTANT (expr
));
684 id
= VN_INFO (PRE_EXPR_NAME (expr
))->value_id
;
687 id
= PRE_EXPR_NARY (expr
)->value_id
;
690 id
= PRE_EXPR_REFERENCE (expr
)->value_id
;
695 /* ??? We cannot assert that expr has a value-id (it can be 0), because
696 we assign value-ids only to expressions that have a result
697 in set_hashtable_value_ids. */
701 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
704 sccvn_valnum_from_value_id (unsigned int val
)
708 bitmap exprset
= value_expressions
[val
];
709 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
711 pre_expr vexpr
= expression_for_id (i
);
712 if (vexpr
->kind
== NAME
)
713 return VN_INFO (PRE_EXPR_NAME (vexpr
))->valnum
;
714 else if (vexpr
->kind
== CONSTANT
)
715 return PRE_EXPR_CONSTANT (vexpr
);
720 /* Remove an expression EXPR from a bitmapped set. */
723 bitmap_remove_from_set (bitmap_set_t set
, pre_expr expr
)
725 unsigned int val
= get_expr_value_id (expr
);
726 if (!value_id_constant_p (val
))
728 bitmap_clear_bit (&set
->values
, val
);
729 bitmap_clear_bit (&set
->expressions
, get_expression_id (expr
));
733 /* Insert an expression EXPR into a bitmapped set. */
736 bitmap_insert_into_set (bitmap_set_t set
, pre_expr expr
)
738 unsigned int val
= get_expr_value_id (expr
);
739 if (! value_id_constant_p (val
))
741 /* Note this is the only function causing multiple expressions
742 for the same value to appear in a set. This is needed for
743 TMP_GEN, PHI_GEN and NEW_SETs. */
744 bitmap_set_bit (&set
->values
, val
);
745 bitmap_set_bit (&set
->expressions
, get_or_alloc_expression_id (expr
));
749 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
752 bitmap_set_copy (bitmap_set_t dest
, bitmap_set_t orig
)
754 bitmap_copy (&dest
->expressions
, &orig
->expressions
);
755 bitmap_copy (&dest
->values
, &orig
->values
);
759 /* Free memory used up by SET. */
761 bitmap_set_free (bitmap_set_t set
)
763 bitmap_clear (&set
->expressions
);
764 bitmap_clear (&set
->values
);
768 /* Generate an topological-ordered array of bitmap set SET. */
771 sorted_array_from_bitmap_set (bitmap_set_t set
)
774 bitmap_iterator bi
, bj
;
775 vec
<pre_expr
> result
;
777 /* Pre-allocate enough space for the array. */
778 result
.create (bitmap_count_bits (&set
->expressions
));
780 FOR_EACH_VALUE_ID_IN_SET (set
, i
, bi
)
782 /* The number of expressions having a given value is usually
783 relatively small. Thus, rather than making a vector of all
784 the expressions and sorting it by value-id, we walk the values
785 and check in the reverse mapping that tells us what expressions
786 have a given value, to filter those in our set. As a result,
787 the expressions are inserted in value-id order, which means
790 If this is somehow a significant lose for some cases, we can
791 choose which set to walk based on the set size. */
792 bitmap exprset
= value_expressions
[i
];
793 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, j
, bj
)
795 if (bitmap_bit_p (&set
->expressions
, j
))
796 result
.quick_push (expression_for_id (j
));
803 /* Perform bitmapped set operation DEST &= ORIG. */
806 bitmap_set_and (bitmap_set_t dest
, bitmap_set_t orig
)
813 bitmap_and_into (&dest
->values
, &orig
->values
);
815 unsigned int to_clear
= -1U;
816 FOR_EACH_EXPR_ID_IN_SET (dest
, i
, bi
)
820 bitmap_clear_bit (&dest
->expressions
, to_clear
);
823 pre_expr expr
= expression_for_id (i
);
824 unsigned int value_id
= get_expr_value_id (expr
);
825 if (!bitmap_bit_p (&dest
->values
, value_id
))
829 bitmap_clear_bit (&dest
->expressions
, to_clear
);
833 /* Subtract all expressions contained in ORIG from DEST. */
836 bitmap_set_subtract (bitmap_set_t dest
, bitmap_set_t orig
)
838 bitmap_set_t result
= bitmap_set_new ();
842 bitmap_and_compl (&result
->expressions
, &dest
->expressions
,
845 FOR_EACH_EXPR_ID_IN_SET (result
, i
, bi
)
847 pre_expr expr
= expression_for_id (i
);
848 unsigned int value_id
= get_expr_value_id (expr
);
849 bitmap_set_bit (&result
->values
, value_id
);
855 /* Subtract all values in bitmap set B from bitmap set A. */
858 bitmap_set_subtract_values (bitmap_set_t a
, bitmap_set_t b
)
862 pre_expr to_remove
= NULL
;
863 FOR_EACH_EXPR_ID_IN_SET (a
, i
, bi
)
867 bitmap_remove_from_set (a
, to_remove
);
870 pre_expr expr
= expression_for_id (i
);
871 if (bitmap_set_contains_value (b
, get_expr_value_id (expr
)))
875 bitmap_remove_from_set (a
, to_remove
);
879 /* Return true if bitmapped set SET contains the value VALUE_ID. */
882 bitmap_set_contains_value (bitmap_set_t set
, unsigned int value_id
)
884 if (value_id_constant_p (value_id
))
887 if (!set
|| bitmap_empty_p (&set
->expressions
))
890 return bitmap_bit_p (&set
->values
, value_id
);
894 bitmap_set_contains_expr (bitmap_set_t set
, const pre_expr expr
)
896 return bitmap_bit_p (&set
->expressions
, get_expression_id (expr
));
899 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
902 bitmap_set_replace_value (bitmap_set_t set
, unsigned int lookfor
,
909 if (value_id_constant_p (lookfor
))
912 if (!bitmap_set_contains_value (set
, lookfor
))
915 /* The number of expressions having a given value is usually
916 significantly less than the total number of expressions in SET.
917 Thus, rather than check, for each expression in SET, whether it
918 has the value LOOKFOR, we walk the reverse mapping that tells us
919 what expressions have a given value, and see if any of those
920 expressions are in our set. For large testcases, this is about
921 5-10x faster than walking the bitmap. If this is somehow a
922 significant lose for some cases, we can choose which set to walk
923 based on the set size. */
924 exprset
= value_expressions
[lookfor
];
925 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
927 if (bitmap_clear_bit (&set
->expressions
, i
))
929 bitmap_set_bit (&set
->expressions
, get_expression_id (expr
));
937 /* Return true if two bitmap sets are equal. */
940 bitmap_set_equal (bitmap_set_t a
, bitmap_set_t b
)
942 return bitmap_equal_p (&a
->values
, &b
->values
);
945 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
946 and add it otherwise. */
949 bitmap_value_replace_in_set (bitmap_set_t set
, pre_expr expr
)
951 unsigned int val
= get_expr_value_id (expr
);
953 if (bitmap_set_contains_value (set
, val
))
954 bitmap_set_replace_value (set
, val
, expr
);
956 bitmap_insert_into_set (set
, expr
);
959 /* Insert EXPR into SET if EXPR's value is not already present in
963 bitmap_value_insert_into_set (bitmap_set_t set
, pre_expr expr
)
965 unsigned int val
= get_expr_value_id (expr
);
967 gcc_checking_assert (expr
->id
== get_or_alloc_expression_id (expr
));
969 /* Constant values are always considered to be part of the set. */
970 if (value_id_constant_p (val
))
973 /* If the value membership changed, add the expression. */
974 if (bitmap_set_bit (&set
->values
, val
))
975 bitmap_set_bit (&set
->expressions
, expr
->id
);
978 /* Print out EXPR to outfile. */
981 print_pre_expr (FILE *outfile
, const pre_expr expr
)
985 fprintf (outfile
, "NULL");
991 print_generic_expr (outfile
, PRE_EXPR_CONSTANT (expr
));
994 print_generic_expr (outfile
, PRE_EXPR_NAME (expr
));
999 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
1000 fprintf (outfile
, "{%s,", get_tree_code_name (nary
->opcode
));
1001 for (i
= 0; i
< nary
->length
; i
++)
1003 print_generic_expr (outfile
, nary
->op
[i
]);
1004 if (i
!= (unsigned) nary
->length
- 1)
1005 fprintf (outfile
, ",");
1007 fprintf (outfile
, "}");
1013 vn_reference_op_t vro
;
1015 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
1016 fprintf (outfile
, "{");
1018 ref
->operands
.iterate (i
, &vro
);
1021 bool closebrace
= false;
1022 if (vro
->opcode
!= SSA_NAME
1023 && TREE_CODE_CLASS (vro
->opcode
) != tcc_declaration
)
1025 fprintf (outfile
, "%s", get_tree_code_name (vro
->opcode
));
1028 fprintf (outfile
, "<");
1034 print_generic_expr (outfile
, vro
->op0
);
1037 fprintf (outfile
, ",");
1038 print_generic_expr (outfile
, vro
->op1
);
1042 fprintf (outfile
, ",");
1043 print_generic_expr (outfile
, vro
->op2
);
1047 fprintf (outfile
, ">");
1048 if (i
!= ref
->operands
.length () - 1)
1049 fprintf (outfile
, ",");
1051 fprintf (outfile
, "}");
1054 fprintf (outfile
, "@");
1055 print_generic_expr (outfile
, ref
->vuse
);
1061 void debug_pre_expr (pre_expr
);
1063 /* Like print_pre_expr but always prints to stderr. */
1065 debug_pre_expr (pre_expr e
)
1067 print_pre_expr (stderr
, e
);
1068 fprintf (stderr
, "\n");
1071 /* Print out SET to OUTFILE. */
1074 print_bitmap_set (FILE *outfile
, bitmap_set_t set
,
1075 const char *setname
, int blockindex
)
1077 fprintf (outfile
, "%s[%d] := { ", setname
, blockindex
);
1084 FOR_EACH_EXPR_ID_IN_SET (set
, i
, bi
)
1086 const pre_expr expr
= expression_for_id (i
);
1089 fprintf (outfile
, ", ");
1091 print_pre_expr (outfile
, expr
);
1093 fprintf (outfile
, " (%04d)", get_expr_value_id (expr
));
1096 fprintf (outfile
, " }\n");
1099 void debug_bitmap_set (bitmap_set_t
);
1102 debug_bitmap_set (bitmap_set_t set
)
1104 print_bitmap_set (stderr
, set
, "debug", 0);
1107 void debug_bitmap_sets_for (basic_block
);
1110 debug_bitmap_sets_for (basic_block bb
)
1112 print_bitmap_set (stderr
, AVAIL_OUT (bb
), "avail_out", bb
->index
);
1113 print_bitmap_set (stderr
, EXP_GEN (bb
), "exp_gen", bb
->index
);
1114 print_bitmap_set (stderr
, PHI_GEN (bb
), "phi_gen", bb
->index
);
1115 print_bitmap_set (stderr
, TMP_GEN (bb
), "tmp_gen", bb
->index
);
1116 print_bitmap_set (stderr
, ANTIC_IN (bb
), "antic_in", bb
->index
);
1117 if (do_partial_partial
)
1118 print_bitmap_set (stderr
, PA_IN (bb
), "pa_in", bb
->index
);
1119 print_bitmap_set (stderr
, NEW_SETS (bb
), "new_sets", bb
->index
);
1122 /* Print out the expressions that have VAL to OUTFILE. */
1125 print_value_expressions (FILE *outfile
, unsigned int val
)
1127 bitmap set
= value_expressions
[val
];
1132 sprintf (s
, "%04d", val
);
1133 x
.expressions
= *set
;
1134 print_bitmap_set (outfile
, &x
, s
, 0);
1140 debug_value_expressions (unsigned int val
)
1142 print_value_expressions (stderr
, val
);
1145 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1149 get_or_alloc_expr_for_constant (tree constant
)
1151 unsigned int result_id
;
1152 unsigned int value_id
;
1153 struct pre_expr_d expr
;
1156 expr
.kind
= CONSTANT
;
1157 PRE_EXPR_CONSTANT (&expr
) = constant
;
1158 result_id
= lookup_expression_id (&expr
);
1160 return expression_for_id (result_id
);
1162 newexpr
= pre_expr_pool
.allocate ();
1163 newexpr
->kind
= CONSTANT
;
1164 PRE_EXPR_CONSTANT (newexpr
) = constant
;
1165 alloc_expression_id (newexpr
);
1166 value_id
= get_or_alloc_constant_value_id (constant
);
1167 add_to_value (value_id
, newexpr
);
1171 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1172 Currently only supports constants and SSA_NAMES. */
1174 get_or_alloc_expr_for (tree t
)
1176 if (TREE_CODE (t
) == SSA_NAME
)
1177 return get_or_alloc_expr_for_name (t
);
1178 else if (is_gimple_min_invariant (t
))
1179 return get_or_alloc_expr_for_constant (t
);
1183 /* Return the folded version of T if T, when folded, is a gimple
1184 min_invariant or an SSA name. Otherwise, return T. */
1187 fully_constant_expression (pre_expr e
)
1195 vn_nary_op_t nary
= PRE_EXPR_NARY (e
);
1196 tree res
= vn_nary_simplify (nary
);
1199 if (is_gimple_min_invariant (res
))
1200 return get_or_alloc_expr_for_constant (res
);
1201 if (TREE_CODE (res
) == SSA_NAME
)
1202 return get_or_alloc_expr_for_name (res
);
1207 vn_reference_t ref
= PRE_EXPR_REFERENCE (e
);
1209 if ((folded
= fully_constant_vn_reference_p (ref
)))
1210 return get_or_alloc_expr_for_constant (folded
);
1219 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1220 it has the value it would have in BLOCK. Set *SAME_VALID to true
1221 in case the new vuse doesn't change the value id of the OPERANDS. */
1224 translate_vuse_through_block (vec
<vn_reference_op_s
> operands
,
1225 alias_set_type set
, tree type
, tree vuse
,
1226 basic_block phiblock
,
1227 basic_block block
, bool *same_valid
)
1229 gimple
*phi
= SSA_NAME_DEF_STMT (vuse
);
1236 if (gimple_bb (phi
) != phiblock
)
1239 use_oracle
= ao_ref_init_from_vn_reference (&ref
, set
, type
, operands
);
1241 /* Use the alias-oracle to find either the PHI node in this block,
1242 the first VUSE used in this block that is equivalent to vuse or
1243 the first VUSE which definition in this block kills the value. */
1244 if (gimple_code (phi
) == GIMPLE_PHI
)
1245 e
= find_edge (block
, phiblock
);
1246 else if (use_oracle
)
1247 while (!stmt_may_clobber_ref_p_1 (phi
, &ref
))
1249 vuse
= gimple_vuse (phi
);
1250 phi
= SSA_NAME_DEF_STMT (vuse
);
1251 if (gimple_bb (phi
) != phiblock
)
1253 if (gimple_code (phi
) == GIMPLE_PHI
)
1255 e
= find_edge (block
, phiblock
);
1266 bitmap visited
= NULL
;
1268 /* Try to find a vuse that dominates this phi node by skipping
1269 non-clobbering statements. */
1270 vuse
= get_continuation_for_phi (phi
, &ref
, &cnt
, &visited
, false,
1273 BITMAP_FREE (visited
);
1279 /* If we didn't find any, the value ID can't stay the same,
1280 but return the translated vuse. */
1281 *same_valid
= false;
1282 vuse
= PHI_ARG_DEF (phi
, e
->dest_idx
);
1284 /* ??? We would like to return vuse here as this is the canonical
1285 upmost vdef that this reference is associated with. But during
1286 insertion of the references into the hash tables we only ever
1287 directly insert with their direct gimple_vuse, hence returning
1288 something else would make us not find the other expression. */
1289 return PHI_ARG_DEF (phi
, e
->dest_idx
);
1295 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1296 SET2 *or* SET3. This is used to avoid making a set consisting of the union
1297 of PA_IN and ANTIC_IN during insert and phi-translation. */
1299 static inline pre_expr
1300 find_leader_in_sets (unsigned int val
, bitmap_set_t set1
, bitmap_set_t set2
,
1301 bitmap_set_t set3
= NULL
)
1305 result
= bitmap_find_leader (set1
, val
);
1306 if (!result
&& set2
)
1307 result
= bitmap_find_leader (set2
, val
);
1308 if (!result
&& set3
)
1309 result
= bitmap_find_leader (set3
, val
);
1313 /* Get the tree type for our PRE expression e. */
1316 get_expr_type (const pre_expr e
)
1321 return TREE_TYPE (PRE_EXPR_NAME (e
));
1323 return TREE_TYPE (PRE_EXPR_CONSTANT (e
));
1325 return PRE_EXPR_REFERENCE (e
)->type
;
1327 return PRE_EXPR_NARY (e
)->type
;
1332 /* Get a representative SSA_NAME for a given expression.
1333 Since all of our sub-expressions are treated as values, we require
1334 them to be SSA_NAME's for simplicity.
1335 Prior versions of GVNPRE used to use "value handles" here, so that
1336 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1337 either case, the operands are really values (IE we do not expect
1338 them to be usable without finding leaders). */
1341 get_representative_for (const pre_expr e
)
1344 unsigned int value_id
= get_expr_value_id (e
);
1349 return VN_INFO (PRE_EXPR_NAME (e
))->valnum
;
1351 return PRE_EXPR_CONSTANT (e
);
1355 /* Go through all of the expressions representing this value
1356 and pick out an SSA_NAME. */
1359 bitmap exprs
= value_expressions
[value_id
];
1360 EXECUTE_IF_SET_IN_BITMAP (exprs
, 0, i
, bi
)
1362 pre_expr rep
= expression_for_id (i
);
1363 if (rep
->kind
== NAME
)
1364 return VN_INFO (PRE_EXPR_NAME (rep
))->valnum
;
1365 else if (rep
->kind
== CONSTANT
)
1366 return PRE_EXPR_CONSTANT (rep
);
1372 /* If we reached here we couldn't find an SSA_NAME. This can
1373 happen when we've discovered a value that has never appeared in
1374 the program as set to an SSA_NAME, as the result of phi translation.
