2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
3 Free Software Foundation, Inc.
4 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
25 #include "coretypes.h"
29 #include "basic-block.h"
30 #include "diagnostic.h"
31 #include "tree-inline.h"
32 #include "tree-flow.h"
34 #include "tree-dump.h"
38 #include "tree-iterator.h"
40 #include "alloc-pool.h"
42 #include "tree-pass.h"
45 #include "langhooks.h"
47 #include "tree-ssa-sccvn.h"
48 #include "tree-scalar-evolution.h"
54 1. Avail sets can be shared by making an avail_find_leader that
55 walks up the dominator tree and looks in those avail sets.
56 This might affect code optimality, it's unclear right now.
57 2. Strength reduction can be performed by anticipating expressions
58 we can repair later on.
59 3. We can do back-substitution or smarter value numbering to catch
60 commutative expressions split up over multiple statements.
63 /* For ease of terminology, "expression node" in the below refers to
64 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
65 represent the actual statement containing the expressions we care about,
66 and we cache the value number by putting it in the expression. */
70 First we walk the statements to generate the AVAIL sets, the
71 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
72 generation of values/expressions by a given block. We use them
73 when computing the ANTIC sets. The AVAIL sets consist of
74 SSA_NAME's that represent values, so we know what values are
75 available in what blocks. AVAIL is a forward dataflow problem. In
76 SSA, values are never killed, so we don't need a kill set, or a
77 fixpoint iteration, in order to calculate the AVAIL sets. In
78 traditional parlance, AVAIL sets tell us the downsafety of the
81 Next, we generate the ANTIC sets. These sets represent the
82 anticipatable expressions. ANTIC is a backwards dataflow
83 problem. An expression is anticipatable in a given block if it could
84 be generated in that block. This means that if we had to perform
85 an insertion in that block, of the value of that expression, we
86 could. Calculating the ANTIC sets requires phi translation of
87 expressions, because the flow goes backwards through phis. We must
88 iterate to a fixpoint of the ANTIC sets, because we have a kill
89 set. Even in SSA form, values are not live over the entire
90 function, only from their definition point onwards. So we have to
91 remove values from the ANTIC set once we go past the definition
92 point of the leaders that make them up.
93 compute_antic/compute_antic_aux performs this computation.
95 Third, we perform insertions to make partially redundant
96 expressions fully redundant.
98 An expression is partially redundant (excluding partial
101 1. It is AVAIL in some, but not all, of the predecessors of a
103 2. It is ANTIC in all the predecessors.
105 In order to make it fully redundant, we insert the expression into
106 the predecessors where it is not available, but is ANTIC.
108 For the partial anticipation case, we only perform insertion if it
109 is partially anticipated in some block, and fully available in all
112 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
113 performs these steps.
115 Fourth, we eliminate fully redundant expressions.
116 This is a simple statement walk that replaces redundant
117 calculations with the now available values. */
119 /* Representations of value numbers:
121 Value numbers are represented by a representative SSA_NAME. We
122 will create fake SSA_NAME's in situations where we need a
123 representative but do not have one (because it is a complex
124 expression). In order to facilitate storing the value numbers in
125 bitmaps, and keep the number of wasted SSA_NAME's down, we also
126 associate a value_id with each value number, and create full blown
127 ssa_name's only where we actually need them (IE in operands of
128 existing expressions).
130 Theoretically you could replace all the value_id's with
131 SSA_NAME_VERSION, but this would allocate a large number of
132 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
133 It would also require an additional indirection at each point we
136 /* Representation of expressions on value numbers:
138 Expressions consisting of value numbers are represented the same
139 way as our VN internally represents them, with an additional
140 "pre_expr" wrapping around them in order to facilitate storing all
141 of the expressions in the same sets. */
143 /* Representation of sets:
145 The dataflow sets do not need to be sorted in any particular order
146 for the majority of their lifetime, are simply represented as two
147 bitmaps, one that keeps track of values present in the set, and one
148 that keeps track of expressions present in the set.
150 When we need them in topological order, we produce it on demand by
151 transforming the bitmap into an array and sorting it into topo
154 /* Type of expression, used to know which member of the PRE_EXPR union
165 typedef union pre_expr_union_d
170 vn_reference_t reference
;
173 typedef struct pre_expr_d
175 enum pre_expr_kind kind
;
180 #define PRE_EXPR_NAME(e) (e)->u.name
181 #define PRE_EXPR_NARY(e) (e)->u.nary
182 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
183 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
186 pre_expr_eq (const void *p1
, const void *p2
)
188 const struct pre_expr_d
*e1
= (const struct pre_expr_d
*) p1
;
189 const struct pre_expr_d
*e2
= (const struct pre_expr_d
*) p2
;
191 if (e1
->kind
!= e2
->kind
)
197 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1
),
198 PRE_EXPR_CONSTANT (e2
));
200 return PRE_EXPR_NAME (e1
) == PRE_EXPR_NAME (e2
);
202 return vn_nary_op_eq (PRE_EXPR_NARY (e1
), PRE_EXPR_NARY (e2
));
204 return vn_reference_eq (PRE_EXPR_REFERENCE (e1
),
205 PRE_EXPR_REFERENCE (e2
));
212 pre_expr_hash (const void *p1
)
214 const struct pre_expr_d
*e
= (const struct pre_expr_d
*) p1
;
218 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e
));
220 return iterative_hash_hashval_t (SSA_NAME_VERSION (PRE_EXPR_NAME (e
)), 0);
222 return PRE_EXPR_NARY (e
)->hashcode
;
224 return PRE_EXPR_REFERENCE (e
)->hashcode
;
231 /* Next global expression id number. */
232 static unsigned int next_expression_id
;
234 /* Mapping from expression to id number we can use in bitmap sets. */
235 DEF_VEC_P (pre_expr
);
236 DEF_VEC_ALLOC_P (pre_expr
, heap
);
237 static VEC(pre_expr
, heap
) *expressions
;
238 static htab_t expression_to_id
;
240 /* Allocate an expression id for EXPR. */
242 static inline unsigned int
243 alloc_expression_id (pre_expr expr
)
246 /* Make sure we won't overflow. */
247 gcc_assert (next_expression_id
+ 1 > next_expression_id
);
248 expr
->id
= next_expression_id
++;
249 VEC_safe_push (pre_expr
, heap
, expressions
, expr
);
250 slot
= htab_find_slot (expression_to_id
, expr
, INSERT
);
253 return next_expression_id
- 1;
256 /* Return the expression id for tree EXPR. */
258 static inline unsigned int
259 get_expression_id (const pre_expr expr
)
264 static inline unsigned int
265 lookup_expression_id (const pre_expr expr
)
269 slot
= htab_find_slot (expression_to_id
, expr
, NO_INSERT
);
272 return ((pre_expr
)*slot
)->id
;
275 /* Return the existing expression id for EXPR, or create one if one
276 does not exist yet. */
278 static inline unsigned int
279 get_or_alloc_expression_id (pre_expr expr
)
281 unsigned int id
= lookup_expression_id (expr
);
283 return alloc_expression_id (expr
);
284 return expr
->id
= id
;
287 /* Return the expression that has expression id ID */
289 static inline pre_expr
290 expression_for_id (unsigned int id
)
292 return VEC_index (pre_expr
, expressions
, id
);
295 /* Free the expression id field in all of our expressions,
296 and then destroy the expressions array. */
299 clear_expression_ids (void)
301 VEC_free (pre_expr
, heap
, expressions
);
304 static alloc_pool pre_expr_pool
;
306 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
309 get_or_alloc_expr_for_name (tree name
)
311 pre_expr result
= (pre_expr
) pool_alloc (pre_expr_pool
);
312 unsigned int result_id
;
316 PRE_EXPR_NAME (result
) = name
;
317 result_id
= lookup_expression_id (result
);
320 pool_free (pre_expr_pool
, result
);
321 result
= expression_for_id (result_id
);
324 get_or_alloc_expression_id (result
);
328 static bool in_fre
= false;
330 /* An unordered bitmap set. One bitmap tracks values, the other,
332 typedef struct bitmap_set
338 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
339 EXECUTE_IF_SET_IN_BITMAP((set)->expressions, 0, (id), (bi))
341 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
342 EXECUTE_IF_SET_IN_BITMAP((set)->values, 0, (id), (bi))
344 /* Mapping from value id to expressions with that value_id. */
345 DEF_VEC_P (bitmap_set_t
);
346 DEF_VEC_ALLOC_P (bitmap_set_t
, heap
);
347 static VEC(bitmap_set_t
, heap
) *value_expressions
;
349 /* Sets that we need to keep track of. */
350 typedef struct bb_bitmap_sets
352 /* The EXP_GEN set, which represents expressions/values generated in
354 bitmap_set_t exp_gen
;
356 /* The PHI_GEN set, which represents PHI results generated in a
358 bitmap_set_t phi_gen
;
360 /* The TMP_GEN set, which represents results/temporaries generated
361 in a basic block. IE the LHS of an expression. */
362 bitmap_set_t tmp_gen
;
364 /* The AVAIL_OUT set, which represents which values are available in
365 a given basic block. */
366 bitmap_set_t avail_out
;
368 /* The ANTIC_IN set, which represents which values are anticipatable
369 in a given basic block. */
370 bitmap_set_t antic_in
;
372 /* The PA_IN set, which represents which values are
373 partially anticipatable in a given basic block. */
376 /* The NEW_SETS set, which is used during insertion to augment the
377 AVAIL_OUT set of blocks with the new insertions performed during
378 the current iteration. */
379 bitmap_set_t new_sets
;
381 /* A cache for value_dies_in_block_x. */
384 /* True if we have visited this block during ANTIC calculation. */
385 unsigned int visited
:1;
387 /* True we have deferred processing this block during ANTIC
388 calculation until its successor is processed. */
389 unsigned int deferred
: 1;
392 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
393 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
394 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
395 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
396 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
397 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
398 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
399 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
400 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
401 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
404 /* Maximal set of values, used to initialize the ANTIC problem, which
405 is an intersection problem. */
406 static bitmap_set_t maximal_set
;
408 /* Basic block list in postorder. */
409 static int *postorder
;
411 /* This structure is used to keep track of statistics on what
412 optimization PRE was able to perform. */
415 /* The number of RHS computations eliminated by PRE. */
418 /* The number of new expressions/temporaries generated by PRE. */
421 /* The number of inserts found due to partial anticipation */
424 /* The number of new PHI nodes added by PRE. */
427 /* The number of values found constant. */
432 static bool do_partial_partial
;
433 static pre_expr
bitmap_find_leader (bitmap_set_t
, unsigned int, gimple
);
434 static void bitmap_value_insert_into_set (bitmap_set_t
, pre_expr
);
435 static void bitmap_value_replace_in_set (bitmap_set_t
, pre_expr
);
436 static void bitmap_set_copy (bitmap_set_t
, bitmap_set_t
);
437 static bool bitmap_set_contains_value (bitmap_set_t
, unsigned int);
438 static void bitmap_insert_into_set (bitmap_set_t
, pre_expr
);
439 static void bitmap_insert_into_set_1 (bitmap_set_t
, pre_expr
, bool);
440 static bitmap_set_t
bitmap_set_new (void);
441 static tree
create_expression_by_pieces (basic_block
, pre_expr
, gimple_seq
*,
443 static tree
find_or_generate_expression (basic_block
, pre_expr
, gimple_seq
*,
445 static unsigned int get_expr_value_id (pre_expr
);
447 /* We can add and remove elements and entries to and from sets
448 and hash tables, so we use alloc pools for them. */
450 static alloc_pool bitmap_set_pool
;
451 static bitmap_obstack grand_bitmap_obstack
;
453 /* To avoid adding 300 temporary variables when we only need one, we
454 only create one temporary variable, on demand, and build ssa names
455 off that. We do have to change the variable if the types don't
456 match the current variable's type. */
458 static tree storetemp
;
459 static tree prephitemp
;
461 /* Set of blocks with statements that have had its EH information
463 static bitmap need_eh_cleanup
;
465 /* Which expressions have been seen during a given phi translation. */
466 static bitmap seen_during_translate
;
468 /* The phi_translate_table caches phi translations for a given
469 expression and predecessor. */
471 static htab_t phi_translate_table
;
473 /* A three tuple {e, pred, v} used to cache phi translations in the
474 phi_translate_table. */
476 typedef struct expr_pred_trans_d
478 /* The expression. */
481 /* The predecessor block along which we translated the expression. */
484 /* The value that resulted from the translation. */
487 /* The hashcode for the expression, pred pair. This is cached for
490 } *expr_pred_trans_t
;
491 typedef const struct expr_pred_trans_d
*const_expr_pred_trans_t
;
493 /* Return the hash value for a phi translation table entry. */
496 expr_pred_trans_hash (const void *p
)
498 const_expr_pred_trans_t
const ve
= (const_expr_pred_trans_t
) p
;
502 /* Return true if two phi translation table entries are the same.
503 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
506 expr_pred_trans_eq (const void *p1
, const void *p2
)
508 const_expr_pred_trans_t
const ve1
= (const_expr_pred_trans_t
) p1
;
509 const_expr_pred_trans_t
const ve2
= (const_expr_pred_trans_t
) p2
;
510 basic_block b1
= ve1
->pred
;
511 basic_block b2
= ve2
->pred
;
513 /* If they are not translations for the same basic block, they can't
517 return pre_expr_eq (ve1
->e
, ve2
->e
);
520 /* Search in the phi translation table for the translation of
521 expression E in basic block PRED.
