1 /* Full and partial redundancy elimination and code hoisting on SSA GIMPLE.
2 Copyright (C) 2001-2016 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 void bitmap_insert_into_set_1 (bitmap_set_t
, pre_expr
,
545 static bitmap_set_t
bitmap_set_new (void);
546 static tree
create_expression_by_pieces (basic_block
, pre_expr
, gimple_seq
*,
548 static tree
find_or_generate_expression (basic_block
, tree
, gimple_seq
*);
549 static unsigned int get_expr_value_id (pre_expr
);
551 /* We can add and remove elements and entries to and from sets
552 and hash tables, so we use alloc pools for them. */
554 static object_allocator
<bitmap_set
> bitmap_set_pool ("Bitmap sets");
555 static bitmap_obstack grand_bitmap_obstack
;
557 /* Set of blocks with statements that have had their EH properties changed. */
558 static bitmap need_eh_cleanup
;
560 /* Set of blocks with statements that have had their AB properties changed. */
561 static bitmap need_ab_cleanup
;
563 /* A three tuple {e, pred, v} used to cache phi translations in the
564 phi_translate_table. */
566 typedef struct expr_pred_trans_d
: free_ptr_hash
<expr_pred_trans_d
>
568 /* The expression. */
571 /* The predecessor block along which we translated the expression. */
574 /* The value that resulted from the translation. */
577 /* The hashcode for the expression, pred pair. This is cached for
581 /* hash_table support. */
582 static inline hashval_t
hash (const expr_pred_trans_d
*);
583 static inline int equal (const expr_pred_trans_d
*, const expr_pred_trans_d
*);
584 } *expr_pred_trans_t
;
585 typedef const struct expr_pred_trans_d
*const_expr_pred_trans_t
;
588 expr_pred_trans_d::hash (const expr_pred_trans_d
*e
)
594 expr_pred_trans_d::equal (const expr_pred_trans_d
*ve1
,
595 const expr_pred_trans_d
*ve2
)
597 basic_block b1
= ve1
->pred
;
598 basic_block b2
= ve2
->pred
;
600 /* If they are not translations for the same basic block, they can't
604 return pre_expr_d::equal (ve1
->e
, ve2
->e
);
607 /* The phi_translate_table caches phi translations for a given
608 expression and predecessor. */
609 static hash_table
<expr_pred_trans_d
> *phi_translate_table
;
611 /* Add the tuple mapping from {expression E, basic block PRED} to
612 the phi translation table and return whether it pre-existed. */
615 phi_trans_add (expr_pred_trans_t
*entry
, pre_expr e
, basic_block pred
)
617 expr_pred_trans_t
*slot
;
618 expr_pred_trans_d tem
;
619 hashval_t hash
= iterative_hash_hashval_t (pre_expr_d::hash (e
),
624 slot
= phi_translate_table
->find_slot_with_hash (&tem
, hash
, INSERT
);
631 *entry
= *slot
= XNEW (struct expr_pred_trans_d
);
633 (*entry
)->pred
= pred
;
634 (*entry
)->hashcode
= hash
;
639 /* Add expression E to the expression set of value id V. */
642 add_to_value (unsigned int v
, pre_expr e
)
646 gcc_checking_assert (get_expr_value_id (e
) == v
);
648 if (v
>= value_expressions
.length ())
650 value_expressions
.safe_grow_cleared (v
+ 1);
653 set
= value_expressions
[v
];
656 set
= BITMAP_ALLOC (&grand_bitmap_obstack
);
657 value_expressions
[v
] = set
;
660 bitmap_set_bit (set
, get_or_alloc_expression_id (e
));
663 /* Create a new bitmap set and return it. */
666 bitmap_set_new (void)
668 bitmap_set_t ret
= bitmap_set_pool
.allocate ();
669 bitmap_initialize (&ret
->expressions
, &grand_bitmap_obstack
);
670 bitmap_initialize (&ret
->values
, &grand_bitmap_obstack
);
674 /* Return the value id for a PRE expression EXPR. */
677 get_expr_value_id (pre_expr expr
)
683 id
= get_constant_value_id (PRE_EXPR_CONSTANT (expr
));
686 id
= VN_INFO (PRE_EXPR_NAME (expr
))->value_id
;
689 id
= PRE_EXPR_NARY (expr
)->value_id
;
692 id
= PRE_EXPR_REFERENCE (expr
)->value_id
;
697 /* ??? We cannot assert that expr has a value-id (it can be 0), because
698 we assign value-ids only to expressions that have a result
699 in set_hashtable_value_ids. */
703 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
706 sccvn_valnum_from_value_id (unsigned int val
)
710 bitmap exprset
= value_expressions
[val
];
711 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
713 pre_expr vexpr
= expression_for_id (i
);
714 if (vexpr
->kind
== NAME
)
715 return VN_INFO (PRE_EXPR_NAME (vexpr
))->valnum
;
716 else if (vexpr
->kind
== CONSTANT
)
717 return PRE_EXPR_CONSTANT (vexpr
);
722 /* Remove an expression EXPR from a bitmapped set. */
725 bitmap_remove_from_set (bitmap_set_t set
, pre_expr expr
)
727 unsigned int val
= get_expr_value_id (expr
);
728 if (!value_id_constant_p (val
))
730 bitmap_clear_bit (&set
->values
, val
);
731 bitmap_clear_bit (&set
->expressions
, get_expression_id (expr
));
736 bitmap_insert_into_set_1 (bitmap_set_t set
, pre_expr expr
,
737 unsigned int val
, bool allow_constants
)
739 if (allow_constants
|| !value_id_constant_p (val
))
741 /* We specifically expect this and only this function to be able to
742 insert constants into a set. */
743 bitmap_set_bit (&set
->values
, val
);
744 bitmap_set_bit (&set
->expressions
, get_or_alloc_expression_id (expr
));
748 /* Insert an expression EXPR into a bitmapped set. */
751 bitmap_insert_into_set (bitmap_set_t set
, pre_expr expr
)
753 bitmap_insert_into_set_1 (set
, expr
, get_expr_value_id (expr
), false);
756 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
759 bitmap_set_copy (bitmap_set_t dest
, bitmap_set_t orig
)
761 bitmap_copy (&dest
->expressions
, &orig
->expressions
);
762 bitmap_copy (&dest
->values
, &orig
->values
);
766 /* Free memory used up by SET. */
768 bitmap_set_free (bitmap_set_t set
)
770 bitmap_clear (&set
->expressions
);
771 bitmap_clear (&set
->values
);
775 /* Generate an topological-ordered array of bitmap set SET. */
778 sorted_array_from_bitmap_set (bitmap_set_t set
)
781 bitmap_iterator bi
, bj
;
782 vec
<pre_expr
> result
;
784 /* Pre-allocate enough space for the array. */
785 result
.create (bitmap_count_bits (&set
->expressions
));
787 FOR_EACH_VALUE_ID_IN_SET (set
, i
, bi
)
789 /* The number of expressions having a given value is usually
790 relatively small. Thus, rather than making a vector of all
791 the expressions and sorting it by value-id, we walk the values
792 and check in the reverse mapping that tells us what expressions
793 have a given value, to filter those in our set. As a result,
794 the expressions are inserted in value-id order, which means
797 If this is somehow a significant lose for some cases, we can
798 choose which set to walk based on the set size. */
799 bitmap exprset
= value_expressions
[i
];
800 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, j
, bj
)
802 if (bitmap_bit_p (&set
->expressions
, j
))
803 result
.quick_push (expression_for_id (j
));
810 /* Perform bitmapped set operation DEST &= ORIG. */
813 bitmap_set_and (bitmap_set_t dest
, bitmap_set_t orig
)
821 bitmap_initialize (&temp
, &grand_bitmap_obstack
);
823 bitmap_and_into (&dest
->values
, &orig
->values
);
824 bitmap_copy (&temp
, &dest
->expressions
);
825 EXECUTE_IF_SET_IN_BITMAP (&temp
, 0, i
, bi
)
827 pre_expr expr
= expression_for_id (i
);
828 unsigned int value_id
= get_expr_value_id (expr
);
829 if (!bitmap_bit_p (&dest
->values
, value_id
))
830 bitmap_clear_bit (&dest
->expressions
, i
);
832 bitmap_clear (&temp
);
836 /* Subtract all values and expressions contained in ORIG from DEST. */
839 bitmap_set_subtract (bitmap_set_t dest
, bitmap_set_t orig
)
841 bitmap_set_t result
= bitmap_set_new ();
845 bitmap_and_compl (&result
->expressions
, &dest
->expressions
,
848 FOR_EACH_EXPR_ID_IN_SET (result
, i
, bi
)
850 pre_expr expr
= expression_for_id (i
);
851 unsigned int value_id
= get_expr_value_id (expr
);
852 bitmap_set_bit (&result
->values
, value_id
);
858 /* Subtract all the values in bitmap set B from bitmap set A. */
861 bitmap_set_subtract_values (bitmap_set_t a
, bitmap_set_t b
)
867 bitmap_initialize (&temp
, &grand_bitmap_obstack
);
869 bitmap_copy (&temp
, &a
->expressions
);
870 EXECUTE_IF_SET_IN_BITMAP (&temp
, 0, i
, bi
)
872 pre_expr expr
= expression_for_id (i
);
873 if (bitmap_set_contains_value (b
, get_expr_value_id (expr
)))
874 bitmap_remove_from_set (a
, expr
);
876 bitmap_clear (&temp
);
880 /* Return true if bitmapped set SET contains the value VALUE_ID. */
883 bitmap_set_contains_value (bitmap_set_t set
, unsigned int value_id
)
885 if (value_id_constant_p (value_id
))
888 if (!set
|| bitmap_empty_p (&set
->expressions
))
891 return bitmap_bit_p (&set
->values
, value_id
);
895 bitmap_set_contains_expr (bitmap_set_t set
, const pre_expr expr
)
897 return bitmap_bit_p (&set
->expressions
, get_expression_id (expr
));
900 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
903 bitmap_set_replace_value (bitmap_set_t set
, unsigned int lookfor
,
910 if (value_id_constant_p (lookfor
))
913 if (!bitmap_set_contains_value (set
, lookfor
))
916 /* The number of expressions having a given value is usually
917 significantly less than the total number of expressions in SET.
918 Thus, rather than check, for each expression in SET, whether it
919 has the value LOOKFOR, we walk the reverse mapping that tells us
920 what expressions have a given value, and see if any of those
921 expressions are in our set. For large testcases, this is about
922 5-10x faster than walking the bitmap. If this is somehow a
923 significant lose for some cases, we can choose which set to walk
924 based on the set size. */
925 exprset
= value_expressions
[lookfor
];
926 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
928 if (bitmap_clear_bit (&set
->expressions
, i
))
930 bitmap_set_bit (&set
->expressions
, get_expression_id (expr
));
938 /* Return true if two bitmap sets are equal. */
941 bitmap_set_equal (bitmap_set_t a
, bitmap_set_t b
)
943 return bitmap_equal_p (&a
->values
, &b
->values
);
946 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
947 and add it otherwise. */
950 bitmap_value_replace_in_set (bitmap_set_t set
, pre_expr expr
)
952 unsigned int val
= get_expr_value_id (expr
);
954 if (bitmap_set_contains_value (set
, val
))
955 bitmap_set_replace_value (set
, val
, expr
);
957 bitmap_insert_into_set (set
, expr
);
960 /* Insert EXPR into SET if EXPR's value is not already present in
964 bitmap_value_insert_into_set (bitmap_set_t set
, pre_expr expr
)
966 unsigned int val
= get_expr_value_id (expr
);
968 gcc_checking_assert (expr
->id
== get_or_alloc_expression_id (expr
));
970 /* Constant values are always considered to be part of the set. */
971 if (value_id_constant_p (val
))
974 /* If the value membership changed, add the expression. */
975 if (bitmap_set_bit (&set
->values
, val
))
976 bitmap_set_bit (&set
->expressions
, expr
->id
);
979 /* Print out EXPR to outfile. */
982 print_pre_expr (FILE *outfile
, const pre_expr expr
)
987 print_generic_expr (outfile
, PRE_EXPR_CONSTANT (expr
), 0);
990 print_generic_expr (outfile
, PRE_EXPR_NAME (expr
), 0);
995 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
996 fprintf (outfile
, "{%s,", get_tree_code_name (nary
->opcode
));
997 for (i
= 0; i
< nary
->length
; i
++)
999 print_generic_expr (outfile
, nary
->op
[i
], 0);
1000 if (i
!= (unsigned) nary
->length
- 1)
1001 fprintf (outfile
, ",");
1003 fprintf (outfile
, "}");
1009 vn_reference_op_t vro
;
1011 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
1012 fprintf (outfile
, "{");
1014 ref
->operands
.iterate (i
, &vro
);
1017 bool closebrace
= false;
1018 if (vro
->opcode
!= SSA_NAME
1019 && TREE_CODE_CLASS (vro
->opcode
) != tcc_declaration
)
1021 fprintf (outfile
, "%s", get_tree_code_name (vro
->opcode
));
1024 fprintf (outfile
, "<");
1030 print_generic_expr (outfile
, vro
->op0
, 0);
1033 fprintf (outfile
, ",");
1034 print_generic_expr (outfile
, vro
->op1
, 0);
1038 fprintf (outfile
, ",");
1039 print_generic_expr (outfile
, vro
->op2
, 0);
1043 fprintf (outfile
, ">");
1044 if (i
!= ref
->operands
.length () - 1)
1045 fprintf (outfile
, ",");
1047 fprintf (outfile
, "}");
1050 fprintf (outfile
, "@");
1051 print_generic_expr (outfile
, ref
->vuse
, 0);
1057 void debug_pre_expr (pre_expr
);
1059 /* Like print_pre_expr but always prints to stderr. */
1061 debug_pre_expr (pre_expr e
)
1063 print_pre_expr (stderr
, e
);
1064 fprintf (stderr
, "\n");
1067 /* Print out SET to OUTFILE. */
1070 print_bitmap_set (FILE *outfile
, bitmap_set_t set
,
1071 const char *setname
, int blockindex
)
1073 fprintf (outfile
, "%s[%d] := { ", setname
, blockindex
);
1080 FOR_EACH_EXPR_ID_IN_SET (set
, i
, bi
)
1082 const pre_expr expr
= expression_for_id (i
);
1085 fprintf (outfile
, ", ");
1087 print_pre_expr (outfile
, expr
);
1089 fprintf (outfile
, " (%04d)", get_expr_value_id (expr
));
1092 fprintf (outfile
, " }\n");
1095 void debug_bitmap_set (bitmap_set_t
);
1098 debug_bitmap_set (bitmap_set_t set
)
1100 print_bitmap_set (stderr
, set
, "debug", 0);
1103 void debug_bitmap_sets_for (basic_block
);
1106 debug_bitmap_sets_for (basic_block bb
)
1108 print_bitmap_set (stderr
, AVAIL_OUT (bb
), "avail_out", bb
->index
);
1109 print_bitmap_set (stderr
, EXP_GEN (bb
), "exp_gen", bb
->index
);
1110 print_bitmap_set (stderr
, PHI_GEN (bb
), "phi_gen", bb
->index
);
1111 print_bitmap_set (stderr
, TMP_GEN (bb
), "tmp_gen", bb
->index
);
1112 print_bitmap_set (stderr
, ANTIC_IN (bb
), "antic_in", bb
->index
);
1113 if (do_partial_partial
)
1114 print_bitmap_set (stderr
, PA_IN (bb
), "pa_in", bb
->index
);
1115 print_bitmap_set (stderr
, NEW_SETS (bb
), "new_sets", bb
->index
);
1118 /* Print out the expressions that have VAL to OUTFILE. */
1121 print_value_expressions (FILE *outfile
, unsigned int val
)
1123 bitmap set
= value_expressions
[val
];
1128 sprintf (s
, "%04d", val
);
1129 x
.expressions
= *set
;
1130 print_bitmap_set (outfile
, &x
, s
, 0);
1136 debug_value_expressions (unsigned int val
)
1138 print_value_expressions (stderr
, val
);
1141 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1145 get_or_alloc_expr_for_constant (tree constant
)
1147 unsigned int result_id
;
1148 unsigned int value_id
;
1149 struct pre_expr_d expr
;
1152 expr
.kind
= CONSTANT
;
1153 PRE_EXPR_CONSTANT (&expr
) = constant
;
1154 result_id
= lookup_expression_id (&expr
);
1156 return expression_for_id (result_id
);
1158 newexpr
= pre_expr_pool
.allocate ();
1159 newexpr
->kind
= CONSTANT
;
1160 PRE_EXPR_CONSTANT (newexpr
) = constant
;
1161 alloc_expression_id (newexpr
);
1162 value_id
= get_or_alloc_constant_value_id (constant
);
1163 add_to_value (value_id
, newexpr
);
1167 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1168 Currently only supports constants and SSA_NAMES. */
1170 get_or_alloc_expr_for (tree t
)
1172 if (TREE_CODE (t
) == SSA_NAME
)
1173 return get_or_alloc_expr_for_name (t
);
1174 else if (is_gimple_min_invariant (t
))
1175 return get_or_alloc_expr_for_constant (t
);
1178 /* More complex expressions can result from SCCVN expression
1179 simplification that inserts values for them. As they all
1180 do not have VOPs the get handled by the nary ops struct. */
1181 vn_nary_op_t result
;
1182 unsigned int result_id
;
1183 vn_nary_op_lookup (t
, &result
);
1186 pre_expr e
= pre_expr_pool
.allocate ();
1188 PRE_EXPR_NARY (e
) = result
;
1189 result_id
= lookup_expression_id (e
);
1192 pre_expr_pool
.remove (e
);
1193 e
= expression_for_id (result_id
);
1196 alloc_expression_id (e
);
1203 /* Return the folded version of T if T, when folded, is a gimple
1204 min_invariant or an SSA name. Otherwise, return T. */
1207 fully_constant_expression (pre_expr e
)
1215 vn_nary_op_t nary
= PRE_EXPR_NARY (e
);
1216 tree res
= vn_nary_simplify (nary
);
1219 if (is_gimple_min_invariant (res
))
1220 return get_or_alloc_expr_for_constant (res
);
1221 if (TREE_CODE (res
) == SSA_NAME
)
1222 return get_or_alloc_expr_for_name (res
);
1227 vn_reference_t ref
= PRE_EXPR_REFERENCE (e
);
1229 if ((folded
= fully_constant_vn_reference_p (ref
)))
1230 return get_or_alloc_expr_for_constant (folded
);
1239 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1240 it has the value it would have in BLOCK. Set *SAME_VALID to true
1241 in case the new vuse doesn't change the value id of the OPERANDS. */
1244 translate_vuse_through_block (vec
<vn_reference_op_s
> operands
,
1245 alias_set_type set
, tree type
, tree vuse
,
1246 basic_block phiblock
,
1247 basic_block block
, bool *same_valid
)
1249 gimple
*phi
= SSA_NAME_DEF_STMT (vuse
);
1256 if (gimple_bb (phi
) != phiblock
)
1259 use_oracle
= ao_ref_init_from_vn_reference (&ref
, set
, type
, operands
);
1261 /* Use the alias-oracle to find either the PHI node in this block,
1262 the first VUSE used in this block that is equivalent to vuse or
1263 the first VUSE which definition in this block kills the value. */
1264 if (gimple_code (phi
) == GIMPLE_PHI
)
1265 e
= find_edge (block
, phiblock
);
1266 else if (use_oracle
)
1267 while (!stmt_may_clobber_ref_p_1 (phi
, &ref
))
1269 vuse
= gimple_vuse (phi
);
1270 phi
= SSA_NAME_DEF_STMT (vuse
);
1271 if (gimple_bb (phi
) != phiblock
)
1273 if (gimple_code (phi
) == GIMPLE_PHI
)
1275 e
= find_edge (block
, phiblock
);
1286 bitmap visited
= NULL
;
1288 /* Try to find a vuse that dominates this phi node by skipping
1289 non-clobbering statements. */
1290 vuse
= get_continuation_for_phi (phi
, &ref
, &cnt
, &visited
, false,
1293 BITMAP_FREE (visited
);
1299 /* If we didn't find any, the value ID can't stay the same,
1300 but return the translated vuse. */
1301 *same_valid
= false;
1302 vuse
= PHI_ARG_DEF (phi
, e
->dest_idx
);
1304 /* ??? We would like to return vuse here as this is the canonical
1305 upmost vdef that this reference is associated with. But during
1306 insertion of the references into the hash tables we only ever
1307 directly insert with their direct gimple_vuse, hence returning
1308 something else would make us not find the other expression. */
1309 return PHI_ARG_DEF (phi
, e
->dest_idx
);
1315 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1316 SET2. This is used to avoid making a set consisting of the union
1317 of PA_IN and ANTIC_IN during insert. */
1319 static inline pre_expr
1320 find_leader_in_sets (unsigned int val
, bitmap_set_t set1
, bitmap_set_t set2
)
1324 result
= bitmap_find_leader (set1
, val
);
1325 if (!result
&& set2
)
1326 result
= bitmap_find_leader (set2
, val
);
1330 /* Get the tree type for our PRE expression e. */
1333 get_expr_type (const pre_expr e
)
1338 return TREE_TYPE (PRE_EXPR_NAME (e
));
1340 return TREE_TYPE (PRE_EXPR_CONSTANT (e
));
1342 return PRE_EXPR_REFERENCE (e
)->type
;
1344 return PRE_EXPR_NARY (e
)->type
;
1349 /* Get a representative SSA_NAME for a given expression.