1376 ??? We should be able to re-use this when we insert the statement
1378 name
= make_temp_ssa_name (get_expr_type (e
), gimple_build_nop (), "pretmp");
1379 VN_INFO_GET (name
)->value_id
= value_id
;
1380 VN_INFO (name
)->valnum
= name
;
1381 /* ??? For now mark this SSA name for release by SCCVN. */
1382 VN_INFO (name
)->needs_insertion
= true;
1383 add_to_value (value_id
, get_or_alloc_expr_for_name (name
));
1384 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1386 fprintf (dump_file
, "Created SSA_NAME representative ");
1387 print_generic_expr (dump_file
, name
);
1388 fprintf (dump_file
, " for expression:");
1389 print_pre_expr (dump_file
, e
);
1390 fprintf (dump_file
, " (%04d)\n", value_id
);
1399 phi_translate (pre_expr expr
, bitmap_set_t set1
, bitmap_set_t set2
,
1400 basic_block pred
, basic_block phiblock
);
1402 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1403 the phis in PRED. Return NULL if we can't find a leader for each part
1404 of the translated expression. */
1407 phi_translate_1 (pre_expr expr
, bitmap_set_t set1
, bitmap_set_t set2
,
1408 basic_block pred
, basic_block phiblock
)
1415 bool changed
= false;
1416 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
1417 vn_nary_op_t newnary
= XALLOCAVAR (struct vn_nary_op_s
,
1418 sizeof_vn_nary_op (nary
->length
));
1419 memcpy (newnary
, nary
, sizeof_vn_nary_op (nary
->length
));
1421 for (i
= 0; i
< newnary
->length
; i
++)
1423 if (TREE_CODE (newnary
->op
[i
]) != SSA_NAME
)
1427 pre_expr leader
, result
;
1428 unsigned int op_val_id
= VN_INFO (newnary
->op
[i
])->value_id
;
1429 leader
= find_leader_in_sets (op_val_id
, set1
, set2
);
1430 result
= phi_translate (leader
, set1
, set2
, pred
, phiblock
);
1431 if (result
&& result
!= leader
)
1432 newnary
->op
[i
] = get_representative_for (result
);
1436 changed
|= newnary
->op
[i
] != nary
->op
[i
];
1442 unsigned int new_val_id
;
1444 PRE_EXPR_NARY (expr
) = newnary
;
1445 constant
= fully_constant_expression (expr
);
1446 PRE_EXPR_NARY (expr
) = nary
;
1447 if (constant
!= expr
)
1449 /* For non-CONSTANTs we have to make sure we can eventually
1450 insert the expression. Which means we need to have a
1452 if (constant
->kind
!= CONSTANT
)
1454 /* Do not allow simplifications to non-constants over
1455 backedges as this will likely result in a loop PHI node
1456 to be inserted and increased register pressure.
1457 See PR77498 - this avoids doing predcoms work in
1458 a less efficient way. */
1459 if (find_edge (pred
, phiblock
)->flags
& EDGE_DFS_BACK
)
1463 unsigned value_id
= get_expr_value_id (constant
);
1464 constant
= find_leader_in_sets (value_id
, set1
, set2
,
1474 tree result
= vn_nary_op_lookup_pieces (newnary
->length
,
1479 if (result
&& is_gimple_min_invariant (result
))
1480 return get_or_alloc_expr_for_constant (result
);
1482 expr
= pre_expr_pool
.allocate ();
1487 PRE_EXPR_NARY (expr
) = nary
;
1488 new_val_id
= nary
->value_id
;
1489 get_or_alloc_expression_id (expr
);
1490 /* When we end up re-using a value number make sure that
1491 doesn't have unrelated (which we can't check here)
1492 range or points-to info on it. */
1494 && INTEGRAL_TYPE_P (TREE_TYPE (result
))
1495 && SSA_NAME_RANGE_INFO (result
)
1496 && ! SSA_NAME_IS_DEFAULT_DEF (result
))
1498 if (! VN_INFO (result
)->info
.range_info
)
1500 VN_INFO (result
)->info
.range_info
1501 = SSA_NAME_RANGE_INFO (result
);
1502 VN_INFO (result
)->range_info_anti_range_p
1503 = SSA_NAME_ANTI_RANGE_P (result
);
1505 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1507 fprintf (dump_file
, "clearing range info of ");
1508 print_generic_expr (dump_file
, result
);
1509 fprintf (dump_file
, "\n");
1511 SSA_NAME_RANGE_INFO (result
) = NULL
;
1514 && POINTER_TYPE_P (TREE_TYPE (result
))
1515 && SSA_NAME_PTR_INFO (result
)
1516 && ! SSA_NAME_IS_DEFAULT_DEF (result
))
1518 if (! VN_INFO (result
)->info
.ptr_info
)
1519 VN_INFO (result
)->info
.ptr_info
1520 = SSA_NAME_PTR_INFO (result
);
1521 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1523 fprintf (dump_file
, "clearing points-to info of ");
1524 print_generic_expr (dump_file
, result
);
1525 fprintf (dump_file
, "\n");
1527 SSA_NAME_PTR_INFO (result
) = NULL
;
1532 new_val_id
= get_next_value_id ();
1533 value_expressions
.safe_grow_cleared (get_max_value_id () + 1);
1534 nary
= vn_nary_op_insert_pieces (newnary
->length
,
1538 result
, new_val_id
);
1539 PRE_EXPR_NARY (expr
) = nary
;
1540 get_or_alloc_expression_id (expr
);
1542 add_to_value (new_val_id
, expr
);
1550 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
1551 vec
<vn_reference_op_s
> operands
= ref
->operands
;
1552 tree vuse
= ref
->vuse
;
1553 tree newvuse
= vuse
;
1554 vec
<vn_reference_op_s
> newoperands
= vNULL
;
1555 bool changed
= false, same_valid
= true;
1557 vn_reference_op_t operand
;
1558 vn_reference_t newref
;
1560 for (i
= 0; operands
.iterate (i
, &operand
); i
++)
1565 tree type
= operand
->type
;
1566 vn_reference_op_s newop
= *operand
;
1567 op
[0] = operand
->op0
;
1568 op
[1] = operand
->op1
;
1569 op
[2] = operand
->op2
;
1570 for (n
= 0; n
< 3; ++n
)
1572 unsigned int op_val_id
;
1575 if (TREE_CODE (op
[n
]) != SSA_NAME
)
1577 /* We can't possibly insert these. */
1579 && !is_gimple_min_invariant (op
[n
]))
1583 op_val_id
= VN_INFO (op
[n
])->value_id
;
1584 leader
= find_leader_in_sets (op_val_id
, set1
, set2
);
1585 opresult
= phi_translate (leader
, set1
, set2
, pred
, phiblock
);
1586 if (opresult
&& opresult
!= leader
)
1588 tree name
= get_representative_for (opresult
);
1589 changed
|= name
!= op
[n
];
1597 newoperands
.release ();
1602 if (!newoperands
.exists ())
1603 newoperands
= operands
.copy ();
1604 /* We may have changed from an SSA_NAME to a constant */
1605 if (newop
.opcode
== SSA_NAME
&& TREE_CODE (op
[0]) != SSA_NAME
)
1606 newop
.opcode
= TREE_CODE (op
[0]);
1611 newoperands
[i
] = newop
;
1613 gcc_checking_assert (i
== operands
.length ());
1617 newvuse
= translate_vuse_through_block (newoperands
.exists ()
1618 ? newoperands
: operands
,
1619 ref
->set
, ref
->type
,
1620 vuse
, phiblock
, pred
,
1622 if (newvuse
== NULL_TREE
)
1624 newoperands
.release ();
1629 if (changed
|| newvuse
!= vuse
)
1631 unsigned int new_val_id
;
1634 tree result
= vn_reference_lookup_pieces (newvuse
, ref
->set
,
1636 newoperands
.exists ()
1637 ? newoperands
: operands
,
1640 newoperands
.release ();
1642 /* We can always insert constants, so if we have a partial
1643 redundant constant load of another type try to translate it
1644 to a constant of appropriate type. */
1645 if (result
&& is_gimple_min_invariant (result
))
1648 if (!useless_type_conversion_p (ref
->type
, TREE_TYPE (result
)))
1650 tem
= fold_unary (VIEW_CONVERT_EXPR
, ref
->type
, result
);
1651 if (tem
&& !is_gimple_min_invariant (tem
))
1655 return get_or_alloc_expr_for_constant (tem
);
1658 /* If we'd have to convert things we would need to validate
1659 if we can insert the translated expression. So fail
1660 here for now - we cannot insert an alias with a different
1661 type in the VN tables either, as that would assert. */
1663 && !useless_type_conversion_p (ref
->type
, TREE_TYPE (result
)))
1665 else if (!result
&& newref
1666 && !useless_type_conversion_p (ref
->type
, newref
->type
))
1668 newoperands
.release ();
1672 expr
= pre_expr_pool
.allocate ();
1673 expr
->kind
= REFERENCE
;
1678 PRE_EXPR_REFERENCE (expr
) = newref
;
1679 constant
= fully_constant_expression (expr
);
1680 if (constant
!= expr
)
1683 new_val_id
= newref
->value_id
;
1684 get_or_alloc_expression_id (expr
);
1688 if (changed
|| !same_valid
)
1690 new_val_id
= get_next_value_id ();
1691 value_expressions
.safe_grow_cleared
1692 (get_max_value_id () + 1);
1695 new_val_id
= ref
->value_id
;
1696 if (!newoperands
.exists ())
1697 newoperands
= operands
.copy ();
1698 newref
= vn_reference_insert_pieces (newvuse
, ref
->set
,
1701 result
, new_val_id
);
1702 newoperands
= vNULL
;
1703 PRE_EXPR_REFERENCE (expr
) = newref
;
1704 constant
= fully_constant_expression (expr
);
1705 if (constant
!= expr
)
1707 get_or_alloc_expression_id (expr
);
1709 add_to_value (new_val_id
, expr
);
1711 newoperands
.release ();
1718 tree name
= PRE_EXPR_NAME (expr
);
1719 gimple
*def_stmt
= SSA_NAME_DEF_STMT (name
);
1720 /* If the SSA name is defined by a PHI node in this block,
1722 if (gimple_code (def_stmt
) == GIMPLE_PHI
1723 && gimple_bb (def_stmt
) == phiblock
)
1725 edge e
= find_edge (pred
, gimple_bb (def_stmt
));
1726 tree def
= PHI_ARG_DEF (def_stmt
, e
->dest_idx
);
1728 /* Handle constant. */
1729 if (is_gimple_min_invariant (def
))
1730 return get_or_alloc_expr_for_constant (def
);
1732 return get_or_alloc_expr_for_name (def
);
1734 /* Otherwise return it unchanged - it will get removed if its
1735 value is not available in PREDs AVAIL_OUT set of expressions
1736 by the subtraction of TMP_GEN. */
1745 /* Wrapper around phi_translate_1 providing caching functionality. */
1748 phi_translate (pre_expr expr
, bitmap_set_t set1
, bitmap_set_t set2
,
1749 basic_block pred
, basic_block phiblock
)
1751 expr_pred_trans_t slot
= NULL
;
1757 /* Constants contain no values that need translation. */
1758 if (expr
->kind
== CONSTANT
)
1761 if (value_id_constant_p (get_expr_value_id (expr
)))
1764 /* Don't add translations of NAMEs as those are cheap to translate. */
1765 if (expr
->kind
!= NAME
)
1767 if (phi_trans_add (&slot
, expr
, pred
))
1769 /* Store NULL for the value we want to return in the case of
1775 phitrans
= phi_translate_1 (expr
, set1
, set2
, pred
, phiblock
);
1782 /* Remove failed translations again, they cause insert
1783 iteration to not pick up new opportunities reliably. */
1784 phi_translate_table
->remove_elt_with_hash (slot
, slot
->hashcode
);
1791 /* For each expression in SET, translate the values through phi nodes
1792 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1793 expressions in DEST. */
1796 phi_translate_set (bitmap_set_t dest
, bitmap_set_t set
, basic_block pred
,
1797 basic_block phiblock
)
1799 vec
<pre_expr
> exprs
;
1803 if (gimple_seq_empty_p (phi_nodes (phiblock
)))
1805 bitmap_set_copy (dest
, set
);
1809 exprs
= sorted_array_from_bitmap_set (set
);
1810 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
1812 pre_expr translated
;
1813 translated
= phi_translate (expr
, set
, NULL
, pred
, phiblock
);
1817 /* We might end up with multiple expressions from SET being
1818 translated to the same value. In this case we do not want
1819 to retain the NARY or REFERENCE expression but prefer a NAME
1820 which would be the leader. */
1821 if (translated
->kind
== NAME
)
1822 bitmap_value_replace_in_set (dest
, translated
);
1824 bitmap_value_insert_into_set (dest
, translated
);
1829 /* Find the leader for a value (i.e., the name representing that
1830 value) in a given set, and return it. Return NULL if no leader
1834 bitmap_find_leader (bitmap_set_t set
, unsigned int val
)
1836 if (value_id_constant_p (val
))
1840 bitmap exprset
= value_expressions
[val
];
1842 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
1844 pre_expr expr
= expression_for_id (i
);
1845 if (expr
->kind
== CONSTANT
)
1849 if (bitmap_set_contains_value (set
, val
))
1851 /* Rather than walk the entire bitmap of expressions, and see
1852 whether any of them has the value we are looking for, we look
1853 at the reverse mapping, which tells us the set of expressions
1854 that have a given value (IE value->expressions with that
1855 value) and see if any of those expressions are in our set.
1856 The number of expressions per value is usually significantly
1857 less than the number of expressions in the set. In fact, for
1858 large testcases, doing it this way is roughly 5-10x faster
1859 than walking the bitmap.
1860 If this is somehow a significant lose for some cases, we can
1861 choose which set to walk based on which set is smaller. */
1864 bitmap exprset
= value_expressions
[val
];
1866 EXECUTE_IF_AND_IN_BITMAP (exprset
, &set
->expressions
, 0, i
, bi
)
1867 return expression_for_id (i
);
1872 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1873 BLOCK by seeing if it is not killed in the block. Note that we are
1874 only determining whether there is a store that kills it. Because
1875 of the order in which clean iterates over values, we are guaranteed
1876 that altered operands will have caused us to be eliminated from the
1877 ANTIC_IN set already. */
1880 value_dies_in_block_x (pre_expr expr
, basic_block block
)
1882 tree vuse
= PRE_EXPR_REFERENCE (expr
)->vuse
;
1883 vn_reference_t refx
= PRE_EXPR_REFERENCE (expr
);
1885 gimple_stmt_iterator gsi
;
1886 unsigned id
= get_expression_id (expr
);
1893 /* Lookup a previously calculated result. */
1894 if (EXPR_DIES (block
)
1895 && bitmap_bit_p (EXPR_DIES (block
), id
* 2))
1896 return bitmap_bit_p (EXPR_DIES (block
), id
* 2 + 1);
1898 /* A memory expression {e, VUSE} dies in the block if there is a
1899 statement that may clobber e. If, starting statement walk from the
1900 top of the basic block, a statement uses VUSE there can be no kill
1901 inbetween that use and the original statement that loaded {e, VUSE},
1902 so we can stop walking. */
1903 ref
.base
= NULL_TREE
;
1904 for (gsi
= gsi_start_bb (block
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1906 tree def_vuse
, def_vdef
;
1907 def
= gsi_stmt (gsi
);
1908 def_vuse
= gimple_vuse (def
);
1909 def_vdef
= gimple_vdef (def
);
1911 /* Not a memory statement. */
1915 /* Not a may-def. */
1918 /* A load with the same VUSE, we're done. */
1919 if (def_vuse
== vuse
)
1925 /* Init ref only if we really need it. */
1926 if (ref
.base
== NULL_TREE
1927 && !ao_ref_init_from_vn_reference (&ref
, refx
->set
, refx
->type
,
1933 /* If the statement may clobber expr, it dies. */
1934 if (stmt_may_clobber_ref_p_1 (def
, &ref
))
1941 /* Remember the result. */
1942 if (!EXPR_DIES (block
))
1943 EXPR_DIES (block
) = BITMAP_ALLOC (&grand_bitmap_obstack
);
1944 bitmap_set_bit (EXPR_DIES (block
), id
* 2);
1946 bitmap_set_bit (EXPR_DIES (block
), id
* 2 + 1);
1952 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1953 contains its value-id. */
1956 op_valid_in_sets (bitmap_set_t set1
, bitmap_set_t set2
, tree op
)
1958 if (op
&& TREE_CODE (op
) == SSA_NAME
)
1960 unsigned int value_id
= VN_INFO (op
)->value_id
;
1961 if (!(bitmap_set_contains_value (set1
, value_id
)
1962 || (set2
&& bitmap_set_contains_value (set2
, value_id
))))
1968 /* Determine if the expression EXPR is valid in SET1 U SET2.
1969 ONLY SET2 CAN BE NULL.
1970 This means that we have a leader for each part of the expression
1971 (if it consists of values), or the expression is an SSA_NAME.