522 Return the translated value, if found, NULL otherwise. */
524 static inline pre_expr
525 phi_trans_lookup (pre_expr e
, basic_block pred
)
528 struct expr_pred_trans_d ept
;
532 ept
.hashcode
= iterative_hash_hashval_t (pre_expr_hash (e
), pred
->index
);
533 slot
= htab_find_slot_with_hash (phi_translate_table
, &ept
, ept
.hashcode
,
538 return ((expr_pred_trans_t
) *slot
)->v
;
542 /* Add the tuple mapping from {expression E, basic block PRED} to
543 value V, to the phi translation table. */
546 phi_trans_add (pre_expr e
, pre_expr v
, basic_block pred
)
549 expr_pred_trans_t new_pair
= XNEW (struct expr_pred_trans_d
);
551 new_pair
->pred
= pred
;
553 new_pair
->hashcode
= iterative_hash_hashval_t (pre_expr_hash (e
),
556 slot
= htab_find_slot_with_hash (phi_translate_table
, new_pair
,
557 new_pair
->hashcode
, INSERT
);
560 *slot
= (void *) new_pair
;
564 /* Add expression E to the expression set of value id V. */
567 add_to_value (unsigned int v
, pre_expr e
)
571 gcc_assert (get_expr_value_id (e
) == v
);
573 if (v
>= VEC_length (bitmap_set_t
, value_expressions
))
575 VEC_safe_grow_cleared (bitmap_set_t
, heap
, value_expressions
,
579 set
= VEC_index (bitmap_set_t
, value_expressions
, v
);
582 set
= bitmap_set_new ();
583 VEC_replace (bitmap_set_t
, value_expressions
, v
, set
);
586 bitmap_insert_into_set_1 (set
, e
, true);
589 /* Create a new bitmap set and return it. */
592 bitmap_set_new (void)
594 bitmap_set_t ret
= (bitmap_set_t
) pool_alloc (bitmap_set_pool
);
595 ret
->expressions
= BITMAP_ALLOC (&grand_bitmap_obstack
);
596 ret
->values
= BITMAP_ALLOC (&grand_bitmap_obstack
);
600 /* Return the value id for a PRE expression EXPR. */
603 get_expr_value_id (pre_expr expr
)
610 id
= get_constant_value_id (PRE_EXPR_CONSTANT (expr
));
613 id
= get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr
));
614 add_to_value (id
, expr
);
619 return VN_INFO (PRE_EXPR_NAME (expr
))->value_id
;
621 return PRE_EXPR_NARY (expr
)->value_id
;
623 return PRE_EXPR_REFERENCE (expr
)->value_id
;
629 /* Remove an expression EXPR from a bitmapped set. */
632 bitmap_remove_from_set (bitmap_set_t set
, pre_expr expr
)
634 unsigned int val
= get_expr_value_id (expr
);
635 if (!value_id_constant_p (val
))
637 bitmap_clear_bit (set
->values
, val
);
638 bitmap_clear_bit (set
->expressions
, get_expression_id (expr
));
643 bitmap_insert_into_set_1 (bitmap_set_t set
, pre_expr expr
,
644 bool allow_constants
)
646 unsigned int val
= get_expr_value_id (expr
);
647 if (allow_constants
|| !value_id_constant_p (val
))
649 /* We specifically expect this and only this function to be able to
650 insert constants into a set. */
651 bitmap_set_bit (set
->values
, val
);
652 bitmap_set_bit (set
->expressions
, get_or_alloc_expression_id (expr
));
656 /* Insert an expression EXPR into a bitmapped set. */
659 bitmap_insert_into_set (bitmap_set_t set
, pre_expr expr
)
661 bitmap_insert_into_set_1 (set
, expr
, false);
664 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
667 bitmap_set_copy (bitmap_set_t dest
, bitmap_set_t orig
)
669 bitmap_copy (dest
->expressions
, orig
->expressions
);
670 bitmap_copy (dest
->values
, orig
->values
);
674 /* Free memory used up by SET. */
676 bitmap_set_free (bitmap_set_t set
)
678 BITMAP_FREE (set
->expressions
);
679 BITMAP_FREE (set
->values
);
683 /* Generate an topological-ordered array of bitmap set SET. */
685 static VEC(pre_expr
, heap
) *
686 sorted_array_from_bitmap_set (bitmap_set_t set
)
689 bitmap_iterator bi
, bj
;
690 VEC(pre_expr
, heap
) *result
= NULL
;
692 FOR_EACH_VALUE_ID_IN_SET (set
, i
, bi
)
694 /* The number of expressions having a given value is usually
695 relatively small. Thus, rather than making a vector of all
696 the expressions and sorting it by value-id, we walk the values
697 and check in the reverse mapping that tells us what expressions
698 have a given value, to filter those in our set. As a result,
699 the expressions are inserted in value-id order, which means
702 If this is somehow a significant lose for some cases, we can
703 choose which set to walk based on the set size. */
704 bitmap_set_t exprset
= VEC_index (bitmap_set_t
, value_expressions
, i
);
705 FOR_EACH_EXPR_ID_IN_SET (exprset
, j
, bj
)
707 if (bitmap_bit_p (set
->expressions
, j
))
708 VEC_safe_push (pre_expr
, heap
, result
, expression_for_id (j
));
715 /* Perform bitmapped set operation DEST &= ORIG. */
718 bitmap_set_and (bitmap_set_t dest
, bitmap_set_t orig
)
725 bitmap temp
= BITMAP_ALLOC (&grand_bitmap_obstack
);
727 bitmap_and_into (dest
->values
, orig
->values
);
728 bitmap_copy (temp
, dest
->expressions
);
729 EXECUTE_IF_SET_IN_BITMAP (temp
, 0, i
, bi
)
731 pre_expr expr
= expression_for_id (i
);
732 unsigned int value_id
= get_expr_value_id (expr
);
733 if (!bitmap_bit_p (dest
->values
, value_id
))
734 bitmap_clear_bit (dest
->expressions
, i
);
740 /* Subtract all values and expressions contained in ORIG from DEST. */
743 bitmap_set_subtract (bitmap_set_t dest
, bitmap_set_t orig
)
745 bitmap_set_t result
= bitmap_set_new ();
749 bitmap_and_compl (result
->expressions
, dest
->expressions
,
752 FOR_EACH_EXPR_ID_IN_SET (result
, i
, bi
)
754 pre_expr expr
= expression_for_id (i
);
755 unsigned int value_id
= get_expr_value_id (expr
);
756 bitmap_set_bit (result
->values
, value_id
);
762 /* Subtract all the values in bitmap set B from bitmap set A. */
765 bitmap_set_subtract_values (bitmap_set_t a
, bitmap_set_t b
)
769 bitmap temp
= BITMAP_ALLOC (&grand_bitmap_obstack
);
771 bitmap_copy (temp
, a
->expressions
);
772 EXECUTE_IF_SET_IN_BITMAP (temp
, 0, i
, bi
)
774 pre_expr expr
= expression_for_id (i
);
775 if (bitmap_set_contains_value (b
, get_expr_value_id (expr
)))
776 bitmap_remove_from_set (a
, expr
);
782 /* Return true if bitmapped set SET contains the value VALUE_ID. */
785 bitmap_set_contains_value (bitmap_set_t set
, unsigned int value_id
)
787 if (value_id_constant_p (value_id
))
790 if (!set
|| bitmap_empty_p (set
->expressions
))
793 return bitmap_bit_p (set
->values
, value_id
);
797 bitmap_set_contains_expr (bitmap_set_t set
, const pre_expr expr
)
799 return bitmap_bit_p (set
->expressions
, get_expression_id (expr
));
802 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
805 bitmap_set_replace_value (bitmap_set_t set
, unsigned int lookfor
,
808 bitmap_set_t exprset
;
812 if (value_id_constant_p (lookfor
))
815 if (!bitmap_set_contains_value (set
, lookfor
))
818 /* The number of expressions having a given value is usually
819 significantly less than the total number of expressions in SET.
820 Thus, rather than check, for each expression in SET, whether it
821 has the value LOOKFOR, we walk the reverse mapping that tells us
822 what expressions have a given value, and see if any of those
823 expressions are in our set. For large testcases, this is about
824 5-10x faster than walking the bitmap. If this is somehow a
825 significant lose for some cases, we can choose which set to walk
826 based on the set size. */
827 exprset
= VEC_index (bitmap_set_t
, value_expressions
, lookfor
);
828 FOR_EACH_EXPR_ID_IN_SET (exprset
, i
, bi
)
830 if (bitmap_bit_p (set
->expressions
, i
))
832 bitmap_clear_bit (set
->expressions
, i
);
833 bitmap_set_bit (set
->expressions
, get_expression_id (expr
));
839 /* Return true if two bitmap sets are equal. */
842 bitmap_set_equal (bitmap_set_t a
, bitmap_set_t b
)
844 return bitmap_equal_p (a
->values
, b
->values
);
847 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
848 and add it otherwise. */
851 bitmap_value_replace_in_set (bitmap_set_t set
, pre_expr expr
)
853 unsigned int val
= get_expr_value_id (expr
);
855 if (bitmap_set_contains_value (set
, val
))
856 bitmap_set_replace_value (set
, val
, expr
);
858 bitmap_insert_into_set (set
, expr
);
861 /* Insert EXPR into SET if EXPR's value is not already present in
865 bitmap_value_insert_into_set (bitmap_set_t set
, pre_expr expr
)
867 unsigned int val
= get_expr_value_id (expr
);
869 if (value_id_constant_p (val
))
872 if (!bitmap_set_contains_value (set
, val
))
873 bitmap_insert_into_set (set
, expr
);
876 /* Print out EXPR to outfile. */
879 print_pre_expr (FILE *outfile
, const pre_expr expr
)
884 print_generic_expr (outfile
, PRE_EXPR_CONSTANT (expr
), 0);
887 print_generic_expr (outfile
, PRE_EXPR_NAME (expr
), 0);
892 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
893 fprintf (outfile
, "{%s,", tree_code_name
[nary
->opcode
]);
894 for (i
= 0; i
< nary
->length
; i
++)
896 print_generic_expr (outfile
, nary
->op
[i
], 0);
897 if (i
!= (unsigned) nary
->length
- 1)
898 fprintf (outfile
, ",");
900 fprintf (outfile
, "}");
906 vn_reference_op_t vro
;
908 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
909 fprintf (outfile
, "{");
911 VEC_iterate (vn_reference_op_s
, ref
->operands
, i
, vro
);
914 bool closebrace
= false;
915 if (vro
->opcode
!= SSA_NAME
916 && TREE_CODE_CLASS (vro
->opcode
) != tcc_declaration
)
918 fprintf (outfile
, "%s", tree_code_name
[vro
->opcode
]);
921 fprintf (outfile
, "<");
927 print_generic_expr (outfile
, vro
->op0
, 0);
930 fprintf (outfile
, ",");
931 print_generic_expr (outfile
, vro
->op1
, 0);
935 fprintf (outfile
, ",");
936 print_generic_expr (outfile
, vro
->op2
, 0);
940 fprintf (outfile
, ">");
941 if (i
!= VEC_length (vn_reference_op_s
, ref
->operands
) - 1)
942 fprintf (outfile
, ",");
944 fprintf (outfile
, "}");
947 fprintf (outfile
, "@");
948 print_generic_expr (outfile
, ref
->vuse
, 0);
954 void debug_pre_expr (pre_expr
);
956 /* Like print_pre_expr but always prints to stderr. */
958 debug_pre_expr (pre_expr e
)
960 print_pre_expr (stderr
, e
);
961 fprintf (stderr
, "\n");
964 /* Print out SET to OUTFILE. */
967 print_bitmap_set (FILE *outfile
, bitmap_set_t set
,
968 const char *setname
, int blockindex
)
970 fprintf (outfile
, "%s[%d] := { ", setname
, blockindex
);
977 FOR_EACH_EXPR_ID_IN_SET (set
, i
, bi
)
979 const pre_expr expr
= expression_for_id (i
);
982 fprintf (outfile
, ", ");
984 print_pre_expr (outfile
, expr
);
986 fprintf (outfile
, " (%04d)", get_expr_value_id (expr
));
989 fprintf (outfile
, " }\n");
992 void debug_bitmap_set (bitmap_set_t
);
995 debug_bitmap_set (bitmap_set_t set
)
997 print_bitmap_set (stderr
, set
, "debug", 0);
1000 /* Print out the expressions that have VAL to OUTFILE. */
1003 print_value_expressions (FILE *outfile
, unsigned int val
)
1005 bitmap_set_t set
= VEC_index (bitmap_set_t
, value_expressions
, val
);
1009 sprintf (s
, "%04d", val
);
1010 print_bitmap_set (outfile
, set
, s
, 0);
1016 debug_value_expressions (unsigned int val
)
1018 print_value_expressions (stderr
, val
);
1021 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1025 get_or_alloc_expr_for_constant (tree constant
)
1027 unsigned int result_id
;
1028 unsigned int value_id
;
1029 pre_expr newexpr
= (pre_expr
) pool_alloc (pre_expr_pool
);
1030 newexpr
->kind
= CONSTANT
;
1031 PRE_EXPR_CONSTANT (newexpr
) = constant
;
1032 result_id
= lookup_expression_id (newexpr
);
1035 pool_free (pre_expr_pool
, newexpr
);
1036 newexpr
= expression_for_id (result_id
);
1039 value_id
= get_or_alloc_constant_value_id (constant
);
1040 get_or_alloc_expression_id (newexpr
);
1041 add_to_value (value_id
, newexpr
);
1045 /* Given a value id V, find the actual tree representing the constant
1046 value if there is one, and return it. Return NULL if we can't find
1050 get_constant_for_value_id (unsigned int v
)
1052 if (value_id_constant_p (v
))
1056 bitmap_set_t exprset
= VEC_index (bitmap_set_t
, value_expressions
, v
);
1058 FOR_EACH_EXPR_ID_IN_SET (exprset
, i
, bi
)
1060 pre_expr expr
= expression_for_id (i
);
1061 if (expr
->kind
== CONSTANT
)
1062 return PRE_EXPR_CONSTANT (expr
);
1068 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1069 Currently only supports constants and SSA_NAMES. */
1071 get_or_alloc_expr_for (tree t
)
1073 if (TREE_CODE (t
) == SSA_NAME
)
1074 return get_or_alloc_expr_for_name (t
);
1075 else if (is_gimple_min_invariant (t
)
1076 || TREE_CODE (t
) == EXC_PTR_EXPR
1077 || TREE_CODE (t
) == FILTER_EXPR
)
1078 return get_or_alloc_expr_for_constant (t
);
1081 /* More complex expressions can result from SCCVN expression
1082 simplification that inserts values for them. As they all
1083 do not have VOPs the get handled by the nary ops struct. */
1084 vn_nary_op_t result
;
1085 unsigned int result_id
;
1086 vn_nary_op_lookup (t
, &result
);
1089 pre_expr e
= (pre_expr
) pool_alloc (pre_expr_pool
);
1091 PRE_EXPR_NARY (e
) = result
;
1092 result_id
= lookup_expression_id (e
);
1095 pool_free (pre_expr_pool
, e
);
1096 e
= expression_for_id (result_id
);
1099 alloc_expression_id (e
);
1106 /* Return the folded version of T if T, when folded, is a gimple
1107 min_invariant. Otherwise, return T. */
1110 fully_constant_expression (pre_expr e
)
1118 vn_nary_op_t nary
= PRE_EXPR_NARY (e
);
1119 switch (TREE_CODE_CLASS (nary
->opcode
))
1121 case tcc_expression
:
1122 if (nary
->opcode
== TRUTH_NOT_EXPR
)
1124 if (nary
->opcode
!= TRUTH_AND_EXPR
1125 && nary
->opcode
!= TRUTH_OR_EXPR
1126 && nary
->opcode
!= TRUTH_XOR_EXPR
)
1130 case tcc_comparison
:
1132 /* We have to go from trees to pre exprs to value ids to
1134 tree naryop0
= nary
->op
[0];
1135 tree naryop1
= nary
->op
[1];
1137 if (!is_gimple_min_invariant (naryop0
))
1139 pre_expr rep0
= get_or_alloc_expr_for (naryop0
);
1140 unsigned int vrep0
= get_expr_value_id (rep0
);
1141 tree const0
= get_constant_for_value_id (vrep0
);
1143 naryop0
= fold_convert (TREE_TYPE (naryop0
), const0
);
1145 if (!is_gimple_min_invariant (naryop1
))
1147 pre_expr rep1
= get_or_alloc_expr_for (naryop1
);
1148 unsigned int vrep1
= get_expr_value_id (rep1
);
1149 tree const1
= get_constant_for_value_id (vrep1
);
1151 naryop1
= fold_convert (TREE_TYPE (naryop1
), const1
);
1153 result
= fold_binary (nary
->opcode
, nary
->type
,
1155 if (result
&& is_gimple_min_invariant (result
))
1156 return get_or_alloc_expr_for_constant (result
);
1157 /* We might have simplified the expression to a
1158 SSA_NAME for example from x_1 * 1. But we cannot
1159 insert a PHI for x_1 unconditionally as x_1 might
1160 not be available readily. */
1164 if (nary
->opcode
!= REALPART_EXPR
1165 && nary
->opcode
!= IMAGPART_EXPR
1166 && nary
->opcode
!= VIEW_CONVERT_EXPR
)
1172 /* We have to go from trees to pre exprs to value ids to
1174 tree naryop0
= nary
->op
[0];
1175 tree const0
, result
;
1176 if (is_gimple_min_invariant (naryop0
))
1180 pre_expr rep0
= get_or_alloc_expr_for (naryop0
);
1181 unsigned int vrep0
= get_expr_value_id (rep0
);
1182 const0
= get_constant_for_value_id (vrep0
);
1187 tree type1
= TREE_TYPE (nary
->op
[0]);
1188 const0
= fold_convert (type1
, const0
);
1189 result
= fold_unary (nary
->opcode
, nary
->type
, const0
);
1191 if (result
&& is_gimple_min_invariant (result
))
1192 return get_or_alloc_expr_for_constant (result
);
1201 vn_reference_t ref
= PRE_EXPR_REFERENCE (e
);
1202 VEC (vn_reference_op_s
, heap
) *operands
= ref
->operands
;
1203 vn_reference_op_t op
;
1205 /* Try to simplify the translated expression if it is
1206 a call to a builtin function with at most two arguments. */
1207 op
= VEC_index (vn_reference_op_s
, operands
, 0);
1208 if (op
->opcode
== CALL_EXPR
1209 && TREE_CODE (op
->op0
) == ADDR_EXPR
1210 && TREE_CODE (TREE_OPERAND (op
->op0
, 0)) == FUNCTION_DECL
1211 && DECL_BUILT_IN (TREE_OPERAND (op
->op0
, 0))
1212 && VEC_length (vn_reference_op_s
, operands
) >= 2
1213 && VEC_length (vn_reference_op_s
, operands
) <= 3)
1215 vn_reference_op_t arg0
, arg1
= NULL
;
1216 bool anyconst
= false;
1217 arg0
= VEC_index (vn_reference_op_s
, operands
, 1);
1218 if (VEC_length (vn_reference_op_s
, operands
) > 2)
1219 arg1
= VEC_index (vn_reference_op_s
, operands
, 2);
1220 if (TREE_CODE_CLASS (arg0
->opcode
) == tcc_constant
1221 || (arg0
->opcode
== ADDR_EXPR
1222 && is_gimple_min_invariant (arg0
->op0
)))
1225 && (TREE_CODE_CLASS (arg1
->opcode
) == tcc_constant
1226 || (arg1
->opcode
== ADDR_EXPR
1227 && is_gimple_min_invariant (arg1
->op0
))))
1231 tree folded
= build_call_expr (TREE_OPERAND (op
->op0
, 0),
1234 arg1
? arg1
->op0
: NULL
);
1236 && TREE_CODE (folded
) == NOP_EXPR
)
1237 folded
= TREE_OPERAND (folded
, 0);
1239 && is_gimple_min_invariant (folded
))
1240 return get_or_alloc_expr_for_constant (folded
);
1251 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1252 it has the value it would have in BLOCK. */
1255 translate_vuse_through_block (VEC (vn_reference_op_s
, heap
) *operands
,
1256 alias_set_type set
, tree type
, tree vuse
,
1257 basic_block phiblock
,
1260 gimple phi
= SSA_NAME_DEF_STMT (vuse
);
1263 if (gimple_bb (phi
) != phiblock
)
1266 if (gimple_code (phi
) == GIMPLE_PHI
)
1268 edge e
= find_edge (block
, phiblock
);
1269 return PHI_ARG_DEF (phi
, e
->dest_idx
);
1272 if (!ao_ref_init_from_vn_reference (&ref
, set
, type
, operands
))
1275 /* Use the alias-oracle to find either the PHI node in this block,
1276 the first VUSE used in this block that is equivalent to vuse or
1277 the first VUSE which definition in this block kills the value. */
1278 while (!stmt_may_clobber_ref_p_1 (phi
, &ref
))
1280 vuse
= gimple_vuse (phi
);
1281 phi
= SSA_NAME_DEF_STMT (vuse
);
1282 if (gimple_bb (phi
) != phiblock
)
1284 if (gimple_code (phi
) == GIMPLE_PHI
)
1286 edge e
= find_edge (block
, phiblock
);
1287 return PHI_ARG_DEF (phi
, e
->dest_idx
);
1294 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1295 SET2. This is used to avoid making a set consisting of the union
1296 of PA_IN and ANTIC_IN during insert. */
1298 static inline pre_expr
1299 find_leader_in_sets (unsigned int val
, bitmap_set_t set1
, bitmap_set_t set2
)
1303 result
= bitmap_find_leader (set1
, val
, NULL
);
1304 if (!result
&& set2
)
1305 result
= bitmap_find_leader (set2
, val
, NULL
);
1309 /* Get the tree type for our PRE expression e. */
1312 get_expr_type (const pre_expr e
)
1317 return TREE_TYPE (PRE_EXPR_NAME (e
));
1319 return TREE_TYPE (PRE_EXPR_CONSTANT (e
));
1321 return PRE_EXPR_REFERENCE (e
)->type
;
1323 return PRE_EXPR_NARY (e
)->type
;
1328 /* Get a representative SSA_NAME for a given expression.
