2 Copyright (C) 2001, 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 2, or (at your option)
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING. If not, write to
20 the Free Software Foundation, 59 Temple Place - Suite 330,
21 Boston, MA 02111-1307, USA. */
25 #include "coretypes.h"
30 #include "basic-block.h"
31 #include "diagnostic.h"
32 #include "tree-inline.h"
33 #include "tree-flow.h"
34 #include "tree-gimple.h"
35 #include "tree-dump.h"
39 #include "tree-iterator.h"
41 #include "alloc-pool.h"
42 #include "tree-pass.h"
45 #include "langhooks.h"
50 1. Avail sets can be shared by making an avail_find_leader that
51 walks up the dominator tree and looks in those avail sets.
52 This might affect code optimality, it's unclear right now.
53 2. Load motion can be performed by value numbering the loads the
54 same as we do other expressions. This requires iterative
55 hashing the vuses into the values. Right now we simply assign
56 a new value every time we see a statement with a vuse.
57 3. Strength reduction can be performed by anticipating expressions
58 we can repair later on.
59 4. 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 MODIFY_EXPR, because MODIFY_EXPR's represent
65 the actual statement containing the expressions we care about, and
66 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.
107 insert/insert_aux performs this insertion.
109 Fourth, we eliminate fully redundant expressions.
110 This is a simple statement walk that replaces redundant
111 calculations with the now available values. */
113 /* Representations of value numbers:
115 Value numbers are represented using the "value handle" approach.
116 This means that each SSA_NAME (and for other reasons to be
117 disclosed in a moment, expression nodes) has a value handle that
118 can be retrieved through get_value_handle. This value handle, *is*
119 the value number of the SSA_NAME. You can pointer compare the
120 value handles for equivalence purposes.
122 For debugging reasons, the value handle is internally more than
123 just a number, it is a VAR_DECL named "value.x", where x is a
124 unique number for each value number in use. This allows
125 expressions with SSA_NAMES replaced by value handles to still be
126 pretty printed in a sane way. They simply print as "value.3 *
129 Expression nodes have value handles associated with them as a
130 cache. Otherwise, we'd have to look them up again in the hash
131 table This makes significant difference (factor of two or more) on
132 some test cases. They can be thrown away after the pass is
135 /* Representation of expressions on value numbers:
137 In some portions of this code, you will notice we allocate "fake"
138 analogues to the expression we are value numbering, and replace the
139 operands with the values of the expression. Since we work on
140 values, and not just names, we canonicalize expressions to value
141 expressions for use in the ANTIC sets, the EXP_GEN set, etc.
143 This is theoretically unnecessary, it just saves a bunch of
144 repeated get_value_handle and find_leader calls in the remainder of
145 the code, trading off temporary memory usage for speed. The tree
146 nodes aren't actually creating more garbage, since they are
147 allocated in a special pools which are thrown away at the end of
150 All of this also means that if you print the EXP_GEN or ANTIC sets,
151 you will see "value.5 + value.7" in the set, instead of "a_55 +
152 b_66" or something. The only thing that actually cares about
153 seeing the value leaders is phi translation, and it needs to be
154 able to find the leader for a value in an arbitrary block, so this
155 "value expression" form is perfect for it (otherwise you'd do
156 get_value_handle->find_leader->translate->get_value_handle->find_leader).*/
159 /* Representation of sets:
161 There are currently two types of sets used, hopefully to be unified soon.
162 The AVAIL sets do not need to be sorted in any particular order,
163 and thus, are simply represented as two bitmaps, one that keeps
164 track of values present in the set, and one that keeps track of
165 expressions present in the set.
167 The other sets are represented as doubly linked lists kept in topological
168 order, with an optional supporting bitmap of values present in the
169 set. The sets represent values, and the elements can be values or
170 expressions. The elements can appear in different sets, but each
171 element can only appear once in each set.