1350 Since all of our sub-expressions are treated as values, we require
1351 them to be SSA_NAME's for simplicity.
1352 Prior versions of GVNPRE used to use "value handles" here, so that
1353 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1354 either case, the operands are really values (IE we do not expect
1355 them to be usable without finding leaders). */
1358 get_representative_for (const pre_expr e
)
1361 unsigned int value_id
= get_expr_value_id (e
);
1366 return VN_INFO (PRE_EXPR_NAME (e
))->valnum
;
1368 return PRE_EXPR_CONSTANT (e
);
1372 /* Go through all of the expressions representing this value
1373 and pick out an SSA_NAME. */
1376 bitmap exprs
= value_expressions
[value_id
];
1377 EXECUTE_IF_SET_IN_BITMAP (exprs
, 0, i
, bi
)
1379 pre_expr rep
= expression_for_id (i
);
1380 if (rep
->kind
== NAME
)
1381 return VN_INFO (PRE_EXPR_NAME (rep
))->valnum
;
1382 else if (rep
->kind
== CONSTANT
)
1383 return PRE_EXPR_CONSTANT (rep
);
1389 /* If we reached here we couldn't find an SSA_NAME. This can
1390 happen when we've discovered a value that has never appeared in
1391 the program as set to an SSA_NAME, as the result of phi translation.
1393 ??? We should be able to re-use this when we insert the statement
1395 name
= make_temp_ssa_name (get_expr_type (e
), gimple_build_nop (), "pretmp");
1396 VN_INFO_GET (name
)->value_id
= value_id
;
1397 VN_INFO (name
)->valnum
= name
;
1398 /* ??? For now mark this SSA name for release by SCCVN. */
1399 VN_INFO (name
)->needs_insertion
= true;
1400 add_to_value (value_id
, get_or_alloc_expr_for_name (name
));
1401 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1403 fprintf (dump_file
, "Created SSA_NAME representative ");
1404 print_generic_expr (dump_file
, name
, 0);
1405 fprintf (dump_file
, " for expression:");
1406 print_pre_expr (dump_file
, e
);
1407 fprintf (dump_file
, " (%04d)\n", value_id
);
1416 phi_translate (pre_expr expr
, bitmap_set_t set1
, bitmap_set_t set2
,
1417 basic_block pred
, basic_block phiblock
);
1419 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1420 the phis in PRED. Return NULL if we can't find a leader for each part
1421 of the translated expression. */
1424 phi_translate_1 (pre_expr expr
, bitmap_set_t set1
, bitmap_set_t set2
,
1425 basic_block pred
, basic_block phiblock
)
1432 bool changed
= false;
1433 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
1434 vn_nary_op_t newnary
= XALLOCAVAR (struct vn_nary_op_s
,
1435 sizeof_vn_nary_op (nary
->length
));
1436 memcpy (newnary
, nary
, sizeof_vn_nary_op (nary
->length
));
1438 for (i
= 0; i
< newnary
->length
; i
++)
1440 if (TREE_CODE (newnary
->op
[i
]) != SSA_NAME
)
1444 pre_expr leader
, result
;
1445 unsigned int op_val_id
= VN_INFO (newnary
->op
[i
])->value_id
;
1446 leader
= find_leader_in_sets (op_val_id
, set1
, set2
);
1447 result
= phi_translate (leader
, set1
, set2
, pred
, phiblock
);
1448 if (result
&& result
!= leader
)
1449 newnary
->op
[i
] = get_representative_for (result
);
1453 changed
|= newnary
->op
[i
] != nary
->op
[i
];
1459 unsigned int new_val_id
;
1461 PRE_EXPR_NARY (expr
) = newnary
;
1462 constant
= fully_constant_expression (expr
);
1463 PRE_EXPR_NARY (expr
) = nary
;
1464 if (constant
!= expr
)
1466 /* For non-CONSTANTs we have to make sure we can eventually
1467 insert the expression. Which means we need to have a
1469 if (constant
->kind
!= CONSTANT
)
1471 unsigned value_id
= get_expr_value_id (constant
);
1472 constant
= find_leader_in_sets (value_id
, set1
, set2
);
1480 tree result
= vn_nary_op_lookup_pieces (newnary
->length
,
1485 if (result
&& is_gimple_min_invariant (result
))
1486 return get_or_alloc_expr_for_constant (result
);
1488 expr
= pre_expr_pool
.allocate ();
1493 PRE_EXPR_NARY (expr
) = nary
;
1494 new_val_id
= nary
->value_id
;
1495 get_or_alloc_expression_id (expr
);
1499 new_val_id
= get_next_value_id ();
1500 value_expressions
.safe_grow_cleared (get_max_value_id () + 1);
1501 nary
= vn_nary_op_insert_pieces (newnary
->length
,
1505 result
, new_val_id
);
1506 PRE_EXPR_NARY (expr
) = nary
;
1507 get_or_alloc_expression_id (expr
);
1509 add_to_value (new_val_id
, expr
);
1517 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
1518 vec
<vn_reference_op_s
> operands
= ref
->operands
;
1519 tree vuse
= ref
->vuse
;
1520 tree newvuse
= vuse
;
1521 vec
<vn_reference_op_s
> newoperands
= vNULL
;
1522 bool changed
= false, same_valid
= true;
1524 vn_reference_op_t operand
;
1525 vn_reference_t newref
;
1527 for (i
= 0; operands
.iterate (i
, &operand
); i
++)
1532 tree type
= operand
->type
;
1533 vn_reference_op_s newop
= *operand
;
1534 op
[0] = operand
->op0
;
1535 op
[1] = operand
->op1
;
1536 op
[2] = operand
->op2
;
1537 for (n
= 0; n
< 3; ++n
)
1539 unsigned int op_val_id
;
1542 if (TREE_CODE (op
[n
]) != SSA_NAME
)
1544 /* We can't possibly insert these. */
1546 && !is_gimple_min_invariant (op
[n
]))
1550 op_val_id
= VN_INFO (op
[n
])->value_id
;
1551 leader
= find_leader_in_sets (op_val_id
, set1
, set2
);
1552 opresult
= phi_translate (leader
, set1
, set2
, pred
, phiblock
);
1553 if (opresult
&& opresult
!= leader
)
1555 tree name
= get_representative_for (opresult
);
1556 changed
|= name
!= op
[n
];
1564 newoperands
.release ();
1569 if (!newoperands
.exists ())
1570 newoperands
= operands
.copy ();
1571 /* We may have changed from an SSA_NAME to a constant */
1572 if (newop
.opcode
== SSA_NAME
&& TREE_CODE (op
[0]) != SSA_NAME
)
1573 newop
.opcode
= TREE_CODE (op
[0]);
1578 newoperands
[i
] = newop
;
1580 gcc_checking_assert (i
== operands
.length ());
1584 newvuse
= translate_vuse_through_block (newoperands
.exists ()
1585 ? newoperands
: operands
,
1586 ref
->set
, ref
->type
,
1587 vuse
, phiblock
, pred
,
1589 if (newvuse
== NULL_TREE
)
1591 newoperands
.release ();
1596 if (changed
|| newvuse
!= vuse
)
1598 unsigned int new_val_id
;
1601 tree result
= vn_reference_lookup_pieces (newvuse
, ref
->set
,
1603 newoperands
.exists ()
1604 ? newoperands
: operands
,
1607 newoperands
.release ();
1609 /* We can always insert constants, so if we have a partial
1610 redundant constant load of another type try to translate it
1611 to a constant of appropriate type. */
1612 if (result
&& is_gimple_min_invariant (result
))
1615 if (!useless_type_conversion_p (ref
->type
, TREE_TYPE (result
)))
1617 tem
= fold_unary (VIEW_CONVERT_EXPR
, ref
->type
, result
);
1618 if (tem
&& !is_gimple_min_invariant (tem
))
1622 return get_or_alloc_expr_for_constant (tem
);
1625 /* If we'd have to convert things we would need to validate
1626 if we can insert the translated expression. So fail
1627 here for now - we cannot insert an alias with a different
1628 type in the VN tables either, as that would assert. */
1630 && !useless_type_conversion_p (ref
->type
, TREE_TYPE (result
)))
1632 else if (!result
&& newref
1633 && !useless_type_conversion_p (ref
->type
, newref
->type
))
1635 newoperands
.release ();
1639 expr
= pre_expr_pool
.allocate ();
1640 expr
->kind
= REFERENCE
;
1645 PRE_EXPR_REFERENCE (expr
) = newref
;
1646 constant
= fully_constant_expression (expr
);
1647 if (constant
!= expr
)
1650 new_val_id
= newref
->value_id
;
1651 get_or_alloc_expression_id (expr
);
1655 if (changed
|| !same_valid
)
1657 new_val_id
= get_next_value_id ();
1658 value_expressions
.safe_grow_cleared
1659 (get_max_value_id () + 1);
1662 new_val_id
= ref
->value_id
;
1663 if (!newoperands
.exists ())
1664 newoperands
= operands
.copy ();
1665 newref
= vn_reference_insert_pieces (newvuse
, ref
->set
,
1668 result
, new_val_id
);
1669 newoperands
= vNULL
;
1670 PRE_EXPR_REFERENCE (expr
) = newref
;
1671 constant
= fully_constant_expression (expr
);
1672 if (constant
!= expr
)
1674 get_or_alloc_expression_id (expr
);
1676 add_to_value (new_val_id
, expr
);
1678 newoperands
.release ();
1685 tree name
= PRE_EXPR_NAME (expr
);
1686 gimple
*def_stmt
= SSA_NAME_DEF_STMT (name
);
1687 /* If the SSA name is defined by a PHI node in this block,
1689 if (gimple_code (def_stmt
) == GIMPLE_PHI
1690 && gimple_bb (def_stmt
) == phiblock
)
1692 edge e
= find_edge (pred
, gimple_bb (def_stmt
));
1693 tree def
= PHI_ARG_DEF (def_stmt
, e
->dest_idx
);
1695 /* Handle constant. */
1696 if (is_gimple_min_invariant (def
))
1697 return get_or_alloc_expr_for_constant (def
);
1699 return get_or_alloc_expr_for_name (def
);
1701 /* Otherwise return it unchanged - it will get removed if its
1702 value is not available in PREDs AVAIL_OUT set of expressions
1703 by the subtraction of TMP_GEN. */
1712 /* Wrapper around phi_translate_1 providing caching functionality. */
1715 phi_translate (pre_expr expr
, bitmap_set_t set1
, bitmap_set_t set2
,
1716 basic_block pred
, basic_block phiblock
)
1718 expr_pred_trans_t slot
= NULL
;
1724 /* Constants contain no values that need translation. */
1725 if (expr
->kind
== CONSTANT
)
1728 if (value_id_constant_p (get_expr_value_id (expr
)))
1731 /* Don't add translations of NAMEs as those are cheap to translate. */
1732 if (expr
->kind
!= NAME
)
1734 if (phi_trans_add (&slot
, expr
, pred
))
1736 /* Store NULL for the value we want to return in the case of
1742 phitrans
= phi_translate_1 (expr
, set1
, set2
, pred
, phiblock
);
1749 /* Remove failed translations again, they cause insert
1750 iteration to not pick up new opportunities reliably. */
1751 phi_translate_table
->remove_elt_with_hash (slot
, slot
->hashcode
);
1758 /* For each expression in SET, translate the values through phi nodes
1759 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1760 expressions in DEST. */
1763 phi_translate_set (bitmap_set_t dest
, bitmap_set_t set
, basic_block pred
,
1764 basic_block phiblock
)
1766 vec
<pre_expr
> exprs
;
1770 if (gimple_seq_empty_p (phi_nodes (phiblock
)))
1772 bitmap_set_copy (dest
, set
);
1776 exprs
= sorted_array_from_bitmap_set (set
);
1777 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
1779 pre_expr translated
;
1780 translated
= phi_translate (expr
, set
, NULL
, pred
, phiblock
);
1784 /* We might end up with multiple expressions from SET being
1785 translated to the same value. In this case we do not want
1786 to retain the NARY or REFERENCE expression but prefer a NAME
1787 which would be the leader. */
1788 if (translated
->kind
== NAME
)
1789 bitmap_value_replace_in_set (dest
, translated
);
1791 bitmap_value_insert_into_set (dest
, translated
);
1796 /* Find the leader for a value (i.e., the name representing that
1797 value) in a given set, and return it. Return NULL if no leader
1801 bitmap_find_leader (bitmap_set_t set
, unsigned int val
)
1803 if (value_id_constant_p (val
))
1807 bitmap exprset
= value_expressions
[val
];
1809 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
1811 pre_expr expr
= expression_for_id (i
);
1812 if (expr
->kind
== CONSTANT
)
1816 if (bitmap_set_contains_value (set
, val
))
1818 /* Rather than walk the entire bitmap of expressions, and see
1819 whether any of them has the value we are looking for, we look
1820 at the reverse mapping, which tells us the set of expressions
1821 that have a given value (IE value->expressions with that
1822 value) and see if any of those expressions are in our set.
1823 The number of expressions per value is usually significantly
1824 less than the number of expressions in the set. In fact, for
1825 large testcases, doing it this way is roughly 5-10x faster
1826 than walking the bitmap.