1329 Since all of our sub-expressions are treated as values, we require
1330 them to be SSA_NAME's for simplicity.
1331 Prior versions of GVNPRE used to use "value handles" here, so that
1332 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1333 either case, the operands are really values (IE we do not expect
1334 them to be usable without finding leaders). */
1337 get_representative_for (const pre_expr e
)
1341 unsigned int value_id
= get_expr_value_id (e
);
1346 return PRE_EXPR_NAME (e
);
1348 return PRE_EXPR_CONSTANT (e
);
1352 /* Go through all of the expressions representing this value
1353 and pick out an SSA_NAME. */
1356 bitmap_set_t exprs
= VEC_index (bitmap_set_t
, value_expressions
,
1358 FOR_EACH_EXPR_ID_IN_SET (exprs
, i
, bi
)
1360 pre_expr rep
= expression_for_id (i
);
1361 if (rep
->kind
== NAME
)
1362 return PRE_EXPR_NAME (rep
);
1367 /* If we reached here we couldn't find an SSA_NAME. This can
1368 happen when we've discovered a value that has never appeared in
1369 the program as set to an SSA_NAME, most likely as the result of
1374 "Could not find SSA_NAME representative for expression:");
1375 print_pre_expr (dump_file
, e
);
1376 fprintf (dump_file
, "\n");
1379 exprtype
= get_expr_type (e
);
1381 /* Build and insert the assignment of the end result to the temporary
1382 that we will return. */
1383 if (!pretemp
|| exprtype
!= TREE_TYPE (pretemp
))
1385 pretemp
= create_tmp_var (exprtype
, "pretmp");
1386 get_var_ann (pretemp
);
1389 name
= make_ssa_name (pretemp
, gimple_build_nop ());
1390 VN_INFO_GET (name
)->value_id
= value_id
;
1391 if (e
->kind
== CONSTANT
)
1392 VN_INFO (name
)->valnum
= PRE_EXPR_CONSTANT (e
);
1394 VN_INFO (name
)->valnum
= name
;
1396 add_to_value (value_id
, get_or_alloc_expr_for_name (name
));
1399 fprintf (dump_file
, "Created SSA_NAME representative ");
1400 print_generic_expr (dump_file
, name
, 0);
1401 fprintf (dump_file
, " for expression:");
1402 print_pre_expr (dump_file
, e
);
1403 fprintf (dump_file
, "\n");
1412 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1413 the phis in PRED. SEEN is a bitmap saying which expression we have
1414 translated since we started translation of the toplevel expression.
1415 Return NULL if we can't find a leader for each part of the
1416 translated expression. */
1419 phi_translate_1 (pre_expr expr
, bitmap_set_t set1
, bitmap_set_t set2
,
1420 basic_block pred
, basic_block phiblock
, bitmap seen
)
1422 pre_expr oldexpr
= expr
;
1428 if (value_id_constant_p (get_expr_value_id (expr
)))
1431 phitrans
= phi_trans_lookup (expr
, pred
);
1435 /* Prevent cycles when we have recursively dependent leaders. This
1436 can only happen when phi translating the maximal set. */
1439 unsigned int expr_id
= get_expression_id (expr
);
1440 if (bitmap_bit_p (seen
, expr_id
))
1442 bitmap_set_bit (seen
, expr_id
);
1447 /* Constants contain no values that need translation. */
1454 bool changed
= false;
1455 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
1456 struct vn_nary_op_s newnary
;
1457 /* The NARY structure is only guaranteed to have been
1458 allocated to the nary->length operands. */
1459 memcpy (&newnary
, nary
, (sizeof (struct vn_nary_op_s
)
1460 - sizeof (tree
) * (4 - nary
->length
)));
1462 for (i
= 0; i
< newnary
.length
; i
++)
1464 if (TREE_CODE (newnary
.op
[i
]) != SSA_NAME
)
1468 unsigned int op_val_id
= VN_INFO (newnary
.op
[i
])->value_id
;
1469 pre_expr leader
= find_leader_in_sets (op_val_id
, set1
, set2
);
1470 pre_expr result
= phi_translate_1 (leader
, set1
, set2
,
1471 pred
, phiblock
, seen
);
1472 if (result
&& result
!= leader
)
1474 tree name
= get_representative_for (result
);
1477 newnary
.op
[i
] = name
;
1482 changed
|= newnary
.op
[i
] != nary
->op
[i
];
1489 tree result
= vn_nary_op_lookup_pieces (newnary
.length
,
1497 unsigned int new_val_id
;
1499 expr
= (pre_expr
) pool_alloc (pre_expr_pool
);
1502 if (result
&& is_gimple_min_invariant (result
))
1503 return get_or_alloc_expr_for_constant (result
);
1508 PRE_EXPR_NARY (expr
) = nary
;
1509 constant
= fully_constant_expression (expr
);
1510 if (constant
!= expr
)
1513 new_val_id
= nary
->value_id
;
1514 get_or_alloc_expression_id (expr
);
1518 new_val_id
= get_next_value_id ();
1519 VEC_safe_grow_cleared (bitmap_set_t
, heap
,
1521 get_max_value_id() + 1);
1522 nary
= vn_nary_op_insert_pieces (newnary
.length
,
1529 result
, new_val_id
);
1530 PRE_EXPR_NARY (expr
) = nary
;
1531 constant
= fully_constant_expression (expr
);
1532 if (constant
!= expr
)
1534 get_or_alloc_expression_id (expr
);
1536 add_to_value (new_val_id
, expr
);
1538 phi_trans_add (oldexpr
, expr
, pred
);
1545 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
1546 VEC (vn_reference_op_s
, heap
) *operands
= ref
->operands
;
1547 tree vuse
= ref
->vuse
;
1548 tree newvuse
= vuse
;
1549 VEC (vn_reference_op_s
, heap
) *newoperands
= NULL
;
1550 bool changed
= false;
1552 vn_reference_op_t operand
;
1553 vn_reference_t newref
;
1556 VEC_iterate (vn_reference_op_s
, operands
, i
, operand
); i
++, j
++)
1560 tree oldop0
= operand
->op0
;
1561 tree oldop1
= operand
->op1
;
1562 tree oldop2
= operand
->op2
;
1566 tree type
= operand
->type
;
1567 vn_reference_op_s newop
= *operand
;
1569 if (op0
&& TREE_CODE (op0
) == SSA_NAME
)
1571 unsigned int op_val_id
= VN_INFO (op0
)->value_id
;
1572 leader
= find_leader_in_sets (op_val_id
, set1
, set2
);
1573 opresult
= phi_translate_1 (leader
, set1
, set2
,
1574 pred
, phiblock
, seen
);
1575 if (opresult
&& opresult
!= leader
)
1577 tree name
= get_representative_for (opresult
);
1585 changed
|= op0
!= oldop0
;
1587 if (op1
&& TREE_CODE (op1
) == SSA_NAME
)
1589 unsigned int op_val_id
= VN_INFO (op1
)->value_id
;
1590 leader
= find_leader_in_sets (op_val_id
, set1
, set2
);
1591 opresult
= phi_translate_1 (leader
, set1
, set2
,
1592 pred
, phiblock
, seen
);
1593 if (opresult
&& opresult
!= leader
)
1595 tree name
= get_representative_for (opresult
);
1603 changed
|= op1
!= oldop1
;
1604 if (op2
&& TREE_CODE (op2
) == SSA_NAME
)
1606 unsigned int op_val_id
= VN_INFO (op2
)->value_id
;
1607 leader
= find_leader_in_sets (op_val_id
, set1
, set2
);
1608 opresult
= phi_translate_1 (leader
, set1
, set2
,
1609 pred
, phiblock
, seen
);
1610 if (opresult
&& opresult
!= leader
)
1612 tree name
= get_representative_for (opresult
);
1620 changed
|= op2
!= oldop2
;
1623 newoperands
= VEC_copy (vn_reference_op_s
, heap
, operands
);
1624 /* We may have changed from an SSA_NAME to a constant */
1625 if (newop
.opcode
== SSA_NAME
&& TREE_CODE (op0
) != SSA_NAME
)
1626 newop
.opcode
= TREE_CODE (op0
);
1631 VEC_replace (vn_reference_op_s
, newoperands
, j
, &newop
);
1632 /* If it transforms from an SSA_NAME to an address, fold with
1633 a preceding indirect reference. */
1634 if (j
> 0 && op0
&& TREE_CODE (op0
) == ADDR_EXPR
1635 && VEC_index (vn_reference_op_s
,
1636 newoperands
, j
- 1)->opcode
== INDIRECT_REF
)
1637 vn_reference_fold_indirect (&newoperands
, &j
);
1639 if (i
!= VEC_length (vn_reference_op_s
, operands
))
1642 VEC_free (vn_reference_op_s
, heap
, newoperands
);
1648 newvuse
= translate_vuse_through_block (newoperands
,
1649 ref
->set
, ref
->type
,
1650 vuse
, phiblock
, pred
);
1651 if (newvuse
== NULL_TREE
)
1653 VEC_free (vn_reference_op_s
, heap
, newoperands
);
1657 changed
|= newvuse
!= vuse
;
1661 unsigned int new_val_id
;
1664 tree result
= vn_reference_lookup_pieces (newvuse
, ref
->set
,
1669 VEC_free (vn_reference_op_s
, heap
, newoperands
);
1671 if (result
&& is_gimple_min_invariant (result
))
1673 gcc_assert (!newoperands
);
1674 return get_or_alloc_expr_for_constant (result
);
1677 expr
= (pre_expr
) pool_alloc (pre_expr_pool
);
1678 expr
->kind
= REFERENCE
;
1683 PRE_EXPR_REFERENCE (expr
) = newref
;
1684 constant
= fully_constant_expression (expr
);
1685 if (constant
!= expr
)
1688 new_val_id
= newref
->value_id
;
1689 get_or_alloc_expression_id (expr
);
1693 new_val_id
= get_next_value_id ();
1694 VEC_safe_grow_cleared (bitmap_set_t
, heap
, value_expressions
,
1695 get_max_value_id() + 1);
1696 newref
= vn_reference_insert_pieces (newvuse
, ref
->set
,
1699 result
, new_val_id
);
1701 PRE_EXPR_REFERENCE (expr
) = newref
;
1702 constant
= fully_constant_expression (expr
);
1703 if (constant
!= expr
)
1705 get_or_alloc_expression_id (expr
);
1707 add_to_value (new_val_id
, expr
);
1709 VEC_free (vn_reference_op_s
, heap
, newoperands
);
1710 phi_trans_add (oldexpr
, expr
, pred
);
1720 tree name
= PRE_EXPR_NAME (expr
);
1722 def_stmt
= SSA_NAME_DEF_STMT (name
);
1723 if (gimple_code (def_stmt
) == GIMPLE_PHI
1724 && gimple_bb (def_stmt
) == phiblock
)
1729 e
= find_edge (pred
, gimple_bb (phi
));
1732 tree def
= PHI_ARG_DEF (phi
, e
->dest_idx
);
1735 if (TREE_CODE (def
) == SSA_NAME
)
1736 def
= VN_INFO (def
)->valnum
;
1738 /* Handle constant. */
1739 if (is_gimple_min_invariant (def
))
1740 return get_or_alloc_expr_for_constant (def
);
1742 if (TREE_CODE (def
) == SSA_NAME
&& ssa_undefined_value_p (def
))
1745 newexpr
= get_or_alloc_expr_for_name (def
);
1756 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1758 Return NULL if we can't find a leader for each part of the
1759 translated expression. */
1762 phi_translate (pre_expr expr
, bitmap_set_t set1
, bitmap_set_t set2
,
1763 basic_block pred
, basic_block phiblock
)
1765 bitmap_clear (seen_during_translate
);
1766 return phi_translate_1 (expr
, set1
, set2
, pred
, phiblock
,
1767 seen_during_translate
);
1770 /* For each expression in SET, translate the values through phi nodes
1771 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1772 expressions in DEST. */
1775 phi_translate_set (bitmap_set_t dest
, bitmap_set_t set
, basic_block pred
,
1776 basic_block phiblock
)
1778 VEC (pre_expr
, heap
) *exprs
;
1782 if (!phi_nodes (phiblock
))
1784 bitmap_set_copy (dest
, set
);
1788 exprs
= sorted_array_from_bitmap_set (set
);
1789 for (i
= 0; VEC_iterate (pre_expr
, exprs
, i
, expr
); i
++)
1791 pre_expr translated
;
1792 translated
= phi_translate (expr
, set
, NULL
, pred
, phiblock
);
1794 /* Don't add empty translations to the cache */
1796 phi_trans_add (expr
, translated
, pred
);
1798 if (translated
!= NULL
)
1799 bitmap_value_insert_into_set (dest
, translated
);
1801 VEC_free (pre_expr
, heap
, exprs
);
1804 /* Find the leader for a value (i.e., the name representing that
1805 value) in a given set, and return it. If STMT is non-NULL it
1806 makes sure the defining statement for the leader dominates it.
1807 Return NULL if no leader is found. */
1810 bitmap_find_leader (bitmap_set_t set
, unsigned int val
, gimple stmt
)
1812 if (value_id_constant_p (val
))
1816 bitmap_set_t exprset
= VEC_index (bitmap_set_t
, value_expressions
, val
);
1818 FOR_EACH_EXPR_ID_IN_SET (exprset
, i
, bi
)
1820 pre_expr expr
= expression_for_id (i
);
1821 if (expr
->kind
== CONSTANT
)
1825 if (bitmap_set_contains_value (set
, val
))
1827 /* Rather than walk the entire bitmap of expressions, and see
1828 whether any of them has the value we are looking for, we look
1829 at the reverse mapping, which tells us the set of expressions
1830 that have a given value (IE value->expressions with that
1831 value) and see if any of those expressions are in our set.
1832 The number of expressions per value is usually significantly
1833 less than the number of expressions in the set. In fact, for
1834 large testcases, doing it this way is roughly 5-10x faster
1835 than walking the bitmap.