173 Since each node in the set represents a value, we also want to be
174 able to map expression, set pairs to something that tells us
175 whether the value is present is a set. We use a per-set bitmap for
176 that. The value handles also point to a linked list of the
177 expressions they represent via a tree annotation. This is mainly
178 useful only for debugging, since we don't do identity lookups. */
181 /* A value set element. Basically a single linked list of
182 expressions/values. */
183 typedef struct value_set_node
188 /* A pointer to the next element of the value set. */
189 struct value_set_node
*next
;
193 /* A value set. This is a singly linked list of value_set_node
194 elements with a possible bitmap that tells us what values exist in
195 the set. This set must be kept in topologically sorted order. */
196 typedef struct value_set
198 /* The head of the list. Used for iterating over the list in
200 value_set_node_t head
;
202 /* The tail of the list. Used for tail insertions, which are
203 necessary to keep the set in topologically sorted order because
204 of how the set is built. */
205 value_set_node_t tail
;
207 /* The length of the list. */
210 /* True if the set is indexed, which means it contains a backing
211 bitmap for quick determination of whether certain values exist in the
215 /* The bitmap of values that exist in the set. May be NULL in an
216 empty or non-indexed set. */
222 /* An unordered bitmap set. One bitmap tracks values, the other,
224 typedef struct bitmap_set
230 /* Sets that we need to keep track of. */
231 typedef struct bb_value_sets
233 /* The EXP_GEN set, which represents expressions/values generated in
237 /* The PHI_GEN set, which represents PHI results generated in a
239 bitmap_set_t phi_gen
;
241 /* The TMP_GEN set, which represents results/temporaries generated
242 in a basic block. IE the LHS of an expression. */
243 bitmap_set_t tmp_gen
;
245 /* The AVAIL_OUT set, which represents which values are available in
246 a given basic block. */
247 bitmap_set_t avail_out
;
249 /* The ANTIC_IN set, which represents which values are anticiptable
250 in a given basic block. */
251 value_set_t antic_in
;
253 /* The NEW_SETS set, which is used during insertion to augment the
254 AVAIL_OUT set of blocks with the new insertions performed during
255 the current iteration. */
256 bitmap_set_t new_sets
;
259 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
260 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
261 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
262 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
263 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
264 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
266 /* This structure is used to keep track of statistics on what
267 optimization PRE was able to perform. */
270 /* The number of RHS computations eliminated by PRE. */
273 /* The number of new expressions/temporaries generated by PRE. */
276 /* The number of new PHI nodes added by PRE. */
279 /* The number of values found constant. */
285 static tree
bitmap_find_leader (bitmap_set_t
, tree
);
286 static tree
find_leader (value_set_t
, tree
);
287 static void value_insert_into_set (value_set_t
, tree
);
288 static void bitmap_value_insert_into_set (bitmap_set_t
, tree
);
289 static void bitmap_value_replace_in_set (bitmap_set_t
, tree
);
290 static void insert_into_set (value_set_t
, tree
);
291 static void bitmap_set_copy (bitmap_set_t
, bitmap_set_t
);
292 static bool bitmap_set_contains_value (bitmap_set_t
, tree
);
293 static bitmap_set_t
bitmap_set_new (void);
294 static value_set_t
set_new (bool);
295 static bool is_undefined_value (tree
);
296 static tree
create_expression_by_pieces (basic_block
, tree
, tree
);
299 /* We can add and remove elements and entries to and from sets
300 and hash tables, so we use alloc pools for them. */
302 static alloc_pool value_set_pool
;
303 static alloc_pool bitmap_set_pool
;
304 static alloc_pool value_set_node_pool
;
305 static alloc_pool binary_node_pool
;
306 static alloc_pool unary_node_pool
;
307 static alloc_pool reference_node_pool
;
308 static bitmap_obstack grand_bitmap_obstack
;
310 /* Set of blocks with statements that have had its EH information
312 static bitmap need_eh_cleanup
;
314 /* The phi_translate_table caches phi translations for a given
315 expression and predecessor. */
317 static htab_t phi_translate_table
;
319 /* A three tuple {e, pred, v} used to cache phi translations in the
320 phi_translate_table. */
322 typedef struct expr_pred_trans_d
324 /* The expression. */
327 /* The predecessor block along which we translated the expression. */
330 /* The value that resulted from the translation. */
333 /* The hashcode for the expression, pred pair. This is cached for
336 } *expr_pred_trans_t
;
338 /* Return the hash value for a phi translation table entry. */
341 expr_pred_trans_hash (const void *p
)
343 const expr_pred_trans_t ve
= (expr_pred_trans_t
) p
;
347 /* Return true if two phi translation table entries are the same.