1827 If this is somehow a significant lose for some cases, we can
1828 choose which set to walk based on which set is smaller. */
1831 bitmap exprset
= value_expressions
[val
];
1833 EXECUTE_IF_AND_IN_BITMAP (exprset
, &set
->expressions
, 0, i
, bi
)
1834 return expression_for_id (i
);
1839 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1840 BLOCK by seeing if it is not killed in the block. Note that we are
1841 only determining whether there is a store that kills it. Because
1842 of the order in which clean iterates over values, we are guaranteed
1843 that altered operands will have caused us to be eliminated from the
1844 ANTIC_IN set already. */
1847 value_dies_in_block_x (pre_expr expr
, basic_block block
)
1849 tree vuse
= PRE_EXPR_REFERENCE (expr
)->vuse
;
1850 vn_reference_t refx
= PRE_EXPR_REFERENCE (expr
);
1852 gimple_stmt_iterator gsi
;
1853 unsigned id
= get_expression_id (expr
);
1860 /* Lookup a previously calculated result. */
1861 if (EXPR_DIES (block
)
1862 && bitmap_bit_p (EXPR_DIES (block
), id
* 2))
1863 return bitmap_bit_p (EXPR_DIES (block
), id
* 2 + 1);
1865 /* A memory expression {e, VUSE} dies in the block if there is a
1866 statement that may clobber e. If, starting statement walk from the
1867 top of the basic block, a statement uses VUSE there can be no kill
1868 inbetween that use and the original statement that loaded {e, VUSE},
1869 so we can stop walking. */
1870 ref
.base
= NULL_TREE
;
1871 for (gsi
= gsi_start_bb (block
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1873 tree def_vuse
, def_vdef
;
1874 def
= gsi_stmt (gsi
);
1875 def_vuse
= gimple_vuse (def
);
1876 def_vdef
= gimple_vdef (def
);
1878 /* Not a memory statement. */
1882 /* Not a may-def. */
1885 /* A load with the same VUSE, we're done. */
1886 if (def_vuse
== vuse
)
1892 /* Init ref only if we really need it. */
1893 if (ref
.base
== NULL_TREE
1894 && !ao_ref_init_from_vn_reference (&ref
, refx
->set
, refx
->type
,
1900 /* If the statement may clobber expr, it dies. */
1901 if (stmt_may_clobber_ref_p_1 (def
, &ref
))
1908 /* Remember the result. */
1909 if (!EXPR_DIES (block
))
1910 EXPR_DIES (block
) = BITMAP_ALLOC (&grand_bitmap_obstack
);
1911 bitmap_set_bit (EXPR_DIES (block
), id
* 2);
1913 bitmap_set_bit (EXPR_DIES (block
), id
* 2 + 1);
1919 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1920 contains its value-id. */
1923 op_valid_in_sets (bitmap_set_t set1
, bitmap_set_t set2
, tree op
)
1925 if (op
&& TREE_CODE (op
) == SSA_NAME
)
1927 unsigned int value_id
= VN_INFO (op
)->value_id
;
1928 if (!(bitmap_set_contains_value (set1
, value_id
)
1929 || (set2
&& bitmap_set_contains_value (set2
, value_id
))))
1935 /* Determine if the expression EXPR is valid in SET1 U SET2.
1936 ONLY SET2 CAN BE NULL.
1937 This means that we have a leader for each part of the expression
1938 (if it consists of values), or the expression is an SSA_NAME.
1939 For loads/calls, we also see if the vuse is killed in this block. */
1942 valid_in_sets (bitmap_set_t set1
, bitmap_set_t set2
, pre_expr expr
)
1947 /* By construction all NAMEs are available. Non-available
1948 NAMEs are removed by subtracting TMP_GEN from the sets. */
1953 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
1954 for (i
= 0; i
< nary
->length
; i
++)
1955 if (!op_valid_in_sets (set1
, set2
, nary
->op
[i
]))
1962 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
1963 vn_reference_op_t vro
;
1966 FOR_EACH_VEC_ELT (ref
->operands
, i
, vro
)
1968 if (!op_valid_in_sets (set1
, set2
, vro
->op0
)
1969 || !op_valid_in_sets (set1
, set2
, vro
->op1
)
1970 || !op_valid_in_sets (set1
, set2
, vro
->op2
))
1980 /* Clean the set of expressions that are no longer valid in SET1 or
1981 SET2. This means expressions that are made up of values we have no
1982 leaders for in SET1 or SET2. This version is used for partial
1983 anticipation, which means it is not valid in either ANTIC_IN or
1987 dependent_clean (bitmap_set_t set1
, bitmap_set_t set2
)
1989 vec
<pre_expr
> exprs
= sorted_array_from_bitmap_set (set1
);
1993 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
1995 if (!valid_in_sets (set1
, set2
, expr
))
1996 bitmap_remove_from_set (set1
, expr
);
2001 /* Clean the set of expressions that are no longer valid in SET. This
2002 means expressions that are made up of values we have no leaders for
2006 clean (bitmap_set_t set
)
2008 vec
<pre_expr
> exprs
= sorted_array_from_bitmap_set (set
);
2012 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
2014 if (!valid_in_sets (set
, NULL
, expr
))
2015 bitmap_remove_from_set (set
, expr
);
2020 /* Clean the set of expressions that are no longer valid in SET because
2021 they are clobbered in BLOCK or because they trap and may not be executed. */
2024 prune_clobbered_mems (bitmap_set_t set
, basic_block block
)
2029 FOR_EACH_EXPR_ID_IN_SET (set
, i
, bi
)
2031 pre_expr expr
= expression_for_id (i
);
2032 if (expr
->kind
== REFERENCE
)
2034 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
2037 gimple
*def_stmt
= SSA_NAME_DEF_STMT (ref
->vuse
);
2038 if (!gimple_nop_p (def_stmt
)
2039 && ((gimple_bb (def_stmt
) != block
2040 && !dominated_by_p (CDI_DOMINATORS
,
2041 block
, gimple_bb (def_stmt
)))
2042 || (gimple_bb (def_stmt
) == block
2043 && value_dies_in_block_x (expr
, block
))))
2044 bitmap_remove_from_set (set
, expr
);
2047 else if (expr
->kind
== NARY
)
2049 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
2050 /* If the NARY may trap make sure the block does not contain
2051 a possible exit point.
2052 ??? This is overly conservative if we translate AVAIL_OUT
2053 as the available expression might be after the exit point. */
2054 if (BB_MAY_NOTRETURN (block
)
2055 && vn_nary_may_trap (nary
))
2056 bitmap_remove_from_set (set
, expr
);
2061 static sbitmap has_abnormal_preds
;
2063 /* Compute the ANTIC set for BLOCK.
2065 If succs(BLOCK) > 1 then
2066 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2067 else if succs(BLOCK) == 1 then
2068 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2070 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2074 compute_antic_aux (basic_block block
, bool block_has_abnormal_pred_edge
)
2076 bool changed
= false;
2077 bitmap_set_t S
, old
, ANTIC_OUT
;
2082 bool was_visited
= BB_VISITED (block
);
2084 old
= ANTIC_OUT
= S
= NULL
;
2085 BB_VISITED (block
) = 1;
2087 /* If any edges from predecessors are abnormal, antic_in is empty,
2089 if (block_has_abnormal_pred_edge
)
2090 goto maybe_dump_sets
;
2092 old
= ANTIC_IN (block
);
2093 ANTIC_OUT
= bitmap_set_new ();
2095 /* If the block has no successors, ANTIC_OUT is empty. */
2096 if (EDGE_COUNT (block
->succs
) == 0)
2098 /* If we have one successor, we could have some phi nodes to
2099 translate through. */
2100 else if (single_succ_p (block
))
2102 basic_block succ_bb
= single_succ (block
);
2103 gcc_assert (BB_VISITED (succ_bb
));
2104 phi_translate_set (ANTIC_OUT
, ANTIC_IN (succ_bb
), block
, succ_bb
);
2106 /* If we have multiple successors, we take the intersection of all of
2107 them. Note that in the case of loop exit phi nodes, we may have
2108 phis to translate through. */
2112 basic_block bprime
, first
= NULL
;
2114 auto_vec
<basic_block
> worklist (EDGE_COUNT (block
->succs
));
2115 FOR_EACH_EDGE (e
, ei
, block
->succs
)
2118 && BB_VISITED (e
->dest
))
2120 else if (BB_VISITED (e
->dest
))
2121 worklist
.quick_push (e
->dest
);
2124 /* Unvisited successors get their ANTIC_IN replaced by the
2125 maximal set to arrive at a maximum ANTIC_IN solution.
2126 We can ignore them in the intersection operation and thus
2127 need not explicitely represent that maximum solution. */
2128 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2129 fprintf (dump_file
, "ANTIC_IN is MAX on %d->%d\n",
2130 e
->src
->index
, e
->dest
->index
);
2134 /* Of multiple successors we have to have visited one already
2135 which is guaranteed by iteration order. */
2136 gcc_assert (first
!= NULL
);
2138 phi_translate_set (ANTIC_OUT
, ANTIC_IN (first
), block
, first
);
2140 FOR_EACH_VEC_ELT (worklist
, i
, bprime
)
2142 if (!gimple_seq_empty_p (phi_nodes (bprime
)))
2144 bitmap_set_t tmp
= bitmap_set_new ();
2145 phi_translate_set (tmp
, ANTIC_IN (bprime
), block
, bprime
);
2146 bitmap_set_and (ANTIC_OUT
, tmp
);
2147 bitmap_set_free (tmp
);
2150 bitmap_set_and (ANTIC_OUT
, ANTIC_IN (bprime
));
2154 /* Prune expressions that are clobbered in block and thus become
2155 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2156 prune_clobbered_mems (ANTIC_OUT
, block
);
2158 /* Generate ANTIC_OUT - TMP_GEN. */
2159 S
= bitmap_set_subtract (ANTIC_OUT
, TMP_GEN (block
));
2161 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2162 ANTIC_IN (block
) = bitmap_set_subtract (EXP_GEN (block
),
2165 /* Then union in the ANTIC_OUT - TMP_GEN values,
2166 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2167 FOR_EACH_EXPR_ID_IN_SET (S
, bii
, bi
)
2168 bitmap_value_insert_into_set (ANTIC_IN (block
),
2169 expression_for_id (bii
));
2171 clean (ANTIC_IN (block
));
2173 if (!was_visited
|| !bitmap_set_equal (old
, ANTIC_IN (block
)))
2177 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2180 print_bitmap_set (dump_file
, ANTIC_OUT
, "ANTIC_OUT", block
->index
);
2183 fprintf (dump_file
, "[changed] ");
2184 print_bitmap_set (dump_file
, ANTIC_IN (block
), "ANTIC_IN",
2188 print_bitmap_set (dump_file
, S
, "S", block
->index
);
2191 bitmap_set_free (old
);
2193 bitmap_set_free (S
);
2195 bitmap_set_free (ANTIC_OUT
);
2199 /* Compute PARTIAL_ANTIC for BLOCK.
2201 If succs(BLOCK) > 1 then
2202 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2203 in ANTIC_OUT for all succ(BLOCK)
2204 else if succs(BLOCK) == 1 then
2205 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2207 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2212 compute_partial_antic_aux (basic_block block
,
2213 bool block_has_abnormal_pred_edge
)
2215 bitmap_set_t old_PA_IN
;
2216 bitmap_set_t PA_OUT
;
2219 unsigned long max_pa
= PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH
);
2221 old_PA_IN
= PA_OUT
= NULL
;
2223 /* If any edges from predecessors are abnormal, antic_in is empty,
2225 if (block_has_abnormal_pred_edge
)
2226 goto maybe_dump_sets
;
2228 /* If there are too many partially anticipatable values in the
2229 block, phi_translate_set can take an exponential time: stop
2230 before the translation starts. */
2232 && single_succ_p (block
)
2233 && bitmap_count_bits (&PA_IN (single_succ (block
))->values
) > max_pa
)
2234 goto maybe_dump_sets
;
2236 old_PA_IN
= PA_IN (block
);
2237 PA_OUT
= bitmap_set_new ();
2239 /* If the block has no successors, ANTIC_OUT is empty. */
2240 if (EDGE_COUNT (block
->succs
) == 0)
2242 /* If we have one successor, we could have some phi nodes to
2243 translate through. Note that we can't phi translate across DFS
2244 back edges in partial antic, because it uses a union operation on
2245 the successors. For recurrences like IV's, we will end up
2246 generating a new value in the set on each go around (i + 3 (VH.1)
2247 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2248 else if (single_succ_p (block
))
2250 basic_block succ
= single_succ (block
);
2251 if (!(single_succ_edge (block
)->flags
& EDGE_DFS_BACK
))
2252 phi_translate_set (PA_OUT
, PA_IN (succ
), block
, succ
);
2254 /* If we have multiple successors, we take the union of all of
2261 auto_vec
<basic_block
> worklist (EDGE_COUNT (block
->succs
));
2262 FOR_EACH_EDGE (e
, ei
, block
->succs
)
2264 if (e
->flags
& EDGE_DFS_BACK
)
2266 worklist
.quick_push (e
->dest
);
2268 if (worklist
.length () > 0)
2270 FOR_EACH_VEC_ELT (worklist
, i
, bprime
)
2275 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime
), i
, bi
)
2276 bitmap_value_insert_into_set (PA_OUT
,
2277 expression_for_id (i
));
2278 if (!gimple_seq_empty_p (phi_nodes (bprime
)))
2280 bitmap_set_t pa_in
= bitmap_set_new ();
2281 phi_translate_set (pa_in
, PA_IN (bprime
), block
, bprime
);
2282 FOR_EACH_EXPR_ID_IN_SET (pa_in
, i
, bi
)
2283 bitmap_value_insert_into_set (PA_OUT
,
2284 expression_for_id (i
));
2285 bitmap_set_free (pa_in
);
2288 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime
), i
, bi
)
2289 bitmap_value_insert_into_set (PA_OUT
,
2290 expression_for_id (i
));
2295 /* Prune expressions that are clobbered in block and thus become
2296 invalid if translated from PA_OUT to PA_IN. */
2297 prune_clobbered_mems (PA_OUT
, block
);
2299 /* PA_IN starts with PA_OUT - TMP_GEN.
2300 Then we subtract things from ANTIC_IN. */
2301 PA_IN (block
) = bitmap_set_subtract (PA_OUT
, TMP_GEN (block
));
2303 /* For partial antic, we want to put back in the phi results, since
2304 we will properly avoid making them partially antic over backedges. */
2305 bitmap_ior_into (&PA_IN (block
)->values
, &PHI_GEN (block
)->values
);
2306 bitmap_ior_into (&PA_IN (block
)->expressions
, &PHI_GEN (block
)->expressions
);
2308 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2309 bitmap_set_subtract_values (PA_IN (block
), ANTIC_IN (block
));
2311 dependent_clean (PA_IN (block
), ANTIC_IN (block
));
2314 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2317 print_bitmap_set (dump_file
, PA_OUT
, "PA_OUT", block
->index
);
2319 print_bitmap_set (dump_file
, PA_IN (block
), "PA_IN", block
->index
);
2322 bitmap_set_free (old_PA_IN
);
2324 bitmap_set_free (PA_OUT
);
2327 /* Compute ANTIC and partial ANTIC sets. */
2330 compute_antic (void)
2332 bool changed
= true;
2333 int num_iterations
= 0;
2339 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2340 We pre-build the map of blocks with incoming abnormal edges here. */
2341 has_abnormal_preds
= sbitmap_alloc (last_basic_block_for_fn (cfun
));
2342 bitmap_clear (has_abnormal_preds
);
2344 FOR_ALL_BB_FN (block
, cfun
)
2346 BB_VISITED (block
) = 0;
2348 FOR_EACH_EDGE (e
, ei
, block
->preds
)
2349 if (e
->flags
& EDGE_ABNORMAL
)
2351 bitmap_set_bit (has_abnormal_preds
, block
->index
);
2353 /* We also anticipate nothing. */
2354 BB_VISITED (block
) = 1;
2358 /* While we are here, give empty ANTIC_IN sets to each block. */
2359 ANTIC_IN (block
) = bitmap_set_new ();
2360 if (do_partial_partial
)
2361 PA_IN (block
) = bitmap_set_new ();
2364 /* At the exit block we anticipate nothing. */
2365 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun
)) = 1;
2367 /* For ANTIC computation we need a postorder that also guarantees that
2368 a block with a single successor is visited after its successor.