1836 If this is somehow a significant lose for some cases, we can
1837 choose which set to walk based on which set is smaller. */
1840 bitmap_set_t exprset
= VEC_index (bitmap_set_t
, value_expressions
, val
);
1842 EXECUTE_IF_AND_IN_BITMAP (exprset
->expressions
,
1843 set
->expressions
, 0, i
, bi
)
1845 pre_expr val
= expression_for_id (i
);
1846 /* At the point where stmt is not null, there should always
1847 be an SSA_NAME first in the list of expressions. */
1850 gimple def_stmt
= SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val
));
1851 if (gimple_code (def_stmt
) != GIMPLE_PHI
1852 && gimple_bb (def_stmt
) == gimple_bb (stmt
)
1853 && gimple_uid (def_stmt
) >= gimple_uid (stmt
))
1862 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1863 BLOCK by seeing if it is not killed in the block. Note that we are
1864 only determining whether there is a store that kills it. Because
1865 of the order in which clean iterates over values, we are guaranteed
1866 that altered operands will have caused us to be eliminated from the
1867 ANTIC_IN set already. */
1870 value_dies_in_block_x (pre_expr expr
, basic_block block
)
1872 tree vuse
= PRE_EXPR_REFERENCE (expr
)->vuse
;
1873 vn_reference_t refx
= PRE_EXPR_REFERENCE (expr
);
1875 gimple_stmt_iterator gsi
;
1876 unsigned id
= get_expression_id (expr
);
1883 /* Lookup a previously calculated result. */
1884 if (EXPR_DIES (block
)
1885 && bitmap_bit_p (EXPR_DIES (block
), id
* 2))
1886 return bitmap_bit_p (EXPR_DIES (block
), id
* 2 + 1);
1888 /* A memory expression {e, VUSE} dies in the block if there is a
1889 statement that may clobber e. If, starting statement walk from the
1890 top of the basic block, a statement uses VUSE there can be no kill
1891 inbetween that use and the original statement that loaded {e, VUSE},
1892 so we can stop walking. */
1893 ref
.base
= NULL_TREE
;
1894 for (gsi
= gsi_start_bb (block
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1896 tree def_vuse
, def_vdef
;
1897 def
= gsi_stmt (gsi
);
1898 def_vuse
= gimple_vuse (def
);
1899 def_vdef
= gimple_vdef (def
);
1901 /* Not a memory statement. */
1905 /* Not a may-def. */
1908 /* A load with the same VUSE, we're done. */
1909 if (def_vuse
== vuse
)
1915 /* Init ref only if we really need it. */
1916 if (ref
.base
== NULL_TREE
1917 && !ao_ref_init_from_vn_reference (&ref
, refx
->set
, refx
->type
,
1923 /* If the statement may clobber expr, it dies. */
1924 if (stmt_may_clobber_ref_p_1 (def
, &ref
))
1931 /* Remember the result. */
1932 if (!EXPR_DIES (block
))
1933 EXPR_DIES (block
) = BITMAP_ALLOC (&grand_bitmap_obstack
);
1934 bitmap_set_bit (EXPR_DIES (block
), id
* 2);
1936 bitmap_set_bit (EXPR_DIES (block
), id
* 2 + 1);
1942 #define union_contains_value(SET1, SET2, VAL) \
1943 (bitmap_set_contains_value ((SET1), (VAL)) \
1944 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
1946 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
1949 vro_valid_in_sets (bitmap_set_t set1
, bitmap_set_t set2
,
1950 vn_reference_op_t vro
)
1952 if (vro
->op0
&& TREE_CODE (vro
->op0
) == SSA_NAME
)
1954 struct pre_expr_d temp
;
1957 PRE_EXPR_NAME (&temp
) = vro
->op0
;
1958 temp
.id
= lookup_expression_id (&temp
);
1961 if (!union_contains_value (set1
, set2
,
1962 get_expr_value_id (&temp
)))
1965 if (vro
->op1
&& TREE_CODE (vro
->op1
) == SSA_NAME
)
1967 struct pre_expr_d temp
;
1970 PRE_EXPR_NAME (&temp
) = vro
->op1
;
1971 temp
.id
= lookup_expression_id (&temp
);
1974 if (!union_contains_value (set1
, set2
,
1975 get_expr_value_id (&temp
)))
1979 if (vro
->op2
&& TREE_CODE (vro
->op2
) == SSA_NAME
)
1981 struct pre_expr_d temp
;
1984 PRE_EXPR_NAME (&temp
) = vro
->op2
;
1985 temp
.id
= lookup_expression_id (&temp
);
1988 if (!union_contains_value (set1
, set2
,
1989 get_expr_value_id (&temp
)))
1996 /* Determine if the expression EXPR is valid in SET1 U SET2.
1997 ONLY SET2 CAN BE NULL.
1998 This means that we have a leader for each part of the expression
1999 (if it consists of values), or the expression is an SSA_NAME.
2000 For loads/calls, we also see if the vuse is killed in this block. */
2003 valid_in_sets (bitmap_set_t set1
, bitmap_set_t set2
, pre_expr expr
,
2009 return bitmap_set_contains_expr (AVAIL_OUT (block
), expr
);
2013 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
2014 for (i
= 0; i
< nary
->length
; i
++)
2016 if (TREE_CODE (nary
->op
[i
]) == SSA_NAME
)
2018 struct pre_expr_d temp
;
2021 PRE_EXPR_NAME (&temp
) = nary
->op
[i
];
2022 temp
.id
= lookup_expression_id (&temp
);
2025 if (!union_contains_value (set1
, set2
,
2026 get_expr_value_id (&temp
)))
2035 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
2036 vn_reference_op_t vro
;
2039 for (i
= 0; VEC_iterate (vn_reference_op_s
, ref
->operands
, i
, vro
); i
++)
2041 if (!vro_valid_in_sets (set1
, set2
, vro
))
2046 gimple def_stmt
= SSA_NAME_DEF_STMT (ref
->vuse
);
2047 if (!gimple_nop_p (def_stmt
)
2048 && gimple_bb (def_stmt
) != block
2049 && !dominated_by_p (CDI_DOMINATORS
,
2050 block
, gimple_bb (def_stmt
)))
2053 return !value_dies_in_block_x (expr
, block
);
2060 /* Clean the set of expressions that are no longer valid in SET1 or
2061 SET2. This means expressions that are made up of values we have no
2062 leaders for in SET1 or SET2. This version is used for partial
2063 anticipation, which means it is not valid in either ANTIC_IN or
2067 dependent_clean (bitmap_set_t set1
, bitmap_set_t set2
, basic_block block
)
2069 VEC (pre_expr
, heap
) *exprs
= sorted_array_from_bitmap_set (set1
);
2073 for (i
= 0; VEC_iterate (pre_expr
, exprs
, i
, expr
); i
++)
2075 if (!valid_in_sets (set1
, set2
, expr
, block
))
2076 bitmap_remove_from_set (set1
, expr
);
2078 VEC_free (pre_expr
, heap
, exprs
);
2081 /* Clean the set of expressions that are no longer valid in SET. This
2082 means expressions that are made up of values we have no leaders for
2086 clean (bitmap_set_t set
, basic_block block
)
2088 VEC (pre_expr
, heap
) *exprs
= sorted_array_from_bitmap_set (set
);
2092 for (i
= 0; VEC_iterate (pre_expr
, exprs
, i
, expr
); i
++)
2094 if (!valid_in_sets (set
, NULL
, expr
, block
))
2095 bitmap_remove_from_set (set
, expr
);
2097 VEC_free (pre_expr
, heap
, exprs
);
2100 static sbitmap has_abnormal_preds
;
2102 /* List of blocks that may have changed during ANTIC computation and
2103 thus need to be iterated over. */
2105 static sbitmap changed_blocks
;
2107 /* Decide whether to defer a block for a later iteration, or PHI
2108 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2109 should defer the block, and true if we processed it. */
2112 defer_or_phi_translate_block (bitmap_set_t dest
, bitmap_set_t source
,
2113 basic_block block
, basic_block phiblock
)
2115 if (!BB_VISITED (phiblock
))
2117 SET_BIT (changed_blocks
, block
->index
);
2118 BB_VISITED (block
) = 0;
2119 BB_DEFERRED (block
) = 1;
2123 phi_translate_set (dest
, source
, block
, phiblock
);
2127 /* Compute the ANTIC set for BLOCK.
2129 If succs(BLOCK) > 1 then
2130 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2131 else if succs(BLOCK) == 1 then
2132 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2134 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2138 compute_antic_aux (basic_block block
, bool block_has_abnormal_pred_edge
)
2140 bool changed
= false;
2141 bitmap_set_t S
, old
, ANTIC_OUT
;
2147 old
= ANTIC_OUT
= S
= NULL
;
2148 BB_VISITED (block
) = 1;
2150 /* If any edges from predecessors are abnormal, antic_in is empty,
2152 if (block_has_abnormal_pred_edge
)
2153 goto maybe_dump_sets
;
2155 old
= ANTIC_IN (block
);
2156 ANTIC_OUT
= bitmap_set_new ();
2158 /* If the block has no successors, ANTIC_OUT is empty. */
2159 if (EDGE_COUNT (block
->succs
) == 0)
2161 /* If we have one successor, we could have some phi nodes to
2162 translate through. */
2163 else if (single_succ_p (block
))
2165 basic_block succ_bb
= single_succ (block
);
2167 /* We trade iterations of the dataflow equations for having to
2168 phi translate the maximal set, which is incredibly slow
2169 (since the maximal set often has 300+ members, even when you
2170 have a small number of blocks).
2171 Basically, we defer the computation of ANTIC for this block
2172 until we have processed it's successor, which will inevitably
2173 have a *much* smaller set of values to phi translate once
2174 clean has been run on it.
2175 The cost of doing this is that we technically perform more
2176 iterations, however, they are lower cost iterations.
2178 Timings for PRE on tramp3d-v4:
2179 without maximal set fix: 11 seconds
2180 with maximal set fix/without deferring: 26 seconds
2181 with maximal set fix/with deferring: 11 seconds
2184 if (!defer_or_phi_translate_block (ANTIC_OUT
, ANTIC_IN (succ_bb
),
2188 goto maybe_dump_sets
;
2191 /* If we have multiple successors, we take the intersection of all of
2192 them. Note that in the case of loop exit phi nodes, we may have
2193 phis to translate through. */
2196 VEC(basic_block
, heap
) * worklist
;
2198 basic_block bprime
, first
;
2200 worklist
= VEC_alloc (basic_block
, heap
, EDGE_COUNT (block
->succs
));
2201 FOR_EACH_EDGE (e
, ei
, block
->succs
)
2202 VEC_quick_push (basic_block
, worklist
, e
->dest
);
2203 first
= VEC_index (basic_block
, worklist
, 0);
2205 if (phi_nodes (first
))
2207 bitmap_set_t from
= ANTIC_IN (first
);
2209 if (!BB_VISITED (first
))
2211 phi_translate_set (ANTIC_OUT
, from
, block
, first
);
2215 if (!BB_VISITED (first
))
2216 bitmap_set_copy (ANTIC_OUT
, maximal_set
);
2218 bitmap_set_copy (ANTIC_OUT
, ANTIC_IN (first
));
2221 for (i
= 1; VEC_iterate (basic_block
, worklist
, i
, bprime
); i
++)
2223 if (phi_nodes (bprime
))
2225 bitmap_set_t tmp
= bitmap_set_new ();
2226 bitmap_set_t from
= ANTIC_IN (bprime
);
2228 if (!BB_VISITED (bprime
))
2230 phi_translate_set (tmp
, from
, block
, bprime
);
2231 bitmap_set_and (ANTIC_OUT
, tmp
);
2232 bitmap_set_free (tmp
);
2236 if (!BB_VISITED (bprime
))
2237 bitmap_set_and (ANTIC_OUT
, maximal_set
);
2239 bitmap_set_and (ANTIC_OUT
, ANTIC_IN (bprime
));
2242 VEC_free (basic_block
, heap
, worklist
);
2245 /* Generate ANTIC_OUT - TMP_GEN. */
2246 S
= bitmap_set_subtract (ANTIC_OUT
, TMP_GEN (block
));
2248 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2249 ANTIC_IN (block
) = bitmap_set_subtract (EXP_GEN (block
),
2252 /* Then union in the ANTIC_OUT - TMP_GEN values,
2253 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2254 FOR_EACH_EXPR_ID_IN_SET (S
, bii
, bi
)
2255 bitmap_value_insert_into_set (ANTIC_IN (block
),
2256 expression_for_id (bii
));
2258 clean (ANTIC_IN (block
), block
);
2260 /* !old->expressions can happen when we deferred a block. */
2261 if (!old
->expressions
|| !bitmap_set_equal (old
, ANTIC_IN (block
)))
2264 SET_BIT (changed_blocks
, block
->index
);
2265 FOR_EACH_EDGE (e
, ei
, block
->preds
)
2266 SET_BIT (changed_blocks
, e
->src
->index
);
2269 RESET_BIT (changed_blocks
, block
->index
);
2272 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2274 if (!BB_DEFERRED (block
) || BB_VISITED (block
))
2277 print_bitmap_set (dump_file
, ANTIC_OUT
, "ANTIC_OUT", block
->index
);
2279 print_bitmap_set (dump_file
, ANTIC_IN (block
), "ANTIC_IN",
2283 print_bitmap_set (dump_file
, S
, "S", block
->index
);
2288 "Block %d was deferred for a future iteration.\n",
2293 bitmap_set_free (old
);
2295 bitmap_set_free (S
);
2297 bitmap_set_free (ANTIC_OUT
);
2301 /* Compute PARTIAL_ANTIC for BLOCK.
2303 If succs(BLOCK) > 1 then
2304 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2305 in ANTIC_OUT for all succ(BLOCK)
2306 else if succs(BLOCK) == 1 then
2307 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2309 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2314 compute_partial_antic_aux (basic_block block
,
2315 bool block_has_abnormal_pred_edge
)
2317 bool changed
= false;
2318 bitmap_set_t old_PA_IN
;
2319 bitmap_set_t PA_OUT
;
2322 unsigned long max_pa
= PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH
);
2324 old_PA_IN
= PA_OUT
= NULL
;
2326 /* If any edges from predecessors are abnormal, antic_in is empty,
2328 if (block_has_abnormal_pred_edge
)
2329 goto maybe_dump_sets
;
2331 /* If there are too many partially anticipatable values in the
2332 block, phi_translate_set can take an exponential time: stop
2333 before the translation starts. */
2335 && single_succ_p (block
)
2336 && bitmap_count_bits (PA_IN (single_succ (block
))->values
) > max_pa
)
2337 goto maybe_dump_sets
;
2339 old_PA_IN
= PA_IN (block
);
2340 PA_OUT
= bitmap_set_new ();
2342 /* If the block has no successors, ANTIC_OUT is empty. */
2343 if (EDGE_COUNT (block
->succs
) == 0)
2345 /* If we have one successor, we could have some phi nodes to
2346 translate through. Note that we can't phi translate across DFS
2347 back edges in partial antic, because it uses a union operation on
2348 the successors. For recurrences like IV's, we will end up
2349 generating a new value in the set on each go around (i + 3 (VH.1)
2350 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2351 else if (single_succ_p (block
))
2353 basic_block succ
= single_succ (block
);
2354 if (!(single_succ_edge (block
)->flags
& EDGE_DFS_BACK
))
2355 phi_translate_set (PA_OUT
, PA_IN (succ
), block
, succ
);
2357 /* If we have multiple successors, we take the union of all of
2361 VEC(basic_block
, heap
) * worklist
;
2365 worklist
= VEC_alloc (basic_block
, heap
, EDGE_COUNT (block
->succs
));
2366 FOR_EACH_EDGE (e
, ei
, block
->succs
)
2368 if (e
->flags
& EDGE_DFS_BACK
)
2370 VEC_quick_push (basic_block
, worklist
, e
->dest
);
2372 if (VEC_length (basic_block
, worklist
) > 0)
2374 for (i
= 0; VEC_iterate (basic_block
, worklist
, i
, bprime
); i
++)
2379 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime
), i
, bi
)
2380 bitmap_value_insert_into_set (PA_OUT
,
2381 expression_for_id (i
));
2382 if (phi_nodes (bprime
))
2384 bitmap_set_t pa_in
= bitmap_set_new ();
2385 phi_translate_set (pa_in
, PA_IN (bprime
), block
, bprime
);
2386 FOR_EACH_EXPR_ID_IN_SET (pa_in
, i
, bi
)
2387 bitmap_value_insert_into_set (PA_OUT
,
2388 expression_for_id (i
));
2389 bitmap_set_free (pa_in
);
2392 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime
), i
, bi
)
2393 bitmap_value_insert_into_set (PA_OUT
,
2394 expression_for_id (i
));
2397 VEC_free (basic_block
, heap
, worklist
);
2400 /* PA_IN starts with PA_OUT - TMP_GEN.