348 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
351 expr_pred_trans_eq (const void *p1
, const void *p2
)
353 const expr_pred_trans_t ve1
= (expr_pred_trans_t
) p1
;
354 const expr_pred_trans_t ve2
= (expr_pred_trans_t
) p2
;
355 basic_block b1
= ve1
->pred
;
356 basic_block b2
= ve2
->pred
;
359 /* If they are not translations for the same basic block, they can't
364 /* If they are for the same basic block, determine if the
365 expressions are equal. */
366 if (expressions_equal_p (ve1
->e
, ve2
->e
))
372 /* Search in the phi translation table for the translation of
373 expression E in basic block PRED. Return the translated value, if
374 found, NULL otherwise. */
377 phi_trans_lookup (tree e
, basic_block pred
)
380 struct expr_pred_trans_d ept
;
383 ept
.hashcode
= vn_compute (e
, (unsigned long) pred
, NULL
);
384 slot
= htab_find_slot_with_hash (phi_translate_table
, &ept
, ept
.hashcode
,
389 return ((expr_pred_trans_t
) *slot
)->v
;
393 /* Add the tuple mapping from {expression E, basic block PRED} to
394 value V, to the phi translation table. */
397 phi_trans_add (tree e
, tree v
, basic_block pred
)
400 expr_pred_trans_t new_pair
= xmalloc (sizeof (*new_pair
));
402 new_pair
->pred
= pred
;
404 new_pair
->hashcode
= vn_compute (e
, (unsigned long) pred
, NULL
);
405 slot
= htab_find_slot_with_hash (phi_translate_table
, new_pair
,
406 new_pair
->hashcode
, INSERT
);
409 *slot
= (void *) new_pair
;
413 /* Add expression E to the expression set of value V. */
416 add_to_value (tree v
, tree e
)
418 /* Constants have no expression sets. */
419 if (is_gimple_min_invariant (v
))
422 if (VALUE_HANDLE_EXPR_SET (v
) == NULL
)
423 VALUE_HANDLE_EXPR_SET (v
) = set_new (false);
425 insert_into_set (VALUE_HANDLE_EXPR_SET (v
), e
);
429 /* Return true if value V exists in the bitmap for SET. */
432 value_exists_in_set_bitmap (value_set_t set
, tree v
)
437 return bitmap_bit_p (set
->values
, VALUE_HANDLE_ID (v
));
441 /* Remove value V from the bitmap for SET. */
444 value_remove_from_set_bitmap (value_set_t set
, tree v
)
446 gcc_assert (set
->indexed
);
451 bitmap_clear_bit (set
->values
, VALUE_HANDLE_ID (v
));
455 /* Insert the value number V into the bitmap of values existing in
459 value_insert_into_set_bitmap (value_set_t set
, tree v
)
461 gcc_assert (set
->indexed
);
463 if (set
->values
== NULL
)
464 set
->values
= BITMAP_ALLOC (&grand_bitmap_obstack
);
466 bitmap_set_bit (set
->values
, VALUE_HANDLE_ID (v
));
470 /* Create a new bitmap set and return it. */
473 bitmap_set_new (void)
475 bitmap_set_t ret
= pool_alloc (bitmap_set_pool
);
476 ret
->expressions
= BITMAP_ALLOC (&grand_bitmap_obstack
);
477 ret
->values
= BITMAP_ALLOC (&grand_bitmap_obstack
);
481 /* Create a new set. */
484 set_new (bool indexed
)
487 ret
= pool_alloc (value_set_pool
);
488 ret
->head
= ret
->tail
= NULL
;
490 ret
->indexed
= indexed
;
495 /* Insert an expression EXPR into a bitmapped set. */
498 bitmap_insert_into_set (bitmap_set_t set
, tree expr
)
501 /* XXX: For now, we only let SSA_NAMES into the bitmap sets. */
502 gcc_assert (TREE_CODE (expr
) == SSA_NAME
);
503 val
= get_value_handle (expr
);
506 if (!is_gimple_min_invariant (val
))
508 bitmap_set_bit (set
->values
, VALUE_HANDLE_ID (val
));
509 bitmap_set_bit (set
->expressions
, SSA_NAME_VERSION (expr
));
513 /* Insert EXPR into SET. */
516 insert_into_set (value_set_t set
, tree expr
)
518 value_set_node_t newnode
= pool_alloc (value_set_node_pool
);
519 tree val
= get_value_handle (expr
);
522 if (is_gimple_min_invariant (val
))
525 /* For indexed sets, insert the value into the set value bitmap.
526 For all sets, add it to the linked list and increment the list
529 value_insert_into_set_bitmap (set
, val
);
531 newnode
->next
= NULL
;
532 newnode
->expr
= expr
;
534 if (set
->head
== NULL
)
536 set
->head
= set
->tail
= newnode
;
540 set
->tail
->next
= newnode
;
545 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
548 bitmap_set_copy (bitmap_set_t dest
, bitmap_set_t orig
)
550 bitmap_copy (dest
->expressions
, orig
->expressions
);
551 bitmap_copy (dest
->values
, orig
->values
);
554 /* Copy the set ORIG to the set DEST. */
557 set_copy (value_set_t dest
, value_set_t orig
)
559 value_set_node_t node
;
561 if (!orig
|| !orig
->head
)
564 for (node
= orig
->head
;
568 insert_into_set (dest
, node
->expr
);
572 /* Remove EXPR from SET. */
575 set_remove (value_set_t set
, tree expr
)
577 value_set_node_t node
, prev
;
579 /* Remove the value of EXPR from the bitmap, decrement the set
580 length, and remove it from the actual double linked list. */
581 value_remove_from_set_bitmap (set
, get_value_handle (expr
));
584 for (node
= set
->head
;
586 prev
= node
, node
= node
->next
)
588 if (node
->expr
== expr
)
591 set
->head
= node
->next
;
593 prev
->next
= node
->next
;
595 if (node
== set
->tail
)
597 pool_free (value_set_node_pool
, node
);
603 /* Return true if SET contains the value VAL. */
606 set_contains_value (value_set_t set
, tree val
)
608 /* All constants are in every set. */
609 if (is_gimple_min_invariant (val
))
612 if (set
->length
== 0)
615 return value_exists_in_set_bitmap (set
, val
);
618 /* Return true if bitmapped set SET contains the expression EXPR. */
620 bitmap_set_contains (bitmap_set_t set
, tree expr
)
622 /* All constants are in every set. */
623 if (is_gimple_min_invariant (get_value_handle (expr
)))
626 /* XXX: Bitmapped sets only contain SSA_NAME's for now. */
627 if (TREE_CODE (expr
) != SSA_NAME
)
629 return bitmap_bit_p (set
->expressions
, SSA_NAME_VERSION (expr
));
633 /* Return true if bitmapped set SET contains the value VAL. */
636 bitmap_set_contains_value (bitmap_set_t set
, tree val
)
638 if (is_gimple_min_invariant (val
))
640 return bitmap_bit_p (set
->values
, VALUE_HANDLE_ID (val
));
643 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
646 bitmap_set_replace_value (bitmap_set_t set
, tree lookfor
, tree expr
)
649 value_set_node_t node
;
650 if (is_gimple_min_invariant (lookfor
))
652 if (!bitmap_set_contains_value (set
, lookfor
))
655 /* The number of expressions having a given value is usually
656 significantly less than the total number of expressions in SET.