2369 RPO on the inverted CFG has this property. */
2370 int *postorder
= XNEWVEC (int, n_basic_blocks_for_fn (cfun
));
2371 int postorder_num
= inverted_post_order_compute (postorder
);
2373 auto_sbitmap
worklist (last_basic_block_for_fn (cfun
) + 1);
2374 bitmap_ones (worklist
);
2377 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2378 fprintf (dump_file
, "Starting iteration %d\n", num_iterations
);
2379 /* ??? We need to clear our PHI translation cache here as the
2380 ANTIC sets shrink and we restrict valid translations to
2381 those having operands with leaders in ANTIC. Same below
2382 for PA ANTIC computation. */
2385 for (i
= postorder_num
- 1; i
>= 0; i
--)
2387 if (bitmap_bit_p (worklist
, postorder
[i
]))
2389 basic_block block
= BASIC_BLOCK_FOR_FN (cfun
, postorder
[i
]);
2390 bitmap_clear_bit (worklist
, block
->index
);
2391 if (compute_antic_aux (block
,
2392 bitmap_bit_p (has_abnormal_preds
,
2395 FOR_EACH_EDGE (e
, ei
, block
->preds
)
2396 bitmap_set_bit (worklist
, e
->src
->index
);
2401 /* Theoretically possible, but *highly* unlikely. */
2402 gcc_checking_assert (num_iterations
< 500);
2405 statistics_histogram_event (cfun
, "compute_antic iterations",
2408 if (do_partial_partial
)
2410 /* For partial antic we ignore backedges and thus we do not need
2411 to perform any iteration when we process blocks in postorder. */
2412 postorder_num
= pre_and_rev_post_order_compute (NULL
, postorder
, false);
2413 for (i
= postorder_num
- 1 ; i
>= 0; i
--)
2415 basic_block block
= BASIC_BLOCK_FOR_FN (cfun
, postorder
[i
]);
2416 compute_partial_antic_aux (block
,
2417 bitmap_bit_p (has_abnormal_preds
,
2422 sbitmap_free (has_abnormal_preds
);
2427 /* Inserted expressions are placed onto this worklist, which is used
2428 for performing quick dead code elimination of insertions we made
2429 that didn't turn out to be necessary. */
2430 static bitmap inserted_exprs
;
2432 /* The actual worker for create_component_ref_by_pieces. */
2435 create_component_ref_by_pieces_1 (basic_block block
, vn_reference_t ref
,
2436 unsigned int *operand
, gimple_seq
*stmts
)
2438 vn_reference_op_t currop
= &ref
->operands
[*operand
];
2441 switch (currop
->opcode
)
2448 tree baseop
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2452 tree offset
= currop
->op0
;
2453 if (TREE_CODE (baseop
) == ADDR_EXPR
2454 && handled_component_p (TREE_OPERAND (baseop
, 0)))
2458 base
= get_addr_base_and_unit_offset (TREE_OPERAND (baseop
, 0),
2461 offset
= int_const_binop (PLUS_EXPR
, offset
,
2462 build_int_cst (TREE_TYPE (offset
),
2464 baseop
= build_fold_addr_expr (base
);
2466 genop
= build2 (MEM_REF
, currop
->type
, baseop
, offset
);
2467 MR_DEPENDENCE_CLIQUE (genop
) = currop
->clique
;
2468 MR_DEPENDENCE_BASE (genop
) = currop
->base
;
2469 REF_REVERSE_STORAGE_ORDER (genop
) = currop
->reverse
;
2473 case TARGET_MEM_REF
:
2475 tree genop0
= NULL_TREE
, genop1
= NULL_TREE
;
2476 vn_reference_op_t nextop
= &ref
->operands
[++*operand
];
2477 tree baseop
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2483 genop0
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2489 genop1
= find_or_generate_expression (block
, nextop
->op0
, stmts
);
2493 genop
= build5 (TARGET_MEM_REF
, currop
->type
,
2494 baseop
, currop
->op2
, genop0
, currop
->op1
, genop1
);
2496 MR_DEPENDENCE_CLIQUE (genop
) = currop
->clique
;
2497 MR_DEPENDENCE_BASE (genop
) = currop
->base
;
2504 gcc_assert (is_gimple_min_invariant (currop
->op0
));
2510 case VIEW_CONVERT_EXPR
:
2512 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2516 return fold_build1 (currop
->opcode
, currop
->type
, genop0
);
2519 case WITH_SIZE_EXPR
:
2521 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2525 tree genop1
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2528 return fold_build2 (currop
->opcode
, currop
->type
, genop0
, genop1
);
2533 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2537 tree op1
= currop
->op0
;
2538 tree op2
= currop
->op1
;
2539 tree t
= build3 (BIT_FIELD_REF
, currop
->type
, genop0
, op1
, op2
);
2540 REF_REVERSE_STORAGE_ORDER (t
) = currop
->reverse
;
2544 /* For array ref vn_reference_op's, operand 1 of the array ref
2545 is op0 of the reference op and operand 3 of the array ref is
2547 case ARRAY_RANGE_REF
:
2551 tree genop1
= currop
->op0
;
2552 tree genop2
= currop
->op1
;
2553 tree genop3
= currop
->op2
;
2554 genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2558 genop1
= find_or_generate_expression (block
, genop1
, stmts
);
2563 tree domain_type
= TYPE_DOMAIN (TREE_TYPE (genop0
));
2564 /* Drop zero minimum index if redundant. */
2565 if (integer_zerop (genop2
)
2567 || integer_zerop (TYPE_MIN_VALUE (domain_type
))))
2571 genop2
= find_or_generate_expression (block
, genop2
, stmts
);
2578 tree elmt_type
= TREE_TYPE (TREE_TYPE (genop0
));
2579 /* We can't always put a size in units of the element alignment
2580 here as the element alignment may be not visible. See
2581 PR43783. Simply drop the element size for constant
2583 if (TREE_CODE (genop3
) == INTEGER_CST
2584 && TREE_CODE (TYPE_SIZE_UNIT (elmt_type
)) == INTEGER_CST
2585 && wi::eq_p (wi::to_offset (TYPE_SIZE_UNIT (elmt_type
)),
2586 (wi::to_offset (genop3
)
2587 * vn_ref_op_align_unit (currop
))))
2591 genop3
= find_or_generate_expression (block
, genop3
, stmts
);
2596 return build4 (currop
->opcode
, currop
->type
, genop0
, genop1
,
2603 tree genop2
= currop
->op1
;
2604 op0
= create_component_ref_by_pieces_1 (block
, ref
, operand
, stmts
);
2607 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2611 genop2
= find_or_generate_expression (block
, genop2
, stmts
);
2615 return fold_build3 (COMPONENT_REF
, TREE_TYPE (op1
), op0
, op1
, genop2
);
2620 genop
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2641 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2642 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2643 trying to rename aggregates into ssa form directly, which is a no no.
2645 Thus, this routine doesn't create temporaries, it just builds a
2646 single access expression for the array, calling
2647 find_or_generate_expression to build the innermost pieces.
2649 This function is a subroutine of create_expression_by_pieces, and
2650 should not be called on it's own unless you really know what you
2654 create_component_ref_by_pieces (basic_block block
, vn_reference_t ref
,
2657 unsigned int op
= 0;
2658 return create_component_ref_by_pieces_1 (block
, ref
, &op
, stmts
);
2661 /* Find a simple leader for an expression, or generate one using
2662 create_expression_by_pieces from a NARY expression for the value.
2663 BLOCK is the basic_block we are looking for leaders in.
2664 OP is the tree expression to find a leader for or generate.
2665 Returns the leader or NULL_TREE on failure. */
2668 find_or_generate_expression (basic_block block
, tree op
, gimple_seq
*stmts
)
2670 pre_expr expr
= get_or_alloc_expr_for (op
);
2671 unsigned int lookfor
= get_expr_value_id (expr
);
2672 pre_expr leader
= bitmap_find_leader (AVAIL_OUT (block
), lookfor
);
2675 if (leader
->kind
== NAME
)
2676 return PRE_EXPR_NAME (leader
);
2677 else if (leader
->kind
== CONSTANT
)
2678 return PRE_EXPR_CONSTANT (leader
);
2684 /* It must be a complex expression, so generate it recursively. Note
2685 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2686 where the insert algorithm fails to insert a required expression. */
2687 bitmap exprset
= value_expressions
[lookfor
];
2690 EXECUTE_IF_SET_IN_BITMAP (exprset
, 0, i
, bi
)
2692 pre_expr temp
= expression_for_id (i
);
2693 /* We cannot insert random REFERENCE expressions at arbitrary
2694 places. We can insert NARYs which eventually re-materializes
2695 its operand values. */
2696 if (temp
->kind
== NARY
)
2697 return create_expression_by_pieces (block
, temp
, stmts
,
2698 get_expr_type (expr
));
2705 #define NECESSARY GF_PLF_1
2707 /* Create an expression in pieces, so that we can handle very complex
2708 expressions that may be ANTIC, but not necessary GIMPLE.
2709 BLOCK is the basic block the expression will be inserted into,
2710 EXPR is the expression to insert (in value form)
2711 STMTS is a statement list to append the necessary insertions into.
2713 This function will die if we hit some value that shouldn't be
2714 ANTIC but is (IE there is no leader for it, or its components).
2715 The function returns NULL_TREE in case a different antic expression
2716 has to be inserted first.
2717 This function may also generate expressions that are themselves
2718 partially or fully redundant. Those that are will be either made
2719 fully redundant during the next iteration of insert (for partially
2720 redundant ones), or eliminated by eliminate (for fully redundant
2724 create_expression_by_pieces (basic_block block
, pre_expr expr
,
2725 gimple_seq
*stmts
, tree type
)
2729 gimple_seq forced_stmts
= NULL
;
2730 unsigned int value_id
;
2731 gimple_stmt_iterator gsi
;
2732 tree exprtype
= type
? type
: get_expr_type (expr
);
2738 /* We may hit the NAME/CONSTANT case if we have to convert types
2739 that value numbering saw through. */
2741 folded
= PRE_EXPR_NAME (expr
);
2742 if (useless_type_conversion_p (exprtype
, TREE_TYPE (folded
)))
2747 folded
= PRE_EXPR_CONSTANT (expr
);
2748 tree tem
= fold_convert (exprtype
, folded
);
2749 if (is_gimple_min_invariant (tem
))
2754 if (PRE_EXPR_REFERENCE (expr
)->operands
[0].opcode
== CALL_EXPR
)
2756 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
2757 unsigned int operand
= 1;
2758 vn_reference_op_t currop
= &ref
->operands
[0];
2759 tree sc
= NULL_TREE
;
2761 if (TREE_CODE (currop
->op0
) == FUNCTION_DECL
)
2764 fn
= find_or_generate_expression (block
, currop
->op0
, stmts
);
2769 sc
= find_or_generate_expression (block
, currop
->op1
, stmts
);
2773 auto_vec
<tree
> args (ref
->operands
.length () - 1);
2774 while (operand
< ref
->operands
.length ())
2776 tree arg
= create_component_ref_by_pieces_1 (block
, ref
,
2780 args
.quick_push (arg
);
2783 = gimple_build_call_vec ((TREE_CODE (fn
) == FUNCTION_DECL
2784 ? build_fold_addr_expr (fn
) : fn
), args
);
2785 gimple_call_set_with_bounds (call
, currop
->with_bounds
);
2787 gimple_call_set_chain (call
, sc
);
2788 tree forcedname
= make_ssa_name (currop
->type
);
2789 gimple_call_set_lhs (call
, forcedname
);
2790 gimple_set_vuse (call
, BB_LIVE_VOP_ON_EXIT (block
));
2791 gimple_seq_add_stmt_without_update (&forced_stmts
, call
);
2792 folded
= forcedname
;
2796 folded
= create_component_ref_by_pieces (block
,
2797 PRE_EXPR_REFERENCE (expr
),
2801 name
= make_temp_ssa_name (exprtype
, NULL
, "pretmp");
2802 newstmt
= gimple_build_assign (name
, folded
);
2803 gimple_seq_add_stmt_without_update (&forced_stmts
, newstmt
);
2804 gimple_set_vuse (newstmt
, BB_LIVE_VOP_ON_EXIT (block
));
2810 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
2811 tree
*genop
= XALLOCAVEC (tree
, nary
->length
);
2813 for (i
= 0; i
< nary
->length
; ++i
)
2815 genop
[i
] = find_or_generate_expression (block
, nary
->op
[i
], stmts
);
2818 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2819 may have conversions stripped. */
2820 if (nary
->opcode
== POINTER_PLUS_EXPR
)
2823 genop
[i
] = gimple_convert (&forced_stmts
,
2824 nary
->type
, genop
[i
]);
2826 genop
[i
] = gimple_convert (&forced_stmts
,
2827 sizetype
, genop
[i
]);
2830 genop
[i
] = gimple_convert (&forced_stmts
,
2831 TREE_TYPE (nary
->op
[i
]), genop
[i
]);
2833 if (nary
->opcode
== CONSTRUCTOR
)
2835 vec
<constructor_elt
, va_gc
> *elts
= NULL
;
2836 for (i
= 0; i
< nary
->length
; ++i
)
2837 CONSTRUCTOR_APPEND_ELT (elts
, NULL_TREE
, genop
[i
]);
2838 folded
= build_constructor (nary
->type
, elts
);
2839 name
= make_temp_ssa_name (exprtype
, NULL
, "pretmp");
2840 newstmt
= gimple_build_assign (name
, folded
);
2841 gimple_seq_add_stmt_without_update (&forced_stmts
, newstmt
);
2846 switch (nary
->length
)
2849 folded
= gimple_build (&forced_stmts
, nary
->opcode
, nary
->type
,
2853 folded
= gimple_build (&forced_stmts
, nary
->opcode
, nary
->type
,
2854 genop
[0], genop
[1]);
2857 folded
= gimple_build (&forced_stmts
, nary
->opcode
, nary
->type
,
2858 genop
[0], genop
[1], genop
[2]);
2870 folded
= gimple_convert (&forced_stmts
, exprtype
, folded
);
2872 /* If there is nothing to insert, return the simplified result. */
2873 if (gimple_seq_empty_p (forced_stmts
))
2875 /* If we simplified to a constant return it and discard eventually
2877 if (is_gimple_min_invariant (folded
))
2879 gimple_seq_discard (forced_stmts
);
2883 gcc_assert (TREE_CODE (folded
) == SSA_NAME
);
2885 /* If we have any intermediate expressions to the value sets, add them
2886 to the value sets and chain them in the instruction stream. */
2889 gsi
= gsi_start (forced_stmts
);
2890 for (; !gsi_end_p (gsi
); gsi_next (&gsi
))
2892 gimple
*stmt
= gsi_stmt (gsi
);
2893 tree forcedname
= gimple_get_lhs (stmt
);
2896 if (forcedname
!= folded
)
2898 VN_INFO_GET (forcedname
)->valnum
= forcedname
;
2899 VN_INFO (forcedname
)->value_id
= get_next_value_id ();
2900 nameexpr
= get_or_alloc_expr_for_name (forcedname
);
2901 add_to_value (VN_INFO (forcedname
)->value_id
, nameexpr
);
2902 bitmap_value_replace_in_set (NEW_SETS (block
), nameexpr
);
2903 bitmap_value_replace_in_set (AVAIL_OUT (block
), nameexpr
);
2906 bitmap_set_bit (inserted_exprs
, SSA_NAME_VERSION (forcedname
));
2907 gimple_set_plf (stmt
, NECESSARY
, false);
2909 gimple_seq_add_seq (stmts
, forced_stmts
);
2914 /* Fold the last statement. */
2915 gsi
= gsi_last (*stmts
);
2916 if (fold_stmt_inplace (&gsi
))
2917 update_stmt (gsi_stmt (gsi
));
2919 /* Add a value number to the temporary.
2920 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2921 we are creating the expression by pieces, and this particular piece of
2922 the expression may have been represented. There is no harm in replacing
2924 value_id
= get_expr_value_id (expr
);
2925 VN_INFO_GET (name
)->value_id
= value_id
;
2926 VN_INFO (name
)->valnum
= sccvn_valnum_from_value_id (value_id
);
2927 if (VN_INFO (name
)->valnum
== NULL_TREE
)
2928 VN_INFO (name
)->valnum
= name
;
2929 gcc_assert (VN_INFO (name
)->valnum
!= NULL_TREE
);
2930 nameexpr
= get_or_alloc_expr_for_name (name
);
2931 add_to_value (value_id
, nameexpr
);
2932 if (NEW_SETS (block
))
2933 bitmap_value_replace_in_set (NEW_SETS (block
), nameexpr
);
2934 bitmap_value_replace_in_set (AVAIL_OUT (block
), nameexpr
);
2936 pre_stats
.insertions
++;
2937 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2939 fprintf (dump_file
, "Inserted ");
2940 print_gimple_stmt (dump_file
, gsi_stmt (gsi_last (*stmts
)), 0, 0);
2941 fprintf (dump_file
, " in predecessor %d (%04d)\n",
2942 block
->index
, value_id
);
2949 /* Insert the to-be-made-available values of expression EXPRNUM for each
2950 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
2951 merge the result with a phi node, given the same value number as
2952 NODE. Return true if we have inserted new stuff. */
2955 insert_into_preds_of_block (basic_block block
, unsigned int exprnum
,
2956 vec
<pre_expr
> avail
)
2958 pre_expr expr
= expression_for_id (exprnum
);
2960 unsigned int val
= get_expr_value_id (expr
);
2962 bool insertions
= false;
2967 tree type
= get_expr_type (expr
);
2971 /* Make sure we aren't creating an induction variable. */
2972 if (bb_loop_depth (block
) > 0 && EDGE_COUNT (block
->preds
) == 2)
2974 bool firstinsideloop
= false;
2975 bool secondinsideloop
= false;
2976 firstinsideloop
= flow_bb_inside_loop_p (block
->loop_father
,
2977 EDGE_PRED (block
, 0)->src
);
2978 secondinsideloop
= flow_bb_inside_loop_p (block
->loop_father
,
2979 EDGE_PRED (block
, 1)->src
);
2980 /* Induction variables only have one edge inside the loop. */
2981 if ((firstinsideloop
^ secondinsideloop
)
2982 && expr
->kind
!= REFERENCE
)
2984 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2985 fprintf (dump_file
, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
2990 /* Make the necessary insertions. */
2991 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
2993 gimple_seq stmts
= NULL
;
2996 eprime
= avail
[pred
->dest_idx
];
2997 builtexpr
= create_expression_by_pieces (bprime
, eprime
,
2999 gcc_assert (!(pred
->flags
& EDGE_ABNORMAL
));
3000 if (!gimple_seq_empty_p (stmts
))
3002 gsi_insert_seq_on_edge (pred
, stmts
);
3007 /* We cannot insert a PHI node if we failed to insert
3012 if (is_gimple_min_invariant (builtexpr
))
3013 avail
[pred
->dest_idx
] = get_or_alloc_expr_for_constant (builtexpr
);
3015 avail
[pred
->dest_idx
] = get_or_alloc_expr_for_name (builtexpr
);
3017 /* If we didn't want a phi node, and we made insertions, we still have
3018 inserted new stuff, and thus return true. If we didn't want a phi node,
3019 and didn't make insertions, we haven't added anything new, so return
3021 if (nophi
&& insertions
)
3023 else if (nophi
&& !insertions
)
3026 /* Now build a phi for the new variable. */
3027 temp
= make_temp_ssa_name (type
, NULL
, "prephitmp");
3028 phi
= create_phi_node (temp
, block
);
3030 gimple_set_plf (phi
, NECESSARY
, false);
3031 VN_INFO_GET (temp
)->value_id
= val
;
3032 VN_INFO (temp
)->valnum
= sccvn_valnum_from_value_id (val
);
3033 if (VN_INFO (temp
)->valnum
== NULL_TREE
)
3034 VN_INFO (temp
)->valnum
= temp
;
3035 bitmap_set_bit (inserted_exprs
, SSA_NAME_VERSION (temp
));
3036 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3038 pre_expr ae
= avail
[pred
->dest_idx
];
3039 gcc_assert (get_expr_type (ae
) == type
3040 || useless_type_conversion_p (type
, get_expr_type (ae
)));
3041 if (ae
->kind
== CONSTANT
)
3042 add_phi_arg (phi
, unshare_expr (PRE_EXPR_CONSTANT (ae
)),
3043 pred
, UNKNOWN_LOCATION
);
3045 add_phi_arg (phi
, PRE_EXPR_NAME (ae
), pred
, UNKNOWN_LOCATION
);
3048 newphi
= get_or_alloc_expr_for_name (temp
);
3049 add_to_value (val
, newphi
);
3051 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3052 this insertion, since we test for the existence of this value in PHI_GEN
3053 before proceeding with the partial redundancy checks in insert_aux.