1972 For loads/calls, we also see if the vuse is killed in this block. */
1975 valid_in_sets (bitmap_set_t set1
, bitmap_set_t set2
, pre_expr expr
)
1980 /* By construction all NAMEs are available. Non-available
1981 NAMEs are removed by subtracting TMP_GEN from the sets. */
1986 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
1987 for (i
= 0; i
< nary
->length
; i
++)
1988 if (!op_valid_in_sets (set1
, set2
, nary
->op
[i
]))
1995 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
1996 vn_reference_op_t vro
;
1999 FOR_EACH_VEC_ELT (ref
->operands
, i
, vro
)
2001 if (!op_valid_in_sets (set1
, set2
, vro
->op0
)
2002 || !op_valid_in_sets (set1
, set2
, vro
->op1
)
2003 || !op_valid_in_sets (set1
, set2
, vro
->op2
))
2013 /* Clean the set of expressions that are no longer valid in SET1 or
2014 SET2. This means expressions that are made up of values we have no
2015 leaders for in SET1 or SET2. This version is used for partial
2016 anticipation, which means it is not valid in either ANTIC_IN or
2020 dependent_clean (bitmap_set_t set1
, bitmap_set_t set2
)
2022 vec
<pre_expr
> exprs
= sorted_array_from_bitmap_set (set1
);
2026 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
2028 if (!valid_in_sets (set1
, set2
, expr
))
2029 bitmap_remove_from_set (set1
, expr
);
2034 /* Clean the set of expressions that are no longer valid in SET. This
2035 means expressions that are made up of values we have no leaders for
2039 clean (bitmap_set_t set
)
2041 vec
<pre_expr
> exprs
= sorted_array_from_bitmap_set (set
);
2045 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
2047 if (!valid_in_sets (set
, NULL
, expr
))
2048 bitmap_remove_from_set (set
, expr
);
2053 /* Clean the set of expressions that are no longer valid in SET because
2054 they are clobbered in BLOCK or because they trap and may not be executed. */
2057 prune_clobbered_mems (bitmap_set_t set
, basic_block block
)
2061 pre_expr to_remove
= NULL
;
2063 FOR_EACH_EXPR_ID_IN_SET (set
, i
, bi
)
2065 /* Remove queued expr. */
2068 bitmap_remove_from_set (set
, to_remove
);
2072 pre_expr expr
= expression_for_id (i
);
2073 if (expr
->kind
== REFERENCE
)
2075 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
2078 gimple
*def_stmt
= SSA_NAME_DEF_STMT (ref
->vuse
);
2079 if (!gimple_nop_p (def_stmt
)
2080 && ((gimple_bb (def_stmt
) != block
2081 && !dominated_by_p (CDI_DOMINATORS
,
2082 block
, gimple_bb (def_stmt
)))
2083 || (gimple_bb (def_stmt
) == block
2084 && value_dies_in_block_x (expr
, block
))))
2088 else if (expr
->kind
== NARY
)
2090 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
2091 /* If the NARY may trap make sure the block does not contain
2092 a possible exit point.
2093 ??? This is overly conservative if we translate AVAIL_OUT
2094 as the available expression might be after the exit point. */
2095 if (BB_MAY_NOTRETURN (block
)
2096 && vn_nary_may_trap (nary
))
2101 /* Remove queued expr. */
2103 bitmap_remove_from_set (set
, to_remove
);
2106 static sbitmap has_abnormal_preds
;
2108 /* Compute the ANTIC set for BLOCK.
2110 If succs(BLOCK) > 1 then
2111 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2112 else if succs(BLOCK) == 1 then
2113 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2115 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2117 Note that clean() is deferred until after the iteration. */
2120 compute_antic_aux (basic_block block
, bool block_has_abnormal_pred_edge
)
2122 bitmap_set_t S
, old
, ANTIC_OUT
;
2128 bool changed
= ! BB_VISITED (block
);
2129 BB_VISITED (block
) = 1;
2130 old
= ANTIC_OUT
= S
= NULL
;
2132 /* If any edges from predecessors are abnormal, antic_in is empty,
2134 if (block_has_abnormal_pred_edge
)
2135 goto maybe_dump_sets
;
2137 old
= ANTIC_IN (block
);
2138 ANTIC_OUT
= bitmap_set_new ();
2140 /* If the block has no successors, ANTIC_OUT is empty. */
2141 if (EDGE_COUNT (block
->succs
) == 0)
2143 /* If we have one successor, we could have some phi nodes to
2144 translate through. */
2145 else if (single_succ_p (block
))
2147 basic_block succ_bb
= single_succ (block
);
2148 gcc_assert (BB_VISITED (succ_bb
));
2149 phi_translate_set (ANTIC_OUT
, ANTIC_IN (succ_bb
), block
, succ_bb
);
2151 /* If we have multiple successors, we take the intersection of all of
2152 them. Note that in the case of loop exit phi nodes, we may have
2153 phis to translate through. */
2157 basic_block bprime
, first
= NULL
;
2159 auto_vec
<basic_block
> worklist (EDGE_COUNT (block
->succs
));
2160 FOR_EACH_EDGE (e
, ei
, block
->succs
)
2163 && BB_VISITED (e
->dest
))
2165 else if (BB_VISITED (e
->dest
))
2166 worklist
.quick_push (e
->dest
);
2169 /* Unvisited successors get their ANTIC_IN replaced by the
2170 maximal set to arrive at a maximum ANTIC_IN solution.
2171 We can ignore them in the intersection operation and thus
2172 need not explicitely represent that maximum solution. */
2173 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2174 fprintf (dump_file
, "ANTIC_IN is MAX on %d->%d\n",
2175 e
->src
->index
, e
->dest
->index
);
2179 /* Of multiple successors we have to have visited one already
2180 which is guaranteed by iteration order. */
2181 gcc_assert (first
!= NULL
);
2183 phi_translate_set (ANTIC_OUT
, ANTIC_IN (first
), block
, first
);
2185 FOR_EACH_VEC_ELT (worklist
, i
, bprime
)
2187 if (!gimple_seq_empty_p (phi_nodes (bprime
)))
2189 bitmap_set_t tmp
= bitmap_set_new ();
2190 phi_translate_set (tmp
, ANTIC_IN (bprime
), block
, bprime
);
2191 bitmap_set_and (ANTIC_OUT
, tmp
);
2192 bitmap_set_free (tmp
);
2195 bitmap_set_and (ANTIC_OUT
, ANTIC_IN (bprime
));
2199 /* Prune expressions that are clobbered in block and thus become
2200 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2201 prune_clobbered_mems (ANTIC_OUT
, block
);
2203 /* Generate ANTIC_OUT - TMP_GEN. */
2204 S
= bitmap_set_subtract (ANTIC_OUT
, TMP_GEN (block
));
2206 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2207 ANTIC_IN (block
) = bitmap_set_subtract (EXP_GEN (block
),
2210 /* Then union in the ANTIC_OUT - TMP_GEN values,
2211 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2212 FOR_EACH_EXPR_ID_IN_SET (S
, bii
, bi
)
2213 bitmap_value_insert_into_set (ANTIC_IN (block
),
2214 expression_for_id (bii
));
2216 /* clean (ANTIC_IN (block)) is defered to after the iteration converged
2217 because it can cause non-convergence, see for example PR81181. */
2219 if (!bitmap_set_equal (old
, ANTIC_IN (block
)))
2223 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2226 print_bitmap_set (dump_file
, ANTIC_OUT
, "ANTIC_OUT", block
->index
);
2229 fprintf (dump_file
, "[changed] ");
2230 print_bitmap_set (dump_file
, ANTIC_IN (block
), "ANTIC_IN",
2234 print_bitmap_set (dump_file
, S
, "S", block
->index
);
2237 bitmap_set_free (old
);
2239 bitmap_set_free (S
);
2241 bitmap_set_free (ANTIC_OUT
);
2245 /* Compute PARTIAL_ANTIC for BLOCK.
2247 If succs(BLOCK) > 1 then
2248 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2249 in ANTIC_OUT for all succ(BLOCK)
2250 else if succs(BLOCK) == 1 then
2251 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2253 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2258 compute_partial_antic_aux (basic_block block
,
2259 bool block_has_abnormal_pred_edge
)
2261 bitmap_set_t old_PA_IN
;
2262 bitmap_set_t PA_OUT
;
2265 unsigned long max_pa
= PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH
);
2267 old_PA_IN
= PA_OUT
= NULL
;
2269 /* If any edges from predecessors are abnormal, antic_in is empty,
2271 if (block_has_abnormal_pred_edge
)
2272 goto maybe_dump_sets
;
2274 /* If there are too many partially anticipatable values in the
2275 block, phi_translate_set can take an exponential time: stop
2276 before the translation starts. */
2278 && single_succ_p (block
)
2279 && bitmap_count_bits (&PA_IN (single_succ (block
))->values
) > max_pa
)
2280 goto maybe_dump_sets
;
2282 old_PA_IN
= PA_IN (block
);
2283 PA_OUT
= bitmap_set_new ();
2285 /* If the block has no successors, ANTIC_OUT is empty. */
2286 if (EDGE_COUNT (block
->succs
) == 0)
2288 /* If we have one successor, we could have some phi nodes to
2289 translate through. Note that we can't phi translate across DFS
2290 back edges in partial antic, because it uses a union operation on
2291 the successors. For recurrences like IV's, we will end up
2292 generating a new value in the set on each go around (i + 3 (VH.1)
2293 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2294 else if (single_succ_p (block
))
2296 basic_block succ
= single_succ (block
);
2297 if (!(single_succ_edge (block
)->flags
& EDGE_DFS_BACK
))
2298 phi_translate_set (PA_OUT
, PA_IN (succ
), block
, succ
);
2300 /* If we have multiple successors, we take the union of all of
2307 auto_vec
<basic_block
> worklist (EDGE_COUNT (block
->succs
));
2308 FOR_EACH_EDGE (e
, ei
, block
->succs
)
2310 if (e
->flags
& EDGE_DFS_BACK
)
2312 worklist
.quick_push (e
->dest
);
2314 if (worklist
.length () > 0)
2316 FOR_EACH_VEC_ELT (worklist
, i
, bprime
)
2321 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime
), i
, bi
)
2322 bitmap_value_insert_into_set (PA_OUT
,
2323 expression_for_id (i
));
2324 if (!gimple_seq_empty_p (phi_nodes (bprime
)))
2326 bitmap_set_t pa_in
= bitmap_set_new ();
2327 phi_translate_set (pa_in
, PA_IN (bprime
), block
, bprime
);
2328 FOR_EACH_EXPR_ID_IN_SET (pa_in
, i
, bi
)
2329 bitmap_value_insert_into_set (PA_OUT
,
2330 expression_for_id (i
));
2331 bitmap_set_free (pa_in
);
2334 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime
), i
, bi
)
2335 bitmap_value_insert_into_set (PA_OUT
,
2336 expression_for_id (i
));
2341 /* Prune expressions that are clobbered in block and thus become
2342 invalid if translated from PA_OUT to PA_IN. */
2343 prune_clobbered_mems (PA_OUT
, block
);
2345 /* PA_IN starts with PA_OUT - TMP_GEN.
2346 Then we subtract things from ANTIC_IN. */
2347 PA_IN (block
) = bitmap_set_subtract (PA_OUT
, TMP_GEN (block
));
2349 /* For partial antic, we want to put back in the phi results, since
2350 we will properly avoid making them partially antic over backedges. */
2351 bitmap_ior_into (&PA_IN (block
)->values
, &PHI_GEN (block
)->values
);
2352 bitmap_ior_into (&PA_IN (block
)->expressions
, &PHI_GEN (block
)->expressions
);
2354 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2355 bitmap_set_subtract_values (PA_IN (block
), ANTIC_IN (block
));
2357 dependent_clean (PA_IN (block
), ANTIC_IN (block
));
2360 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2363 print_bitmap_set (dump_file
, PA_OUT
, "PA_OUT", block
->index
);
2365 print_bitmap_set (dump_file
, PA_IN (block
), "PA_IN", block
->index
);
2368 bitmap_set_free (old_PA_IN
);
2370 bitmap_set_free (PA_OUT
);
2373 /* Compute ANTIC and partial ANTIC sets. */
2376 compute_antic (void)
2378 bool changed
= true;
2379 int num_iterations
= 0;
2385 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2386 We pre-build the map of blocks with incoming abnormal edges here. */
2387 has_abnormal_preds
= sbitmap_alloc (last_basic_block_for_fn (cfun
));
2388 bitmap_clear (has_abnormal_preds
);
2390 FOR_ALL_BB_FN (block
, cfun
)
2392 BB_VISITED (block
) = 0;
2394 FOR_EACH_EDGE (e
, ei
, block
->preds
)
2395 if (e
->flags
& EDGE_ABNORMAL
)
2397 bitmap_set_bit (has_abnormal_preds
, block
->index
);
2401 /* While we are here, give empty ANTIC_IN sets to each block. */
2402 ANTIC_IN (block
) = bitmap_set_new ();
2403 if (do_partial_partial
)
2404 PA_IN (block
) = bitmap_set_new ();
2407 /* At the exit block we anticipate nothing. */
2408 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun
)) = 1;
2410 /* For ANTIC computation we need a postorder that also guarantees that
2411 a block with a single successor is visited after its successor.
2412 RPO on the inverted CFG has this property. */
2413 auto_vec
<int, 20> postorder
;
2414 inverted_post_order_compute (&postorder
);
2416 auto_sbitmap
worklist (last_basic_block_for_fn (cfun
) + 1);
2417 bitmap_clear (worklist
);
2418 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR_FOR_FN (cfun
)->preds
)
2419 bitmap_set_bit (worklist
, e
->src
->index
);
2422 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2423 fprintf (dump_file
, "Starting iteration %d\n", num_iterations
);
2424 /* ??? We need to clear our PHI translation cache here as the
2425 ANTIC sets shrink and we restrict valid translations to
2426 those having operands with leaders in ANTIC. Same below
2427 for PA ANTIC computation. */
2430 for (i
= postorder
.length () - 1; i
>= 0; i
--)
2432 if (bitmap_bit_p (worklist
, postorder
[i
]))
2434 basic_block block
= BASIC_BLOCK_FOR_FN (cfun
, postorder
[i
]);
2435 bitmap_clear_bit (worklist
, block
->index
);
2436 if (compute_antic_aux (block
,
2437 bitmap_bit_p (has_abnormal_preds
,
2440 FOR_EACH_EDGE (e
, ei
, block
->preds
)
2441 bitmap_set_bit (worklist
, e
->src
->index
);
2446 /* Theoretically possible, but *highly* unlikely. */
2447 gcc_checking_assert (num_iterations
< 500);
2450 /* We have to clean after the dataflow problem converged as cleaning
2451 can cause non-convergence because it is based on expressions
2452 rather than values. */
2453 FOR_EACH_BB_FN (block
, cfun
)
2454 clean (ANTIC_IN (block
));
2456 statistics_histogram_event (cfun
, "compute_antic iterations",
2459 if (do_partial_partial
)
2461 /* For partial antic we ignore backedges and thus we do not need
2462 to perform any iteration when we process blocks in postorder. */
2464 = pre_and_rev_post_order_compute (NULL
, postorder
.address (), false);
2465 for (i
= postorder_num
- 1 ; i
>= 0; i
--)
2467 basic_block block
= BASIC_BLOCK_FOR_FN (cfun
, postorder
[i
]);
2468 compute_partial_antic_aux (block
,
2469 bitmap_bit_p (has_abnormal_preds
,
2474 sbitmap_free (has_abnormal_preds
);
2478 /* Inserted expressions are placed onto this worklist, which is used
2479 for performing quick dead code elimination of insertions we made
2480 that didn't turn out to be necessary. */
2481 static bitmap inserted_exprs
;
2483 /* The actual worker for create_component_ref_by_pieces. */
2486 create_component_ref_by_pieces_1 (basic_block block
, vn_reference_t ref
,
2487 unsigned int *operand
, gimple_seq
*stmts
)
2489 vn_reference_op_t currop
= &ref
->operands
[*operand
];
2492 switch (currop
->opcode
)
2499 tree baseop
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2503 tree offset
= currop
->op0
;
2504 if (TREE_CODE (baseop
) == ADDR_EXPR
2505 && handled_component_p (TREE_OPERAND (baseop
, 0)))
2509 base
= get_addr_base_and_unit_offset (TREE_OPERAND (baseop
, 0),
2512 offset
= int_const_binop (PLUS_EXPR
, offset
,
2513 build_int_cst (TREE_TYPE (offset
),
2515 baseop
= build_fold_addr_expr (base
);
2517 genop
= build2 (MEM_REF
, currop
->type
, baseop
, offset
);
2518 MR_DEPENDENCE_CLIQUE (genop
) = currop
->clique
;
2519 MR_DEPENDENCE_BASE (genop
) = currop
->base
;
2520 REF_REVERSE_STORAGE_ORDER (genop
) = currop
->reverse
;
2524 case TARGET_MEM_REF
:
2526 tree genop0
= NULL_TREE
, genop1
= NULL_TREE
;
2527 vn_reference_op_t nextop
= &ref
->operands
[++*operand
];
2528 tree baseop
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2534 genop0
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2540 genop1
= find_or_generate_expression (block
, nextop
->op0
, stmts
);
2544 genop
= build5 (TARGET_MEM_REF
, currop
->type
,
2545 baseop
, currop
->op2
, genop0
, currop
->op1
, genop1
);
2547 MR_DEPENDENCE_CLIQUE (genop
) = currop
->clique
;
2548 MR_DEPENDENCE_BASE (genop
) = currop
->base
;
2555 gcc_assert (is_gimple_min_invariant (currop
->op0
));
2561 case VIEW_CONVERT_EXPR
:
2563 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2567 return fold_build1 (currop
->opcode
, currop
->type
, genop0
);
2570 case WITH_SIZE_EXPR
:
2572 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2576 tree genop1
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2579 return fold_build2 (currop
->opcode
, currop
->type
, genop0
, genop1
);
2584 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2588 tree op1
= currop
->op0
;
2589 tree op2
= currop
->op1
;
2590 tree t
= build3 (BIT_FIELD_REF
, currop
->type
, genop0
, op1
, op2
);
2591 REF_REVERSE_STORAGE_ORDER (t
) = currop
->reverse
;
2595 /* For array ref vn_reference_op's, operand 1 of the array ref
2596 is op0 of the reference op and operand 3 of the array ref is
2598 case ARRAY_RANGE_REF
:
2602 tree genop1
= currop
->op0
;
2603 tree genop2
= currop
->op1
;
2604 tree genop3
= currop
->op2
;
2605 genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2609 genop1
= find_or_generate_expression (block
, genop1
, stmts
);
2614 tree domain_type
= TYPE_DOMAIN (TREE_TYPE (genop0
));
2615 /* Drop zero minimum index if redundant. */
2616 if (integer_zerop (genop2
)
2618 || integer_zerop (TYPE_MIN_VALUE (domain_type
))))
2622 genop2
= find_or_generate_expression (block
, genop2
, stmts
);
2629 tree elmt_type
= TREE_TYPE (TREE_TYPE (genop0
));
2630 /* We can't always put a size in units of the element alignment
2631 here as the element alignment may be not visible. See
2632 PR43783. Simply drop the element size for constant
2634 if (TREE_CODE (genop3
) == INTEGER_CST
2635 && TREE_CODE (TYPE_SIZE_UNIT (elmt_type
)) == INTEGER_CST
2636 && wi::eq_p (wi::to_offset (TYPE_SIZE_UNIT (elmt_type
)),
2637 (wi::to_offset (genop3
)
2638 * vn_ref_op_align_unit (currop
))))
2642 genop3
= find_or_generate_expression (block
, genop3
, stmts
);
2647 return build4 (currop
->opcode
, currop
->type
, genop0
, genop1
,
2654 tree genop2
= currop
->op1
;
2655 op0
= create_component_ref_by_pieces_1 (block
, ref
, operand
, stmts
);
2658 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2662 genop2
= find_or_generate_expression (block
, genop2
, stmts
);
2666 return fold_build3 (COMPONENT_REF
, TREE_TYPE (op1
), op0
, op1
, genop2
);
2671 genop
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2692 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2693 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2694 trying to rename aggregates into ssa form directly, which is a no no.