657 Thus, rather than check, for each expression in SET, whether it
658 has the value LOOKFOR, we walk the reverse mapping that tells us
659 what expressions have a given value, and see if any of those
660 expressions are in our set. For large testcases, this is about
661 5-10x faster than walking the bitmap. If this is somehow a
662 significant lose for some cases, we can choose which set to walk
663 based on the set size. */
664 exprset
= VALUE_HANDLE_EXPR_SET (lookfor
);
665 for (node
= exprset
->head
; node
; node
= node
->next
)
667 if (TREE_CODE (node
->expr
) == SSA_NAME
)
669 if (bitmap_bit_p (set
->expressions
, SSA_NAME_VERSION (node
->expr
)))
671 bitmap_clear_bit (set
->expressions
, SSA_NAME_VERSION (node
->expr
));
672 bitmap_set_bit (set
->expressions
, SSA_NAME_VERSION (expr
));
679 /* Subtract bitmapped set B from value set A, and return the new set. */
682 bitmap_set_subtract_from_value_set (value_set_t a
, bitmap_set_t b
,
685 value_set_t ret
= set_new (indexed
);
686 value_set_node_t node
;
691 if (!bitmap_set_contains (b
, node
->expr
))
692 insert_into_set (ret
, node
->expr
);
697 /* Return true if two sets are equal. */
700 set_equal (value_set_t a
, value_set_t b
)
702 value_set_node_t node
;
704 if (a
->length
!= b
->length
)
710 if (!set_contains_value (b
, get_value_handle (node
->expr
)))
716 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
717 and add it otherwise. */
720 bitmap_value_replace_in_set (bitmap_set_t set
, tree expr
)
722 tree val
= get_value_handle (expr
);
723 if (bitmap_set_contains_value (set
, val
))
724 bitmap_set_replace_value (set
, val
, expr
);
726 bitmap_insert_into_set (set
, expr
);
729 /* Insert EXPR into SET if EXPR's value is not already present in
733 bitmap_value_insert_into_set (bitmap_set_t set
, tree expr
)
735 tree val
= get_value_handle (expr
);
737 if (is_gimple_min_invariant (val
))
740 if (!bitmap_set_contains_value (set
, val
))
741 bitmap_insert_into_set (set
, expr
);
744 /* Insert the value for EXPR into SET, if it doesn't exist already. */
747 value_insert_into_set (value_set_t set
, tree expr
)
749 tree val
= get_value_handle (expr
);
751 /* Constant and invariant values exist everywhere, and thus,
752 actually keeping them in the sets is pointless. */
753 if (is_gimple_min_invariant (val
))
756 if (!set_contains_value (set
, val
))
757 insert_into_set (set
, expr
);
761 /* Print out SET to OUTFILE. */
764 bitmap_print_value_set (FILE *outfile
, bitmap_set_t set
,
765 const char *setname
, int blockindex
)
767 fprintf (outfile
, "%s[%d] := { ", setname
, blockindex
);
774 EXECUTE_IF_SET_IN_BITMAP (set
->expressions
, 0, i
, bi
)
777 fprintf (outfile
, ", ");
779 print_generic_expr (outfile
, ssa_name (i
), 0);
781 fprintf (outfile
, " (");
782 print_generic_expr (outfile
, get_value_handle (ssa_name (i
)), 0);
783 fprintf (outfile
, ") ");
786 fprintf (outfile
, " }\n");
788 /* Print out the value_set SET to OUTFILE. */
791 print_value_set (FILE *outfile
, value_set_t set
,
792 const char *setname
, int blockindex
)
794 value_set_node_t node
;
795 fprintf (outfile
, "%s[%d] := { ", setname
, blockindex
);
798 for (node
= set
->head
;
802 print_generic_expr (outfile
, node
->expr
, 0);
804 fprintf (outfile
, " (");
805 print_generic_expr (outfile
, get_value_handle (node
->expr
), 0);
806 fprintf (outfile
, ") ");
809 fprintf (outfile
, ", ");
813 fprintf (outfile
, " }\n");
816 /* Print out the expressions that have VAL to OUTFILE. */
819 print_value_expressions (FILE *outfile
, tree val
)
821 if (VALUE_HANDLE_EXPR_SET (val
))
824 sprintf (s
, "VH.%04d", VALUE_HANDLE_ID (val
));
825 print_value_set (outfile
, VALUE_HANDLE_EXPR_SET (val
), s
, 0);
831 debug_value_expressions (tree val
)
833 print_value_expressions (stderr
, val
);
837 void debug_value_set (value_set_t