3055 The value may exist in AVAIL_OUT, in particular, it could be represented
3056 by the expression we are trying to eliminate, in which case we want the
3057 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3060 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3061 this block, because if it did, it would have existed in our dominator's
3062 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3065 bitmap_insert_into_set (PHI_GEN (block
), newphi
);
3066 bitmap_value_replace_in_set (AVAIL_OUT (block
),
3068 bitmap_insert_into_set (NEW_SETS (block
),
3071 /* If we insert a PHI node for a conversion of another PHI node
3072 in the same basic-block try to preserve range information.
3073 This is important so that followup loop passes receive optimal
3074 number of iteration analysis results. See PR61743. */
3075 if (expr
->kind
== NARY
3076 && CONVERT_EXPR_CODE_P (expr
->u
.nary
->opcode
)
3077 && TREE_CODE (expr
->u
.nary
->op
[0]) == SSA_NAME
3078 && gimple_bb (SSA_NAME_DEF_STMT (expr
->u
.nary
->op
[0])) == block
3079 && INTEGRAL_TYPE_P (type
)
3080 && INTEGRAL_TYPE_P (TREE_TYPE (expr
->u
.nary
->op
[0]))
3081 && (TYPE_PRECISION (type
)
3082 >= TYPE_PRECISION (TREE_TYPE (expr
->u
.nary
->op
[0])))
3083 && SSA_NAME_RANGE_INFO (expr
->u
.nary
->op
[0]))
3086 if (get_range_info (expr
->u
.nary
->op
[0], &min
, &max
) == VR_RANGE
3087 && !wi::neg_p (min
, SIGNED
)
3088 && !wi::neg_p (max
, SIGNED
))
3089 /* Just handle extension and sign-changes of all-positive ranges. */
3090 set_range_info (temp
,
3091 SSA_NAME_RANGE_TYPE (expr
->u
.nary
->op
[0]),
3092 wide_int_storage::from (min
, TYPE_PRECISION (type
),
3094 wide_int_storage::from (max
, TYPE_PRECISION (type
),
3098 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3100 fprintf (dump_file
, "Created phi ");
3101 print_gimple_stmt (dump_file
, phi
, 0, 0);
3102 fprintf (dump_file
, " in block %d (%04d)\n", block
->index
, val
);
3110 /* Perform insertion of partially redundant or hoistable values.
3111 For BLOCK, do the following:
3112 1. Propagate the NEW_SETS of the dominator into the current block.
3113 If the block has multiple predecessors,
3114 2a. Iterate over the ANTIC expressions for the block to see if
3115 any of them are partially redundant.
3116 2b. If so, insert them into the necessary predecessors to make
3117 the expression fully redundant.
3118 2c. Insert a new PHI merging the values of the predecessors.
3119 2d. Insert the new PHI, and the new expressions, into the
3121 If the block has multiple successors,
3122 3a. Iterate over the ANTIC values for the block to see if
3123 any of them are good candidates for hoisting.
3124 3b. If so, insert expressions computing the values in BLOCK,
3125 and add the new expressions into the NEW_SETS set.
3126 4. Recursively call ourselves on the dominator children of BLOCK.
3128 Steps 1, 2a, and 4 are done by insert_aux. 2b, 2c and 2d are done by
3129 do_pre_regular_insertion and do_partial_insertion. 3a and 3b are
3130 done in do_hoist_insertion.
3134 do_pre_regular_insertion (basic_block block
, basic_block dom
)
3136 bool new_stuff
= false;
3137 vec
<pre_expr
> exprs
;
3139 auto_vec
<pre_expr
> avail
;
3142 exprs
= sorted_array_from_bitmap_set (ANTIC_IN (block
));
3143 avail
.safe_grow (EDGE_COUNT (block
->preds
));
3145 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
3147 if (expr
->kind
== NARY
3148 || expr
->kind
== REFERENCE
)
3151 bool by_some
= false;
3152 bool cant_insert
= false;
3153 bool all_same
= true;
3154 pre_expr first_s
= NULL
;
3157 pre_expr eprime
= NULL
;
3159 pre_expr edoubleprime
= NULL
;
3160 bool do_insertion
= false;
3162 val
= get_expr_value_id (expr
);
3163 if (bitmap_set_contains_value (PHI_GEN (block
), val
))
3165 if (bitmap_set_contains_value (AVAIL_OUT (dom
), val
))
3167 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3169 fprintf (dump_file
, "Found fully redundant value: ");
3170 print_pre_expr (dump_file
, expr
);
3171 fprintf (dump_file
, "\n");
3176 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3178 unsigned int vprime
;
3180 /* We should never run insertion for the exit block
3181 and so not come across fake pred edges. */
3182 gcc_assert (!(pred
->flags
& EDGE_FAKE
));
3184 /* We are looking at ANTIC_OUT of bprime. */
3185 eprime
= phi_translate (expr
, ANTIC_IN (block
), NULL
,
3188 /* eprime will generally only be NULL if the
3189 value of the expression, translated
3190 through the PHI for this predecessor, is
3191 undefined. If that is the case, we can't
3192 make the expression fully redundant,
3193 because its value is undefined along a
3194 predecessor path. We can thus break out
3195 early because it doesn't matter what the
3196 rest of the results are. */
3199 avail
[pred
->dest_idx
] = NULL
;
3204 vprime
= get_expr_value_id (eprime
);
3205 edoubleprime
= bitmap_find_leader (AVAIL_OUT (bprime
),
3207 if (edoubleprime
== NULL
)
3209 avail
[pred
->dest_idx
] = eprime
;
3214 avail
[pred
->dest_idx
] = edoubleprime
;
3216 /* We want to perform insertions to remove a redundancy on
3217 a path in the CFG we want to optimize for speed. */
3218 if (optimize_edge_for_speed_p (pred
))
3219 do_insertion
= true;
3220 if (first_s
== NULL
)
3221 first_s
= edoubleprime
;
3222 else if (!pre_expr_d::equal (first_s
, edoubleprime
))
3226 /* If we can insert it, it's not the same value
3227 already existing along every predecessor, and
3228 it's defined by some predecessor, it is
3229 partially redundant. */
3230 if (!cant_insert
&& !all_same
&& by_some
)
3234 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3236 fprintf (dump_file
, "Skipping partial redundancy for "
3238 print_pre_expr (dump_file
, expr
);
3239 fprintf (dump_file
, " (%04d), no redundancy on to be "
3240 "optimized for speed edge\n", val
);
3243 else if (dbg_cnt (treepre_insert
))
3245 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3247 fprintf (dump_file
, "Found partial redundancy for "
3249 print_pre_expr (dump_file
, expr
);
3250 fprintf (dump_file
, " (%04d)\n",
3251 get_expr_value_id (expr
));
3253 if (insert_into_preds_of_block (block
,
3254 get_expression_id (expr
),
3259 /* If all edges produce the same value and that value is
3260 an invariant, then the PHI has the same value on all
3261 edges. Note this. */
3262 else if (!cant_insert
&& all_same
)
3264 gcc_assert (edoubleprime
->kind
== CONSTANT
3265 || edoubleprime
->kind
== NAME
);
3267 tree temp
= make_temp_ssa_name (get_expr_type (expr
),
3270 = gimple_build_assign (temp
,
3271 edoubleprime
->kind
== CONSTANT
?
3272 PRE_EXPR_CONSTANT (edoubleprime
) :
3273 PRE_EXPR_NAME (edoubleprime
));
3274 gimple_stmt_iterator gsi
= gsi_after_labels (block
);
3275 gsi_insert_before (&gsi
, assign
, GSI_NEW_STMT
);
3277 gimple_set_plf (assign
, NECESSARY
, false);
3278 VN_INFO_GET (temp
)->value_id
= val
;
3279 VN_INFO (temp
)->valnum
= sccvn_valnum_from_value_id (val
);
3280 if (VN_INFO (temp
)->valnum
== NULL_TREE
)
3281 VN_INFO (temp
)->valnum
= temp
;
3282 bitmap_set_bit (inserted_exprs
, SSA_NAME_VERSION (temp
));
3283 pre_expr newe
= get_or_alloc_expr_for_name (temp
);
3284 add_to_value (val
, newe
);
3285 bitmap_value_replace_in_set (AVAIL_OUT (block
), newe
);
3286 bitmap_insert_into_set (NEW_SETS (block
), newe
);
3296 /* Perform insertion for partially anticipatable expressions. There
3297 is only one case we will perform insertion for these. This case is
3298 if the expression is partially anticipatable, and fully available.
3299 In this case, we know that putting it earlier will enable us to
3300 remove the later computation. */
3303 do_pre_partial_partial_insertion (basic_block block
, basic_block dom
)
3305 bool new_stuff
= false;
3306 vec
<pre_expr
> exprs
;
3308 auto_vec
<pre_expr
> avail
;
3311 exprs
= sorted_array_from_bitmap_set (PA_IN (block
));
3312 avail
.safe_grow (EDGE_COUNT (block
->preds
));
3314 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
3316 if (expr
->kind
== NARY
3317 || expr
->kind
== REFERENCE
)
3321 bool cant_insert
= false;
3324 pre_expr eprime
= NULL
;
3327 val
= get_expr_value_id (expr
);
3328 if (bitmap_set_contains_value (PHI_GEN (block
), val
))
3330 if (bitmap_set_contains_value (AVAIL_OUT (dom
), val
))
3333 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3335 unsigned int vprime
;
3336 pre_expr edoubleprime
;
3338 /* We should never run insertion for the exit block
3339 and so not come across fake pred edges. */
3340 gcc_assert (!(pred
->flags
& EDGE_FAKE
));
3342 eprime
= phi_translate (expr
, ANTIC_IN (block
),
3346 /* eprime will generally only be NULL if the
3347 value of the expression, translated
3348 through the PHI for this predecessor, is
3349 undefined. If that is the case, we can't
3350 make the expression fully redundant,
3351 because its value is undefined along a
3352 predecessor path. We can thus break out
3353 early because it doesn't matter what the
3354 rest of the results are. */
3357 avail
[pred
->dest_idx
] = NULL
;
3362 vprime
= get_expr_value_id (eprime
);
3363 edoubleprime
= bitmap_find_leader (AVAIL_OUT (bprime
), vprime
);
3364 avail
[pred
->dest_idx
] = edoubleprime
;
3365 if (edoubleprime
== NULL
)
3372 /* If we can insert it, it's not the same value
3373 already existing along every predecessor, and
3374 it's defined by some predecessor, it is
3375 partially redundant. */
3376 if (!cant_insert
&& by_all
)
3379 bool do_insertion
= false;
3381 /* Insert only if we can remove a later expression on a path
3382 that we want to optimize for speed.
3383 The phi node that we will be inserting in BLOCK is not free,
3384 and inserting it for the sake of !optimize_for_speed successor
3385 may cause regressions on the speed path. */
3386 FOR_EACH_EDGE (succ
, ei
, block
->succs
)
3388 if (bitmap_set_contains_value (PA_IN (succ
->dest
), val
)
3389 || bitmap_set_contains_value (ANTIC_IN (succ
->dest
), val
))
3391 if (optimize_edge_for_speed_p (succ
))
3392 do_insertion
= true;
3398 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3400 fprintf (dump_file
, "Skipping partial partial redundancy "
3402 print_pre_expr (dump_file
, expr
);
3403 fprintf (dump_file
, " (%04d), not (partially) anticipated "
3404 "on any to be optimized for speed edges\n", val
);
3407 else if (dbg_cnt (treepre_insert
))
3409 pre_stats
.pa_insert
++;
3410 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3412 fprintf (dump_file
, "Found partial partial redundancy "
3414 print_pre_expr (dump_file
, expr
);
3415 fprintf (dump_file
, " (%04d)\n",
3416 get_expr_value_id (expr
));
3418 if (insert_into_preds_of_block (block
,
3419 get_expression_id (expr
),
3431 /* Insert expressions in BLOCK to compute hoistable values up.
3432 Return TRUE if something was inserted, otherwise return FALSE.
3433 The caller has to make sure that BLOCK has at least two successors. */
3436 do_hoist_insertion (basic_block block
)
3440 bool new_stuff
= false;
3442 gimple_stmt_iterator last
;
3444 /* At least two successors, or else... */
3445 gcc_assert (EDGE_COUNT (block
->succs
) >= 2);
3447 /* Check that all successors of BLOCK are dominated by block.