2401 Then we subtract things from ANTIC_IN. */
2402 PA_IN (block
) = bitmap_set_subtract (PA_OUT
, TMP_GEN (block
));
2404 /* For partial antic, we want to put back in the phi results, since
2405 we will properly avoid making them partially antic over backedges. */
2406 bitmap_ior_into (PA_IN (block
)->values
, PHI_GEN (block
)->values
);
2407 bitmap_ior_into (PA_IN (block
)->expressions
, PHI_GEN (block
)->expressions
);
2409 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2410 bitmap_set_subtract_values (PA_IN (block
), ANTIC_IN (block
));
2412 dependent_clean (PA_IN (block
), ANTIC_IN (block
), block
);
2414 if (!bitmap_set_equal (old_PA_IN
, PA_IN (block
)))
2417 SET_BIT (changed_blocks
, block
->index
);
2418 FOR_EACH_EDGE (e
, ei
, block
->preds
)
2419 SET_BIT (changed_blocks
, e
->src
->index
);
2422 RESET_BIT (changed_blocks
, block
->index
);
2425 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2428 print_bitmap_set (dump_file
, PA_OUT
, "PA_OUT", block
->index
);
2430 print_bitmap_set (dump_file
, PA_IN (block
), "PA_IN", block
->index
);
2433 bitmap_set_free (old_PA_IN
);
2435 bitmap_set_free (PA_OUT
);
2439 /* Compute ANTIC and partial ANTIC sets. */
2442 compute_antic (void)
2444 bool changed
= true;
2445 int num_iterations
= 0;
2449 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2450 We pre-build the map of blocks with incoming abnormal edges here. */
2451 has_abnormal_preds
= sbitmap_alloc (last_basic_block
);
2452 sbitmap_zero (has_abnormal_preds
);
2459 FOR_EACH_EDGE (e
, ei
, block
->preds
)
2461 e
->flags
&= ~EDGE_DFS_BACK
;
2462 if (e
->flags
& EDGE_ABNORMAL
)
2464 SET_BIT (has_abnormal_preds
, block
->index
);
2469 BB_VISITED (block
) = 0;
2470 BB_DEFERRED (block
) = 0;
2471 /* While we are here, give empty ANTIC_IN sets to each block. */
2472 ANTIC_IN (block
) = bitmap_set_new ();
2473 PA_IN (block
) = bitmap_set_new ();
2476 /* At the exit block we anticipate nothing. */
2477 ANTIC_IN (EXIT_BLOCK_PTR
) = bitmap_set_new ();
2478 BB_VISITED (EXIT_BLOCK_PTR
) = 1;
2479 PA_IN (EXIT_BLOCK_PTR
) = bitmap_set_new ();
2481 changed_blocks
= sbitmap_alloc (last_basic_block
+ 1);
2482 sbitmap_ones (changed_blocks
);
2485 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2486 fprintf (dump_file
, "Starting iteration %d\n", num_iterations
);
2489 for (i
= 0; i
< n_basic_blocks
- NUM_FIXED_BLOCKS
; i
++)
2491 if (TEST_BIT (changed_blocks
, postorder
[i
]))
2493 basic_block block
= BASIC_BLOCK (postorder
[i
]);
2494 changed
|= compute_antic_aux (block
,
2495 TEST_BIT (has_abnormal_preds
,
2499 #ifdef ENABLE_CHECKING
2500 /* Theoretically possible, but *highly* unlikely. */
2501 gcc_assert (num_iterations
< 500);
2505 statistics_histogram_event (cfun
, "compute_antic iterations",
2508 if (do_partial_partial
)
2510 sbitmap_ones (changed_blocks
);
2511 mark_dfs_back_edges ();
2516 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2517 fprintf (dump_file
, "Starting iteration %d\n", num_iterations
);
2520 for (i
= 0; i
< n_basic_blocks
- NUM_FIXED_BLOCKS
; i
++)
2522 if (TEST_BIT (changed_blocks
, postorder
[i
]))
2524 basic_block block
= BASIC_BLOCK (postorder
[i
]);
2526 |= compute_partial_antic_aux (block
,
2527 TEST_BIT (has_abnormal_preds
,
2531 #ifdef ENABLE_CHECKING
2532 /* Theoretically possible, but *highly* unlikely. */
2533 gcc_assert (num_iterations
< 500);
2536 statistics_histogram_event (cfun
, "compute_partial_antic iterations",
2539 sbitmap_free (has_abnormal_preds
);
2540 sbitmap_free (changed_blocks
);
2543 /* Return true if we can value number the call in STMT. This is true
2544 if we have a pure or constant call. */
2547 can_value_number_call (gimple stmt
)
2549 if (gimple_call_flags (stmt
) & (ECF_PURE
| ECF_CONST
))
2554 /* Return true if OP is an exception handler related operation, such as
2555 FILTER_EXPR or EXC_PTR_EXPR. */
2558 is_exception_related (gimple stmt
)
2560 return (is_gimple_assign (stmt
)
2561 && (gimple_assign_rhs_code (stmt
) == FILTER_EXPR
2562 || gimple_assign_rhs_code (stmt
) == EXC_PTR_EXPR
));
2565 /* Return true if OP is a tree which we can perform PRE on.
2566 This may not match the operations we can value number, but in
2567 a perfect world would. */
2570 can_PRE_operation (tree op
)
2572 return UNARY_CLASS_P (op
)
2573 || BINARY_CLASS_P (op
)
2574 || COMPARISON_CLASS_P (op
)
2575 || TREE_CODE (op
) == INDIRECT_REF
2576 || TREE_CODE (op
) == COMPONENT_REF
2577 || TREE_CODE (op
) == VIEW_CONVERT_EXPR
2578 || TREE_CODE (op
) == CALL_EXPR
2579 || TREE_CODE (op
) == ARRAY_REF
;
2583 /* Inserted expressions are placed onto this worklist, which is used
2584 for performing quick dead code elimination of insertions we made
2585 that didn't turn out to be necessary. */
2586 static VEC(gimple
,heap
) *inserted_exprs
;
2587 static bitmap inserted_phi_names
;
2589 /* Pool allocated fake store expressions are placed onto this
2590 worklist, which, after performing dead code elimination, is walked
2591 to see which expressions need to be put into GC'able memory */
2592 static VEC(gimple
, heap
) *need_creation
;
2594 /* The actual worker for create_component_ref_by_pieces. */
2597 create_component_ref_by_pieces_1 (basic_block block
, vn_reference_t ref
,
2598 unsigned int *operand
, gimple_seq
*stmts
,
2601 vn_reference_op_t currop
= VEC_index (vn_reference_op_s
, ref
->operands
,
2605 switch (currop
->opcode
)
2609 tree folded
, sc
= currop
->op1
;
2610 unsigned int nargs
= 0;
2611 tree
*args
= XNEWVEC (tree
, VEC_length (vn_reference_op_s
,
2612 ref
->operands
) - 1);
2613 while (*operand
< VEC_length (vn_reference_op_s
, ref
->operands
))
2615 args
[nargs
] = create_component_ref_by_pieces_1 (block
, ref
,
2620 folded
= build_call_array (currop
->type
,
2621 TREE_CODE (currop
->op0
) == FUNCTION_DECL
2622 ? build_fold_addr_expr (currop
->op0
)
2628 pre_expr scexpr
= get_or_alloc_expr_for (sc
);
2629 sc
= find_or_generate_expression (block
, scexpr
, stmts
, domstmt
);
2632 CALL_EXPR_STATIC_CHAIN (folded
) = sc
;
2637 case TARGET_MEM_REF
:
2639 vn_reference_op_t nextop
= VEC_index (vn_reference_op_s
, ref
->operands
,
2642 tree genop0
= NULL_TREE
;
2643 tree baseop
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2649 op0expr
= get_or_alloc_expr_for (currop
->op0
);
2650 genop0
= find_or_generate_expression (block
, op0expr
,
2655 if (DECL_P (baseop
))
2656 return build6 (TARGET_MEM_REF
, currop
->type
,
2658 genop0
, currop
->op1
, currop
->op2
,
2659 unshare_expr (nextop
->op1
));
2661 return build6 (TARGET_MEM_REF
, currop
->type
,
2663 genop0
, currop
->op1
, currop
->op2
,
2664 unshare_expr (nextop
->op1
));
2670 gcc_assert (is_gimple_min_invariant (currop
->op0
));
2676 case VIEW_CONVERT_EXPR
:
2679 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
,
2684 folded
= fold_build1 (currop
->opcode
, currop
->type
,
2689 case ALIGN_INDIRECT_REF
:
2690 case MISALIGNED_INDIRECT_REF
:
2694 tree genop1
= create_component_ref_by_pieces_1 (block
, ref
,
2699 genop1
= fold_convert (build_pointer_type (currop
->type
),
2702 if (currop
->opcode
== MISALIGNED_INDIRECT_REF
)
2703 folded
= fold_build2 (currop
->opcode
, currop
->type
,
2704 genop1
, currop
->op1
);
2706 folded
= fold_build1 (currop
->opcode
, currop
->type
,
2714 tree genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2716 pre_expr op1expr
= get_or_alloc_expr_for (currop
->op0
);
2717 pre_expr op2expr
= get_or_alloc_expr_for (currop
->op1
);
2723 genop1
= find_or_generate_expression (block
, op1expr
, stmts
, domstmt
);
2726 genop2
= find_or_generate_expression (block
, op2expr
, stmts
, domstmt
);
2729 folded
= fold_build3 (BIT_FIELD_REF
, currop
->type
, genop0
, genop1
,
2734 /* For array ref vn_reference_op's, operand 1 of the array ref
2735 is op0 of the reference op and operand 3 of the array ref is
2737 case ARRAY_RANGE_REF
:
2741 tree genop1
= currop
->op0
;
2743 tree genop2
= currop
->op1
;
2746 genop0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2750 op1expr
= get_or_alloc_expr_for (genop1
);
2751 genop1
= find_or_generate_expression (block
, op1expr
, stmts
, domstmt
);
2756 op2expr
= get_or_alloc_expr_for (genop2
);
2757 genop2
= find_or_generate_expression (block
, op2expr
, stmts
,
2763 genop3
= currop
->op2
;
2764 return build4 (currop
->opcode
, currop
->type
, genop0
, genop1
,
2771 tree genop2
= currop
->op1
;
2773 op0
= create_component_ref_by_pieces_1 (block
, ref
, operand
,
2777 /* op1 should be a FIELD_DECL, which are represented by
2782 op2expr
= get_or_alloc_expr_for (genop2
);
2783 genop2
= find_or_generate_expression (block
, op2expr
, stmts
,
2789 return fold_build3 (COMPONENT_REF
, TREE_TYPE (op1
), op0
, op1
,
2795 pre_expr op0expr
= get_or_alloc_expr_for (currop
->op0
);
2796 genop
= find_or_generate_expression (block
, op0expr
, stmts
, domstmt
);
2817 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2818 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2819 trying to rename aggregates into ssa form directly, which is a no no.
2821 Thus, this routine doesn't create temporaries, it just builds a
2822 single access expression for the array, calling
2823 find_or_generate_expression to build the innermost pieces.
2825 This function is a subroutine of create_expression_by_pieces, and
2826 should not be called on it's own unless you really know what you
2830 create_component_ref_by_pieces (basic_block block
, vn_reference_t ref
,
2831 gimple_seq
*stmts
, gimple domstmt
)
2833 unsigned int op
= 0;
2834 return create_component_ref_by_pieces_1 (block
, ref
, &op
, stmts
, domstmt
);
2837 /* Find a leader for an expression, or generate one using
2838 create_expression_by_pieces if it's ANTIC but
2840 BLOCK is the basic_block we are looking for leaders in.
2841 EXPR is the expression to find a leader or generate for.
2842 STMTS is the statement list to put the inserted expressions on.
2843 Returns the SSA_NAME of the LHS of the generated expression or the
2845 DOMSTMT if non-NULL is a statement that should be dominated by
2846 all uses in the generated expression. If DOMSTMT is non-NULL this
2847 routine can fail and return NULL_TREE. Otherwise it will assert
2851 find_or_generate_expression (basic_block block
, pre_expr expr
,
2852 gimple_seq
*stmts
, gimple domstmt
)
2854 pre_expr leader
= bitmap_find_leader (AVAIL_OUT (block
),
2855 get_expr_value_id (expr
), domstmt
);
2859 if (leader
->kind
== NAME
)
2860 genop
= PRE_EXPR_NAME (leader
);
2861 else if (leader
->kind
== CONSTANT
)
2862 genop
= PRE_EXPR_CONSTANT (leader
);
2865 /* If it's still NULL, it must be a complex expression, so generate
2866 it recursively. Not so for FRE though. */
2870 bitmap_set_t exprset
;
2871 unsigned int lookfor
= get_expr_value_id (expr
);
2872 bool handled
= false;
2876 exprset
= VEC_index (bitmap_set_t
, value_expressions
, lookfor
);
2877 FOR_EACH_EXPR_ID_IN_SET (exprset
, i
, bi
)
2879 pre_expr temp
= expression_for_id (i
);
2880 if (temp
->kind
!= NAME
)
2883 genop
= create_expression_by_pieces (block
, temp
, stmts
,
2885 get_expr_type (expr
));
2889 if (!handled
&& domstmt
)
2892 gcc_assert (handled
);
2897 #define NECESSARY GF_PLF_1
2899 /* Create an expression in pieces, so that we can handle very complex
2900 expressions that may be ANTIC, but not necessary GIMPLE.
2901 BLOCK is the basic block the expression will be inserted into,
2902 EXPR is the expression to insert (in value form)
2903 STMTS is a statement list to append the necessary insertions into.
2905 This function will die if we hit some value that shouldn't be
2906 ANTIC but is (IE there is no leader for it, or its components).
2907 This function may also generate expressions that are themselves
2908 partially or fully redundant. Those that are will be either made
2909 fully redundant during the next iteration of insert (for partially
2910 redundant ones), or eliminated by eliminate (for fully redundant
2913 If DOMSTMT is non-NULL then we make sure that all uses in the
2914 expressions dominate that statement. In this case the function
2915 can return NULL_TREE to signal failure. */
2918 create_expression_by_pieces (basic_block block
, pre_expr expr
,
2919 gimple_seq
*stmts
, gimple domstmt
, tree type
)
2923 gimple_seq forced_stmts
= NULL
;
2924 unsigned int value_id
;
2925 gimple_stmt_iterator gsi
;
2926 tree exprtype
= type
? type
: get_expr_type (expr
);
2932 /* We may hit the NAME/CONSTANT case if we have to convert types
2933 that value numbering saw through. */
2935 folded
= PRE_EXPR_NAME (expr
);
2938 folded
= PRE_EXPR_CONSTANT (expr
);
2942 vn_reference_t ref
= PRE_EXPR_REFERENCE (expr
);
2943 folded
= create_component_ref_by_pieces (block
, ref
, stmts
, domstmt
);
2948 vn_nary_op_t nary
= PRE_EXPR_NARY (expr
);
2949 switch (nary
->length
)
2953 pre_expr op1
= get_or_alloc_expr_for (nary
->op
[0]);
2954 pre_expr op2
= get_or_alloc_expr_for (nary
->op
[1]);
2955 tree genop1
= find_or_generate_expression (block
, op1
,
2957 tree genop2
= find_or_generate_expression (block
, op2
,
2959 if (!genop1
|| !genop2
)
2961 genop1
= fold_convert (TREE_TYPE (nary
->op
[0]),
2963 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
2964 may be a constant with the wrong type. */
2965 if (nary
->opcode
== POINTER_PLUS_EXPR
)
2966 genop2
= fold_convert (sizetype
, genop2
);
2968 genop2
= fold_convert (TREE_TYPE (nary
->op
[1]), genop2
);
2970 folded
= fold_build2 (nary
->opcode
, nary
->type
,
2976 pre_expr op1
= get_or_alloc_expr_for (nary
->op
[0]);
2977 tree genop1
= find_or_generate_expression (block
, op1
,
2981 genop1
= fold_convert (TREE_TYPE (nary
->op
[0]), genop1
);
2983 folded
= fold_build1 (nary
->opcode
, nary
->type
,
2996 if (!useless_type_conversion_p (exprtype
, TREE_TYPE (folded
)))
2997 folded
= fold_convert (exprtype
, folded
);
2999 /* Force the generated expression to be a sequence of GIMPLE
3001 We have to call unshare_expr because force_gimple_operand may
3002 modify the tree we pass to it. */
3003 folded
= force_gimple_operand (unshare_expr (folded
), &forced_stmts
,
3006 /* If we have any intermediate expressions to the value sets, add them
3007 to the value sets and chain them in the instruction stream. */
3010 gsi
= gsi_start (forced_stmts
);
3011 for (; !gsi_end_p (gsi
); gsi_next (&gsi
))
3013 gimple stmt
= gsi_stmt (gsi
);
3014 tree forcedname
= gimple_get_lhs (stmt
);
3017 VEC_safe_push (gimple
, heap
, inserted_exprs
, stmt
);
3018 if (TREE_CODE (forcedname
) == SSA_NAME
)
3020 VN_INFO_GET (forcedname
)->valnum
= forcedname
;
3021 VN_INFO (forcedname
)->value_id
= get_next_value_id ();
3022 nameexpr
= get_or_alloc_expr_for_name (forcedname
);
3023 add_to_value (VN_INFO (forcedname
)->value_id
, nameexpr
);
3025 bitmap_value_replace_in_set (NEW_SETS (block
), nameexpr
);
3026 bitmap_value_replace_in_set (AVAIL_OUT (block
), nameexpr
);
3028 mark_symbols_for_renaming (stmt
);
3030 gimple_seq_add_seq (stmts
, forced_stmts
);
3033 /* Build and insert the assignment of the end result to the temporary
3034 that we will return. */
3035 if (!pretemp
|| exprtype
!= TREE_TYPE (pretemp
))
3037 pretemp
= create_tmp_var (exprtype
, "pretmp");
3038 get_var_ann (pretemp
);
3042 add_referenced_var (temp
);
3044 if (TREE_CODE (exprtype
) == COMPLEX_TYPE
3045 || TREE_CODE (exprtype
) == VECTOR_TYPE
)
3046 DECL_GIMPLE_REG_P (temp
) = 1;
3048 newstmt
= gimple_build_assign (temp
, folded
);
3049 name
= make_ssa_name (temp
, newstmt
);
3050 gimple_assign_set_lhs (newstmt
, name
);
3051 gimple_set_plf (newstmt
, NECESSARY
, false);
3053 gimple_seq_add_stmt (stmts
, newstmt
);
3054 VEC_safe_push (gimple
, heap
, inserted_exprs
, newstmt
);
3056 /* All the symbols in NEWEXPR should be put into SSA form. */
3057 mark_symbols_for_renaming (newstmt
);
3059 /* Add a value number to the temporary.