2696 Thus, this routine doesn't create temporaries, it just builds a
2697 single access expression for the array, calling
2698 find_or_generate_expression to build the innermost pieces.
2700 This function is a subroutine of create_expression_by_pieces, and
2701 should not be called on it's own unless you really know what you
2705 create_component_ref_by_pieces (basic_block block
, vn_reference_t ref
,
2708 unsigned int op
= 0;
2709 return create_component_ref_by_pieces_1 (block
, ref
, &op
, stmts
);
2712 /* Find a simple leader for an expression, or generate one using
2713 create_expression_by_pieces from a NARY expression for the value.
2714 BLOCK is the basic_block we are looking for leaders in.
2715 OP is the tree expression to find a leader for or generate.
2716 Returns the leader or NULL_TREE on failure. */
2719 find_or_generate_expression (basic_block block
, tree op
, gimple_seq
*stmts
)
2721 pre_expr expr
= get_or_alloc_expr_for (op
);
2722 unsigned int lookfor
= get_expr_value_id (expr
);
2723 pre_expr leader
= bitmap_find_leader (AVAIL_OUT (block
), lookfor
);
2726 if (leader
->kind
== NAME
)
2727 return PRE_EXPR_NAME (leader
);
2728 else if (leader
->kind
== CONSTANT
)
2729 return PRE_EXPR_CONSTANT (leader
);
2735 /* It must be a complex expression, so generate it recursively. Note
2736 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2737 where the insert algorithm fails to insert a required expression. */
2738 bitmap exprset
= value_expressions
[lookfor
];
2741 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
2743 pre_expr temp
= expression_for_id (i
);
2744 /* We cannot insert random REFERENCE expressions at arbitrary
2745 places. We can insert NARYs which eventually re-materializes
2746 its operand values. */
2747 if (temp
->kind
== NARY
)
2748 return create_expression_by_pieces (block
, temp
, stmts
,
2749 get_expr_type (expr
));
2756 #define NECESSARY GF_PLF_1
2758 /* Create an expression in pieces, so that we can handle very complex
2759 expressions that may be ANTIC, but not necessary GIMPLE.
2760 BLOCK is the basic block the expression will be inserted into,
2761 EXPR is the expression to insert (in value form)
2762 STMTS is a statement list to append the necessary insertions into.
2764 This function will die if we hit some value that shouldn't be
2765 ANTIC but is (IE there is no leader for it, or its components).
2766 The function returns NULL_TREE in case a different antic expression
2767 has to be inserted first.
2768 This function may also generate expressions that are themselves
2769 partially or fully redundant. Those that are will be either made
2770 fully redundant during the next iteration of insert (for partially
2771 redundant ones), or eliminated by eliminate (for fully redundant
2775 create_expression_by_pieces (basic_block block
, pre_expr expr
,
2776 gimple_seq
*stmts
, tree type
)
2780 gimple_seq forced_stmts
= NULL
;
2781 unsigned int value_id
;
2782 gimple_stmt_iterator gsi
;
2783 tree exprtype
= type
? type
: get_expr_type (expr
);
2789 /* We may hit the NAME/CONSTANT case if we have to convert types
2790 that value numbering saw through. */
2792 folded
= PRE_EXPR_NAME (expr
);
2793 if (useless_type_conversion_p (exprtype
, TREE_TYPE (folded
)))
2798 folded
= PRE_EXPR_CONSTANT (expr
);
2799 tree tem
= fold_convert (exprtype
, folded
);
2800 if (is_gimple_min_invariant (tem
))
2805 if (PRE_EXPR_REFERENCE (expr
)->operands
[0].opcode
== CALL_EXPR
)
2807 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
2808 unsigned int operand
= 1;
2809 vn_reference_op_t currop
= &ref
->operands
[0];
2810 tree sc
= NULL_TREE
;
2812 if (TREE_CODE (currop
->op0
) == FUNCTION_DECL
)
2815 fn
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2820 sc
= find_or_generate_expression (block
, currop
->op1
, stmts
);
2824 auto_vec
<tree
> args (ref
->operands
.length () - 1);
2825 while (operand
< ref
->operands
.length ())
2827 tree arg
= create_component_ref_by_pieces_1 (block
, ref
,
2831 args
.quick_push (arg
);
2834 = gimple_build_call_vec ((TREE_CODE (fn
) == FUNCTION_DECL
2835 ? build_fold_addr_expr (fn
) : fn
), args
);
2836 gimple_call_set_with_bounds (call
, currop
->with_bounds
);
2838 gimple_call_set_chain (call
, sc
);
2839 tree forcedname
= make_ssa_name (currop
->type
);
2840 gimple_call_set_lhs (call
, forcedname
);
2841 gimple_set_vuse (call
, BB_LIVE_VOP_ON_EXIT (block
));
2842 gimple_seq_add_stmt_without_update (&forced_stmts
, call
);
2843 folded
= forcedname
;
2847 folded
= create_component_ref_by_pieces (block
,
2848 PRE_EXPR_REFERENCE (expr
),
2852 name
= make_temp_ssa_name (exprtype
, NULL
, "pretmp");
2853 newstmt
= gimple_build_assign (name
, folded
);
2854 gimple_seq_add_stmt_without_update (&forced_stmts
, newstmt
);
2855 gimple_set_vuse (newstmt
, BB_LIVE_VOP_ON_EXIT (block
));
2861 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
2862 tree
*genop
= XALLOCAVEC (tree
, nary
->length
);
2864 for (i
= 0; i
< nary
->length
; ++i
)
2866 genop
[i
] = find_or_generate_expression (block
, nary
->op
[i
], stmts
);
2869 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2870 may have conversions stripped. */
2871 if (nary
->opcode
== POINTER_PLUS_EXPR
)
2874 genop
[i
] = gimple_convert (&forced_stmts
,
2875 nary
->type
, genop
[i
]);
2877 genop
[i
] = gimple_convert (&forced_stmts
,
2878 sizetype
, genop
[i
]);
2881 genop
[i
] = gimple_convert (&forced_stmts
,
2882 TREE_TYPE (nary
->op
[i
]), genop
[i
]);
2884 if (nary
->opcode
== CONSTRUCTOR
)
2886 vec
<constructor_elt
, va_gc
> *elts
= NULL
;
2887 for (i
= 0; i
< nary
->length
; ++i
)
2888 CONSTRUCTOR_APPEND_ELT (elts
, NULL_TREE
, genop
[i
]);
2889 folded
= build_constructor (nary
->type
, elts
);
2890 name
= make_temp_ssa_name (exprtype
, NULL
, "pretmp");
2891 newstmt
= gimple_build_assign (name
, folded
);
2892 gimple_seq_add_stmt_without_update (&forced_stmts
, newstmt
);
2897 switch (nary
->length
)
2900 folded
= gimple_build (&forced_stmts
, nary
->opcode
, nary
->type
,
2904 folded
= gimple_build (&forced_stmts
, nary
->opcode
, nary
->type
,
2905 genop
[0], genop
[1]);
2908 folded
= gimple_build (&forced_stmts
, nary
->opcode
, nary
->type
,
2909 genop
[0], genop
[1], genop
[2]);
2921 folded
= gimple_convert (&forced_stmts
, exprtype
, folded
);
2923 /* If there is nothing to insert, return the simplified result. */
2924 if (gimple_seq_empty_p (forced_stmts
))
2926 /* If we simplified to a constant return it and discard eventually
2928 if (is_gimple_min_invariant (folded
))
2930 gimple_seq_discard (forced_stmts
);
2933 /* Likewise if we simplified to sth not queued for insertion. */
2935 gsi
= gsi_last (forced_stmts
);
2936 for (; !gsi_end_p (gsi
); gsi_prev (&gsi
))
2938 gimple
*stmt
= gsi_stmt (gsi
);
2939 tree forcedname
= gimple_get_lhs (stmt
);
2940 if (forcedname
== folded
)
2948 gimple_seq_discard (forced_stmts
);
2951 gcc_assert (TREE_CODE (folded
) == SSA_NAME
);
2953 /* If we have any intermediate expressions to the value sets, add them
2954 to the value sets and chain them in the instruction stream. */
2957 gsi
= gsi_start (forced_stmts
);
2958 for (; !gsi_end_p (gsi
); gsi_next (&gsi
))
2960 gimple
*stmt
= gsi_stmt (gsi
);
2961 tree forcedname
= gimple_get_lhs (stmt
);
2964 if (forcedname
!= folded
)
2966 VN_INFO_GET (forcedname
)->valnum
= forcedname
;
2967 VN_INFO (forcedname
)->value_id
= get_next_value_id ();
2968 nameexpr
= get_or_alloc_expr_for_name (forcedname
);
2969 add_to_value (VN_INFO (forcedname
)->value_id
, nameexpr
);
2970 bitmap_value_replace_in_set (NEW_SETS (block
), nameexpr
);
2971 bitmap_value_replace_in_set (AVAIL_OUT (block
), nameexpr
);
2974 bitmap_set_bit (inserted_exprs
, SSA_NAME_VERSION (forcedname
));
2975 gimple_set_plf (stmt
, NECESSARY
, false);
2977 gimple_seq_add_seq (stmts
, forced_stmts
);
2982 /* Fold the last statement. */
2983 gsi
= gsi_last (*stmts
);
2984 if (fold_stmt_inplace (&gsi
))
2985 update_stmt (gsi_stmt (gsi
));
2987 /* Add a value number to the temporary.
2988 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2989 we are creating the expression by pieces, and this particular piece of
2990 the expression may have been represented. There is no harm in replacing
2992 value_id
= get_expr_value_id (expr
);
2993 VN_INFO_GET (name
)->value_id
= value_id
;
2994 VN_INFO (name
)->valnum
= sccvn_valnum_from_value_id (value_id
);
2995 if (VN_INFO (name
)->valnum
== NULL_TREE
)
2996 VN_INFO (name
)->valnum
= name
;
2997 gcc_assert (VN_INFO (name
)->valnum
!= NULL_TREE
);
2998 nameexpr
= get_or_alloc_expr_for_name (name
);
2999 add_to_value (value_id
, nameexpr
);
3000 if (NEW_SETS (block
))
3001 bitmap_value_replace_in_set (NEW_SETS (block
), nameexpr
);
3002 bitmap_value_replace_in_set (AVAIL_OUT (block
), nameexpr
);
3004 pre_stats
.insertions
++;
3005 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3007 fprintf (dump_file
, "Inserted ");
3008 print_gimple_stmt (dump_file
, gsi_stmt (gsi_last (*stmts
)), 0);
3009 fprintf (dump_file
, " in predecessor %d (%04d)\n",
3010 block
->index
, value_id
);
3017 /* Insert the to-be-made-available values of expression EXPRNUM for each
3018 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3019 merge the result with a phi node, given the same value number as
3020 NODE. Return true if we have inserted new stuff. */
3023 insert_into_preds_of_block (basic_block block
, unsigned int exprnum
,
3024 vec
<pre_expr
> avail
)
3026 pre_expr expr
= expression_for_id (exprnum
);
3028 unsigned int val
= get_expr_value_id (expr
);
3030 bool insertions
= false;
3035 tree type
= get_expr_type (expr
);
3039 /* Make sure we aren't creating an induction variable. */
3040 if (bb_loop_depth (block
) > 0 && EDGE_COUNT (block
->preds
) == 2)
3042 bool firstinsideloop
= false;
3043 bool secondinsideloop
= false;
3044 firstinsideloop
= flow_bb_inside_loop_p (block
->loop_father
,
3045 EDGE_PRED (block
, 0)->src
);
3046 secondinsideloop
= flow_bb_inside_loop_p (block
->loop_father
,
3047 EDGE_PRED (block
, 1)->src
);
3048 /* Induction variables only have one edge inside the loop. */
3049 if ((firstinsideloop
^ secondinsideloop
)
3050 && expr
->kind
!= REFERENCE
)
3052 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3053 fprintf (dump_file
, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3058 /* Make the necessary insertions. */
3059 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3061 gimple_seq stmts
= NULL
;
3064 eprime
= avail
[pred
->dest_idx
];
3065 builtexpr
= create_expression_by_pieces (bprime
, eprime
,
3067 gcc_assert (!(pred
->flags
& EDGE_ABNORMAL
));
3068 if (!gimple_seq_empty_p (stmts
))
3070 gsi_insert_seq_on_edge (pred
, stmts
);
3075 /* We cannot insert a PHI node if we failed to insert
3080 if (is_gimple_min_invariant (builtexpr
))
3081 avail
[pred
->dest_idx
] = get_or_alloc_expr_for_constant (builtexpr
);
3083 avail
[pred
->dest_idx
] = get_or_alloc_expr_for_name (builtexpr
);
3085 /* If we didn't want a phi node, and we made insertions, we still have
3086 inserted new stuff, and thus return true. If we didn't want a phi node,
3087 and didn't make insertions, we haven't added anything new, so return
3089 if (nophi
&& insertions
)
3091 else if (nophi
&& !insertions
)
3094 /* Now build a phi for the new variable. */
3095 temp
= make_temp_ssa_name (type
, NULL
, "prephitmp");
3096 phi
= create_phi_node (temp
, block
);
3098 gimple_set_plf (phi
, NECESSARY
, false);
3099 VN_INFO_GET (temp
)->value_id
= val
;
3100 VN_INFO (temp
)->valnum
= sccvn_valnum_from_value_id (val
);
3101 if (VN_INFO (temp
)->valnum
== NULL_TREE
)
3102 VN_INFO (temp
)->valnum
= temp
;
3103 bitmap_set_bit (inserted_exprs
, SSA_NAME_VERSION (temp
));
3104 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3106 pre_expr ae
= avail
[pred
->dest_idx
];
3107 gcc_assert (get_expr_type (ae
) == type
3108 || useless_type_conversion_p (type
, get_expr_type (ae
)));
3109 if (ae
->kind
== CONSTANT
)
3110 add_phi_arg (phi
, unshare_expr (PRE_EXPR_CONSTANT (ae
)),
3111 pred
, UNKNOWN_LOCATION
);
3113 add_phi_arg (phi
, PRE_EXPR_NAME (ae
), pred
, UNKNOWN_LOCATION
);
3116 newphi
= get_or_alloc_expr_for_name (temp
);
3117 add_to_value (val
, newphi
);
3119 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3120 this insertion, since we test for the existence of this value in PHI_GEN
3121 before proceeding with the partial redundancy checks in insert_aux.
3123 The value may exist in AVAIL_OUT, in particular, it could be represented
3124 by the expression we are trying to eliminate, in which case we want the
3125 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3128 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3129 this block, because if it did, it would have existed in our dominator's
3130 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3133 bitmap_insert_into_set (PHI_GEN (block
), newphi
);
3134 bitmap_value_replace_in_set (AVAIL_OUT (block
),
3136 bitmap_insert_into_set (NEW_SETS (block
),
3139 /* If we insert a PHI node for a conversion of another PHI node
3140 in the same basic-block try to preserve range information.