3448 We could use dominated_by_p() for this, but actually there is a much
3449 quicker check: any successor that is dominated by BLOCK can't have
3450 more than one predecessor edge. */
3451 FOR_EACH_EDGE (e
, ei
, block
->succs
)
3452 if (! single_pred_p (e
->dest
))
3455 /* Determine the insertion point. If we cannot safely insert before
3456 the last stmt if we'd have to, bail out. */
3457 last
= gsi_last_bb (block
);
3458 if (!gsi_end_p (last
)
3459 && !is_ctrl_stmt (gsi_stmt (last
))
3460 && stmt_ends_bb_p (gsi_stmt (last
)))
3463 /* Compute the set of hoistable expressions from ANTIC_IN. First compute
3464 hoistable values. */
3465 bitmap_set hoistable_set
;
3467 /* A hoistable value must be in ANTIC_IN(block)
3468 but not in AVAIL_OUT(BLOCK). */
3469 bitmap_initialize (&hoistable_set
.values
, &grand_bitmap_obstack
);
3470 bitmap_and_compl (&hoistable_set
.values
,
3471 &ANTIC_IN (block
)->values
, &AVAIL_OUT (block
)->values
);
3473 /* Short-cut for a common case: hoistable_set is empty. */
3474 if (bitmap_empty_p (&hoistable_set
.values
))
3477 /* Compute which of the hoistable values is in AVAIL_OUT of
3478 at least one of the successors of BLOCK. */
3479 bitmap_head availout_in_some
;
3480 bitmap_initialize (&availout_in_some
, &grand_bitmap_obstack
);
3481 FOR_EACH_EDGE (e
, ei
, block
->succs
)
3482 /* Do not consider expressions solely because their availability
3483 on loop exits. They'd be ANTIC-IN throughout the whole loop
3484 and thus effectively hoisted across loops by combination of
3485 PRE and hoisting. */
3486 if (! loop_exit_edge_p (block
->loop_father
, e
))
3487 bitmap_ior_and_into (&availout_in_some
, &hoistable_set
.values
,
3488 &AVAIL_OUT (e
->dest
)->values
);
3489 bitmap_clear (&hoistable_set
.values
);
3491 /* Short-cut for a common case: availout_in_some is empty. */
3492 if (bitmap_empty_p (&availout_in_some
))
3495 /* Hack hoitable_set in-place so we can use sorted_array_from_bitmap_set. */
3496 hoistable_set
.values
= availout_in_some
;
3497 hoistable_set
.expressions
= ANTIC_IN (block
)->expressions
;
3499 /* Now finally construct the topological-ordered expression set. */
3500 vec
<pre_expr
> exprs
= sorted_array_from_bitmap_set (&hoistable_set
);
3502 bitmap_clear (&hoistable_set
.values
);
3504 /* If there are candidate values for hoisting, insert expressions
3505 strategically to make the hoistable expressions fully redundant. */
3507 FOR_EACH_VEC_ELT (exprs
, i
, expr
)
3509 /* While we try to sort expressions topologically above the
3510 sorting doesn't work out perfectly. Catch expressions we
3511 already inserted. */
3512 unsigned int value_id
= get_expr_value_id (expr
);
3513 if (bitmap_set_contains_value (AVAIL_OUT (block
), value_id
))
3515 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3518 "Already inserted expression for ");
3519 print_pre_expr (dump_file
, expr
);
3520 fprintf (dump_file
, " (%04d)\n", value_id
);
3525 /* OK, we should hoist this value. Perform the transformation. */
3526 pre_stats
.hoist_insert
++;
3527 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3530 "Inserting expression in block %d for code hoisting: ",
3532 print_pre_expr (dump_file
, expr
);
3533 fprintf (dump_file
, " (%04d)\n", value_id
);
3536 gimple_seq stmts
= NULL
;
3537 tree res
= create_expression_by_pieces (block
, expr
, &stmts
,
3538 get_expr_type (expr
));
3540 /* Do not return true if expression creation ultimately
3541 did not insert any statements. */
3542 if (gimple_seq_empty_p (stmts
))
3546 if (gsi_end_p (last
) || is_ctrl_stmt (gsi_stmt (last
)))
3547 gsi_insert_seq_before (&last
, stmts
, GSI_SAME_STMT
);
3549 gsi_insert_seq_after (&last
, stmts
, GSI_NEW_STMT
);
3552 /* Make sure to not return true if expression creation ultimately
3553 failed but also make sure to insert any stmts produced as they
3554 are tracked in inserted_exprs. */
3566 /* Do a dominator walk on the control flow graph, and insert computations
3567 of values as necessary for PRE and hoisting. */
3570 insert_aux (basic_block block
, bool do_pre
, bool do_hoist
)
3573 bool new_stuff
= false;
3578 dom
= get_immediate_dominator (CDI_DOMINATORS
, block
);
3583 bitmap_set_t newset
;
3585 /* First, update the AVAIL_OUT set with anything we may have
3586 inserted higher up in the dominator tree. */
3587 newset
= NEW_SETS (dom
);
3590 /* Note that we need to value_replace both NEW_SETS, and
3591 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3592 represented by some non-simple expression here that we want
3593 to replace it with. */
3594 FOR_EACH_EXPR_ID_IN_SET (newset
, i
, bi
)
3596 pre_expr expr
= expression_for_id (i
);
3597 bitmap_value_replace_in_set (NEW_SETS (block
), expr
);
3598 bitmap_value_replace_in_set (AVAIL_OUT (block
), expr
);
3602 /* Insert expressions for partial redundancies. */
3603 if (do_pre
&& !single_pred_p (block
))
3605 new_stuff
|= do_pre_regular_insertion (block
, dom
);
3606 if (do_partial_partial
)
3607 new_stuff
|= do_pre_partial_partial_insertion (block
, dom
);
3610 /* Insert expressions for hoisting. */
3611 if (do_hoist
&& EDGE_COUNT (block
->succs
) >= 2)
3612 new_stuff
|= do_hoist_insertion (block
);
3615 for (son
= first_dom_son (CDI_DOMINATORS
, block
);
3617 son
= next_dom_son (CDI_DOMINATORS
, son
))
3619 new_stuff
|= insert_aux (son
, do_pre
, do_hoist
);
3625 /* Perform insertion of partially redundant and hoistable values. */
3630 bool new_stuff
= true;
3632 int num_iterations
= 0;
3634 FOR_ALL_BB_FN (bb
, cfun
)
3635 NEW_SETS (bb
) = bitmap_set_new ();
3640 if (dump_file
&& dump_flags
& TDF_DETAILS
)
3641 fprintf (dump_file
, "Starting insert iteration %d\n", num_iterations
);
3642 new_stuff
= insert_aux (ENTRY_BLOCK_PTR_FOR_FN (cfun
), flag_tree_pre
,
3643 flag_code_hoisting
);
3645 /* Clear the NEW sets before the next iteration. We have already
3646 fully propagated its contents. */
3648 FOR_ALL_BB_FN (bb
, cfun
)
3649 bitmap_set_free (NEW_SETS (bb
));
3651 statistics_histogram_event (cfun
, "insert iterations", num_iterations
);
3655 /* Compute the AVAIL set for all basic blocks.
3657 This function performs value numbering of the statements in each basic
3658 block. The AVAIL sets are built from information we glean while doing
3659 this value numbering, since the AVAIL sets contain only one entry per
3662 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3663 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3666 compute_avail (void)
3669 basic_block block
, son
;
3670 basic_block
*worklist
;
3674 /* We pretend that default definitions are defined in the entry block.
3675 This includes function arguments and the static chain decl. */
3676 for (i
= 1; i
< num_ssa_names
; ++i
)
3678 tree name
= ssa_name (i
);
3681 || !SSA_NAME_IS_DEFAULT_DEF (name
)
3682 || has_zero_uses (name
)
3683 || virtual_operand_p (name
))
3686 e
= get_or_alloc_expr_for_name (name
);
3687 add_to_value (get_expr_value_id (e
), e
);
3688 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun
)), e
);
3689 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun
)),
3693 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3695 print_bitmap_set (dump_file
, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun
)),
3696 "tmp_gen", ENTRY_BLOCK
);
3697 print_bitmap_set (dump_file
, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun
)),
3698 "avail_out", ENTRY_BLOCK
);
3701 /* Allocate the worklist. */
3702 worklist
= XNEWVEC (basic_block
, n_basic_blocks_for_fn (cfun
));
3704 /* Seed the algorithm by putting the dominator children of the entry
3705 block on the worklist. */
3706 for (son
= first_dom_son (CDI_DOMINATORS
, ENTRY_BLOCK_PTR_FOR_FN (cfun
));
3708 son
= next_dom_son (CDI_DOMINATORS
, son
))
3709 worklist
[sp
++] = son
;
3711 BB_LIVE_VOP_ON_EXIT (ENTRY_BLOCK_PTR_FOR_FN (cfun
))
3712 = ssa_default_def (cfun
, gimple_vop (cfun
));
3714 /* Loop until the worklist is empty. */
3720 /* Pick a block from the worklist. */
3721 block
= worklist
[--sp
];
3723 /* Initially, the set of available values in BLOCK is that of
3724 its immediate dominator. */
3725 dom
= get_immediate_dominator (CDI_DOMINATORS
, block
);
3728 bitmap_set_copy (AVAIL_OUT (block
), AVAIL_OUT (dom
));
3729 BB_LIVE_VOP_ON_EXIT (block
) = BB_LIVE_VOP_ON_EXIT (dom
);
3732 /* Generate values for PHI nodes. */
3733 for (gphi_iterator gsi
= gsi_start_phis (block
); !gsi_end_p (gsi
);
3736 tree result
= gimple_phi_result (gsi
.phi ());
3738 /* We have no need for virtual phis, as they don't represent
3739 actual computations. */
3740 if (virtual_operand_p (result
))
3742 BB_LIVE_VOP_ON_EXIT (block
) = result
;
3746 pre_expr e
= get_or_alloc_expr_for_name (result
);
3747 add_to_value (get_expr_value_id (e
), e
);
3748 bitmap_value_insert_into_set (AVAIL_OUT (block
), e
);
3749 bitmap_insert_into_set (PHI_GEN (block
), e
);
3752 BB_MAY_NOTRETURN (block
) = 0;
3754 /* Now compute value numbers and populate value sets with all
3755 the expressions computed in BLOCK. */
3756 for (gimple_stmt_iterator gsi
= gsi_start_bb (block
); !gsi_end_p (gsi
);
3762 stmt
= gsi_stmt (gsi
);
3764 /* Cache whether the basic-block has any non-visible side-effect
3766 If this isn't a call or it is the last stmt in the
3767 basic-block then the CFG represents things correctly. */
3768 if (is_gimple_call (stmt
) && !stmt_ends_bb_p (stmt
))
3770 /* Non-looping const functions always return normally.
3771 Otherwise the call might not return or have side-effects
3772 that forbids hoisting possibly trapping expressions
3774 int flags
= gimple_call_flags (stmt
);
3775 if (!(flags
& ECF_CONST
)
3776 || (flags
& ECF_LOOPING_CONST_OR_PURE
))
3777 BB_MAY_NOTRETURN (block
) = 1;
3780 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_DEF
)
3782 pre_expr e
= get_or_alloc_expr_for_name (op
);
3784 add_to_value (get_expr_value_id (e
), e
);
3785 bitmap_insert_into_set (TMP_GEN (block
), e
);
3786 bitmap_value_insert_into_set (AVAIL_OUT (block
), e
);
3789 if (gimple_vdef (stmt
))
3790 BB_LIVE_VOP_ON_EXIT (block
) = gimple_vdef (stmt
);
3792 if (gimple_has_side_effects (stmt
)
3793 || stmt_could_throw_p (stmt
)
3794 || is_gimple_debug (stmt
))
3797 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_USE
)
3799 if (ssa_undefined_value_p (op
))
3801 pre_expr e
= get_or_alloc_expr_for_name (op
);
3802 bitmap_value_insert_into_set (EXP_GEN (block
), e
);
3805 switch (gimple_code (stmt
))
3813 vn_reference_s ref1
;
3814 pre_expr result
= NULL
;
3816 /* We can value number only calls to real functions. */
3817 if (gimple_call_internal_p (stmt
))
3820 vn_reference_lookup_call (as_a
<gcall
*> (stmt
), &ref
, &ref1
);
3824 /* If the value of the call is not invalidated in
3825 this block until it is computed, add the expression
3827 if (!gimple_vuse (stmt
)
3829 (SSA_NAME_DEF_STMT (gimple_vuse (stmt
))) == GIMPLE_PHI
3830 || gimple_bb (SSA_NAME_DEF_STMT
3831 (gimple_vuse (stmt
))) != block
)
3833 result
= pre_expr_pool
.allocate ();
3834 result
->kind
= REFERENCE
;
3836 PRE_EXPR_REFERENCE (result
) = ref
;
3838 get_or_alloc_expression_id (result
);
3839 add_to_value (get_expr_value_id (result
), result
);
3840 bitmap_value_insert_into_set (EXP_GEN (block
), result
);
3847 pre_expr result
= NULL
;
3848 switch (vn_get_stmt_kind (stmt
))
3852 enum tree_code code
= gimple_assign_rhs_code (stmt
);
3855 /* COND_EXPR and VEC_COND_EXPR are awkward in
3856 that they contain an embedded complex expression.
3857 Don't even try to shove those through PRE. */
3858 if (code
== COND_EXPR
3859 || code
== VEC_COND_EXPR
)
3862 vn_nary_op_lookup_stmt (stmt
, &nary
);
3866 /* If the NARY traps and there was a preceding
3867 point in the block that might not return avoid
3868 adding the nary to EXP_GEN. */
3869 if (BB_MAY_NOTRETURN (block
)
3870 && vn_nary_may_trap (nary
))
3873 result
= pre_expr_pool
.allocate ();
3874 result
->kind
= NARY
;
3876 PRE_EXPR_NARY (result
) = nary
;
3882 tree rhs1
= gimple_assign_rhs1 (stmt
);
3883 alias_set_type set
= get_alias_set (rhs1
);
3884 vec
<vn_reference_op_s
> operands
3885 = vn_reference_operands_for_lookup (rhs1
);
3887 vn_reference_lookup_pieces (gimple_vuse (stmt
), set
,
3889 operands
, &ref
, VN_WALK
);
3892 operands
.release ();
3896 /* If the value of the reference is not invalidated in
3897 this block until it is computed, add the expression
3899 if (gimple_vuse (stmt
))
3903 def_stmt
= SSA_NAME_DEF_STMT (gimple_vuse (stmt
));
3904 while (!gimple_nop_p (def_stmt
)
3905 && gimple_code (def_stmt
) != GIMPLE_PHI
3906 && gimple_bb (def_stmt
) == block
)
3908 if (stmt_may_clobber_ref_p
3909 (def_stmt
, gimple_assign_rhs1 (stmt
)))
3915 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt
));
3919 operands
.release ();
3924 /* If the load was value-numbered to another
3925 load make sure we do not use its expression
3926 for insertion if it wouldn't be a valid
3928 /* At the momemt we have a testcase
3929 for hoist insertion of aligned vs. misaligned
3930 variants in gcc.dg/torture/pr65270-1.c thus
3931 with just alignment to be considered we can
3932 simply replace the expression in the hashtable
3933 with the most conservative one. */
3934 vn_reference_op_t ref1
= &ref
->operands
.last ();
3935 while (ref1
->opcode
!= TARGET_MEM_REF
3936 && ref1
->opcode
!= MEM_REF
3937 && ref1
!= &ref
->operands
[0])
3939 vn_reference_op_t ref2
= &operands
.last ();
3940 while (ref2
->opcode
!= TARGET_MEM_REF
3941 && ref2
->opcode
!= MEM_REF
3942 && ref2
!= &operands
[0])
3944 if ((ref1
->opcode
== TARGET_MEM_REF
3945 || ref1
->opcode
== MEM_REF
)
3946 && (TYPE_ALIGN (ref1
->type
)
3947 > TYPE_ALIGN (ref2
->type
)))
3949 = build_aligned_type (ref1
->type
,
3950 TYPE_ALIGN (ref2
->type
));
3951 /* TBAA behavior is an obvious part so make sure
3952 that the hashtable one covers this as well
3953 by adjusting the ref alias set and its base. */
3955 || alias_set_subset_of (set
, ref
->set
))
3957 else if (alias_set_subset_of (ref
->set
, set
))
3960 if (ref1
->opcode
== MEM_REF
)
3961 ref1
->op0
= fold_convert (TREE_TYPE (ref2
->op0
),
3964 ref1
->op2
= fold_convert (TREE_TYPE (ref2
->op2
),
3970 if (ref1
->opcode
== MEM_REF
)
3971 ref1
->op0
= fold_convert (ptr_type_node
,
3974 ref1
->op2
= fold_convert (ptr_type_node
,
3977 operands
.release ();
3979 result
= pre_expr_pool
.allocate ();
3980 result
->kind
= REFERENCE
;
3982 PRE_EXPR_REFERENCE (result
) = ref
;
3990 get_or_alloc_expression_id (result
);
3991 add_to_value (get_expr_value_id (result
), result
);
3992 bitmap_value_insert_into_set (EXP_GEN (block
), result
);
4000 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4002 print_bitmap_set (dump_file
, EXP_GEN (block
),
4003 "exp_gen", block
->index
);
4004 print_bitmap_set (dump_file
, PHI_GEN (block
),
4005 "phi_gen", block
->index
);
4006 print_bitmap_set (dump_file
, TMP_GEN (block
),
4007 "tmp_gen", block
->index
);
4008 print_bitmap_set (dump_file
, AVAIL_OUT (block
),
4009 "avail_out", block
->index
);
4012 /* Put the dominator children of BLOCK on the worklist of blocks
4013 to compute available sets for. */
4014 for (son
= first_dom_son (CDI_DOMINATORS
, block
);
4016 son
= next_dom_son (CDI_DOMINATORS
, son
))
4017 worklist
[sp
++] = son
;
4024 /* Local state for the eliminate domwalk. */
4025 static vec
<gimple
*> el_to_remove
;
4026 static vec
<gimple
*> el_to_fixup
;
4027 static unsigned int el_todo
;
4028 static vec
<tree
> el_avail
;
4029 static vec
<tree
> el_avail_stack
;
4031 /* Return a leader for OP that is available at the current point of the
4032 eliminate domwalk. */
4035 eliminate_avail (tree op
)
4037 tree valnum
= VN_INFO (op
)->valnum
;
4038 if (TREE_CODE (valnum
) == SSA_NAME
)
4040 if (SSA_NAME_IS_DEFAULT_DEF (valnum
))
4042 if (el_avail
.length () > SSA_NAME_VERSION (valnum
))
4043 return el_avail
[SSA_NAME_VERSION (valnum
)];
4045 else if (is_gimple_min_invariant (valnum
))
4050 /* At the current point of the eliminate domwalk make OP available. */
4053 eliminate_push_avail (tree op
)
4055 tree valnum
= VN_INFO (op
)->valnum
;
4056 if (TREE_CODE (valnum
) == SSA_NAME
)
4058 if (el_avail
.length () <= SSA_NAME_VERSION (valnum
))
4059 el_avail
.safe_grow_cleared (SSA_NAME_VERSION (valnum
) + 1);
4061 if (el_avail
[SSA_NAME_VERSION (valnum
)])
4062 pushop
= el_avail
[SSA_NAME_VERSION (valnum
)];
4063 el_avail_stack
.safe_push (pushop
);
4064 el_avail
[SSA_NAME_VERSION (valnum
)] = op
;
4068 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
4069 the leader for the expression if insertion was successful. */
4072 eliminate_insert (gimple_stmt_iterator
*gsi
, tree val
)
4074 gimple
*stmt
= gimple_seq_first_stmt (VN_INFO (val
)->expr
);
4075 if (!is_gimple_assign (stmt
)
4076 || (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt
))
4077 && gimple_assign_rhs_code (stmt
) != VIEW_CONVERT_EXPR
4078 && gimple_assign_rhs_code (stmt
) != BIT_FIELD_REF
))
4081 tree op
= gimple_assign_rhs1 (stmt
);
4082 if (gimple_assign_rhs_code (stmt
) == VIEW_CONVERT_EXPR
4083 || gimple_assign_rhs_code (stmt
) == BIT_FIELD_REF
)
4084 op
= TREE_OPERAND (op
, 0);
4085 tree leader
= TREE_CODE (op
) == SSA_NAME
? eliminate_avail (op
) : op
;
4089 gimple_seq stmts
= NULL
;
4091 if (gimple_assign_rhs_code (stmt
) == BIT_FIELD_REF
)
4092 res
= gimple_build (&stmts
, BIT_FIELD_REF
,
4093 TREE_TYPE (val
), leader
,
4094 TREE_OPERAND (gimple_assign_rhs1 (stmt
), 1),
4095 TREE_OPERAND (gimple_assign_rhs1 (stmt
), 2));
4097 res
= gimple_build (&stmts
, gimple_assign_rhs_code (stmt
),
4098 TREE_TYPE (val
), leader
);
4099 gsi_insert_seq_before (gsi
, stmts
, GSI_SAME_STMT
);
4100 VN_INFO_GET (res
)->valnum
= val
;
4102 if (TREE_CODE (leader
) == SSA_NAME
)
4103 gimple_set_plf (SSA_NAME_DEF_STMT (leader
), NECESSARY
, true);
4105 pre_stats
.insertions
++;
4106 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4108 fprintf (dump_file
, "Inserted ");
4109 print_gimple_stmt (dump_file
, SSA_NAME_DEF_STMT (res
), 0, 0);
4115 class eliminate_dom_walker
: public dom_walker
4118 eliminate_dom_walker (cdi_direction direction
, bool do_pre_
)
4119 : dom_walker (direction
), do_pre (do_pre_
) {}
4121 virtual edge
before_dom_children (basic_block
);
4122 virtual void after_dom_children (basic_block
);
4127 /* Perform elimination for the basic-block B during the domwalk. */
4130 eliminate_dom_walker::before_dom_children (basic_block b
)
4133 el_avail_stack
.safe_push (NULL_TREE
);
4135 /* ??? If we do nothing for unreachable blocks then this will confuse
4136 tailmerging. Eventually we can reduce its reliance on SCCVN now
4137 that we fully copy/constant-propagate (most) things. */
4139 for (gphi_iterator gsi
= gsi_start_phis (b
); !gsi_end_p (gsi
);)
4141 gphi
*phi
= gsi
.phi ();
4142 tree res
= PHI_RESULT (phi
);
4144 if (virtual_operand_p (res
))
4150 tree sprime
= eliminate_avail (res
);
4154 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4156 fprintf (dump_file
, "Replaced redundant PHI node defining ");
4157 print_generic_expr (dump_file
, res
, 0);
4158 fprintf (dump_file
, " with ");
4159 print_generic_expr (dump_file
, sprime
, 0);
4160 fprintf (dump_file
, "\n");
4163 /* If we inserted this PHI node ourself, it's not an elimination. */
4165 && bitmap_bit_p (inserted_exprs
, SSA_NAME_VERSION (res
)))
4168 pre_stats
.eliminations
++;
4170 /* If we will propagate into all uses don't bother to do
4172 if (may_propagate_copy (res
, sprime
))
4174 /* Mark the PHI for removal. */
4175 el_to_remove
.safe_push (phi
);
4180 remove_phi_node (&gsi
, false);
4183 && !bitmap_bit_p (inserted_exprs
, SSA_NAME_VERSION (res
))
4184 && TREE_CODE (sprime
) == SSA_NAME
)
4185 gimple_set_plf (SSA_NAME_DEF_STMT (sprime
), NECESSARY
, true);
4187 if (!useless_type_conversion_p (TREE_TYPE (res
), TREE_TYPE (sprime
)))
4188 sprime
= fold_convert (TREE_TYPE (res
), sprime
);
4189 gimple
*stmt
= gimple_build_assign (res
, sprime
);
4190 /* ??? It cannot yet be necessary (DOM walk). */
4191 gimple_set_plf (stmt
, NECESSARY
, gimple_plf (phi
, NECESSARY
));
4193 gimple_stmt_iterator gsi2
= gsi_after_labels (b
);
4194 gsi_insert_before (&gsi2
, stmt
, GSI_NEW_STMT
);
4198 eliminate_push_avail (res
);
4202 for (gimple_stmt_iterator gsi
= gsi_start_bb (b
);
4206 tree sprime
= NULL_TREE
;
4207 gimple
*stmt
= gsi_stmt (gsi
);
4208 tree lhs
= gimple_get_lhs (stmt
);
4209 if (lhs
&& TREE_CODE (lhs
) == SSA_NAME
4210 && !gimple_has_volatile_ops (stmt
)
4211 /* See PR43491. Do not replace a global register variable when
4212 it is a the RHS of an assignment. Do replace local register
4213 variables since gcc does not guarantee a local variable will
4214 be allocated in register.