, const char *, int);
840 debug_value_set (value_set_t set
, const char *setname
, int blockindex
)
842 print_value_set (stderr
, set
, setname
, blockindex
);
845 /* Return the folded version of T if T, when folded, is a gimple
846 min_invariant. Otherwise, return T. */
849 fully_constant_expression (tree t
)
853 if (folded
&& is_gimple_min_invariant (folded
))
858 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
859 the phis in PRED. Return NULL if we can't find a leader for each
860 part of the translated expression. */
863 phi_translate (tree expr
, value_set_t set
, basic_block pred
,
864 basic_block phiblock
)
866 tree phitrans
= NULL
;
872 if (is_gimple_min_invariant (expr
))
875 /* Phi translations of a given expression don't change. */
876 phitrans
= phi_trans_lookup (expr
, pred
);
880 switch (TREE_CODE_CLASS (TREE_CODE (expr
)))
883 /* XXX: Until we have PRE of loads working, none will be ANTIC. */
889 tree oldop1
= TREE_OPERAND (expr
, 0);
890 tree oldop2
= TREE_OPERAND (expr
, 1);
895 newop1
= phi_translate (find_leader (set
, oldop1
),
896 set
, pred
, phiblock
);
899 newop2
= phi_translate (find_leader (set
, oldop2
),
900 set
, pred
, phiblock
);
903 if (newop1
!= oldop1
|| newop2
!= oldop2
)
906 newexpr
= pool_alloc (binary_node_pool
);
907 memcpy (newexpr
, expr
, tree_size (expr
));
908 TREE_OPERAND (newexpr
, 0) = newop1
== oldop1
? oldop1
: get_value_handle (newop1
);
909 TREE_OPERAND (newexpr
, 1) = newop2
== oldop2
? oldop2
: get_value_handle (newop2
);
910 t
= fully_constant_expression (newexpr
);
913 pool_free (binary_node_pool
, newexpr
);
918 create_tree_ann (newexpr
);
919 vn_lookup_or_add (newexpr
, NULL
);
922 phi_trans_add (oldexpr
, newexpr
, pred
);
929 tree oldop1
= TREE_OPERAND (expr
, 0);
933 newop1
= phi_translate (find_leader (set
, oldop1
),
934 set
, pred
, phiblock
);
937 if (newop1
!= oldop1
)
940 newexpr
= pool_alloc (unary_node_pool
);
941 memcpy (newexpr
, expr
, tree_size (expr
));
942 TREE_OPERAND (newexpr
, 0) = get_value_handle (newop1
);
943 t
= fully_constant_expression (newexpr
);
946 pool_free (unary_node_pool
, newexpr
);
951 create_tree_ann (newexpr
);
952 vn_lookup_or_add (newexpr
, NULL
);
955 phi_trans_add (oldexpr
, newexpr
, pred
);
960 case tcc_exceptional
:
964 gcc_assert (TREE_CODE (expr
) == SSA_NAME
);
965 if (TREE_CODE (SSA_NAME_DEF_STMT (expr
)) == PHI_NODE
)
966 phi
= SSA_NAME_DEF_STMT (expr
);
970 e
= find_edge (pred
, bb_for_stmt (phi
));
973 if (is_undefined_value (PHI_ARG_DEF (phi
, e
->dest_idx
)))
975 vn_lookup_or_add (PHI_ARG_DEF (phi
, e
->dest_idx
), NULL
);
976 return PHI_ARG_DEF (phi
, e
->dest_idx
);
986 /* For each expression in SET, translate the value handles through phi nodes
987 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
988 expressions in DEST. */
991 phi_translate_set (value_set_t dest
, value_set_t set
, basic_block pred
,
992 basic_block phiblock
)
994 value_set_node_t node
;
995 for (node
= set
->head
;
1000 translated
= phi_translate (node
->expr
, set
, pred
, phiblock
);
1001 phi_trans_add (node
->expr
, translated
, pred
);
1003 if (translated
!= NULL
)
1004 value_insert_into_set (dest
, translated
);
1008 /* Find the leader for a value (i.e., the name representing that
1009 value) in a given set, and return it. Return NULL if no leader is
1013 bitmap_find_leader (bitmap_set_t set
, tree val
)
1018 if (is_gimple_min_invariant (val
))
1020 if (bitmap_set_contains_value (set
, val
))
1022 /* Rather than walk the entire bitmap of expressions, and see
1023 whether any of them has the value we are looking for, we look
1024 at the reverse mapping, which tells us the set of expressions
1025 that have a given value (IE value->expressions with that
1026 value) and see if any of those expressions are in our set.