3141 This is important so that followup loop passes receive optimal
3142 number of iteration analysis results. See PR61743. */
3143 if (expr
->kind
== NARY
3144 && CONVERT_EXPR_CODE_P (expr
->u
.nary
->opcode
)
3145 && TREE_CODE (expr
->u
.nary
->op
[0]) == SSA_NAME
3146 && gimple_bb (SSA_NAME_DEF_STMT (expr
->u
.nary
->op
[0])) == block
3147 && INTEGRAL_TYPE_P (type
)
3148 && INTEGRAL_TYPE_P (TREE_TYPE (expr
->u
.nary
->op
[0]))
3149 && (TYPE_PRECISION (type
)
3150 >= TYPE_PRECISION (TREE_TYPE (expr
->u
.nary
->op
[0])))
3151 && SSA_NAME_RANGE_INFO (expr
->u
.nary
->op
[0]))
3154 if (get_range_info (expr
->u
.nary
->op
[0], &min
, &max
) == VR_RANGE
3155 && !wi::neg_p (min
, SIGNED
)
3156 && !wi::neg_p (max
, SIGNED
))
3157 /* Just handle extension and sign-changes of all-positive ranges. */
3158 set_range_info (temp
,
3159 SSA_NAME_RANGE_TYPE (expr
->u
.nary
->op
[0]),
3160 wide_int_storage::from (min
, TYPE_PRECISION (type
),
3162 wide_int_storage::from (max
, TYPE_PRECISION (type
),
3166 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3168 fprintf (dump_file
, "Created phi ");
3169 print_gimple_stmt (dump_file
, phi
, 0);
3170 fprintf (dump_file
, " in block %d (%04d)\n", block
->index
, val
);
3178 /* Perform insertion of partially redundant or hoistable values.
3179 For BLOCK, do the following:
3180 1. Propagate the NEW_SETS of the dominator into the current block.
3181 If the block has multiple predecessors,
3182 2a. Iterate over the ANTIC expressions for the block to see if
3183 any of them are partially redundant.
3184 2b. If so, insert them into the necessary predecessors to make
3185 the expression fully redundant.
3186 2c. Insert a new PHI merging the values of the predecessors.
3187 2d. Insert the new PHI, and the new expressions, into the
3189 If the block has multiple successors,
3190 3a. Iterate over the ANTIC values for the block to see if
3191 any of them are good candidates for hoisting.
3192 3b. If so, insert expressions computing the values in BLOCK,
3193 and add the new expressions into the NEW_SETS set.
3194 4. Recursively call ourselves on the dominator children of BLOCK.
3196 Steps 1, 2a, and 4 are done by insert_aux. 2b, 2c and 2d are done by
3197 do_pre_regular_insertion and do_partial_insertion. 3a and 3b are
3198 done in do_hoist_insertion.
3202 do_pre_regular_insertion (basic_block block
, basic_block dom
)
3204 bool new_stuff
= false;
3205 vec
<pre_expr
> exprs
;
3207 auto_vec
<pre_expr
> avail
;
3210 exprs
= sorted_array_from_bitmap_set (ANTIC_IN (block
));
3211 avail
.safe_grow (EDGE_COUNT (block
->preds
));
3213 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
3215 if (expr
->kind
== NARY
3216 || expr
->kind
== REFERENCE
)
3219 bool by_some
= false;
3220 bool cant_insert
= false;
3221 bool all_same
= true;
3222 pre_expr first_s
= NULL
;
3225 pre_expr eprime
= NULL
;
3227 pre_expr edoubleprime
= NULL
;
3228 bool do_insertion
= false;
3230 val
= get_expr_value_id (expr
);
3231 if (bitmap_set_contains_value (PHI_GEN (block
), val
))
3233 if (bitmap_set_contains_value (AVAIL_OUT (dom
), val
))
3235 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3237 fprintf (dump_file
, "Found fully redundant value: ");
3238 print_pre_expr (dump_file
, expr
);
3239 fprintf (dump_file
, "\n");
3244 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3246 unsigned int vprime
;
3248 /* We should never run insertion for the exit block
3249 and so not come across fake pred edges. */
3250 gcc_assert (!(pred
->flags
& EDGE_FAKE
));
3252 /* We are looking at ANTIC_OUT of bprime. */
3253 eprime
= phi_translate (expr
, ANTIC_IN (block
), NULL
,
3256 /* eprime will generally only be NULL if the
3257 value of the expression, translated
3258 through the PHI for this predecessor, is
3259 undefined. If that is the case, we can't
3260 make the expression fully redundant,
3261 because its value is undefined along a
3262 predecessor path. We can thus break out
3263 early because it doesn't matter what the
3264 rest of the results are. */
3267 avail
[pred
->dest_idx
] = NULL
;
3272 vprime
= get_expr_value_id (eprime
);
3273 edoubleprime
= bitmap_find_leader (AVAIL_OUT (bprime
),
3275 if (edoubleprime
== NULL
)
3277 avail
[pred
->dest_idx
] = eprime
;
3282 avail
[pred
->dest_idx
] = edoubleprime
;
3284 /* We want to perform insertions to remove a redundancy on
3285 a path in the CFG we want to optimize for speed. */
3286 if (optimize_edge_for_speed_p (pred
))
3287 do_insertion
= true;
3288 if (first_s
== NULL
)
3289 first_s
= edoubleprime
;
3290 else if (!pre_expr_d::equal (first_s
, edoubleprime
))
3294 /* If we can insert it, it's not the same value
3295 already existing along every predecessor, and
3296 it's defined by some predecessor, it is
3297 partially redundant. */
3298 if (!cant_insert
&& !all_same
&& by_some
)
3302 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3304 fprintf (dump_file
, "Skipping partial redundancy for "
3306 print_pre_expr (dump_file
, expr
);
3307 fprintf (dump_file
, " (%04d), no redundancy on to be "
3308 "optimized for speed edge\n", val
);
3311 else if (dbg_cnt (treepre_insert
))
3313 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3315 fprintf (dump_file
, "Found partial redundancy for "
3317 print_pre_expr (dump_file
, expr
);
3318 fprintf (dump_file
, " (%04d)\n",
3319 get_expr_value_id (expr
));
3321 if (insert_into_preds_of_block (block
,
3322 get_expression_id (expr
),
3327 /* If all edges produce the same value and that value is
3328 an invariant, then the PHI has the same value on all
3329 edges. Note this. */
3330 else if (!cant_insert
&& all_same
)
3332 gcc_assert (edoubleprime
->kind
== CONSTANT
3333 || edoubleprime
->kind
== NAME
);
3335 tree temp
= make_temp_ssa_name (get_expr_type (expr
),
3338 = gimple_build_assign (temp
,
3339 edoubleprime
->kind
== CONSTANT
?
3340 PRE_EXPR_CONSTANT (edoubleprime
) :
3341 PRE_EXPR_NAME (edoubleprime
));
3342 gimple_stmt_iterator gsi
= gsi_after_labels (block
);
3343 gsi_insert_before (&gsi
, assign
, GSI_NEW_STMT
);
3345 gimple_set_plf (assign
, NECESSARY
, false);
3346 VN_INFO_GET (temp
)->value_id
= val
;
3347 VN_INFO (temp
)->valnum
= sccvn_valnum_from_value_id (val
);
3348 if (VN_INFO (temp
)->valnum
== NULL_TREE
)
3349 VN_INFO (temp
)->valnum
= temp
;
3350 bitmap_set_bit (inserted_exprs
, SSA_NAME_VERSION (temp
));
3351 pre_expr newe
= get_or_alloc_expr_for_name (temp
);
3352 add_to_value (val
, newe
);
3353 bitmap_value_replace_in_set (AVAIL_OUT (block
), newe
);
3354 bitmap_insert_into_set (NEW_SETS (block
), newe
);
3364 /* Perform insertion for partially anticipatable expressions. There
3365 is only one case we will perform insertion for these. This case is
3366 if the expression is partially anticipatable, and fully available.
3367 In this case, we know that putting it earlier will enable us to
3368 remove the later computation. */
3371 do_pre_partial_partial_insertion (basic_block block
, basic_block dom
)
3373 bool new_stuff
= false;
3374 vec
<pre_expr
> exprs
;
3376 auto_vec
<pre_expr
> avail
;
3379 exprs
= sorted_array_from_bitmap_set (PA_IN (block
));
3380 avail
.safe_grow (EDGE_COUNT (block
->preds
));
3382 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
3384 if (expr
->kind
== NARY
3385 || expr
->kind
== REFERENCE
)
3389 bool cant_insert
= false;
3392 pre_expr eprime
= NULL
;
3395 val
= get_expr_value_id (expr
);
3396 if (bitmap_set_contains_value (PHI_GEN (block
), val
))
3398 if (bitmap_set_contains_value (AVAIL_OUT (dom
), val
))
3401 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3403 unsigned int vprime
;
3404 pre_expr edoubleprime
;
3406 /* We should never run insertion for the exit block
3407 and so not come across fake pred edges. */
3408 gcc_assert (!(pred
->flags
& EDGE_FAKE
));
3410 eprime
= phi_translate (expr
, ANTIC_IN (block
),
3414 /* eprime will generally only be NULL if the
3415 value of the expression, translated
3416 through the PHI for this predecessor, is
3417 undefined. If that is the case, we can't
3418 make the expression fully redundant,
3419 because its value is undefined along a
3420 predecessor path. We can thus break out
3421 early because it doesn't matter what the
3422 rest of the results are. */
3425 avail
[pred
->dest_idx
] = NULL
;
3430 vprime
= get_expr_value_id (eprime
);
3431 edoubleprime
= bitmap_find_leader (AVAIL_OUT (bprime
), vprime
);
3432 avail
[pred
->dest_idx
] = edoubleprime
;
3433 if (edoubleprime
== NULL
)
3440 /* If we can insert it, it's not the same value
3441 already existing along every predecessor, and
3442 it's defined by some predecessor, it is
3443 partially redundant. */
3444 if (!cant_insert
&& by_all
)
3447 bool do_insertion
= false;
3449 /* Insert only if we can remove a later expression on a path
3450 that we want to optimize for speed.
3451 The phi node that we will be inserting in BLOCK is not free,
3452 and inserting it for the sake of !optimize_for_speed successor
3453 may cause regressions on the speed path. */
3454 FOR_EACH_EDGE (succ
, ei
, block
->succs
)
3456 if (bitmap_set_contains_value (PA_IN (succ
->dest
), val
)
3457 || bitmap_set_contains_value (ANTIC_IN (succ
->dest
), val
))
3459 if (optimize_edge_for_speed_p (succ
))
3460 do_insertion
= true;
3466 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3468 fprintf (dump_file
, "Skipping partial partial redundancy "
3470 print_pre_expr (dump_file
, expr
);
3471 fprintf (dump_file
, " (%04d), not (partially) anticipated "
3472 "on any to be optimized for speed edges\n", val
);
3475 else if (dbg_cnt (treepre_insert
))
3477 pre_stats
.pa_insert
++;
3478 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3480 fprintf (dump_file
, "Found partial partial redundancy "
3482 print_pre_expr (dump_file
, expr
);
3483 fprintf (dump_file
, " (%04d)\n",
3484 get_expr_value_id (expr
));
3486 if (insert_into_preds_of_block (block
,
3487 get_expression_id (expr
),
3499 /* Insert expressions in BLOCK to compute hoistable values up.
3500 Return TRUE if something was inserted, otherwise return FALSE.
3501 The caller has to make sure that BLOCK has at least two successors. */
3504 do_hoist_insertion (basic_block block
)
3508 bool new_stuff
= false;
3510 gimple_stmt_iterator last
;
3512 /* At least two successors, or else... */
3513 gcc_assert (EDGE_COUNT (block
->succs
) >= 2);
3515 /* Check that all successors of BLOCK are dominated by block.
3516 We could use dominated_by_p() for this, but actually there is a much
3517 quicker check: any successor that is dominated by BLOCK can't have
3518 more than one predecessor edge. */
3519 FOR_EACH_EDGE (e
, ei
, block
->succs
)
3520 if (! single_pred_p (e
->dest
))
3523 /* Determine the insertion point. If we cannot safely insert before
3524 the last stmt if we'd have to, bail out. */
3525 last
= gsi_last_bb (block
);
3526 if (!gsi_end_p (last
)
3527 && !is_ctrl_stmt (gsi_stmt (last
))
3528 && stmt_ends_bb_p (gsi_stmt (last
)))
3531 /* Compute the set of hoistable expressions from ANTIC_IN. First compute
3532 hoistable values. */
3533 bitmap_set hoistable_set
;
3535 /* A hoistable value must be in ANTIC_IN(block)
3536 but not in AVAIL_OUT(BLOCK). */
3537 bitmap_initialize (&hoistable_set
.values
, &grand_bitmap_obstack
);
3538 bitmap_and_compl (&hoistable_set
.values
,
3539 &ANTIC_IN (block
)->values
, &AVAIL_OUT (block
)->values
);
3541 /* Short-cut for a common case: hoistable_set is empty. */
3542 if (bitmap_empty_p (&hoistable_set
.values
))
3545 /* Compute which of the hoistable values is in AVAIL_OUT of
3546 at least one of the successors of BLOCK. */
3547 bitmap_head availout_in_some
;
3548 bitmap_initialize (&availout_in_some
, &grand_bitmap_obstack
);
3549 FOR_EACH_EDGE (e
, ei
, block
->succs
)
3550 /* Do not consider expressions solely because their availability
3551 on loop exits. They'd be ANTIC-IN throughout the whole loop
3552 and thus effectively hoisted across loops by combination of
3553 PRE and hoisting. */
3554 if (! loop_exit_edge_p (block
->loop_father
, e
))
3555 bitmap_ior_and_into (&availout_in_some
, &hoistable_set
.values
,
3556 &AVAIL_OUT (e
->dest
)->values
);
3557 bitmap_clear (&hoistable_set
.values
);
3559 /* Short-cut for a common case: availout_in_some is empty. */
3560 if (bitmap_empty_p (&availout_in_some
))
3563 /* Hack hoitable_set in-place so we can use sorted_array_from_bitmap_set. */
3564 hoistable_set
.values
= availout_in_some
;
3565 hoistable_set
.expressions
= ANTIC_IN (block
)->expressions
;
3567 /* Now finally construct the topological-ordered expression set. */
3568 vec
<pre_expr
> exprs
= sorted_array_from_bitmap_set (&hoistable_set
);
3570 bitmap_clear (&hoistable_set
.values
);
3572 /* If there are candidate values for hoisting, insert expressions
3573 strategically to make the hoistable expressions fully redundant. */
3575 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
3577 /* While we try to sort expressions topologically above the
3578 sorting doesn't work out perfectly. Catch expressions we
3579 already inserted. */
3580 unsigned int value_id
= get_expr_value_id (expr
);
3581 if (bitmap_set_contains_value (AVAIL_OUT (block
), value_id
))
3583 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3586 "Already inserted expression for ");
3587 print_pre_expr (dump_file
, expr
);
3588 fprintf (dump_file
, " (%04d)\n", value_id
);
3593 /* OK, we should hoist this value. Perform the transformation. */
3594 pre_stats
.hoist_insert
++;
3595 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3598 "Inserting expression in block %d for code hoisting: ",
3600 print_pre_expr (dump_file
, expr
);
3601 fprintf (dump_file
, " (%04d)\n", value_id
);
3604 gimple_seq stmts
= NULL
;
3605 tree res
= create_expression_by_pieces (block
, expr
, &stmts
,
3606 get_expr_type (expr
));
3608 /* Do not return true if expression creation ultimately
3609 did not insert any statements. */
3610 if (gimple_seq_empty_p (stmts
))
3614 if (gsi_end_p (last
) || is_ctrl_stmt (gsi_stmt (last
)))
3615 gsi_insert_seq_before (&last
, stmts
, GSI_SAME_STMT
);
3617 gsi_insert_seq_after (&last
, stmts
, GSI_NEW_STMT
);
3620 /* Make sure to not return true if expression creation ultimately
3621 failed but also make sure to insert any stmts produced as they
3622 are tracked in inserted_exprs. */
3634 /* Do a dominator walk on the control flow graph, and insert computations
3635 of values as necessary for PRE and hoisting. */
3638 insert_aux (basic_block block
, bool do_pre
, bool do_hoist
)
3641 bool new_stuff
= false;
3646 dom
= get_immediate_dominator (CDI_DOMINATORS
, block
);
3651 bitmap_set_t newset
;
3653 /* First, update the AVAIL_OUT set with anything we may have
3654 inserted higher up in the dominator tree. */
3655 newset
= NEW_SETS (dom
);
3658 /* Note that we need to value_replace both NEW_SETS, and
3659 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3660 represented by some non-simple expression here that we want
3661 to replace it with. */
3662 FOR_EACH_EXPR_ID_IN_SET (newset
, i
, bi
)
3664 pre_expr expr
= expression_for_id (i
);
3665 bitmap_value_replace_in_set (NEW_SETS (block
), expr
);
3666 bitmap_value_replace_in_set (AVAIL_OUT (block
), expr
);
3670 /* Insert expressions for partial redundancies. */
3671 if (do_pre
&& !single_pred_p (block
))
3673 new_stuff
|= do_pre_regular_insertion (block
, dom
);
3674 if (do_partial_partial
)
3675 new_stuff
|= do_pre_partial_partial_insertion (block
, dom
);
3678 /* Insert expressions for hoisting. */
3679 if (do_hoist
&& EDGE_COUNT (block
->succs
) >= 2)
3680 new_stuff
|= do_hoist_insertion (block
);
3683 for (son
= first_dom_son (CDI_DOMINATORS
, block
);
3685 son
= next_dom_son (CDI_DOMINATORS
, son
))
3687 new_stuff
|= insert_aux (son
, do_pre
, do_hoist
);
3693 /* Perform insertion of partially redundant and hoistable values. */
3698 bool new_stuff
= true;
3700 int num_iterations
= 0;
3702 FOR_ALL_BB_FN (bb
, cfun
)
3703 NEW_SETS (bb
) = bitmap_set_new ();
3708 if (dump_file
&& dump_flags
& TDF_DETAILS
)
3709 fprintf (dump_file
, "Starting insert iteration %d\n", num_iterations
);
3710 new_stuff
= insert_aux (ENTRY_BLOCK_PTR_FOR_FN (cfun
), flag_tree_pre
,
3711 flag_code_hoisting
);
3713 /* Clear the NEW sets before the next iteration. We have already
3714 fully propagated its contents. */
3716 FOR_ALL_BB_FN (bb
, cfun
)
3717 bitmap_set_free (NEW_SETS (bb
));
3719 statistics_histogram_event (cfun
, "insert iterations", num_iterations
);
3723 /* Compute the AVAIL set for all basic blocks.