3060 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3061 we are creating the expression by pieces, and this particular piece of
3062 the expression may have been represented. There is no harm in replacing
3064 VN_INFO_GET (name
)->valnum
= name
;
3065 value_id
= get_expr_value_id (expr
);
3066 VN_INFO (name
)->value_id
= value_id
;
3067 nameexpr
= get_or_alloc_expr_for_name (name
);
3068 add_to_value (value_id
, nameexpr
);
3070 bitmap_value_replace_in_set (NEW_SETS (block
), nameexpr
);
3071 bitmap_value_replace_in_set (AVAIL_OUT (block
), nameexpr
);
3073 pre_stats
.insertions
++;
3074 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3076 fprintf (dump_file
, "Inserted ");
3077 print_gimple_stmt (dump_file
, newstmt
, 0, 0);
3078 fprintf (dump_file
, " in predecessor %d\n", block
->index
);
3085 /* Returns true if we want to inhibit the insertions of PHI nodes
3086 for the given EXPR for basic block BB (a member of a loop).
3087 We want to do this, when we fear that the induction variable we
3088 create might inhibit vectorization. */
3091 inhibit_phi_insertion (basic_block bb
, pre_expr expr
)
3093 vn_reference_t vr
= PRE_EXPR_REFERENCE (expr
);
3094 VEC (vn_reference_op_s
, heap
) *ops
= vr
->operands
;
3095 vn_reference_op_t op
;
3098 /* If we aren't going to vectorize we don't inhibit anything. */
3099 if (!flag_tree_vectorize
)
3102 /* Otherwise we inhibit the insertion when the address of the
3103 memory reference is a simple induction variable. In other
3104 cases the vectorizer won't do anything anyway (either it's
3105 loop invariant or a complicated expression). */
3106 for (i
= 0; VEC_iterate (vn_reference_op_s
, ops
, i
, op
); ++i
)
3111 case ARRAY_RANGE_REF
:
3112 if (TREE_CODE (op
->op0
) != SSA_NAME
)
3117 basic_block defbb
= gimple_bb (SSA_NAME_DEF_STMT (op
->op0
));
3119 /* Default defs are loop invariant. */
3122 /* Defined outside this loop, also loop invariant. */
3123 if (!flow_bb_inside_loop_p (bb
->loop_father
, defbb
))
3125 /* If it's a simple induction variable inhibit insertion,
3126 the vectorizer might be interested in this one. */
3127 if (simple_iv (bb
->loop_father
, bb
->loop_father
,
3128 op
->op0
, &iv
, true))
3130 /* No simple IV, vectorizer can't do anything, hence no
3131 reason to inhibit the transformation for this operand. */
3141 /* Insert the to-be-made-available values of expression EXPRNUM for each
3142 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3143 merge the result with a phi node, given the same value number as
3144 NODE. Return true if we have inserted new stuff. */
3147 insert_into_preds_of_block (basic_block block
, unsigned int exprnum
,
3150 pre_expr expr
= expression_for_id (exprnum
);
3152 unsigned int val
= get_expr_value_id (expr
);
3154 bool insertions
= false;
3159 tree type
= get_expr_type (expr
);
3163 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3165 fprintf (dump_file
, "Found partial redundancy for expression ");
3166 print_pre_expr (dump_file
, expr
);
3167 fprintf (dump_file
, " (%04d)\n", val
);
3170 /* Make sure we aren't creating an induction variable. */
3171 if (block
->loop_depth
> 0 && EDGE_COUNT (block
->preds
) == 2)
3173 bool firstinsideloop
= false;
3174 bool secondinsideloop
= false;
3175 firstinsideloop
= flow_bb_inside_loop_p (block
->loop_father
,
3176 EDGE_PRED (block
, 0)->src
);
3177 secondinsideloop
= flow_bb_inside_loop_p (block
->loop_father
,
3178 EDGE_PRED (block
, 1)->src
);
3179 /* Induction variables only have one edge inside the loop. */
3180 if ((firstinsideloop
^ secondinsideloop
)
3181 && (expr
->kind
!= REFERENCE
3182 || inhibit_phi_insertion (block
, expr
)))
3184 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3185 fprintf (dump_file
, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3190 /* Make sure we are not inserting trapping expressions. */
3191 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3194 eprime
= avail
[bprime
->index
];
3195 if (eprime
->kind
== NARY
3196 && vn_nary_may_trap (PRE_EXPR_NARY (eprime
)))
3200 /* Make the necessary insertions. */
3201 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3203 gimple_seq stmts
= NULL
;
3206 eprime
= avail
[bprime
->index
];
3208 if (eprime
->kind
!= NAME
&& eprime
->kind
!= CONSTANT
)
3210 builtexpr
= create_expression_by_pieces (bprime
,
3214 gcc_assert (!(pred
->flags
& EDGE_ABNORMAL
));
3215 gsi_insert_seq_on_edge (pred
, stmts
);
3216 avail
[bprime
->index
] = get_or_alloc_expr_for_name (builtexpr
);
3219 else if (eprime
->kind
== CONSTANT
)
3221 /* Constants may not have the right type, fold_convert
3222 should give us back a constant with the right type.
3224 tree constant
= PRE_EXPR_CONSTANT (eprime
);
3225 if (!useless_type_conversion_p (type
, TREE_TYPE (constant
)))
3227 tree builtexpr
= fold_convert (type
, constant
);
3228 if (!is_gimple_min_invariant (builtexpr
))
3230 tree forcedexpr
= force_gimple_operand (builtexpr
,
3233 if (!is_gimple_min_invariant (forcedexpr
))
3235 if (forcedexpr
!= builtexpr
)
3237 VN_INFO_GET (forcedexpr
)->valnum
= PRE_EXPR_CONSTANT (eprime
);
3238 VN_INFO (forcedexpr
)->value_id
= get_expr_value_id (eprime
);
3242 gimple_stmt_iterator gsi
;
3243 gsi
= gsi_start (stmts
);
3244 for (; !gsi_end_p (gsi
); gsi_next (&gsi
))
3246 gimple stmt
= gsi_stmt (gsi
);
3247 VEC_safe_push (gimple
, heap
, inserted_exprs
, stmt
);
3248 gimple_set_plf (stmt
, NECESSARY
, false);
3250 gsi_insert_seq_on_edge (pred
, stmts
);
3252 avail
[bprime
->index
] = get_or_alloc_expr_for_name (forcedexpr
);
3257 else if (eprime
->kind
== NAME
)
3259 /* We may have to do a conversion because our value
3260 numbering can look through types in certain cases, but
3261 our IL requires all operands of a phi node have the same
3263 tree name
= PRE_EXPR_NAME (eprime
);
3264 if (!useless_type_conversion_p (type
, TREE_TYPE (name
)))
3268 builtexpr
= fold_convert (type
, name
);
3269 forcedexpr
= force_gimple_operand (builtexpr
,
3273 if (forcedexpr
!= name
)
3275 VN_INFO_GET (forcedexpr
)->valnum
= VN_INFO (name
)->valnum
;
3276 VN_INFO (forcedexpr
)->value_id
= VN_INFO (name
)->value_id
;
3281 gimple_stmt_iterator gsi
;
3282 gsi
= gsi_start (stmts
);
3283 for (; !gsi_end_p (gsi
); gsi_next (&gsi
))
3285 gimple stmt
= gsi_stmt (gsi
);
3286 VEC_safe_push (gimple
, heap
, inserted_exprs
, stmt
);
3287 gimple_set_plf (stmt
, NECESSARY
, false);
3289 gsi_insert_seq_on_edge (pred
, stmts
);
3291 avail
[bprime
->index
] = get_or_alloc_expr_for_name (forcedexpr
);
3295 /* If we didn't want a phi node, and we made insertions, we still have
3296 inserted new stuff, and thus return true. If we didn't want a phi node,
3297 and didn't make insertions, we haven't added anything new, so return
3299 if (nophi
&& insertions
)
3301 else if (nophi
&& !insertions
)
3304 /* Now build a phi for the new variable. */
3305 if (!prephitemp
|| TREE_TYPE (prephitemp
) != type
)
3307 prephitemp
= create_tmp_var (type
, "prephitmp");
3308 get_var_ann (prephitemp
);
3312 add_referenced_var (temp
);
3314 if (TREE_CODE (type
) == COMPLEX_TYPE
3315 || TREE_CODE (type
) == VECTOR_TYPE
)
3316 DECL_GIMPLE_REG_P (temp
) = 1;
3317 phi
= create_phi_node (temp
, block
);
3319 gimple_set_plf (phi
, NECESSARY
, false);
3320 VN_INFO_GET (gimple_phi_result (phi
))->valnum
= gimple_phi_result (phi
);
3321 VN_INFO (gimple_phi_result (phi
))->value_id
= val
;
3322 VEC_safe_push (gimple
, heap
, inserted_exprs
, phi
);
3323 bitmap_set_bit (inserted_phi_names
,
3324 SSA_NAME_VERSION (gimple_phi_result (phi
)));
3325 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3327 pre_expr ae
= avail
[pred
->src
->index
];
3328 gcc_assert (get_expr_type (ae
) == type
3329 || useless_type_conversion_p (type
, get_expr_type (ae
)));
3330 if (ae
->kind
== CONSTANT
)
3331 add_phi_arg (phi
, PRE_EXPR_CONSTANT (ae
), pred
);
3333 add_phi_arg (phi
, PRE_EXPR_NAME (avail
[pred
->src
->index
]), pred
);
3336 newphi
= get_or_alloc_expr_for_name (gimple_phi_result (phi
));
3337 add_to_value (val
, newphi
);
3339 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3340 this insertion, since we test for the existence of this value in PHI_GEN
3341 before proceeding with the partial redundancy checks in insert_aux.
3343 The value may exist in AVAIL_OUT, in particular, it could be represented
3344 by the expression we are trying to eliminate, in which case we want the
3345 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3348 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3349 this block, because if it did, it would have existed in our dominator's
3350 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3353 bitmap_insert_into_set (PHI_GEN (block
), newphi
);
3354 bitmap_value_replace_in_set (AVAIL_OUT (block
),
3356 bitmap_insert_into_set (NEW_SETS (block
),
3359 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3361 fprintf (dump_file
, "Created phi ");
3362 print_gimple_stmt (dump_file
, phi
, 0, 0);
3363 fprintf (dump_file
, " in block %d\n", block
->index
);
3371 /* Perform insertion of partially redundant values.
3372 For BLOCK, do the following:
3373 1. Propagate the NEW_SETS of the dominator into the current block.
3374 If the block has multiple predecessors,
3375 2a. Iterate over the ANTIC expressions for the block to see if
3376 any of them are partially redundant.
3377 2b. If so, insert them into the necessary predecessors to make
3378 the expression fully redundant.
3379 2c. Insert a new PHI merging the values of the predecessors.
3380 2d. Insert the new PHI, and the new expressions, into the
3382 3. Recursively call ourselves on the dominator children of BLOCK.
3384 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3385 do_regular_insertion and do_partial_insertion.
3390 do_regular_insertion (basic_block block
, basic_block dom
)
3392 bool new_stuff
= false;
3393 VEC (pre_expr
, heap
) *exprs
= sorted_array_from_bitmap_set (ANTIC_IN (block
));
3397 for (i
= 0; VEC_iterate (pre_expr
, exprs
, i
, expr
); i
++)
3399 if (expr
->kind
!= NAME
)
3403 bool by_some
= false;
3404 bool cant_insert
= false;
3405 bool all_same
= true;
3406 pre_expr first_s
= NULL
;
3409 pre_expr eprime
= NULL
;
3411 pre_expr edoubleprime
= NULL
;
3413 val
= get_expr_value_id (expr
);
3414 if (bitmap_set_contains_value (PHI_GEN (block
), val
))
3416 if (bitmap_set_contains_value (AVAIL_OUT (dom
), val
))
3418 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3419 fprintf (dump_file
, "Found fully redundant value\n");
3423 avail
= XCNEWVEC (pre_expr
, last_basic_block
);
3424 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3426 unsigned int vprime
;
3428 /* We should never run insertion for the exit block
3429 and so not come across fake pred edges. */
3430 gcc_assert (!(pred
->flags
& EDGE_FAKE
));
3432 eprime
= phi_translate (expr
, ANTIC_IN (block
), NULL
,
3435 /* eprime will generally only be NULL if the
3436 value of the expression, translated
3437 through the PHI for this predecessor, is
3438 undefined. If that is the case, we can't
3439 make the expression fully redundant,
3440 because its value is undefined along a
3441 predecessor path. We can thus break out
3442 early because it doesn't matter what the
3443 rest of the results are. */
3450 eprime
= fully_constant_expression (eprime
);
3451 vprime
= get_expr_value_id (eprime
);
3452 edoubleprime
= bitmap_find_leader (AVAIL_OUT (bprime
),
3454 if (edoubleprime
== NULL
)
3456 avail
[bprime
->index
] = eprime
;
3461 avail
[bprime
->index
] = edoubleprime
;
3463 if (first_s
== NULL
)
3464 first_s
= edoubleprime
;
3465 else if (!pre_expr_eq (first_s
, edoubleprime
))
3469 /* If we can insert it, it's not the same value
3470 already existing along every predecessor, and
3471 it's defined by some predecessor, it is
3472 partially redundant. */
3473 if (!cant_insert
&& !all_same
&& by_some
&& dbg_cnt (treepre_insert
))
3475 if (insert_into_preds_of_block (block
, get_expression_id (expr
),
3479 /* If all edges produce the same value and that value is
3480 an invariant, then the PHI has the same value on all
3481 edges. Note this. */
3482 else if (!cant_insert
&& all_same
&& eprime
3483 && (edoubleprime
->kind
== CONSTANT
3484 || edoubleprime
->kind
== NAME
)
3485 && !value_id_constant_p (val
))
3489 bitmap_set_t exprset
= VEC_index (bitmap_set_t
,
3490 value_expressions
, val
);
3492 unsigned int new_val
= get_expr_value_id (edoubleprime
);
3493 FOR_EACH_EXPR_ID_IN_SET (exprset
, j
, bi
)
3495 pre_expr expr
= expression_for_id (j
);
3497 if (expr
->kind
== NAME
)
3499 vn_ssa_aux_t info
= VN_INFO (PRE_EXPR_NAME (expr
));
3500 /* Just reset the value id and valnum so it is
3501 the same as the constant we have discovered. */
3502 if (edoubleprime
->kind
== CONSTANT
)
3504 info
->valnum
= PRE_EXPR_CONSTANT (edoubleprime
);
3505 pre_stats
.constified
++;
3508 info
->valnum
= VN_INFO (PRE_EXPR_NAME (edoubleprime
))->valnum
;
3509 info
->value_id
= new_val
;
3517 VEC_free (pre_expr
, heap
, exprs
);
3522 /* Perform insertion for partially anticipatable expressions. There
3523 is only one case we will perform insertion for these. This case is
3524 if the expression is partially anticipatable, and fully available.