4215 ??? The fix isn't effective here. This should instead
4216 be ensured by not value-numbering them the same but treating
4217 them like volatiles? */
4218 && !(gimple_assign_single_p (stmt
)
4219 && (TREE_CODE (gimple_assign_rhs1 (stmt
)) == VAR_DECL
4220 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt
))
4221 && is_global_var (gimple_assign_rhs1 (stmt
)))))
4223 sprime
= eliminate_avail (lhs
);
4226 /* If there is no existing usable leader but SCCVN thinks
4227 it has an expression it wants to use as replacement,
4229 tree val
= VN_INFO (lhs
)->valnum
;
4231 && TREE_CODE (val
) == SSA_NAME
4232 && VN_INFO (val
)->needs_insertion
4233 && VN_INFO (val
)->expr
!= NULL
4234 && (sprime
= eliminate_insert (&gsi
, val
)) != NULL_TREE
)
4235 eliminate_push_avail (sprime
);
4238 /* If this now constitutes a copy duplicate points-to
4239 and range info appropriately. This is especially
4240 important for inserted code. See tree-ssa-copy.c
4241 for similar code. */
4243 && TREE_CODE (sprime
) == SSA_NAME
)
4245 basic_block sprime_b
= gimple_bb (SSA_NAME_DEF_STMT (sprime
));
4246 if (POINTER_TYPE_P (TREE_TYPE (lhs
))
4247 && VN_INFO_PTR_INFO (lhs
)
4248 && ! VN_INFO_PTR_INFO (sprime
))
4250 duplicate_ssa_name_ptr_info (sprime
,
4251 VN_INFO_PTR_INFO (lhs
));
4253 mark_ptr_info_alignment_unknown
4254 (SSA_NAME_PTR_INFO (sprime
));
4256 else if (INTEGRAL_TYPE_P (TREE_TYPE (lhs
))
4257 && VN_INFO_RANGE_INFO (lhs
)
4258 && ! VN_INFO_RANGE_INFO (sprime
)
4260 duplicate_ssa_name_range_info (sprime
,
4261 VN_INFO_RANGE_TYPE (lhs
),
4262 VN_INFO_RANGE_INFO (lhs
));
4265 /* Inhibit the use of an inserted PHI on a loop header when
4266 the address of the memory reference is a simple induction
4267 variable. In other cases the vectorizer won't do anything
4268 anyway (either it's loop invariant or a complicated
4271 && TREE_CODE (sprime
) == SSA_NAME
4273 && flag_tree_loop_vectorize
4274 && loop_outer (b
->loop_father
)
4275 && has_zero_uses (sprime
)
4276 && bitmap_bit_p (inserted_exprs
, SSA_NAME_VERSION (sprime
))
4277 && gimple_assign_load_p (stmt
))
4279 gimple
*def_stmt
= SSA_NAME_DEF_STMT (sprime
);
4280 basic_block def_bb
= gimple_bb (def_stmt
);
4281 if (gimple_code (def_stmt
) == GIMPLE_PHI
4282 && def_bb
->loop_father
->header
== def_bb
)
4284 loop_p loop
= def_bb
->loop_father
;
4288 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_USE
)
4291 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (op
));
4293 && flow_bb_inside_loop_p (loop
, def_bb
)
4294 && simple_iv (loop
, loop
, op
, &iv
, true))
4302 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4304 fprintf (dump_file
, "Not replacing ");
4305 print_gimple_expr (dump_file
, stmt
, 0, 0);
4306 fprintf (dump_file
, " with ");
4307 print_generic_expr (dump_file
, sprime
, 0);
4308 fprintf (dump_file
, " which would add a loop"
4309 " carried dependence to loop %d\n",
4312 /* Don't keep sprime available. */
4320 /* If we can propagate the value computed for LHS into
4321 all uses don't bother doing anything with this stmt. */
4322 if (may_propagate_copy (lhs
, sprime
))
4324 /* Mark it for removal. */
4325 el_to_remove
.safe_push (stmt
);
4327 /* ??? Don't count copy/constant propagations. */
4328 if (gimple_assign_single_p (stmt
)
4329 && (TREE_CODE (gimple_assign_rhs1 (stmt
)) == SSA_NAME
4330 || gimple_assign_rhs1 (stmt
) == sprime
))
4333 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4335 fprintf (dump_file
, "Replaced ");
4336 print_gimple_expr (dump_file
, stmt
, 0, 0);
4337 fprintf (dump_file
, " with ");
4338 print_generic_expr (dump_file
, sprime
, 0);
4339 fprintf (dump_file
, " in all uses of ");
4340 print_gimple_stmt (dump_file
, stmt
, 0, 0);
4343 pre_stats
.eliminations
++;
4347 /* If this is an assignment from our leader (which
4348 happens in the case the value-number is a constant)
4349 then there is nothing to do. */
4350 if (gimple_assign_single_p (stmt
)
4351 && sprime
== gimple_assign_rhs1 (stmt
))
4354 /* Else replace its RHS. */
4355 bool can_make_abnormal_goto
4356 = is_gimple_call (stmt
)
4357 && stmt_can_make_abnormal_goto (stmt
);
4359 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4361 fprintf (dump_file
, "Replaced ");
4362 print_gimple_expr (dump_file
, stmt
, 0, 0);
4363 fprintf (dump_file
, " with ");
4364 print_generic_expr (dump_file
, sprime
, 0);
4365 fprintf (dump_file
, " in ");
4366 print_gimple_stmt (dump_file
, stmt
, 0, 0);
4369 if (TREE_CODE (sprime
) == SSA_NAME
)
4370 gimple_set_plf (SSA_NAME_DEF_STMT (sprime
),
4373 pre_stats
.eliminations
++;
4374 gimple
*orig_stmt
= stmt
;
4375 if (!useless_type_conversion_p (TREE_TYPE (lhs
),
4376 TREE_TYPE (sprime
)))
4377 sprime
= fold_convert (TREE_TYPE (lhs
), sprime
);
4378 tree vdef
= gimple_vdef (stmt
);
4379 tree vuse
= gimple_vuse (stmt
);
4380 propagate_tree_value_into_stmt (&gsi
, sprime
);
4381 stmt
= gsi_stmt (gsi
);
4383 if (vdef
!= gimple_vdef (stmt
))
4384 VN_INFO (vdef
)->valnum
= vuse
;
4386 /* If we removed EH side-effects from the statement, clean
4387 its EH information. */
4388 if (maybe_clean_or_replace_eh_stmt (orig_stmt
, stmt
))
4390 bitmap_set_bit (need_eh_cleanup
,
4391 gimple_bb (stmt
)->index
);
4392 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4393 fprintf (dump_file
, " Removed EH side-effects.\n");
4396 /* Likewise for AB side-effects. */
4397 if (can_make_abnormal_goto
4398 && !stmt_can_make_abnormal_goto (stmt
))
4400 bitmap_set_bit (need_ab_cleanup
,
4401 gimple_bb (stmt
)->index
);
4402 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4403 fprintf (dump_file
, " Removed AB side-effects.\n");
4410 /* If the statement is a scalar store, see if the expression
4411 has the same value number as its rhs. If so, the store is
4413 if (gimple_assign_single_p (stmt
)
4414 && !gimple_has_volatile_ops (stmt
)
4415 && !is_gimple_reg (gimple_assign_lhs (stmt
))
4416 && (TREE_CODE (gimple_assign_rhs1 (stmt
)) == SSA_NAME
4417 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt
))))
4420 tree rhs
= gimple_assign_rhs1 (stmt
);
4421 val
= vn_reference_lookup (gimple_assign_lhs (stmt
),
4422 gimple_vuse (stmt
), VN_WALK
, NULL
, false);
4423 if (TREE_CODE (rhs
) == SSA_NAME
)
4424 rhs
= VN_INFO (rhs
)->valnum
;
4426 && operand_equal_p (val
, rhs
, 0))
4428 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4430 fprintf (dump_file
, "Deleted redundant store ");
4431 print_gimple_stmt (dump_file
, stmt
, 0, 0);
4434 /* Queue stmt for removal. */
4435 el_to_remove
.safe_push (stmt
);
4440 /* If this is a control statement value numbering left edges
4441 unexecuted on force the condition in a way consistent with
4443 if (gcond
*cond
= dyn_cast
<gcond
*> (stmt
))
4445 if ((EDGE_SUCC (b
, 0)->flags
& EDGE_EXECUTABLE
)
4446 ^ (EDGE_SUCC (b
, 1)->flags
& EDGE_EXECUTABLE
))
4448 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4450 fprintf (dump_file
, "Removing unexecutable edge from ");
4451 print_gimple_stmt (dump_file
, stmt
, 0, 0);
4453 if (((EDGE_SUCC (b
, 0)->flags
& EDGE_TRUE_VALUE
) != 0)
4454 == ((EDGE_SUCC (b
, 0)->flags
& EDGE_EXECUTABLE
) != 0))
4455 gimple_cond_make_true (cond
);
4457 gimple_cond_make_false (cond
);
4459 el_todo
|= TODO_cleanup_cfg
;
4464 bool can_make_abnormal_goto
= stmt_can_make_abnormal_goto (stmt
);
4465 bool was_noreturn
= (is_gimple_call (stmt
)
4466 && gimple_call_noreturn_p (stmt
));
4467 tree vdef
= gimple_vdef (stmt
);
4468 tree vuse
= gimple_vuse (stmt
);
4470 /* If we didn't replace the whole stmt (or propagate the result
4471 into all uses), replace all uses on this stmt with their
4473 use_operand_p use_p
;
4475 FOR_EACH_SSA_USE_OPERAND (use_p
, stmt
, iter
, SSA_OP_USE
)
4477 tree use
= USE_FROM_PTR (use_p
);
4478 /* ??? The call code above leaves stmt operands un-updated. */
4479 if (TREE_CODE (use
) != SSA_NAME
)
4481 tree sprime
= eliminate_avail (use
);
4482 if (sprime
&& sprime
!= use
4483 && may_propagate_copy (use
, sprime
)
4484 /* We substitute into debug stmts to avoid excessive
4485 debug temporaries created by removed stmts, but we need
4486 to avoid doing so for inserted sprimes as we never want
4487 to create debug temporaries for them. */
4489 || TREE_CODE (sprime
) != SSA_NAME
4490 || !is_gimple_debug (stmt
)
4491 || !bitmap_bit_p (inserted_exprs
, SSA_NAME_VERSION (sprime
))))
4493 propagate_value (use_p
, sprime
);
4494 gimple_set_modified (stmt
, true);
4495 if (TREE_CODE (sprime
) == SSA_NAME
4496 && !is_gimple_debug (stmt
))
4497 gimple_set_plf (SSA_NAME_DEF_STMT (sprime
),
4502 /* Visit indirect calls and turn them into direct calls if
4503 possible using the devirtualization machinery. */
4504 if (gcall
*call_stmt
= dyn_cast
<gcall
*> (stmt
))
4506 tree fn
= gimple_call_fn (call_stmt
);
4508 && flag_devirtualize
4509 && virtual_method_call_p (fn
))
4511 tree otr_type
= obj_type_ref_class (fn
);
4513 ipa_polymorphic_call_context
context (current_function_decl
, fn
, stmt
, &instance
);
4516 context
.get_dynamic_type (instance
, OBJ_TYPE_REF_OBJECT (fn
), otr_type
, stmt
);
4518 vec
<cgraph_node
*>targets
4519 = possible_polymorphic_call_targets (obj_type_ref_class (fn
),
4521 (OBJ_TYPE_REF_TOKEN (fn
)),
4525 dump_possible_polymorphic_call_targets (dump_file
,
4526 obj_type_ref_class (fn
),
4528 (OBJ_TYPE_REF_TOKEN (fn
)),
4530 if (final
&& targets
.length () <= 1 && dbg_cnt (devirt
))
4533 if (targets
.length () == 1)
4534 fn
= targets
[0]->decl
;
4536 fn
= builtin_decl_implicit (BUILT_IN_UNREACHABLE
);
4537 if (dump_enabled_p ())
4539 location_t loc
= gimple_location_safe (stmt
);
4540 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, loc
,
4541 "converting indirect call to "
4543 lang_hooks
.decl_printable_name (fn
, 2));
4545 gimple_call_set_fndecl (call_stmt
, fn
);
4546 maybe_remove_unused_call_args (cfun
, call_stmt
);
4547 gimple_set_modified (stmt
, true);
4552 if (gimple_modified_p (stmt
))
4554 /* If a formerly non-invariant ADDR_EXPR is turned into an
4555 invariant one it was on a separate stmt. */
4556 if (gimple_assign_single_p (stmt
)
4557 && TREE_CODE (gimple_assign_rhs1 (stmt
)) == ADDR_EXPR
)
4558 recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt
));
4559 gimple
*old_stmt
= stmt
;
4560 if (is_gimple_call (stmt
))
4562 /* ??? Only fold calls inplace for now, this may create new
4563 SSA names which in turn will confuse free_scc_vn SSA name
4565 fold_stmt_inplace (&gsi
);
4566 /* When changing a call into a noreturn call, cfg cleanup
4567 is needed to fix up the noreturn call. */
4568 if (!was_noreturn
&& gimple_call_noreturn_p (stmt
))
4569 el_to_fixup
.safe_push (stmt
);
4574 stmt
= gsi_stmt (gsi
);
4575 if ((gimple_code (stmt
) == GIMPLE_COND
4576 && (gimple_cond_true_p (as_a
<gcond
*> (stmt
))
4577 || gimple_cond_false_p (as_a
<gcond
*> (stmt
))))
4578 || (gimple_code (stmt
) == GIMPLE_SWITCH
4579 && TREE_CODE (gimple_switch_index (
4580 as_a
<gswitch
*> (stmt
)))
4582 el_todo
|= TODO_cleanup_cfg
;
4584 /* If we removed EH side-effects from the statement, clean
4585 its EH information. */
4586 if (maybe_clean_or_replace_eh_stmt (old_stmt
, stmt
))
4588 bitmap_set_bit (need_eh_cleanup
,
4589 gimple_bb (stmt
)->index
);
4590 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4591 fprintf (dump_file
, " Removed EH side-effects.\n");
4593 /* Likewise for AB side-effects. */
4594 if (can_make_abnormal_goto
4595 && !stmt_can_make_abnormal_goto (stmt
))
4597 bitmap_set_bit (need_ab_cleanup
,
4598 gimple_bb (stmt
)->index
);
4599 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4600 fprintf (dump_file
, " Removed AB side-effects.\n");
4603 if (vdef
!= gimple_vdef (stmt
))
4604 VN_INFO (vdef
)->valnum
= vuse
;
4607 /* Make new values available - for fully redundant LHS we
4608 continue with the next stmt above and skip this. */
4610 FOR_EACH_SSA_DEF_OPERAND (defp
, stmt
, iter
, SSA_OP_DEF
)
4611 eliminate_push_avail (DEF_FROM_PTR (defp
));
4614 /* Replace destination PHI arguments. */
4617 FOR_EACH_EDGE (e
, ei
, b
->succs
)
4619 for (gphi_iterator gsi
= gsi_start_phis (e
->dest
);
4623 gphi
*phi
= gsi
.phi ();
4624 use_operand_p use_p
= PHI_ARG_DEF_PTR_FROM_EDGE (phi
, e
);
4625 tree arg
= USE_FROM_PTR (use_p
);
4626 if (TREE_CODE (arg
) != SSA_NAME
4627 || virtual_operand_p (arg
))
4629 tree sprime
= eliminate_avail (arg
);
4630 if (sprime
&& may_propagate_copy (arg
, sprime
))
4632 propagate_value (use_p
, sprime
);
4633 if (TREE_CODE (sprime
) == SSA_NAME
)
4634 gimple_set_plf (SSA_NAME_DEF_STMT (sprime
), NECESSARY
, true);
4641 /* Make no longer available leaders no longer available. */
4644 eliminate_dom_walker::after_dom_children (basic_block
)
4647 while ((entry
= el_avail_stack
.pop ()) != NULL_TREE
)
4649 tree valnum
= VN_INFO (entry
)->valnum
;
4650 tree old
= el_avail
[SSA_NAME_VERSION (valnum
)];
4652 el_avail
[SSA_NAME_VERSION (valnum
)] = NULL_TREE
;
4654 el_avail
[SSA_NAME_VERSION (valnum
)] = entry
;
4658 /* Eliminate fully redundant computations. */
4661 eliminate (bool do_pre
)
4663 gimple_stmt_iterator gsi
;
4666 need_eh_cleanup
= BITMAP_ALLOC (NULL
);
4667 need_ab_cleanup
= BITMAP_ALLOC (NULL
);
4669 el_to_remove
.create (0);
4670 el_to_fixup
.create (0);
4672 el_avail
.create (num_ssa_names
);
4673 el_avail_stack
.create (0);
4675 eliminate_dom_walker (CDI_DOMINATORS
,
4676 do_pre
).walk (cfun
->cfg
->x_entry_block_ptr
);
4678 el_avail
.release ();
4679 el_avail_stack
.release ();
4681 /* We cannot remove stmts during BB walk, especially not release SSA
4682 names there as this confuses the VN machinery. The stmts ending
4683 up in el_to_remove are either stores or simple copies.