3725 This function performs value numbering of the statements in each basic
3726 block. The AVAIL sets are built from information we glean while doing
3727 this value numbering, since the AVAIL sets contain only one entry per
3730 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3731 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3734 compute_avail (void)
3737 basic_block block
, son
;
3738 basic_block
*worklist
;
3743 /* We pretend that default definitions are defined in the entry block.
3744 This includes function arguments and the static chain decl. */
3745 FOR_EACH_SSA_NAME (i
, name
, cfun
)
3748 if (!SSA_NAME_IS_DEFAULT_DEF (name
)
3749 || has_zero_uses (name
)
3750 || virtual_operand_p (name
))
3753 e
= get_or_alloc_expr_for_name (name
);
3754 add_to_value (get_expr_value_id (e
), e
);
3755 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun
)), e
);
3756 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun
)),
3760 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3762 print_bitmap_set (dump_file
, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun
)),
3763 "tmp_gen", ENTRY_BLOCK
);
3764 print_bitmap_set (dump_file
, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun
)),
3765 "avail_out", ENTRY_BLOCK
);
3768 /* Allocate the worklist. */
3769 worklist
= XNEWVEC (basic_block
, n_basic_blocks_for_fn (cfun
));
3771 /* Seed the algorithm by putting the dominator children of the entry
3772 block on the worklist. */
3773 for (son
= first_dom_son (CDI_DOMINATORS
, ENTRY_BLOCK_PTR_FOR_FN (cfun
));
3775 son
= next_dom_son (CDI_DOMINATORS
, son
))
3776 worklist
[sp
++] = son
;
3778 BB_LIVE_VOP_ON_EXIT (ENTRY_BLOCK_PTR_FOR_FN (cfun
))
3779 = ssa_default_def (cfun
, gimple_vop (cfun
));
3781 /* Loop until the worklist is empty. */
3787 /* Pick a block from the worklist. */
3788 block
= worklist
[--sp
];
3790 /* Initially, the set of available values in BLOCK is that of
3791 its immediate dominator. */
3792 dom
= get_immediate_dominator (CDI_DOMINATORS
, block
);
3795 bitmap_set_copy (AVAIL_OUT (block
), AVAIL_OUT (dom
));
3796 BB_LIVE_VOP_ON_EXIT (block
) = BB_LIVE_VOP_ON_EXIT (dom
);
3799 /* Generate values for PHI nodes. */
3800 for (gphi_iterator gsi
= gsi_start_phis (block
); !gsi_end_p (gsi
);
3803 tree result
= gimple_phi_result (gsi
.phi ());
3805 /* We have no need for virtual phis, as they don't represent
3806 actual computations. */
3807 if (virtual_operand_p (result
))
3809 BB_LIVE_VOP_ON_EXIT (block
) = result
;
3813 pre_expr e
= get_or_alloc_expr_for_name (result
);
3814 add_to_value (get_expr_value_id (e
), e
);
3815 bitmap_value_insert_into_set (AVAIL_OUT (block
), e
);
3816 bitmap_insert_into_set (PHI_GEN (block
), e
);
3819 BB_MAY_NOTRETURN (block
) = 0;
3821 /* Now compute value numbers and populate value sets with all
3822 the expressions computed in BLOCK. */
3823 for (gimple_stmt_iterator gsi
= gsi_start_bb (block
); !gsi_end_p (gsi
);
3829 stmt
= gsi_stmt (gsi
);
3831 /* Cache whether the basic-block has any non-visible side-effect
3833 If this isn't a call or it is the last stmt in the
3834 basic-block then the CFG represents things correctly. */
3835 if (is_gimple_call (stmt
) && !stmt_ends_bb_p (stmt
))
3837 /* Non-looping const functions always return normally.
3838 Otherwise the call might not return or have side-effects
3839 that forbids hoisting possibly trapping expressions
3841 int flags
= gimple_call_flags (stmt
);
3842 if (!(flags
& ECF_CONST
)
3843 || (flags
& ECF_LOOPING_CONST_OR_PURE
))
3844 BB_MAY_NOTRETURN (block
) = 1;
3847 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_DEF
)
3849 pre_expr e
= get_or_alloc_expr_for_name (op
);
3851 add_to_value (get_expr_value_id (e
), e
);
3852 bitmap_insert_into_set (TMP_GEN (block
), e
);
3853 bitmap_value_insert_into_set (AVAIL_OUT (block
), e
);
3856 if (gimple_vdef (stmt
))
3857 BB_LIVE_VOP_ON_EXIT (block
) = gimple_vdef (stmt
);
3859 if (gimple_has_side_effects (stmt
)
3860 || stmt_could_throw_p (stmt
)
3861 || is_gimple_debug (stmt
))
3864 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_USE
)
3866 if (ssa_undefined_value_p (op
))
3868 pre_expr e
= get_or_alloc_expr_for_name (op
);
3869 bitmap_value_insert_into_set (EXP_GEN (block
), e
);
3872 switch (gimple_code (stmt
))
3880 vn_reference_s ref1
;
3881 pre_expr result
= NULL
;
3883 /* We can value number only calls to real functions. */
3884 if (gimple_call_internal_p (stmt
))
3887 vn_reference_lookup_call (as_a
<gcall
*> (stmt
), &ref
, &ref1
);
3891 /* If the value of the call is not invalidated in
3892 this block until it is computed, add the expression
3894 if (!gimple_vuse (stmt
)
3896 (SSA_NAME_DEF_STMT (gimple_vuse (stmt
))) == GIMPLE_PHI
3897 || gimple_bb (SSA_NAME_DEF_STMT
3898 (gimple_vuse (stmt
))) != block
)
3900 result
= pre_expr_pool
.allocate ();
3901 result
->kind
= REFERENCE
;
3903 PRE_EXPR_REFERENCE (result
) = ref
;
3905 get_or_alloc_expression_id (result
);
3906 add_to_value (get_expr_value_id (result
), result
);
3907 bitmap_value_insert_into_set (EXP_GEN (block
), result
);
3914 pre_expr result
= NULL
;
3915 switch (vn_get_stmt_kind (stmt
))
3919 enum tree_code code
= gimple_assign_rhs_code (stmt
);
3922 /* COND_EXPR and VEC_COND_EXPR are awkward in
3923 that they contain an embedded complex expression.
3924 Don't even try to shove those through PRE. */
3925 if (code
== COND_EXPR
3926 || code
== VEC_COND_EXPR
)
3929 vn_nary_op_lookup_stmt (stmt
, &nary
);
3933 /* If the NARY traps and there was a preceding
3934 point in the block that might not return avoid
3935 adding the nary to EXP_GEN. */
3936 if (BB_MAY_NOTRETURN (block
)
3937 && vn_nary_may_trap (nary
))
3940 result
= pre_expr_pool
.allocate ();
3941 result
->kind
= NARY
;
3943 PRE_EXPR_NARY (result
) = nary
;
3949 tree rhs1
= gimple_assign_rhs1 (stmt
);
3950 alias_set_type set
= get_alias_set (rhs1
);
3951 vec
<vn_reference_op_s
> operands
3952 = vn_reference_operands_for_lookup (rhs1
);
3954 vn_reference_lookup_pieces (gimple_vuse (stmt
), set
,
3956 operands
, &ref
, VN_WALK
);
3959 operands
.release ();
3963 /* If the value of the reference is not invalidated in
3964 this block until it is computed, add the expression
3966 if (gimple_vuse (stmt
))
3970 def_stmt
= SSA_NAME_DEF_STMT (gimple_vuse (stmt
));
3971 while (!gimple_nop_p (def_stmt
)
3972 && gimple_code (def_stmt
) != GIMPLE_PHI
3973 && gimple_bb (def_stmt
) == block
)
3975 if (stmt_may_clobber_ref_p
3976 (def_stmt
, gimple_assign_rhs1 (stmt
)))
3982 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt
));
3986 operands
.release ();
3991 /* If the load was value-numbered to another
3992 load make sure we do not use its expression
3993 for insertion if it wouldn't be a valid
3995 /* At the momemt we have a testcase
3996 for hoist insertion of aligned vs. misaligned
3997 variants in gcc.dg/torture/pr65270-1.c thus
3998 with just alignment to be considered we can
3999 simply replace the expression in the hashtable
4000 with the most conservative one. */
4001 vn_reference_op_t ref1
= &ref
->operands
.last ();
4002 while (ref1
->opcode
!= TARGET_MEM_REF
4003 && ref1
->opcode
!= MEM_REF
4004 && ref1
!= &ref
->operands
[0])
4006 vn_reference_op_t ref2
= &operands
.last ();
4007 while (ref2
->opcode
!= TARGET_MEM_REF
4008 && ref2
->opcode
!= MEM_REF
4009 && ref2
!= &operands
[0])
4011 if ((ref1
->opcode
== TARGET_MEM_REF
4012 || ref1
->opcode
== MEM_REF
)
4013 && (TYPE_ALIGN (ref1
->type
)
4014 > TYPE_ALIGN (ref2
->type
)))
4016 = build_aligned_type (ref1
->type
,
4017 TYPE_ALIGN (ref2
->type
));
4018 /* TBAA behavior is an obvious part so make sure
4019 that the hashtable one covers this as well
4020 by adjusting the ref alias set and its base. */
4022 || alias_set_subset_of (set
, ref
->set
))
4024 else if (alias_set_subset_of (ref
->set
, set
))
4027 if (ref1
->opcode
== MEM_REF
)
4028 ref1
->op0
= wide_int_to_tree (TREE_TYPE (ref2
->op0
),
4031 ref1
->op2
= wide_int_to_tree (TREE_TYPE (ref2
->op2
),
4037 if (ref1
->opcode
== MEM_REF
)
4038 ref1
->op0
= wide_int_to_tree (ptr_type_node
,
4041 ref1
->op2
= wide_int_to_tree (ptr_type_node
,
4044 operands
.release ();
4046 result
= pre_expr_pool
.allocate ();
4047 result
->kind
= REFERENCE
;
4049 PRE_EXPR_REFERENCE (result
) = ref
;
4057 get_or_alloc_expression_id (result
);
4058 add_to_value (get_expr_value_id (result
), result
);
4059 bitmap_value_insert_into_set (EXP_GEN (block
), result
);
4067 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4069 print_bitmap_set (dump_file
, EXP_GEN (block
),
4070 "exp_gen", block
->index
);
4071 print_bitmap_set (dump_file
, PHI_GEN (block
),
4072 "phi_gen", block
->index
);
4073 print_bitmap_set (dump_file
, TMP_GEN (block
),
4074 "tmp_gen", block
->index
);
4075 print_bitmap_set (dump_file
, AVAIL_OUT (block
),
4076 "avail_out", block
->index
);
4079 /* Put the dominator children of BLOCK on the worklist of blocks
4080 to compute available sets for. */
4081 for (son
= first_dom_son (CDI_DOMINATORS
, block
);
4083 son
= next_dom_son (CDI_DOMINATORS
, son
))
4084 worklist
[sp
++] = son
;
4091 /* Local state for the eliminate domwalk. */
4092 static vec
<gimple
*> el_to_remove
;
4093 static vec
<gimple
*> el_to_fixup
;
4094 static unsigned int el_todo
;
4095 static vec
<tree
> el_avail
;
4096 static vec
<tree
> el_avail_stack
;
4098 /* Return a leader for OP that is available at the current point of the
4099 eliminate domwalk. */
4102 eliminate_avail (tree op
)
4104 tree valnum
= VN_INFO (op
)->valnum
;
4105 if (TREE_CODE (valnum
) == SSA_NAME
)
4107 if (SSA_NAME_IS_DEFAULT_DEF (valnum
))
4109 if (el_avail
.length () > SSA_NAME_VERSION (valnum
))
4110 return el_avail
[SSA_NAME_VERSION (valnum
)];
4112 else if (is_gimple_min_invariant (valnum
))
4117 /* At the current point of the eliminate domwalk make OP available. */
4120 eliminate_push_avail (tree op
)
4122 tree valnum
= VN_INFO (op
)->valnum
;
4123 if (TREE_CODE (valnum
) == SSA_NAME
)
4125 if (el_avail
.length () <= SSA_NAME_VERSION (valnum
))
4126 el_avail
.safe_grow_cleared (SSA_NAME_VERSION (valnum
) + 1);
4128 if (el_avail
[SSA_NAME_VERSION (valnum
)])
4129 pushop
= el_avail
[SSA_NAME_VERSION (valnum
)];
4130 el_avail_stack
.safe_push (pushop
);
4131 el_avail
[SSA_NAME_VERSION (valnum
)] = op
;
4135 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
4136 the leader for the expression if insertion was successful. */
4139 eliminate_insert (gimple_stmt_iterator
*gsi
, tree val
)
4141 /* We can insert a sequence with a single assignment only. */
4142 gimple_seq stmts
= VN_INFO (val
)->expr
;
4143 if (!gimple_seq_singleton_p (stmts
))
4145 gassign
*stmt
= dyn_cast
<gassign
*> (gimple_seq_first_stmt (stmts
));
4147 || (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt
))
4148 && gimple_assign_rhs_code (stmt
) != VIEW_CONVERT_EXPR
4149 && gimple_assign_rhs_code (stmt
) != BIT_FIELD_REF
4150 && (gimple_assign_rhs_code (stmt
) != BIT_AND_EXPR
4151 || TREE_CODE (gimple_assign_rhs2 (stmt
)) != INTEGER_CST
)))
4154 tree op
= gimple_assign_rhs1 (stmt
);
4155 if (gimple_assign_rhs_code (stmt
) == VIEW_CONVERT_EXPR
4156 || gimple_assign_rhs_code (stmt
) == BIT_FIELD_REF
)
4157 op
= TREE_OPERAND (op
, 0);
4158 tree leader
= TREE_CODE (op
) == SSA_NAME
? eliminate_avail (op
) : op
;
4164 if (gimple_assign_rhs_code (stmt
) == BIT_FIELD_REF
)
4165 res
= gimple_build (&stmts
, BIT_FIELD_REF
,
4166 TREE_TYPE (val
), leader
,
4167 TREE_OPERAND (gimple_assign_rhs1 (stmt
), 1),
4168 TREE_OPERAND (gimple_assign_rhs1 (stmt
), 2));
4169 else if (gimple_assign_rhs_code (stmt
) == BIT_AND_EXPR
)
4170 res
= gimple_build (&stmts
, BIT_AND_EXPR
,
4171 TREE_TYPE (val
), leader
, gimple_assign_rhs2 (stmt
));
4173 res
= gimple_build (&stmts
, gimple_assign_rhs_code (stmt
),
4174 TREE_TYPE (val
), leader
);
4175 if (TREE_CODE (res
) != SSA_NAME
4176 || SSA_NAME_IS_DEFAULT_DEF (res
)
4177 || gimple_bb (SSA_NAME_DEF_STMT (res
)))
4179 gimple_seq_discard (stmts
);
4181 /* During propagation we have to treat SSA info conservatively
4182 and thus we can end up simplifying the inserted expression
4183 at elimination time to sth not defined in stmts. */
4184 /* But then this is a redundancy we failed to detect. Which means
4185 res now has two values. That doesn't play well with how
4186 we track availability here, so give up. */
4187 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4189 if (TREE_CODE (res
) == SSA_NAME
)
4190 res
= eliminate_avail (res
);
4193 fprintf (dump_file
, "Failed to insert expression for value ");
4194 print_generic_expr (dump_file
, val
);
4195 fprintf (dump_file
, " which is really fully redundant to ");
4196 print_generic_expr (dump_file
, res
);
4197 fprintf (dump_file
, "\n");
4205 gsi_insert_seq_before (gsi
, stmts
, GSI_SAME_STMT
);
4206 VN_INFO_GET (res
)->valnum
= val
;
4208 if (TREE_CODE (leader
) == SSA_NAME
)
4209 gimple_set_plf (SSA_NAME_DEF_STMT (leader
), NECESSARY
, true);
4212 pre_stats
.insertions
++;
4213 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4215 fprintf (dump_file
, "Inserted ");
4216 print_gimple_stmt (dump_file
, SSA_NAME_DEF_STMT (res
), 0);
4222 class eliminate_dom_walker
: public dom_walker
4225 eliminate_dom_walker (cdi_direction direction
, bool do_pre_
)
4226 : dom_walker (direction
), do_pre (do_pre_
) {}
4228 virtual edge
before_dom_children (basic_block
);
4229 virtual void after_dom_children (basic_block
);
4234 /* Perform elimination for the basic-block B during the domwalk. */
4237 eliminate_dom_walker::before_dom_children (basic_block b
)
4240 el_avail_stack
.safe_push (NULL_TREE
);
4242 /* Skip unreachable blocks marked unreachable during the SCCVN domwalk. */
4245 FOR_EACH_EDGE (e
, ei
, b
->preds
)
4246 if (e
->flags
& EDGE_EXECUTABLE
)
4251 for (gphi_iterator gsi
= gsi_start_phis (b
); !gsi_end_p (gsi
);)
4253 gphi
*phi
= gsi
.phi ();
4254 tree res
= PHI_RESULT (phi
);
4256 if (virtual_operand_p (res
))
4262 tree sprime
= eliminate_avail (res
);
4266 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4268 fprintf (dump_file
, "Replaced redundant PHI node defining ");
4269 print_generic_expr (dump_file
, res
);
4270 fprintf (dump_file
, " with ");
4271 print_generic_expr (dump_file
, sprime
);
4272 fprintf (dump_file
, "\n");
4275 /* If we inserted this PHI node ourself, it's not an elimination. */
4277 && bitmap_bit_p (inserted_exprs
, SSA_NAME_VERSION (res
)))
4280 pre_stats
.eliminations
++;
4282 /* If we will propagate into all uses don't bother to do
4284 if (may_propagate_copy (res
, sprime
))
4286 /* Mark the PHI for removal. */
4287 el_to_remove
.safe_push (phi
);
4292 remove_phi_node (&gsi
, false);
4295 && !bitmap_bit_p (inserted_exprs
, SSA_NAME_VERSION (res
))
4296 && TREE_CODE (sprime
) == SSA_NAME
)
4297 gimple_set_plf (SSA_NAME_DEF_STMT (sprime
), NECESSARY
, true);
4299 if (!useless_type_conversion_p (TREE_TYPE (res
), TREE_TYPE (sprime
)))
4300 sprime
= fold_convert (TREE_TYPE (res
), sprime
);
4301 gimple
*stmt
= gimple_build_assign (res
, sprime
);
4302 /* ??? It cannot yet be necessary (DOM walk). */
4303 gimple_set_plf (stmt
, NECESSARY
, gimple_plf (phi
, NECESSARY
));
4305 gimple_stmt_iterator gsi2
= gsi_after_labels (b
);
4306 gsi_insert_before (&gsi2
, stmt
, GSI_NEW_STMT
);
4310 eliminate_push_avail (res
);
4314 for (gimple_stmt_iterator gsi
= gsi_start_bb (b
);
4318 tree sprime
= NULL_TREE
;
4319 gimple
*stmt
= gsi_stmt (gsi
);
4320 tree lhs
= gimple_get_lhs (stmt
);
4321 if (lhs
&& TREE_CODE (lhs
) == SSA_NAME
4322 && !gimple_has_volatile_ops (stmt
)
4323 /* See PR43491. Do not replace a global register variable when
4324 it is a the RHS of an assignment. Do replace local register
4325 variables since gcc does not guarantee a local variable will
4326 be allocated in register.