3525 In this case, we know that putting it earlier will enable us to
3526 remove the later computation. */
3530 do_partial_partial_insertion (basic_block block
, basic_block dom
)
3532 bool new_stuff
= false;
3533 VEC (pre_expr
, heap
) *exprs
= sorted_array_from_bitmap_set (PA_IN (block
));
3537 for (i
= 0; VEC_iterate (pre_expr
, exprs
, i
, expr
); i
++)
3539 if (expr
->kind
!= NAME
)
3544 bool cant_insert
= false;
3547 pre_expr eprime
= NULL
;
3550 val
= get_expr_value_id (expr
);
3551 if (bitmap_set_contains_value (PHI_GEN (block
), val
))
3553 if (bitmap_set_contains_value (AVAIL_OUT (dom
), val
))
3556 avail
= XCNEWVEC (pre_expr
, last_basic_block
);
3557 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
3559 unsigned int vprime
;
3560 pre_expr edoubleprime
;
3562 /* We should never run insertion for the exit block
3563 and so not come across fake pred edges. */
3564 gcc_assert (!(pred
->flags
& EDGE_FAKE
));
3566 eprime
= phi_translate (expr
, ANTIC_IN (block
),
3570 /* eprime will generally only be NULL if the
3571 value of the expression, translated
3572 through the PHI for this predecessor, is
3573 undefined. If that is the case, we can't
3574 make the expression fully redundant,
3575 because its value is undefined along a
3576 predecessor path. We can thus break out
3577 early because it doesn't matter what the
3578 rest of the results are. */
3585 eprime
= fully_constant_expression (eprime
);
3586 vprime
= get_expr_value_id (eprime
);
3587 edoubleprime
= bitmap_find_leader (AVAIL_OUT (bprime
),
3589 if (edoubleprime
== NULL
)
3595 avail
[bprime
->index
] = edoubleprime
;
3599 /* If we can insert it, it's not the same value
3600 already existing along every predecessor, and
3601 it's defined by some predecessor, it is
3602 partially redundant. */
3603 if (!cant_insert
&& by_all
&& dbg_cnt (treepre_insert
))
3605 pre_stats
.pa_insert
++;
3606 if (insert_into_preds_of_block (block
, get_expression_id (expr
),
3614 VEC_free (pre_expr
, heap
, exprs
);
3619 insert_aux (basic_block block
)
3622 bool new_stuff
= false;
3627 dom
= get_immediate_dominator (CDI_DOMINATORS
, block
);
3632 bitmap_set_t newset
= NEW_SETS (dom
);
3635 /* Note that we need to value_replace both NEW_SETS, and
3636 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3637 represented by some non-simple expression here that we want
3638 to replace it with. */
3639 FOR_EACH_EXPR_ID_IN_SET (newset
, i
, bi
)
3641 pre_expr expr
= expression_for_id (i
);
3642 bitmap_value_replace_in_set (NEW_SETS (block
), expr
);
3643 bitmap_value_replace_in_set (AVAIL_OUT (block
), expr
);
3646 if (!single_pred_p (block
))
3648 new_stuff
|= do_regular_insertion (block
, dom
);
3649 if (do_partial_partial
)
3650 new_stuff
|= do_partial_partial_insertion (block
, dom
);
3654 for (son
= first_dom_son (CDI_DOMINATORS
, block
);
3656 son
= next_dom_son (CDI_DOMINATORS
, son
))
3658 new_stuff
|= insert_aux (son
);
3664 /* Perform insertion of partially redundant values. */
3669 bool new_stuff
= true;
3671 int num_iterations
= 0;
3674 NEW_SETS (bb
) = bitmap_set_new ();
3679 new_stuff
= insert_aux (ENTRY_BLOCK_PTR
);
3681 statistics_histogram_event (cfun
, "insert iterations", num_iterations
);
3685 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3688 add_to_exp_gen (basic_block block
, tree op
)
3693 if (TREE_CODE (op
) == SSA_NAME
&& ssa_undefined_value_p (op
))
3695 result
= get_or_alloc_expr_for_name (op
);
3696 bitmap_value_insert_into_set (EXP_GEN (block
), result
);
3697 bitmap_value_insert_into_set (maximal_set
, result
);
3701 /* Create value ids for PHI in BLOCK. */
3704 make_values_for_phi (gimple phi
, basic_block block
)
3706 tree result
= gimple_phi_result (phi
);
3708 /* We have no need for virtual phis, as they don't represent
3709 actual computations. */
3710 if (is_gimple_reg (result
))
3712 pre_expr e
= get_or_alloc_expr_for_name (result
);
3713 add_to_value (get_expr_value_id (e
), e
);
3714 bitmap_insert_into_set (PHI_GEN (block
), e
);
3715 bitmap_value_insert_into_set (AVAIL_OUT (block
), e
);
3719 for (i
= 0; i
< gimple_phi_num_args (phi
); ++i
)
3721 tree arg
= gimple_phi_arg_def (phi
, i
);
3722 if (TREE_CODE (arg
) == SSA_NAME
)
3724 e
= get_or_alloc_expr_for_name (arg
);
3725 add_to_value (get_expr_value_id (e
), e
);
3726 bitmap_value_insert_into_set (maximal_set
, e
);
3733 /* Compute the AVAIL set for all basic blocks.
3735 This function performs value numbering of the statements in each basic
3736 block. The AVAIL sets are built from information we glean while doing
3737 this value numbering, since the AVAIL sets contain only one entry per
3740 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3741 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3744 compute_avail (void)
3747 basic_block block
, son
;
3748 basic_block
*worklist
;
3752 /* We pretend that default definitions are defined in the entry block.
3753 This includes function arguments and the static chain decl. */
3754 for (i
= 1; i
< num_ssa_names
; ++i
)
3756 tree name
= ssa_name (i
);
3759 || !SSA_NAME_IS_DEFAULT_DEF (name
)
3760 || has_zero_uses (name
)
3761 || !is_gimple_reg (name
))
3764 e
= get_or_alloc_expr_for_name (name
);
3765 add_to_value (get_expr_value_id (e
), e
);
3768 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR
), e
);
3769 bitmap_value_insert_into_set (maximal_set
, e
);
3771 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR
), e
);
3774 /* Allocate the worklist. */
3775 worklist
= XNEWVEC (basic_block
, n_basic_blocks
);
3777 /* Seed the algorithm by putting the dominator children of the entry
3778 block on the worklist. */
3779 for (son
= first_dom_son (CDI_DOMINATORS
, ENTRY_BLOCK_PTR
);
3781 son
= next_dom_son (CDI_DOMINATORS
, son
))
3782 worklist
[sp
++] = son
;
3784 /* Loop until the worklist is empty. */
3787 gimple_stmt_iterator gsi
;
3790 unsigned int stmt_uid
= 1;
3792 /* Pick a block from the worklist. */
3793 block
= worklist
[--sp
];
3795 /* Initially, the set of available values in BLOCK is that of
3796 its immediate dominator. */
3797 dom
= get_immediate_dominator (CDI_DOMINATORS
, block
);
3799 bitmap_set_copy (AVAIL_OUT (block
), AVAIL_OUT (dom
));
3801 /* Generate values for PHI nodes. */
3802 for (gsi
= gsi_start_phis (block
); !gsi_end_p (gsi
); gsi_next (&gsi
))
3803 make_values_for_phi (gsi_stmt (gsi
), block
);
3805 /* Now compute value numbers and populate value sets with all
3806 the expressions computed in BLOCK. */
3807 for (gsi
= gsi_start_bb (block
); !gsi_end_p (gsi
); gsi_next (&gsi
))
3812 stmt
= gsi_stmt (gsi
);
3813 gimple_set_uid (stmt
, stmt_uid
++);
3815 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_DEF
)
3817 pre_expr e
= get_or_alloc_expr_for_name (op
);
3819 add_to_value (get_expr_value_id (e
), e
);
3821 bitmap_insert_into_set (TMP_GEN (block
), e
);
3822 bitmap_value_insert_into_set (AVAIL_OUT (block
), e
);
3825 if (gimple_has_volatile_ops (stmt
)
3826 || stmt_could_throw_p (stmt
))
3829 switch (gimple_code (stmt
))
3832 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_USE
)
3833 add_to_exp_gen (block
, op
);
3840 vn_reference_op_t vro
;
3841 pre_expr result
= NULL
;
3842 VEC(vn_reference_op_s
, heap
) *ops
= NULL
;
3844 if (!can_value_number_call (stmt
))
3847 copy_reference_ops_from_call (stmt
, &ops
);
3848 vn_reference_lookup_pieces (gimple_vuse (stmt
), 0,
3849 gimple_expr_type (stmt
),
3851 VEC_free (vn_reference_op_s
, heap
, ops
);
3855 for (i
= 0; VEC_iterate (vn_reference_op_s
,
3859 if (vro
->op0
&& TREE_CODE (vro
->op0
) == SSA_NAME
)
3860 add_to_exp_gen (block
, vro
->op0
);
3861 if (vro
->op1
&& TREE_CODE (vro
->op1
) == SSA_NAME
)
3862 add_to_exp_gen (block
, vro
->op1
);
3863 if (vro
->op2
&& TREE_CODE (vro
->op2
) == SSA_NAME
)
3864 add_to_exp_gen (block
, vro
->op2
);
3866 result
= (pre_expr
) pool_alloc (pre_expr_pool
);
3867 result
->kind
= REFERENCE
;
3869 PRE_EXPR_REFERENCE (result
) = ref
;
3871 get_or_alloc_expression_id (result
);
3872 add_to_value (get_expr_value_id (result
), result
);
3875 bitmap_value_insert_into_set (EXP_GEN (block
),
3877 bitmap_value_insert_into_set (maximal_set
, result
);
3884 pre_expr result
= NULL
;
3885 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt
)))
3888 if (is_exception_related (stmt
))
3891 case tcc_comparison
:
3896 vn_nary_op_lookup_pieces (gimple_num_ops (stmt
) - 1,
3897 gimple_assign_rhs_code (stmt
),
3898 gimple_expr_type (stmt
),
3899 gimple_assign_rhs1 (stmt
),
3900 gimple_assign_rhs2 (stmt
),
3901 NULL_TREE
, NULL_TREE
, &nary
);
3906 for (i
= 0; i
< nary
->length
; i
++)
3907 if (TREE_CODE (nary
->op
[i
]) == SSA_NAME
)
3908 add_to_exp_gen (block
, nary
->op
[i
]);
3910 result
= (pre_expr
) pool_alloc (pre_expr_pool
);
3911 result
->kind
= NARY
;
3913 PRE_EXPR_NARY (result
) = nary
;
3917 case tcc_declaration
:
3922 vn_reference_op_t vro
;
3924 vn_reference_lookup (gimple_assign_rhs1 (stmt
),
3930 for (i
= 0; VEC_iterate (vn_reference_op_s
,
3934 if (vro
->op0
&& TREE_CODE (vro
->op0
) == SSA_NAME
)
3935 add_to_exp_gen (block
, vro
->op0
);
3936 if (vro
->op1
&& TREE_CODE (vro
->op1
) == SSA_NAME
)
3937 add_to_exp_gen (block
, vro
->op1
);
3938 if (vro
->op2
&& TREE_CODE (vro
->op2
) == SSA_NAME
)
3939 add_to_exp_gen (block
, vro
->op2
);
3941 result
= (pre_expr
) pool_alloc (pre_expr_pool
);
3942 result
->kind
= REFERENCE
;
3944 PRE_EXPR_REFERENCE (result
) = ref
;
3949 /* For any other statement that we don't
3950 recognize, simply add all referenced
3951 SSA_NAMEs to EXP_GEN. */
3952 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_USE
)
3953 add_to_exp_gen (block
, op
);
3957 get_or_alloc_expression_id (result
);
3958 add_to_value (get_expr_value_id (result
), result
);
3961 bitmap_value_insert_into_set (EXP_GEN (block
), result
);
3962 bitmap_value_insert_into_set (maximal_set
, result
);
3972 /* Put the dominator children of BLOCK on the worklist of blocks
3973 to compute available sets for. */
3974 for (son
= first_dom_son (CDI_DOMINATORS
, block
);
3976 son
= next_dom_son (CDI_DOMINATORS
, son
))
3977 worklist
[sp
++] = son
;
3983 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
3984 than the available expressions for it. The insertion point is
3985 right before the first use in STMT. Returns the SSA_NAME that should
3986 be used for replacement. */
3989 do_SCCVN_insertion (gimple stmt
, tree ssa_vn
)
3991 basic_block bb
= gimple_bb (stmt
);
3992 gimple_stmt_iterator gsi
;
3993 gimple_seq stmts
= NULL
;
3997 /* First create a value expression from the expression we want
3998 to insert and associate it with the value handle for SSA_VN. */
3999 e
= get_or_alloc_expr_for (vn_get_expr_for (ssa_vn
));
4003 /* Then use create_expression_by_pieces to generate a valid
4004 expression to insert at this point of the IL stream. */
4005 expr
= create_expression_by_pieces (bb
, e
, &stmts
, stmt
, NULL
);
4006 if (expr
== NULL_TREE
)
4008 gsi
= gsi_for_stmt (stmt
);
4009 gsi_insert_seq_before (&gsi
, stmts
, GSI_SAME_STMT
);
4014 /* Eliminate fully redundant computations. */
4019 VEC (gimple
, heap
) *to_remove
= NULL
;
4021 unsigned int todo
= 0;
4022 gimple_stmt_iterator gsi
;
4028 for (gsi
= gsi_start_bb (b
); !gsi_end_p (gsi
); gsi_next (&gsi
))
4030 stmt
= gsi_stmt (gsi
);
4032 /* Lookup the RHS of the expression, see if we have an
4033 available computation for it. If so, replace the RHS with
4034 the available computation. */
4035 if (gimple_has_lhs (stmt
)
4036 && TREE_CODE (gimple_get_lhs (stmt
)) == SSA_NAME
4037 && !gimple_assign_ssa_name_copy_p (stmt
)
4038 && (!gimple_assign_single_p (stmt
)
4039 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt
)))
4040 && !gimple_has_volatile_ops (stmt
)
4041 && !has_zero_uses (gimple_get_lhs (stmt
)))
4043 tree lhs
= gimple_get_lhs (stmt
);
4044 tree rhs
= NULL_TREE
;
4046 pre_expr lhsexpr
= get_or_alloc_expr_for_name (lhs
);
4047 pre_expr sprimeexpr
;
4049 if (gimple_assign_single_p (stmt
))
4050 rhs
= gimple_assign_rhs1 (stmt
);
4052 sprimeexpr
= bitmap_find_leader (AVAIL_OUT (b
),
4053 get_expr_value_id (lhsexpr
),
4058 if (sprimeexpr
->kind
== CONSTANT
)
4059 sprime
= PRE_EXPR_CONSTANT (sprimeexpr
);
4060 else if (sprimeexpr
->kind
== NAME
)
4061 sprime
= PRE_EXPR_NAME (sprimeexpr
);
4066 /* If there is no existing leader but SCCVN knows this
4067 value is constant, use that constant. */
4068 if (!sprime
&& is_gimple_min_invariant (VN_INFO (lhs
)->valnum
))
4070 sprime
= VN_INFO (lhs
)->valnum
;
4071 if (!useless_type_conversion_p (TREE_TYPE (lhs
),
4072 TREE_TYPE (sprime
)))
4073 sprime
= fold_convert (TREE_TYPE (lhs
), sprime
);
4075 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4077 fprintf (dump_file
, "Replaced ");
4078 print_gimple_expr (dump_file
, stmt
, 0, 0);
4079 fprintf (dump_file
, " with ");
4080 print_generic_expr (dump_file
, sprime
, 0);
4081 fprintf (dump_file
, " in ");
4082 print_gimple_stmt (dump_file
, stmt
, 0, 0);
4084 pre_stats
.eliminations
++;
4085 propagate_tree_value_into_stmt (&gsi
, sprime
);
4086 stmt
= gsi_stmt (gsi
);
4091 /* If there is no existing usable leader but SCCVN thinks
4092 it has an expression it wants to use as replacement,
4094 if (!sprime
|| sprime
== lhs
)
4096 tree val
= VN_INFO (lhs
)->valnum
;
4098 && TREE_CODE (val
) == SSA_NAME
4099 && VN_INFO (val
)->needs_insertion
4100 && can_PRE_operation (vn_get_expr_for (val
)))
4101 sprime
= do_SCCVN_insertion (stmt
, val
);
4105 && (rhs
== NULL_TREE
4106 || TREE_CODE (rhs
) != SSA_NAME
4107 || may_propagate_copy (rhs
, sprime
)))
4109 gcc_assert (sprime
!= rhs
);
4111 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4113 fprintf (dump_file
, "Replaced ");
4114 print_gimple_expr (dump_file
, stmt
, 0, 0);
4115 fprintf (dump_file
, " with ");
4116 print_generic_expr (dump_file
, sprime
, 0);
4117 fprintf (dump_file
, " in ");
4118 print_gimple_stmt (dump_file
, stmt
, 0, 0);
4121 if (TREE_CODE (sprime
) == SSA_NAME
)
4122 gimple_set_plf (SSA_NAME_DEF_STMT (sprime
),
4124 /* We need to make sure the new and old types actually match,
4125 which may require adding a simple cast, which fold_convert
4127 if ((!rhs
|| TREE_CODE (rhs
) != SSA_NAME
)
4128 && !useless_type_conversion_p (gimple_expr_type (stmt
),
4129 TREE_TYPE (sprime
)))
4130 sprime
= fold_convert (gimple_expr_type (stmt
), sprime
);
4132 pre_stats
.eliminations
++;
4133 propagate_tree_value_into_stmt (&gsi
, sprime
);
4134 stmt
= gsi_stmt (gsi
);
4137 /* If we removed EH side effects from the statement, clean
4138 its EH information. */
4139 if (maybe_clean_or_replace_eh_stmt (stmt
, stmt
))
4141 bitmap_set_bit (need_eh_cleanup
,
4142 gimple_bb (stmt
)->index
);
4143 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4144 fprintf (dump_file
, " Removed EH side effects.\n");
4148 /* If the statement is a scalar store, see if the expression
4149 has the same value number as its rhs. If so, the store is
4151 else if (gimple_assign_single_p (stmt
)
4152 && !is_gimple_reg (gimple_assign_lhs (stmt
))
4153 && (TREE_CODE (gimple_assign_rhs1 (stmt
)) == SSA_NAME
4154 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt
))))
4156 tree rhs
= gimple_assign_rhs1 (stmt
);
4158 val
= vn_reference_lookup (gimple_assign_lhs (stmt
),
4159 gimple_vuse (stmt
), true, NULL
);
4160 if (TREE_CODE (rhs
) == SSA_NAME
)
4161 rhs
= VN_INFO (rhs
)->valnum
;
4163 && operand_equal_p (val
, rhs
, 0))
4165 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4167 fprintf (dump_file
, "Deleted redundant store ");
4168 print_gimple_stmt (dump_file
, stmt
, 0, 0);
4171 /* Queue stmt for removal. */
4172 VEC_safe_push (gimple
, heap
, to_remove
, stmt
);
4175 /* Visit COND_EXPRs and fold the comparison with the
4176 available value-numbers. */
4177 else if (gimple_code (stmt
) == GIMPLE_COND
)
4179 tree op0
= gimple_cond_lhs (stmt
);
4180 tree op1
= gimple_cond_rhs (stmt
);
4183 if (TREE_CODE (op0
) == SSA_NAME
)
4184 op0
= VN_INFO (op0
)->valnum
;
4185 if (TREE_CODE (op1
) == SSA_NAME
)
4186 op1
= VN_INFO (op1
)->valnum
;
4187 result
= fold_binary (gimple_cond_code (stmt
), boolean_type_node
,
4189 if (result
&& TREE_CODE (result
) == INTEGER_CST
)
4191 if (integer_zerop (result
))
4192 gimple_cond_make_false (stmt
);
4194 gimple_cond_make_true (stmt
);
4196 todo
= TODO_cleanup_cfg
;
4199 /* Visit indirect calls and turn them into direct calls if
4201 if (gimple_code (stmt
) == GIMPLE_CALL
4202 && TREE_CODE (gimple_call_fn (stmt
)) == SSA_NAME
)
4204 tree fn
= VN_INFO (gimple_call_fn (stmt
))->valnum
;
4205 if (TREE_CODE (fn
) == ADDR_EXPR
4206 && TREE_CODE (TREE_OPERAND (fn
, 0)) == FUNCTION_DECL
)
4208 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4210 fprintf (dump_file
, "Replacing call target with ");
4211 print_generic_expr (dump_file
, fn
, 0);
4212 fprintf (dump_file
, " in ");
4213 print_gimple_stmt (dump_file
, stmt
, 0, 0);
4216 gimple_call_set_fn (stmt
, fn
);
4218 if (maybe_clean_or_replace_eh_stmt (stmt
, stmt
))
4219 gimple_purge_dead_eh_edges (b
);
4221 /* Changing an indirect call to a direct call may
4222 have exposed different semantics. This may
4223 require an SSA update. */
4224 todo
|= TODO_update_ssa_only_virtuals
;
4229 for (gsi
= gsi_start_phis (b
); !gsi_end_p (gsi
);)
4231 gimple stmt
, phi
= gsi_stmt (gsi
);
4232 tree sprime
= NULL_TREE
, res
= PHI_RESULT (phi
);
4233 pre_expr sprimeexpr
, resexpr
;
4234 gimple_stmt_iterator gsi2
;
4236 /* We want to perform redundant PHI elimination. Do so by
4237 replacing the PHI with a single copy if possible.