1027 The number of expressions per value is usually significantly
1028 less than the number of expressions in the set. In fact, for
1029 large testcases, doing it this way is roughly 5-10x faster
1030 than walking the bitmap.
1031 If this is somehow a significant lose for some cases, we can
1032 choose which set to walk based on which set is smaller. */
1033 value_set_t exprset
;
1034 value_set_node_t node
;
1035 exprset
= VALUE_HANDLE_EXPR_SET (val
);
1036 for (node
= exprset
->head
; node
; node
= node
->next
)
1038 if (TREE_CODE (node
->expr
) == SSA_NAME
)
1040 if (bitmap_bit_p (set
->expressions
,
1041 SSA_NAME_VERSION (node
->expr
)))
1050 /* Find the leader for a value (i.e., the name representing that
1051 value) in a given set, and return it. Return NULL if no leader is
1055 find_leader (value_set_t set
, tree val
)
1057 value_set_node_t node
;
1062 /* Constants represent themselves. */
1063 if (is_gimple_min_invariant (val
))
1066 if (set
->length
== 0)
1069 if (value_exists_in_set_bitmap (set
, val
))
1071 for (node
= set
->head
;
1075 if (get_value_handle (node
->expr
) == val
)
1083 /* Determine if the expression EXPR is valid in SET. This means that
1084 we have a leader for each part of the expression (if it consists of
1085 values), or the expression is an SSA_NAME.
1087 NB: We never should run into a case where we have SSA_NAME +
1088 SSA_NAME or SSA_NAME + value. The sets valid_in_set is called on,
1089 the ANTIC sets, will only ever have SSA_NAME's or binary value
1090 expression (IE VALUE1 + VALUE2) */
1093 valid_in_set (value_set_t set
, tree expr
)
1095 switch (TREE_CODE_CLASS (TREE_CODE (expr
)))
1098 case tcc_comparison
:
1100 tree op1
= TREE_OPERAND (expr
, 0);
1101 tree op2
= TREE_OPERAND (expr
, 1);
1102 return set_contains_value (set
, op1
) && set_contains_value (set
, op2
);
1107 tree op1
= TREE_OPERAND (expr
, 0);
1108 return set_contains_value (set
, op1
);
1112 /* XXX: Until PRE of loads works, no reference nodes are ANTIC. */
1115 case tcc_exceptional
:
1116 gcc_assert (TREE_CODE (expr
) == SSA_NAME
);
1119 case tcc_declaration
:
1120 /* VAR_DECL and PARM_DECL are never anticipatable. */
1124 /* No other cases should be encountered. */
1129 /* Clean the set of expressions that are no longer valid in SET. This
1130 means expressions that are made up of values we have no leaders for
1134 clean (value_set_t set
)
1136 value_set_node_t node
;
1137 value_set_node_t next
;
1142 if (!valid_in_set (set
, node
->expr
))
1143 set_remove (set
, node
->expr
);
1148 DEF_VEC_P (basic_block
);
1149 DEF_VEC_ALLOC_P (basic_block
, heap
);
1150 static sbitmap has_abnormal_preds
;
1152 /* Compute the ANTIC set for BLOCK.