4327 ??? The fix isn't effective here. This should instead
4328 be ensured by not value-numbering them the same but treating
4329 them like volatiles? */
4330 && !(gimple_assign_single_p (stmt
)
4331 && (TREE_CODE (gimple_assign_rhs1 (stmt
)) == VAR_DECL
4332 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt
))
4333 && is_global_var (gimple_assign_rhs1 (stmt
)))))
4335 sprime
= eliminate_avail (lhs
);
4338 /* If there is no existing usable leader but SCCVN thinks
4339 it has an expression it wants to use as replacement,
4341 tree val
= VN_INFO (lhs
)->valnum
;
4343 && TREE_CODE (val
) == SSA_NAME
4344 && VN_INFO (val
)->needs_insertion
4345 && VN_INFO (val
)->expr
!= NULL
4346 && (sprime
= eliminate_insert (&gsi
, val
)) != NULL_TREE
)
4347 eliminate_push_avail (sprime
);
4350 /* If this now constitutes a copy duplicate points-to
4351 and range info appropriately. This is especially
4352 important for inserted code. See tree-ssa-copy.c
4353 for similar code. */
4355 && TREE_CODE (sprime
) == SSA_NAME
)
4357 basic_block sprime_b
= gimple_bb (SSA_NAME_DEF_STMT (sprime
));
4358 if (POINTER_TYPE_P (TREE_TYPE (lhs
))
4359 && VN_INFO_PTR_INFO (lhs
)
4360 && ! VN_INFO_PTR_INFO (sprime
))
4362 duplicate_ssa_name_ptr_info (sprime
,
4363 VN_INFO_PTR_INFO (lhs
));
4365 mark_ptr_info_alignment_unknown
4366 (SSA_NAME_PTR_INFO (sprime
));
4368 else if (INTEGRAL_TYPE_P (TREE_TYPE (lhs
))
4369 && VN_INFO_RANGE_INFO (lhs
)
4370 && ! VN_INFO_RANGE_INFO (sprime
)
4372 duplicate_ssa_name_range_info (sprime
,
4373 VN_INFO_RANGE_TYPE (lhs
),
4374 VN_INFO_RANGE_INFO (lhs
));
4377 /* Inhibit the use of an inserted PHI on a loop header when
4378 the address of the memory reference is a simple induction
4379 variable. In other cases the vectorizer won't do anything
4380 anyway (either it's loop invariant or a complicated
4383 && TREE_CODE (sprime
) == SSA_NAME
4385 && (flag_tree_loop_vectorize
|| flag_tree_parallelize_loops
> 1)
4386 && loop_outer (b
->loop_father
)
4387 && has_zero_uses (sprime
)
4388 && bitmap_bit_p (inserted_exprs
, SSA_NAME_VERSION (sprime
))
4389 && gimple_assign_load_p (stmt
))
4391 gimple
*def_stmt
= SSA_NAME_DEF_STMT (sprime
);
4392 basic_block def_bb
= gimple_bb (def_stmt
);
4393 if (gimple_code (def_stmt
) == GIMPLE_PHI
4394 && def_bb
->loop_father
->header
== def_bb
)
4396 loop_p loop
= def_bb
->loop_father
;
4400 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_USE
)
4403 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (op
));
4405 && flow_bb_inside_loop_p (loop
, def_bb
)
4406 && simple_iv (loop
, loop
, op
, &iv
, true))
4414 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4416 fprintf (dump_file
, "Not replacing ");
4417 print_gimple_expr (dump_file
, stmt
, 0);
4418 fprintf (dump_file
, " with ");
4419 print_generic_expr (dump_file
, sprime
);
4420 fprintf (dump_file
, " which would add a loop"
4421 " carried dependence to loop %d\n",
4424 /* Don't keep sprime available. */
4432 /* If we can propagate the value computed for LHS into
4433 all uses don't bother doing anything with this stmt. */
4434 if (may_propagate_copy (lhs
, sprime
))
4436 /* Mark it for removal. */
4437 el_to_remove
.safe_push (stmt
);
4439 /* ??? Don't count copy/constant propagations. */
4440 if (gimple_assign_single_p (stmt
)
4441 && (TREE_CODE (gimple_assign_rhs1 (stmt
)) == SSA_NAME
4442 || gimple_assign_rhs1 (stmt
) == sprime
))
4445 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4447 fprintf (dump_file
, "Replaced ");
4448 print_gimple_expr (dump_file
, stmt
, 0);
4449 fprintf (dump_file
, " with ");
4450 print_generic_expr (dump_file
, sprime
);
4451 fprintf (dump_file
, " in all uses of ");
4452 print_gimple_stmt (dump_file
, stmt
, 0);
4455 pre_stats
.eliminations
++;
4459 /* If this is an assignment from our leader (which
4460 happens in the case the value-number is a constant)
4461 then there is nothing to do. */
4462 if (gimple_assign_single_p (stmt
)
4463 && sprime
== gimple_assign_rhs1 (stmt
))
4466 /* Else replace its RHS. */
4467 bool can_make_abnormal_goto
4468 = is_gimple_call (stmt
)
4469 && stmt_can_make_abnormal_goto (stmt
);
4471 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4473 fprintf (dump_file
, "Replaced ");
4474 print_gimple_expr (dump_file
, stmt
, 0);
4475 fprintf (dump_file
, " with ");
4476 print_generic_expr (dump_file
, sprime
);
4477 fprintf (dump_file
, " in ");
4478 print_gimple_stmt (dump_file
, stmt
, 0);
4481 if (TREE_CODE (sprime
) == SSA_NAME
)
4482 gimple_set_plf (SSA_NAME_DEF_STMT (sprime
),
4485 pre_stats
.eliminations
++;
4486 gimple
*orig_stmt
= stmt
;
4487 if (!useless_type_conversion_p (TREE_TYPE (lhs
),
4488 TREE_TYPE (sprime
)))
4489 sprime
= fold_convert (TREE_TYPE (lhs
), sprime
);
4490 tree vdef
= gimple_vdef (stmt
);
4491 tree vuse
= gimple_vuse (stmt
);
4492 propagate_tree_value_into_stmt (&gsi
, sprime
);
4493 stmt
= gsi_stmt (gsi
);
4495 if (vdef
!= gimple_vdef (stmt
))
4496 VN_INFO (vdef
)->valnum
= vuse
;
4498 /* If we removed EH side-effects from the statement, clean
4499 its EH information. */
4500 if (maybe_clean_or_replace_eh_stmt (orig_stmt
, stmt
))
4502 bitmap_set_bit (need_eh_cleanup
,
4503 gimple_bb (stmt
)->index
);
4504 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4505 fprintf (dump_file
, " Removed EH side-effects.\n");
4508 /* Likewise for AB side-effects. */
4509 if (can_make_abnormal_goto
4510 && !stmt_can_make_abnormal_goto (stmt
))
4512 bitmap_set_bit (need_ab_cleanup
,
4513 gimple_bb (stmt
)->index
);
4514 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4515 fprintf (dump_file
, " Removed AB side-effects.\n");
4522 /* If the statement is a scalar store, see if the expression
4523 has the same value number as its rhs. If so, the store is
4525 if (gimple_assign_single_p (stmt
)
4526 && !gimple_has_volatile_ops (stmt
)
4527 && !is_gimple_reg (gimple_assign_lhs (stmt
))
4528 && (TREE_CODE (gimple_assign_rhs1 (stmt
)) == SSA_NAME
4529 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt
))))
4532 tree rhs
= gimple_assign_rhs1 (stmt
);
4533 vn_reference_t vnresult
;
4534 val
= vn_reference_lookup (lhs
, gimple_vuse (stmt
), VN_WALKREWRITE
,
4536 if (TREE_CODE (rhs
) == SSA_NAME
)
4537 rhs
= VN_INFO (rhs
)->valnum
;
4539 && operand_equal_p (val
, rhs
, 0))
4541 /* We can only remove the later store if the former aliases
4542 at least all accesses the later one does or if the store
4543 was to readonly memory storing the same value. */
4544 alias_set_type set
= get_alias_set (lhs
);
4546 || vnresult
->set
== set
4547 || alias_set_subset_of (set
, vnresult
->set
))
4549 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4551 fprintf (dump_file
, "Deleted redundant store ");
4552 print_gimple_stmt (dump_file
, stmt
, 0);
4555 /* Queue stmt for removal. */
4556 el_to_remove
.safe_push (stmt
);
4562 /* If this is a control statement value numbering left edges
4563 unexecuted on force the condition in a way consistent with
4565 if (gcond
*cond
= dyn_cast
<gcond
*> (stmt
))
4567 if ((EDGE_SUCC (b
, 0)->flags
& EDGE_EXECUTABLE
)
4568 ^ (EDGE_SUCC (b
, 1)->flags
& EDGE_EXECUTABLE
))
4570 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4572 fprintf (dump_file
, "Removing unexecutable edge from ");
4573 print_gimple_stmt (dump_file
, stmt
, 0);
4575 if (((EDGE_SUCC (b
, 0)->flags
& EDGE_TRUE_VALUE
) != 0)
4576 == ((EDGE_SUCC (b
, 0)->flags
& EDGE_EXECUTABLE
) != 0))
4577 gimple_cond_make_true (cond
);
4579 gimple_cond_make_false (cond
);
4581 el_todo
|= TODO_cleanup_cfg
;
4586 bool can_make_abnormal_goto
= stmt_can_make_abnormal_goto (stmt
);
4587 bool was_noreturn
= (is_gimple_call (stmt
)
4588 && gimple_call_noreturn_p (stmt
));
4589 tree vdef
= gimple_vdef (stmt
);
4590 tree vuse
= gimple_vuse (stmt
);
4592 /* If we didn't replace the whole stmt (or propagate the result
4593 into all uses), replace all uses on this stmt with their
4595 use_operand_p use_p
;
4597 FOR_EACH_SSA_USE_OPERAND (use_p
, stmt
, iter
, SSA_OP_USE
)
4599 tree use
= USE_FROM_PTR (use_p
);
4600 /* ??? The call code above leaves stmt operands un-updated. */
4601 if (TREE_CODE (use
) != SSA_NAME
)
4603 tree sprime
= eliminate_avail (use
);
4604 if (sprime
&& sprime
!= use
4605 && may_propagate_copy (use
, sprime
)
4606 /* We substitute into debug stmts to avoid excessive
4607 debug temporaries created by removed stmts, but we need
4608 to avoid doing so for inserted sprimes as we never want
4609 to create debug temporaries for them. */
4611 || TREE_CODE (sprime
) != SSA_NAME
4612 || !is_gimple_debug (stmt
)
4613 || !bitmap_bit_p (inserted_exprs
, SSA_NAME_VERSION (sprime
))))
4615 propagate_value (use_p
, sprime
);
4616 gimple_set_modified (stmt
, true);
4617 if (TREE_CODE (sprime
) == SSA_NAME
4618 && !is_gimple_debug (stmt
))
4619 gimple_set_plf (SSA_NAME_DEF_STMT (sprime
),
4624 /* Visit indirect calls and turn them into direct calls if
4625 possible using the devirtualization machinery. */
4626 if (gcall
*call_stmt
= dyn_cast
<gcall
*> (stmt
))
4628 tree fn
= gimple_call_fn (call_stmt
);
4630 && flag_devirtualize
4631 && virtual_method_call_p (fn
))
4633 tree otr_type
= obj_type_ref_class (fn
);
4635 ipa_polymorphic_call_context
context (current_function_decl
, fn
, stmt
, &instance
);
4638 context
.get_dynamic_type (instance
, OBJ_TYPE_REF_OBJECT (fn
), otr_type
, stmt
);
4640 vec
<cgraph_node
*>targets
4641 = possible_polymorphic_call_targets (obj_type_ref_class (fn
),
4643 (OBJ_TYPE_REF_TOKEN (fn
)),
4647 dump_possible_polymorphic_call_targets (dump_file
,
4648 obj_type_ref_class (fn
),
4650 (OBJ_TYPE_REF_TOKEN (fn
)),
4652 if (final
&& targets
.length () <= 1 && dbg_cnt (devirt
))
4655 if (targets
.length () == 1)
4656 fn
= targets
[0]->decl
;
4658 fn
= builtin_decl_implicit (BUILT_IN_UNREACHABLE
);
4659 if (dump_enabled_p ())
4661 location_t loc
= gimple_location_safe (stmt
);
4662 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, loc
,
4663 "converting indirect call to "
4665 lang_hooks
.decl_printable_name (fn
, 2));
4667 gimple_call_set_fndecl (call_stmt
, fn
);
4668 /* If changing the call to __builtin_unreachable
4669 or similar noreturn function, adjust gimple_call_fntype
4671 if (gimple_call_noreturn_p (call_stmt
)
4672 && VOID_TYPE_P (TREE_TYPE (TREE_TYPE (fn
)))
4673 && TYPE_ARG_TYPES (TREE_TYPE (fn
))
4674 && (TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fn
)))
4676 gimple_call_set_fntype (call_stmt
, TREE_TYPE (fn
));
4677 maybe_remove_unused_call_args (cfun
, call_stmt
);
4678 gimple_set_modified (stmt
, true);
4683 if (gimple_modified_p (stmt
))
4685 /* If a formerly non-invariant ADDR_EXPR is turned into an
4686 invariant one it was on a separate stmt. */
4687 if (gimple_assign_single_p (stmt
)
4688 && TREE_CODE (gimple_assign_rhs1 (stmt
)) == ADDR_EXPR
)
4689 recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt
));
4690 gimple
*old_stmt
= stmt
;
4691 gimple_stmt_iterator prev
= gsi
;
4693 if (fold_stmt (&gsi
))
4695 /* fold_stmt may have created new stmts inbetween
4696 the previous stmt and the folded stmt. Mark
4697 all defs created there as varying to not confuse
4698 the SCCVN machinery as we're using that even during
4700 if (gsi_end_p (prev
))
4701 prev
= gsi_start_bb (b
);
4704 if (gsi_stmt (prev
) != gsi_stmt (gsi
))
4709 FOR_EACH_SSA_TREE_OPERAND (def
, gsi_stmt (prev
),
4710 dit
, SSA_OP_ALL_DEFS
)
4711 /* As existing DEFs may move between stmts
4712 we have to guard VN_INFO_GET. */
4713 if (! has_VN_INFO (def
))
4714 VN_INFO_GET (def
)->valnum
= def
;
4715 if (gsi_stmt (prev
) == gsi_stmt (gsi
))
4721 stmt
= gsi_stmt (gsi
);
4722 /* When changing a call into a noreturn call, cfg cleanup
4723 is needed to fix up the noreturn call. */
4725 && is_gimple_call (stmt
) && gimple_call_noreturn_p (stmt
))
4726 el_to_fixup
.safe_push (stmt
);
4727 /* When changing a condition or switch into one we know what
4728 edge will be executed, schedule a cfg cleanup. */
4729 if ((gimple_code (stmt
) == GIMPLE_COND
4730 && (gimple_cond_true_p (as_a
<gcond
*> (stmt
))
4731 || gimple_cond_false_p (as_a
<gcond
*> (stmt
))))
4732 || (gimple_code (stmt
) == GIMPLE_SWITCH
4733 && TREE_CODE (gimple_switch_index
4734 (as_a
<gswitch
*> (stmt
))) == INTEGER_CST
))
4735 el_todo
|= TODO_cleanup_cfg
;
4736 /* If we removed EH side-effects from the statement, clean
4737 its EH information. */
4738 if (maybe_clean_or_replace_eh_stmt (old_stmt
, stmt
))
4740 bitmap_set_bit (need_eh_cleanup
,
4741 gimple_bb (stmt
)->index
);
4742 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4743 fprintf (dump_file
, " Removed EH side-effects.\n");
4745 /* Likewise for AB side-effects. */
4746 if (can_make_abnormal_goto
4747 && !stmt_can_make_abnormal_goto (stmt
))
4749 bitmap_set_bit (need_ab_cleanup
,
4750 gimple_bb (stmt
)->index
);
4751 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4752 fprintf (dump_file
, " Removed AB side-effects.\n");
4755 if (vdef
!= gimple_vdef (stmt
))
4756 VN_INFO (vdef
)->valnum
= vuse
;
4759 /* Make new values available - for fully redundant LHS we
4760 continue with the next stmt above and skip this. */
4762 FOR_EACH_SSA_DEF_OPERAND (defp
, stmt
, iter
, SSA_OP_DEF
)
4763 eliminate_push_avail (DEF_FROM_PTR (defp
));
4766 /* Replace destination PHI arguments. */
4767 FOR_EACH_EDGE (e
, ei
, b
->succs
)
4768 if (e
->flags
& EDGE_EXECUTABLE
)
4769 for (gphi_iterator gsi
= gsi_start_phis (e
->dest
);
4773 gphi
*phi
= gsi
.phi ();
4774 use_operand_p use_p
= PHI_ARG_DEF_PTR_FROM_EDGE (phi
, e
);
4775 tree arg
= USE_FROM_PTR (use_p
);
4776 if (TREE_CODE (arg
) != SSA_NAME
4777 || virtual_operand_p (arg
))
4779 tree sprime
= eliminate_avail (arg
);
4780 if (sprime
&& may_propagate_copy (arg
, sprime
))
4782 propagate_value (use_p
, sprime
);
4783 if (TREE_CODE (sprime
) == SSA_NAME
)
4784 gimple_set_plf (SSA_NAME_DEF_STMT (sprime
), NECESSARY
, true);
4790 /* Make no longer available leaders no longer available. */
4793 eliminate_dom_walker::after_dom_children (basic_block
)
4796 while ((entry
= el_avail_stack
.pop ()) != NULL_TREE
)
4798 tree valnum
= VN_INFO (entry
)->valnum
;
4799 tree old
= el_avail
[SSA_NAME_VERSION (valnum
)];
4801 el_avail
[SSA_NAME_VERSION (valnum
)] = NULL_TREE
;
4803 el_avail
[SSA_NAME_VERSION (valnum
)] = entry
;
4807 /* Eliminate fully redundant computations. */
4810 eliminate (bool do_pre
)
4812 need_eh_cleanup
= BITMAP_ALLOC (NULL
);
4813 need_ab_cleanup
= BITMAP_ALLOC (NULL
);
4815 el_to_remove
.create (0);
4816 el_to_fixup
.create (0);
4818 el_avail
.create (num_ssa_names
);
4819 el_avail_stack
.create (0);
4821 eliminate_dom_walker (CDI_DOMINATORS
,
4822 do_pre
).walk (cfun
->cfg
->x_entry_block_ptr
);
4824 el_avail
.release ();
4825 el_avail_stack
.release ();
4830 /* Perform CFG cleanups made necessary by elimination. */
4833 fini_eliminate (void)
4835 gimple_stmt_iterator gsi
;
4839 /* We cannot remove stmts during BB walk, especially not release SSA
4840 names there as this confuses the VN machinery. The stmts ending
4841 up in el_to_remove are either stores or simple copies.