4684 Remove stmts in reverse order to make debug stmt creation possible. */
4685 while (!el_to_remove
.is_empty ())
4687 stmt
= el_to_remove
.pop ();
4689 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4691 fprintf (dump_file
, "Removing dead stmt ");
4692 print_gimple_stmt (dump_file
, stmt
, 0, 0);
4696 if (gimple_code (stmt
) == GIMPLE_PHI
)
4697 lhs
= gimple_phi_result (stmt
);
4699 lhs
= gimple_get_lhs (stmt
);
4702 && TREE_CODE (lhs
) == SSA_NAME
)
4703 bitmap_clear_bit (inserted_exprs
, SSA_NAME_VERSION (lhs
));
4705 gsi
= gsi_for_stmt (stmt
);
4706 if (gimple_code (stmt
) == GIMPLE_PHI
)
4707 remove_phi_node (&gsi
, true);
4710 basic_block bb
= gimple_bb (stmt
);
4711 unlink_stmt_vdef (stmt
);
4712 if (gsi_remove (&gsi
, true))
4713 bitmap_set_bit (need_eh_cleanup
, bb
->index
);
4714 if (is_gimple_call (stmt
) && stmt_can_make_abnormal_goto (stmt
))
4715 bitmap_set_bit (need_ab_cleanup
, bb
->index
);
4716 release_defs (stmt
);
4719 /* Removing a stmt may expose a forwarder block. */
4720 el_todo
|= TODO_cleanup_cfg
;
4722 el_to_remove
.release ();
4724 /* Fixup stmts that became noreturn calls. This may require splitting
4725 blocks and thus isn't possible during the dominator walk. Do this
4726 in reverse order so we don't inadvertedly remove a stmt we want to
4727 fixup by visiting a dominating now noreturn call first. */
4728 while (!el_to_fixup
.is_empty ())
4730 stmt
= el_to_fixup
.pop ();
4732 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4734 fprintf (dump_file
, "Fixing up noreturn call ");
4735 print_gimple_stmt (dump_file
, stmt
, 0, 0);
4738 if (fixup_noreturn_call (stmt
))
4739 el_todo
|= TODO_cleanup_cfg
;
4741 el_to_fixup
.release ();
4746 /* Perform CFG cleanups made necessary by elimination. */
4749 fini_eliminate (void)
4751 bool do_eh_cleanup
= !bitmap_empty_p (need_eh_cleanup
);
4752 bool do_ab_cleanup
= !bitmap_empty_p (need_ab_cleanup
);
4755 gimple_purge_all_dead_eh_edges (need_eh_cleanup
);
4758 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup
);
4760 BITMAP_FREE (need_eh_cleanup
);
4761 BITMAP_FREE (need_ab_cleanup
);
4763 if (do_eh_cleanup
|| do_ab_cleanup
)
4764 return TODO_cleanup_cfg
;
4768 /* Borrow a bit of tree-ssa-dce.c for the moment.
4769 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4770 this may be a bit faster, and we may want critical edges kept split. */
4772 /* If OP's defining statement has not already been determined to be necessary,
4773 mark that statement necessary. Return the stmt, if it is newly
4776 static inline gimple
*
4777 mark_operand_necessary (tree op
)
4783 if (TREE_CODE (op
) != SSA_NAME
)
4786 stmt
= SSA_NAME_DEF_STMT (op
);
4789 if (gimple_plf (stmt
, NECESSARY
)
4790 || gimple_nop_p (stmt
))
4793 gimple_set_plf (stmt
, NECESSARY
, true);
4797 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4798 to insert PHI nodes sometimes, and because value numbering of casts isn't
4799 perfect, we sometimes end up inserting dead code. This simple DCE-like
4800 pass removes any insertions we made that weren't actually used. */
4803 remove_dead_inserted_code (void)
4810 worklist
= BITMAP_ALLOC (NULL
);
4811 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs
, 0, i
, bi
)
4813 t
= SSA_NAME_DEF_STMT (ssa_name (i
));
4814 if (gimple_plf (t
, NECESSARY
))
4815 bitmap_set_bit (worklist
, i
);
4817 while (!bitmap_empty_p (worklist
))
4819 i
= bitmap_first_set_bit (worklist
);
4820 bitmap_clear_bit (worklist
, i
);
4821 t
= SSA_NAME_DEF_STMT (ssa_name (i
));
4823 /* PHI nodes are somewhat special in that each PHI alternative has
4824 data and control dependencies. All the statements feeding the
4825 PHI node's arguments are always necessary. */
4826 if (gimple_code (t
) == GIMPLE_PHI
)
4830 for (k
= 0; k
< gimple_phi_num_args (t
); k
++)
4832 tree arg
= PHI_ARG_DEF (t
, k
);
4833 if (TREE_CODE (arg
) == SSA_NAME
)
4835 gimple
*n
= mark_operand_necessary (arg
);
4837 bitmap_set_bit (worklist
, SSA_NAME_VERSION (arg
));
4843 /* Propagate through the operands. Examine all the USE, VUSE and
4844 VDEF operands in this statement. Mark all the statements
4845 which feed this statement's uses as necessary. */
4849 /* The operands of VDEF expressions are also needed as they
4850 represent potential definitions that may reach this
4851 statement (VDEF operands allow us to follow def-def
4854 FOR_EACH_SSA_TREE_OPERAND (use
, t
, iter
, SSA_OP_ALL_USES
)
4856 gimple
*n
= mark_operand_necessary (use
);
4858 bitmap_set_bit (worklist
, SSA_NAME_VERSION (use
));
4863 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs
, 0, i
, bi
)
4865 t
= SSA_NAME_DEF_STMT (ssa_name (i
));
4866 if (!gimple_plf (t
, NECESSARY
))
4868 gimple_stmt_iterator gsi
;
4870 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4872 fprintf (dump_file
, "Removing unnecessary insertion:");
4873 print_gimple_stmt (dump_file
, t
, 0, 0);
4876 gsi
= gsi_for_stmt (t
);
4877 if (gimple_code (t
) == GIMPLE_PHI
)
4878 remove_phi_node (&gsi
, true);
4881 gsi_remove (&gsi
, true);
4886 BITMAP_FREE (worklist
);
4890 /* Initialize data structures used by PRE. */
4897 next_expression_id
= 1;
4898 expressions
.create (0);
4899 expressions
.safe_push (NULL
);
4900 value_expressions
.create (get_max_value_id () + 1);
4901 value_expressions
.safe_grow_cleared (get_max_value_id () + 1);
4902 name_to_id
.create (0);
4904 inserted_exprs
= BITMAP_ALLOC (NULL
);
4906 connect_infinite_loops_to_exit ();
4907 memset (&pre_stats
, 0, sizeof (pre_stats
));
4909 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets
));
4911 calculate_dominance_info (CDI_DOMINATORS
);
4913 bitmap_obstack_initialize (&grand_bitmap_obstack
);
4914 phi_translate_table
= new hash_table
<expr_pred_trans_d
> (5110);
4915 expression_to_id
= new hash_table
<pre_expr_d
> (num_ssa_names
* 3);
4916 FOR_ALL_BB_FN (bb
, cfun
)
4918 EXP_GEN (bb
) = bitmap_set_new ();
4919 PHI_GEN (bb
) = bitmap_set_new ();
4920 TMP_GEN (bb
) = bitmap_set_new ();
4921 AVAIL_OUT (bb
) = bitmap_set_new ();
4926 /* Deallocate data structures used by PRE. */
4931 value_expressions
.release ();
4932 BITMAP_FREE (inserted_exprs
);
4933 bitmap_obstack_release (&grand_bitmap_obstack
);
4934 bitmap_set_pool
.release ();
4935 pre_expr_pool
.release ();
4936 delete phi_translate_table
;
4937 phi_translate_table
= NULL
;
4938 delete expression_to_id
;
4939 expression_to_id
= NULL
;
4940 name_to_id
.release ();
4942 free_aux_for_blocks ();
4947 const pass_data pass_data_pre
=
4949 GIMPLE_PASS
, /* type */
4951 OPTGROUP_NONE
, /* optinfo_flags */
4952 TV_TREE_PRE
, /* tv_id */
4953 /* PROP_no_crit_edges is ensured by placing pass_split_crit_edges before
4955 ( PROP_no_crit_edges
| PROP_cfg
| PROP_ssa
), /* properties_required */
4956 0, /* properties_provided */
4957 PROP_no_crit_edges
, /* properties_destroyed */
4958 TODO_rebuild_alias
, /* todo_flags_start */
4959 0, /* todo_flags_finish */
4962 class pass_pre
: public gimple_opt_pass
4965 pass_pre (gcc::context
*ctxt
)
4966 : gimple_opt_pass (pass_data_pre
, ctxt
)
4969 /* opt_pass methods: */
4970 virtual bool gate (function
*)
4971 { return flag_tree_pre
!= 0 || flag_code_hoisting
!= 0; }
4972 virtual unsigned int execute (function
*);
4974 }; // class pass_pre
4977 pass_pre::execute (function
*fun
)
4979 unsigned int todo
= 0;
4981 do_partial_partial
=
4982 flag_tree_partial_pre
&& optimize_function_for_speed_p (fun
);
4984 /* This has to happen before SCCVN runs because
4985 loop_optimizer_init may create new phis, etc. */
4986 loop_optimizer_init (LOOPS_NORMAL
);
4988 if (!run_scc_vn (VN_WALK
))
4990 loop_optimizer_finalize ();
4997 /* Collect and value number expressions computed in each basic block. */
5000 /* Insert can get quite slow on an incredibly large number of basic
5001 blocks due to some quadratic behavior. Until this behavior is
5002 fixed, don't run it when he have an incredibly large number of
5003 bb's. If we aren't going to run insert, there is no point in
5004 computing ANTIC, either, even though it's plenty fast. */
5005 if (n_basic_blocks_for_fn (fun
) < 4000)
5011 /* Make sure to remove fake edges before committing our inserts.
5012 This makes sure we don't end up with extra critical edges that
5013 we would need to split. */
5014 remove_fake_exit_edges ();
5015 gsi_commit_edge_inserts ();
5017 /* Eliminate folds statements which might (should not...) end up
5018 not keeping virtual operands up-to-date. */
5019 gcc_assert (!need_ssa_update_p (fun
));
5021 /* Remove all the redundant expressions. */
5022 todo
|= eliminate (true);
5024 statistics_counter_event (fun
, "Insertions", pre_stats
.insertions
);
5025 statistics_counter_event (fun
, "PA inserted", pre_stats
.pa_insert
);
5026 statistics_counter_event (fun
, "HOIST inserted", pre_stats
.hoist_insert
);
5027 statistics_counter_event (fun
, "New PHIs", pre_stats
.phis
);
5028 statistics_counter_event (fun
, "Eliminated", pre_stats
.eliminations
);
5030 clear_expression_ids ();
5031 remove_dead_inserted_code ();
5035 todo
|= fini_eliminate ();
5036 loop_optimizer_finalize ();
5038 /* Restore SSA info before tail-merging as that resets it as well. */
5039 scc_vn_restore_ssa_info ();
5041 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
5042 case we can merge the block with the remaining predecessor of the block.
5044 - call merge_blocks after each tail merge iteration
5045 - call merge_blocks after all tail merge iterations
5046 - mark TODO_cleanup_cfg when necessary
5047 - share the cfg cleanup with fini_pre. */
5048 todo
|= tail_merge_optimize (todo
);
5052 /* Tail merging invalidates the virtual SSA web, together with
5053 cfg-cleanup opportunities exposed by PRE this will wreck the
5054 SSA updating machinery. So make sure to run update-ssa
5055 manually, before eventually scheduling cfg-cleanup as part of
5057 update_ssa (TODO_update_ssa_only_virtuals
);
5065 make_pass_pre (gcc::context
*ctxt
)
5067 return new pass_pre (ctxt
);
5072 const pass_data pass_data_fre
=
5074 GIMPLE_PASS
, /* type */
5076 OPTGROUP_NONE
, /* optinfo_flags */
5077 TV_TREE_FRE
, /* tv_id */
5078 ( PROP_cfg
| PROP_ssa
), /* properties_required */
5079 0, /* properties_provided */
5080 0, /* properties_destroyed */
5081 0, /* todo_flags_start */
5082 0, /* todo_flags_finish */
5085 class pass_fre
: public gimple_opt_pass
5088 pass_fre (gcc::context
*ctxt
)
5089 : gimple_opt_pass (pass_data_fre
, ctxt
)
5092 /* opt_pass methods: */
5093 opt_pass
* clone () { return new pass_fre (m_ctxt
); }
5094 virtual bool gate (function
*) { return flag_tree_fre
!= 0; }
5095 virtual unsigned int execute (function
*);
5097 }; // class pass_fre
5100 pass_fre::execute (function
*fun
)
5102 unsigned int todo
= 0;
5104 if (!run_scc_vn (VN_WALKREWRITE
))
5107 memset (&pre_stats
, 0, sizeof (pre_stats
));
5109 /* Remove all the redundant expressions. */
5110 todo
|= eliminate (false);
5112 todo
|= fini_eliminate ();
5114 scc_vn_restore_ssa_info ();
5117 statistics_counter_event (fun
, "Insertions", pre_stats
.insertions
);
5118 statistics_counter_event (fun
, "Eliminated", pre_stats
.eliminations
);
5126 make_pass_fre (gcc::context
*ctxt
)
5128 return new pass_fre (ctxt
);