4238 Do not touch inserted, single-argument or virtual PHIs. */
4239 if (gimple_phi_num_args (phi
) == 1
4240 || !is_gimple_reg (res
)
4241 || bitmap_bit_p (inserted_phi_names
, SSA_NAME_VERSION (res
)))
4247 resexpr
= get_or_alloc_expr_for_name (res
);
4248 sprimeexpr
= bitmap_find_leader (AVAIL_OUT (b
),
4249 get_expr_value_id (resexpr
), NULL
);
4252 if (sprimeexpr
->kind
== CONSTANT
)
4253 sprime
= PRE_EXPR_CONSTANT (sprimeexpr
);
4254 else if (sprimeexpr
->kind
== NAME
)
4255 sprime
= PRE_EXPR_NAME (sprimeexpr
);
4266 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4268 fprintf (dump_file
, "Replaced redundant PHI node defining ");
4269 print_generic_expr (dump_file
, res
, 0);
4270 fprintf (dump_file
, " with ");
4271 print_generic_expr (dump_file
, sprime
, 0);
4272 fprintf (dump_file
, "\n");
4275 remove_phi_node (&gsi
, false);
4277 if (!useless_type_conversion_p (TREE_TYPE (res
), TREE_TYPE (sprime
)))
4278 sprime
= fold_convert (TREE_TYPE (res
), sprime
);
4279 stmt
= gimple_build_assign (res
, sprime
);
4280 SSA_NAME_DEF_STMT (res
) = stmt
;
4281 if (TREE_CODE (sprime
) == SSA_NAME
)
4282 gimple_set_plf (SSA_NAME_DEF_STMT (sprime
),
4284 gsi2
= gsi_after_labels (b
);
4285 gsi_insert_before (&gsi2
, stmt
, GSI_NEW_STMT
);
4286 /* Queue the copy for eventual removal. */
4287 VEC_safe_push (gimple
, heap
, to_remove
, stmt
);
4288 pre_stats
.eliminations
++;
4292 /* We cannot remove stmts during BB walk, especially not release SSA
4293 names there as this confuses the VN machinery. The stmts ending
4294 up in to_remove are either stores or simple copies. */
4295 for (i
= 0; VEC_iterate (gimple
, to_remove
, i
, stmt
); ++i
)
4297 tree lhs
= gimple_assign_lhs (stmt
);
4298 use_operand_p use_p
;
4301 /* If there is a single use only, propagate the equivalency
4302 instead of keeping the copy. */
4303 if (TREE_CODE (lhs
) == SSA_NAME
4304 && single_imm_use (lhs
, &use_p
, &use_stmt
)
4305 && may_propagate_copy (USE_FROM_PTR (use_p
),
4306 gimple_assign_rhs1 (stmt
)))
4308 SET_USE (use_p
, gimple_assign_rhs1 (stmt
));
4309 update_stmt (use_stmt
);
4312 /* If this is a store or a now unused copy, remove it. */
4313 if (TREE_CODE (lhs
) != SSA_NAME
4314 || has_zero_uses (lhs
))
4316 gsi
= gsi_for_stmt (stmt
);
4317 unlink_stmt_vdef (stmt
);
4318 gsi_remove (&gsi
, true);
4319 release_defs (stmt
);
4322 VEC_free (gimple
, heap
, to_remove
);
4327 /* Borrow a bit of tree-ssa-dce.c for the moment.
4328 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4329 this may be a bit faster, and we may want critical edges kept split. */
4331 /* If OP's defining statement has not already been determined to be necessary,
4332 mark that statement necessary. Return the stmt, if it is newly
4335 static inline gimple
4336 mark_operand_necessary (tree op
)
4342 if (TREE_CODE (op
) != SSA_NAME
)
4345 stmt
= SSA_NAME_DEF_STMT (op
);
4348 if (gimple_plf (stmt
, NECESSARY
)
4349 || gimple_nop_p (stmt
))
4352 gimple_set_plf (stmt
, NECESSARY
, true);
4356 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4357 to insert PHI nodes sometimes, and because value numbering of casts isn't
4358 perfect, we sometimes end up inserting dead code. This simple DCE-like
4359 pass removes any insertions we made that weren't actually used. */
4362 remove_dead_inserted_code (void)
4364 VEC(gimple
,heap
) *worklist
= NULL
;
4368 worklist
= VEC_alloc (gimple
, heap
, VEC_length (gimple
, inserted_exprs
));
4369 for (i
= 0; VEC_iterate (gimple
, inserted_exprs
, i
, t
); i
++)
4371 if (gimple_plf (t
, NECESSARY
))
4372 VEC_quick_push (gimple
, worklist
, t
);
4374 while (VEC_length (gimple
, worklist
) > 0)
4376 t
= VEC_pop (gimple
, worklist
);
4378 /* PHI nodes are somewhat special in that each PHI alternative has
4379 data and control dependencies. All the statements feeding the
4380 PHI node's arguments are always necessary. */
4381 if (gimple_code (t
) == GIMPLE_PHI
)
4385 VEC_reserve (gimple
, heap
, worklist
, gimple_phi_num_args (t
));
4386 for (k
= 0; k
< gimple_phi_num_args (t
); k
++)
4388 tree arg
= PHI_ARG_DEF (t
, k
);
4389 if (TREE_CODE (arg
) == SSA_NAME
)
4391 gimple n
= mark_operand_necessary (arg
);
4393 VEC_quick_push (gimple
, worklist
, n
);
4399 /* Propagate through the operands. Examine all the USE, VUSE and
4400 VDEF operands in this statement. Mark all the statements
4401 which feed this statement's uses as necessary. */
4405 /* The operands of VDEF expressions are also needed as they
4406 represent potential definitions that may reach this
4407 statement (VDEF operands allow us to follow def-def
4410 FOR_EACH_SSA_TREE_OPERAND (use
, t
, iter
, SSA_OP_ALL_USES
)
4412 gimple n
= mark_operand_necessary (use
);
4414 VEC_safe_push (gimple
, heap
, worklist
, n
);
4419 for (i
= 0; VEC_iterate (gimple
, inserted_exprs
, i
, t
); i
++)
4421 if (!gimple_plf (t
, NECESSARY
))
4423 gimple_stmt_iterator gsi
;
4425 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4427 fprintf (dump_file
, "Removing unnecessary insertion:");
4428 print_gimple_stmt (dump_file
, t
, 0, 0);
4431 gsi
= gsi_for_stmt (t
);
4432 if (gimple_code (t
) == GIMPLE_PHI
)
4433 remove_phi_node (&gsi
, true);
4435 gsi_remove (&gsi
, true);
4439 VEC_free (gimple
, heap
, worklist
);
4442 /* Initialize data structures used by PRE. */
4445 init_pre (bool do_fre
)
4449 next_expression_id
= 1;
4451 VEC_safe_push (pre_expr
, heap
, expressions
, NULL
);
4452 value_expressions
= VEC_alloc (bitmap_set_t
, heap
, get_max_value_id () + 1);
4453 VEC_safe_grow_cleared (bitmap_set_t
, heap
, value_expressions
,
4454 get_max_value_id() + 1);
4458 inserted_exprs
= NULL
;
4459 need_creation
= NULL
;
4460 pretemp
= NULL_TREE
;
4461 storetemp
= NULL_TREE
;
4462 prephitemp
= NULL_TREE
;
4464 connect_infinite_loops_to_exit ();
4465 memset (&pre_stats
, 0, sizeof (pre_stats
));
4468 postorder
= XNEWVEC (int, n_basic_blocks
- NUM_FIXED_BLOCKS
);
4469 post_order_compute (postorder
, false, false);
4472 bb
->aux
= XCNEWVEC (struct bb_bitmap_sets
, 1);
4474 calculate_dominance_info (CDI_POST_DOMINATORS
);
4475 calculate_dominance_info (CDI_DOMINATORS
);
4477 bitmap_obstack_initialize (&grand_bitmap_obstack
);
4478 inserted_phi_names
= BITMAP_ALLOC (&grand_bitmap_obstack
);
4479 phi_translate_table
= htab_create (5110, expr_pred_trans_hash
,
4480 expr_pred_trans_eq
, free
);
4481 expression_to_id
= htab_create (num_ssa_names
* 3,
4484 seen_during_translate
= BITMAP_ALLOC (&grand_bitmap_obstack
);
4485 bitmap_set_pool
= create_alloc_pool ("Bitmap sets",
4486 sizeof (struct bitmap_set
), 30);
4487 pre_expr_pool
= create_alloc_pool ("pre_expr nodes",
4488 sizeof (struct pre_expr_d
), 30);
4491 EXP_GEN (bb
) = bitmap_set_new ();
4492 PHI_GEN (bb
) = bitmap_set_new ();
4493 TMP_GEN (bb
) = bitmap_set_new ();
4494 AVAIL_OUT (bb
) = bitmap_set_new ();
4496 maximal_set
= in_fre
? NULL
: bitmap_set_new ();
4498 need_eh_cleanup
= BITMAP_ALLOC (NULL
);
4502 /* Deallocate data structures used by PRE. */
4505 fini_pre (bool do_fre
)
4510 VEC_free (bitmap_set_t
, heap
, value_expressions
);
4511 VEC_free (gimple
, heap
, inserted_exprs
);
4512 VEC_free (gimple
, heap
, need_creation
);
4513 bitmap_obstack_release (&grand_bitmap_obstack
);
4514 free_alloc_pool (bitmap_set_pool
);
4515 free_alloc_pool (pre_expr_pool
);
4516 htab_delete (phi_translate_table
);
4517 htab_delete (expression_to_id
);
4525 free_dominance_info (CDI_POST_DOMINATORS
);
4527 if (!bitmap_empty_p (need_eh_cleanup
))
4529 gimple_purge_all_dead_eh_edges (need_eh_cleanup
);
4530 cleanup_tree_cfg ();
4533 BITMAP_FREE (need_eh_cleanup
);
4536 loop_optimizer_finalize ();
4539 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4540 only wants to do full redundancy elimination. */
4543 execute_pre (bool do_fre ATTRIBUTE_UNUSED
)
4545 unsigned int todo
= 0;
4547 do_partial_partial
= optimize
> 2;
4549 /* This has to happen before SCCVN runs because
4550 loop_optimizer_init may create new phis, etc. */
4552 loop_optimizer_init (LOOPS_NORMAL
);
4554 if (!run_scc_vn (do_fre
))
4558 remove_dead_inserted_code ();
4559 loop_optimizer_finalize ();
4568 /* Collect and value number expressions computed in each basic block. */
4571 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4577 print_bitmap_set (dump_file
, EXP_GEN (bb
), "exp_gen", bb
->index
);
4578 print_bitmap_set (dump_file
, PHI_GEN (bb
), "phi_gen", bb
->index
);
4579 print_bitmap_set (dump_file
, TMP_GEN (bb
), "tmp_gen", bb
->index
);
4580 print_bitmap_set (dump_file
, AVAIL_OUT (bb
), "avail_out", bb
->index
);
4583 print_bitmap_set (dump_file
, maximal_set
, "maximal", 0);
4586 /* Insert can get quite slow on an incredibly large number of basic
4587 blocks due to some quadratic behavior. Until this behavior is
4588 fixed, don't run it when he have an incredibly large number of
4589 bb's. If we aren't going to run insert, there is no point in
4590 computing ANTIC, either, even though it's plenty fast. */
4591 if (!do_fre
&& n_basic_blocks
< 4000)
4597 /* Remove all the redundant expressions. */
4598 todo
|= eliminate ();
4600 statistics_counter_event (cfun
, "Insertions", pre_stats
.insertions
);
4601 statistics_counter_event (cfun
, "PA inserted", pre_stats
.pa_insert
);
4602 statistics_counter_event (cfun
, "New PHIs", pre_stats
.phis
);
4603 statistics_counter_event (cfun
, "Eliminated", pre_stats
.eliminations
);
4604 statistics_counter_event (cfun
, "Constified", pre_stats
.constified
);
4606 /* Make sure to remove fake edges before committing our inserts.
4607 This makes sure we don't end up with extra critical edges that
4608 we would need to split. */
4609 remove_fake_exit_edges ();
4610 gsi_commit_edge_inserts ();
4612 clear_expression_ids ();
4615 remove_dead_inserted_code ();
4623 /* Gate and execute functions for PRE. */
4628 return execute_pre (false);
4634 /* PRE tends to generate bigger code. */
4635 return flag_tree_pre
!= 0 && optimize_function_for_speed_p (cfun
);
4638 struct gimple_opt_pass pass_pre
=
4643 gate_pre
, /* gate */
4644 do_pre
, /* execute */
4647 0, /* static_pass_number */
4648 TV_TREE_PRE
, /* tv_id */
4649 PROP_no_crit_edges
| PROP_cfg
4650 | PROP_ssa
, /* properties_required */
4651 0, /* properties_provided */
4652 0, /* properties_destroyed */
4653 TODO_rebuild_alias
, /* todo_flags_start */
4654 TODO_update_ssa_only_virtuals
| TODO_dump_func
| TODO_ggc_collect
4655 | TODO_verify_ssa
/* todo_flags_finish */
4660 /* Gate and execute functions for FRE. */
4665 return execute_pre (true);
4671 return flag_tree_fre
!= 0;
4674 struct gimple_opt_pass pass_fre
=
4679 gate_fre
, /* gate */
4680 execute_fre
, /* execute */
4683 0, /* static_pass_number */
4684 TV_TREE_FRE
, /* tv_id */
4685 PROP_cfg
| PROP_ssa
, /* properties_required */
4686 0, /* properties_provided */
4687 0, /* properties_destroyed */
4688 0, /* todo_flags_start */
4689 TODO_dump_func
| TODO_ggc_collect
| TODO_verify_ssa
/* todo_flags_finish */