4842 Remove stmts in reverse order to make debug stmt creation possible. */
4843 while (!el_to_remove
.is_empty ())
4845 stmt
= el_to_remove
.pop ();
4848 if (gimple_code (stmt
) == GIMPLE_PHI
)
4849 lhs
= gimple_phi_result (stmt
);
4851 lhs
= gimple_get_lhs (stmt
);
4854 && TREE_CODE (lhs
) == SSA_NAME
4855 && bitmap_bit_p (inserted_exprs
, SSA_NAME_VERSION (lhs
)))
4858 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4860 fprintf (dump_file
, "Removing dead stmt ");
4861 print_gimple_stmt (dump_file
, stmt
, 0, 0);
4864 gsi
= gsi_for_stmt (stmt
);
4865 if (gimple_code (stmt
) == GIMPLE_PHI
)
4866 remove_phi_node (&gsi
, true);
4869 basic_block bb
= gimple_bb (stmt
);
4870 unlink_stmt_vdef (stmt
);
4871 if (gsi_remove (&gsi
, true))
4872 bitmap_set_bit (need_eh_cleanup
, bb
->index
);
4873 if (is_gimple_call (stmt
) && stmt_can_make_abnormal_goto (stmt
))
4874 bitmap_set_bit (need_ab_cleanup
, bb
->index
);
4875 release_defs (stmt
);
4878 /* Removing a stmt may expose a forwarder block. */
4879 todo
|= TODO_cleanup_cfg
;
4881 el_to_remove
.release ();
4883 /* Fixup stmts that became noreturn calls. This may require splitting
4884 blocks and thus isn't possible during the dominator walk. Do this
4885 in reverse order so we don't inadvertedly remove a stmt we want to
4886 fixup by visiting a dominating now noreturn call first. */
4887 while (!el_to_fixup
.is_empty ())
4889 stmt
= el_to_fixup
.pop ();
4891 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4893 fprintf (dump_file
, "Fixing up noreturn call ");
4894 print_gimple_stmt (dump_file
, stmt
, 0);
4897 if (fixup_noreturn_call (stmt
))
4898 todo
|= TODO_cleanup_cfg
;
4900 el_to_fixup
.release ();
4902 bool do_eh_cleanup
= !bitmap_empty_p (need_eh_cleanup
);
4903 bool do_ab_cleanup
= !bitmap_empty_p (need_ab_cleanup
);
4906 gimple_purge_all_dead_eh_edges (need_eh_cleanup
);
4909 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup
);
4911 BITMAP_FREE (need_eh_cleanup
);
4912 BITMAP_FREE (need_ab_cleanup
);
4914 if (do_eh_cleanup
|| do_ab_cleanup
)
4915 todo
|= TODO_cleanup_cfg
;
4919 /* Borrow a bit of tree-ssa-dce.c for the moment.
4920 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4921 this may be a bit faster, and we may want critical edges kept split. */
4923 /* If OP's defining statement has not already been determined to be necessary,
4924 mark that statement necessary. Return the stmt, if it is newly
4927 static inline gimple
*
4928 mark_operand_necessary (tree op
)
4934 if (TREE_CODE (op
) != SSA_NAME
)
4937 stmt
= SSA_NAME_DEF_STMT (op
);
4940 if (gimple_plf (stmt
, NECESSARY
)
4941 || gimple_nop_p (stmt
))
4944 gimple_set_plf (stmt
, NECESSARY
, true);
4948 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4949 to insert PHI nodes sometimes, and because value numbering of casts isn't
4950 perfect, we sometimes end up inserting dead code. This simple DCE-like
4951 pass removes any insertions we made that weren't actually used. */
4954 remove_dead_inserted_code (void)
4960 auto_bitmap worklist
;
4961 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs
, 0, i
, bi
)
4963 t
= SSA_NAME_DEF_STMT (ssa_name (i
));
4964 if (gimple_plf (t
, NECESSARY
))
4965 bitmap_set_bit (worklist
, i
);
4967 while (!bitmap_empty_p (worklist
))
4969 i
= bitmap_first_set_bit (worklist
);
4970 bitmap_clear_bit (worklist
, i
);
4971 t
= SSA_NAME_DEF_STMT (ssa_name (i
));
4973 /* PHI nodes are somewhat special in that each PHI alternative has
4974 data and control dependencies. All the statements feeding the
4975 PHI node's arguments are always necessary. */
4976 if (gimple_code (t
) == GIMPLE_PHI
)
4980 for (k
= 0; k
< gimple_phi_num_args (t
); k
++)
4982 tree arg
= PHI_ARG_DEF (t
, k
);
4983 if (TREE_CODE (arg
) == SSA_NAME
)
4985 gimple
*n
= mark_operand_necessary (arg
);
4987 bitmap_set_bit (worklist
, SSA_NAME_VERSION (arg
));
4993 /* Propagate through the operands. Examine all the USE, VUSE and
4994 VDEF operands in this statement. Mark all the statements
4995 which feed this statement's uses as necessary. */
4999 /* The operands of VDEF expressions are also needed as they
5000 represent potential definitions that may reach this
5001 statement (VDEF operands allow us to follow def-def
5004 FOR_EACH_SSA_TREE_OPERAND (use
, t
, iter
, SSA_OP_ALL_USES
)
5006 gimple
*n
= mark_operand_necessary (use
);
5008 bitmap_set_bit (worklist
, SSA_NAME_VERSION (use
));
5013 unsigned int to_clear
= -1U;
5014 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs
, 0, i
, bi
)
5016 if (to_clear
!= -1U)
5018 bitmap_clear_bit (inserted_exprs
, to_clear
);
5021 t
= SSA_NAME_DEF_STMT (ssa_name (i
));
5022 if (!gimple_plf (t
, NECESSARY
))
5024 gimple_stmt_iterator gsi
;
5026 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
5028 fprintf (dump_file
, "Removing unnecessary insertion:");
5029 print_gimple_stmt (dump_file
, t
, 0);
5032 gsi
= gsi_for_stmt (t
);
5033 if (gimple_code (t
) == GIMPLE_PHI
)
5034 remove_phi_node (&gsi
, true);
5037 gsi_remove (&gsi
, true);
5042 /* eliminate_fini will skip stmts marked for removal if we
5043 already removed it and uses inserted_exprs for this, so
5044 clear those we didn't end up removing. */
5047 if (to_clear
!= -1U)
5048 bitmap_clear_bit (inserted_exprs
, to_clear
);
5052 /* Initialize data structures used by PRE. */
5059 next_expression_id
= 1;
5060 expressions
.create (0);
5061 expressions
.safe_push (NULL
);
5062 value_expressions
.create (get_max_value_id () + 1);
5063 value_expressions
.safe_grow_cleared (get_max_value_id () + 1);
5064 name_to_id
.create (0);
5066 inserted_exprs
= BITMAP_ALLOC (NULL
);
5068 connect_infinite_loops_to_exit ();
5069 memset (&pre_stats
, 0, sizeof (pre_stats
));
5071 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets
));
5073 calculate_dominance_info (CDI_DOMINATORS
);
5075 bitmap_obstack_initialize (&grand_bitmap_obstack
);
5076 phi_translate_table
= new hash_table
<expr_pred_trans_d
> (5110);
5077 expression_to_id
= new hash_table
<pre_expr_d
> (num_ssa_names
* 3);
5078 FOR_ALL_BB_FN (bb
, cfun
)
5080 EXP_GEN (bb
) = bitmap_set_new ();
5081 PHI_GEN (bb
) = bitmap_set_new ();
5082 TMP_GEN (bb
) = bitmap_set_new ();
5083 AVAIL_OUT (bb
) = bitmap_set_new ();
5088 /* Deallocate data structures used by PRE. */
5093 value_expressions
.release ();
5094 BITMAP_FREE (inserted_exprs
);
5095 bitmap_obstack_release (&grand_bitmap_obstack
);
5096 bitmap_set_pool
.release ();
5097 pre_expr_pool
.release ();
5098 delete phi_translate_table
;
5099 phi_translate_table
= NULL
;
5100 delete expression_to_id
;
5101 expression_to_id
= NULL
;
5102 name_to_id
.release ();
5104 free_aux_for_blocks ();
5109 const pass_data pass_data_pre
=
5111 GIMPLE_PASS
, /* type */
5113 OPTGROUP_NONE
, /* optinfo_flags */
5114 TV_TREE_PRE
, /* tv_id */
5115 /* PROP_no_crit_edges is ensured by placing pass_split_crit_edges before
5117 ( PROP_no_crit_edges
| PROP_cfg
| PROP_ssa
), /* properties_required */
5118 0, /* properties_provided */
5119 PROP_no_crit_edges
, /* properties_destroyed */
5120 TODO_rebuild_alias
, /* todo_flags_start */
5121 0, /* todo_flags_finish */
5124 class pass_pre
: public gimple_opt_pass
5127 pass_pre (gcc::context
*ctxt
)
5128 : gimple_opt_pass (pass_data_pre
, ctxt
)
5131 /* opt_pass methods: */
5132 virtual bool gate (function
*)
5133 { return flag_tree_pre
!= 0 || flag_code_hoisting
!= 0; }
5134 virtual unsigned int execute (function
*);
5136 }; // class pass_pre
5139 pass_pre::execute (function
*fun
)
5141 unsigned int todo
= 0;
5143 do_partial_partial
=
5144 flag_tree_partial_pre
&& optimize_function_for_speed_p (fun
);
5146 /* This has to happen before SCCVN runs because
5147 loop_optimizer_init may create new phis, etc. */
5148 loop_optimizer_init (LOOPS_NORMAL
);
5150 run_scc_vn (VN_WALK
);
5155 /* Collect and value number expressions computed in each basic block. */
5158 /* Insert can get quite slow on an incredibly large number of basic
5159 blocks due to some quadratic behavior. Until this behavior is
5160 fixed, don't run it when he have an incredibly large number of
5161 bb's. If we aren't going to run insert, there is no point in
5162 computing ANTIC, either, even though it's plenty fast. */
5163 if (n_basic_blocks_for_fn (fun
) < 4000)
5169 /* Make sure to remove fake edges before committing our inserts.
5170 This makes sure we don't end up with extra critical edges that
5171 we would need to split. */
5172 remove_fake_exit_edges ();
5173 gsi_commit_edge_inserts ();
5175 /* Eliminate folds statements which might (should not...) end up
5176 not keeping virtual operands up-to-date. */
5177 gcc_assert (!need_ssa_update_p (fun
));
5179 /* Remove all the redundant expressions. */
5180 todo
|= eliminate (true);
5182 statistics_counter_event (fun
, "Insertions", pre_stats
.insertions
);
5183 statistics_counter_event (fun
, "PA inserted", pre_stats
.pa_insert
);
5184 statistics_counter_event (fun
, "HOIST inserted", pre_stats
.hoist_insert
);
5185 statistics_counter_event (fun
, "New PHIs", pre_stats
.phis
);
5186 statistics_counter_event (fun
, "Eliminated", pre_stats
.eliminations
);
5188 clear_expression_ids ();
5189 remove_dead_inserted_code ();
5192 todo
|= fini_eliminate ();
5194 loop_optimizer_finalize ();
5196 /* Restore SSA info before tail-merging as that resets it as well. */
5197 scc_vn_restore_ssa_info ();
5199 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
5200 case we can merge the block with the remaining predecessor of the block.
5202 - call merge_blocks after each tail merge iteration
5203 - call merge_blocks after all tail merge iterations
5204 - mark TODO_cleanup_cfg when necessary
5205 - share the cfg cleanup with fini_pre. */
5206 todo
|= tail_merge_optimize (todo
);
5210 /* Tail merging invalidates the virtual SSA web, together with
5211 cfg-cleanup opportunities exposed by PRE this will wreck the
5212 SSA updating machinery. So make sure to run update-ssa
5213 manually, before eventually scheduling cfg-cleanup as part of
5215 update_ssa (TODO_update_ssa_only_virtuals
);
5223 make_pass_pre (gcc::context
*ctxt
)
5225 return new pass_pre (ctxt
);
5230 const pass_data pass_data_fre
=
5232 GIMPLE_PASS
, /* type */
5234 OPTGROUP_NONE
, /* optinfo_flags */
5235 TV_TREE_FRE
, /* tv_id */
5236 ( PROP_cfg
| PROP_ssa
), /* properties_required */
5237 0, /* properties_provided */
5238 0, /* properties_destroyed */
5239 0, /* todo_flags_start */
5240 0, /* todo_flags_finish */
5243 class pass_fre
: public gimple_opt_pass
5246 pass_fre (gcc::context
*ctxt
)
5247 : gimple_opt_pass (pass_data_fre
, ctxt
)
5250 /* opt_pass methods: */
5251 opt_pass
* clone () { return new pass_fre (m_ctxt
); }
5252 virtual bool gate (function
*) { return flag_tree_fre
!= 0; }
5253 virtual unsigned int execute (function
*);
5255 }; // class pass_fre
5258 pass_fre::execute (function
*fun
)
5260 unsigned int todo
= 0;
5262 run_scc_vn (VN_WALKREWRITE
);
5264 memset (&pre_stats
, 0, sizeof (pre_stats
));
5266 /* Remove all the redundant expressions. */
5267 todo
|= eliminate (false);
5269 todo
|= fini_eliminate ();
5271 scc_vn_restore_ssa_info ();
5274 statistics_counter_event (fun
, "Insertions", pre_stats
.insertions
);
5275 statistics_counter_event (fun
, "Eliminated", pre_stats
.eliminations
);
5283 make_pass_fre (gcc::context
*ctxt
)
5285 return new pass_fre (ctxt
);