1154 If succs(BLOCK) > 1 then
1155 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
1156 else if succs(BLOCK) == 1 then
1157 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
1159 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
1161 XXX: It would be nice to either write a set_clear, and use it for
1162 ANTIC_OUT, or to mark the antic_out set as deleted at the end
1163 of this routine, so that the pool can hand the same memory back out
1164 again for the next ANTIC_OUT. */
1167 compute_antic_aux (basic_block block
, bool block_has_abnormal_pred_edge
)
1170 bool changed
= false;
1171 value_set_t S
, old
, ANTIC_OUT
;
1172 value_set_node_t node
;
1174 ANTIC_OUT
= S
= NULL
;
1176 /* If any edges from predecessors are abnormal, antic_in is empty,
1178 if (block_has_abnormal_pred_edge
)
1179 goto maybe_dump_sets
;
1181 old
= set_new (false);
1182 set_copy (old
, ANTIC_IN (block
));
1183 ANTIC_OUT
= set_new (true);
1185 /* If the block has no successors, ANTIC_OUT is empty. */
1186 if (EDGE_COUNT (block
->succs
) == 0)
1188 /* If we have one successor, we could have some phi nodes to
1189 translate through. */
1190 else if (single_succ_p (block
))
1192 phi_translate_set (ANTIC_OUT
, ANTIC_IN(single_succ (block
)),
1193 block
, single_succ (block
));
1195 /* If we have multiple successors, we take the intersection of all of
1199 VEC(basic_block
, heap
) * worklist
;
1202 basic_block bprime
, first
;
1205 worklist
= VEC_alloc (basic_block
, heap
, EDGE_COUNT (block
->succs
));
1206 FOR_EACH_EDGE (e
, ei
, block
->succs
)
1207 VEC_quick_push (basic_block
, worklist
, e
->dest
);
1208 first
= VEC_index (basic_block
, worklist
, 0);
1209 set_copy (ANTIC_OUT
, ANTIC_IN (first
));
1211 for (i
= 1; VEC_iterate (basic_block
, worklist
, i
, bprime
); i
++)
1213 node
= ANTIC_OUT
->head
;
1217 value_set_node_t next
= node
->next
;
1218 val
= get_value_handle (node
->expr
);
1219 if (!set_contains_value (ANTIC_IN (bprime
), val
))
1220 set_remove (ANTIC_OUT
, node
->expr
);
1224 VEC_free (basic_block
, heap
, worklist
);
1227 /* Generate ANTIC_OUT - TMP_GEN. */
1228 S
= bitmap_set_subtract_from_value_set (ANTIC_OUT
, TMP_GEN (block
), false);
1230 /* Start ANTIC_IN with EXP_GEN - TMP_GEN */
1231 ANTIC_IN (block
) = bitmap_set_subtract_from_value_set (EXP_GEN (block
),
1235 /* Then union in the ANTIC_OUT - TMP_GEN values,
1236 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
1237 for (node
= S
->head
; node
; node
= node
->next
)
1238 value_insert_into_set (ANTIC_IN (block
), node
->expr
);
1240 clean (ANTIC_IN (block
));
1241 if (!set_equal (old
, ANTIC_IN (block
)))
1245 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1248 print_value_set (dump_file
, ANTIC_OUT
, "ANTIC_OUT", block
->index
);
1249 print_value_set (dump_file
, ANTIC_IN (block
), "ANTIC_IN", block
->index
);
1251 print_value_set (dump_file
, S
, "S", block
->index
);
1254 for (son
= first_dom_son (CDI_POST_DOMINATORS
, block
);
1256 son
= next_dom_son (CDI_POST_DOMINATORS
, son
))
1258 changed
|= compute_antic_aux (son
,
1259 TEST_BIT (has_abnormal_preds
, son
->index
));
1264 /* Compute ANTIC sets. */
1267 compute_antic (void)
1269 bool changed
= true;
1270 int num_iterations
= 0;
1273 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
1274 We pre-build the map of blocks with incoming abnormal edges here. */
1275 has_abnormal_preds
= sbitmap_alloc (last_basic_block
);
1276 sbitmap_zero (has_abnormal_preds
);
1282 FOR_EACH_EDGE (e
, ei
, block
->preds
)
1283 if (e
->flags
& EDGE_ABNORMAL
)
1285 SET_BIT (has_abnormal_preds
, block
->index
);
1289 /* While we are here, give empty ANTIC_IN sets to each block. */
1290 ANTIC_IN (block
) = set_new (true);
1292 /* At the exit block we anticipate nothing. */
1293 ANTIC_IN (EXIT_BLOCK_PTR
) = set_new (true);
1299 changed
= compute_antic_aux (EXIT_BLOCK_PTR
, false);
1302 sbitmap_free (has_abnormal_preds
);
1304 if (dump_file
&& (dump_flags
& TDF_STATS
))
1305 fprintf (dump_file
, "compute_antic required %d iterations\n", num_iterations
);
1308 static VEC(tree
,heap
) *inserted_exprs
;
1309 /* Find a leader for an expression, or generate one using
1310 create_expression_by_pieces if it's ANTIC but
1312 BLOCK is the basic_block we are looking for leaders in.
1313 EXPR is the expression to find a leader or generate for.
1314 STMTS is the statement list to put the inserted expressions on.
1315 Returns the SSA_NAME of the LHS of the generated expression or the
1319 find_or_generate_expression (basic_block block
, tree expr
, tree stmts
)
1321 tree genop
= bitmap_find_leader (AVAIL_OUT (block
), expr
);
1323 /* If it's still NULL, it must be a complex expression, so generate
1327 genop
= VALUE_HANDLE_EXPR_SET (expr
)->head
->expr
;
1328 gcc_assert (UNARY_CLASS_P (genop
)
1329 || BINARY_CLASS_P (genop
)
1330 || COMPARISON_CLASS_P (genop
)
1331 || REFERENCE_CLASS_P (genop
));
1332 genop
= create_expression_by_pieces (block
, genop
, stmts
);
1337 #define NECESSARY(stmt) stmt->common.asm_written_flag
1338 /* Create an expression in pieces, so that we can handle very complex
1339 expressions that may be ANTIC, but not necessary GIMPLE.
1340 BLOCK is the basic block the expression will be inserted into,
1341 EXPR is the expression to insert (in value form)
1342 STMTS is a statement list to append the necessary insertions into.
1344 This function will die if we hit some value that shouldn't be
1345 ANTIC but is (IE there is no leader for it, or its components).
1346 This function may also generate expressions that are themselves
1347 partially or fully redundant. Those that are will be either made
1348 fully redundant during the next iteration of insert (for partially
1349 redundant ones), or eliminated by eliminate (for fully redundant
1353 create_expression_by_pieces (basic_block block
, tree expr
, tree stmts
)
1356 tree folded
, forced_stmts
, newexpr
;
1358 tree_stmt_iterator tsi
;
1360 switch (TREE_CODE_CLASS (TREE_CODE (expr
)))
1363 case tcc_comparison
:
1365 tree op1
= TREE_OPERAND (expr
, 0);
1366 tree op2
= TREE_OPERAND (expr
, 1);
1367 tree genop1
= find_or_generate_expression (block
, op1
, stmts
);
1368 tree genop2
= find_or_generate_expression (block
, op2
, stmts
);
1369 folded
= fold (build (TREE_CODE (expr
), TREE_TYPE (expr
),
1376 tree op1
= TREE_OPERAND (expr
, 0);
1377 tree genop1
= find_or_generate_expression (block
, op1
, stmts
);
1378 folded
= fold (build (TREE_CODE (expr
), TREE_TYPE (expr
),
1387 /* Force the generated expression to be a sequence of GIMPLE
1389 We have to call unshare_expr because force_gimple_operand may
1390 modify the tree we pass to it. */
1391 newexpr
= force_gimple_operand (unshare_expr (folded
), &forced_stmts
,
1394 /* If we have any intermediate expressions to the value sets, add them
1395 to the value sets and chain them on in the instruction stream. */
1398 tsi
= tsi_start (forced_stmts
);
1399 for (; !tsi_end_p (tsi
); tsi_next (&tsi
))
1401 tree stmt
= tsi_stmt (tsi
);
1402 tree forcedname
= TREE_OPERAND (stmt
, 0);
1403 tree forcedexpr
= TREE_OPERAND (stmt
, 1);
1404 tree val
= vn_lookup_or_add (forcedexpr
, NULL
);
1406 VEC_safe_push (tree
, heap
, inserted_exprs
, stmt
);
1407 vn_add (forcedname
, val
, NULL
);
1408 bitmap_value_replace_in_set (NEW_SETS (block
), forcedname
);
1409 bitmap_value_replace_in_set (AVAIL_OUT (block
), forcedname
);
1411 tsi
= tsi_last (stmts
);
1412 tsi_link_after (&tsi
, forced_stmts
, TSI_CONTINUE_LINKING
);
1415 /* Build and insert the assignment of the end result to the temporary
1416 that we will return. */
1417 temp
= create_tmp_var (TREE_TYPE (expr
), "pretmp");
1418 add_referenced_tmp_var (temp
);
1419 newexpr
= build (MODIFY_EXPR
, TREE_TYPE (expr
), temp
, newexpr
);
1420 name
= make_ssa_name (temp
, newexpr
);
1421 TREE_OPERAND (newexpr
, 0) = name
;
1422 NECESSARY (newexpr
) = 0;
1423 tsi
= tsi_last (stmts
);
1424 tsi_link_after (&tsi
, newexpr
, TSI_CONTINUE_LINKING
);
1425 VEC_safe_push (tree
, heap
, inserted_exprs
, newexpr
);
1427 /* Add a value handle to the temporary.
1428 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
1429 we are creating the expression by pieces, and this particular piece of
1430 the expression may have been represented. There is no harm in replacing
1432 v
= get_value_handle (expr
);
1433 vn_add (name
, v
, NULL
);
1434 bitmap_value_replace_in_set (NEW_SETS (block
), name
);
1435 bitmap_value_replace_in_set (AVAIL_OUT (block
), name
);
1437 pre_stats
.insertions
++;
1438 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1440 fprintf (dump_file
, "Inserted ");
1441 print_generic_expr (dump_file
, newexpr
, 0);
1442 fprintf (dump_file
, " in predecessor %d\n", block
->index
);
1448 /* Insert the to-be-made-available values of NODE for each predecessor, stored
1449 in AVAIL, into the predecessors of BLOCK, and merge the result with a phi
1450 node, given the same value handle as NODE. The prefix of the phi node is
1451 given with TMPNAME. Return true if we have inserted new stuff. */
1454 insert_into_preds_of_block (basic_block block
, value_set_node_t node
,
1455 tree
*avail
, const char *tmpname
)
1457 tree val
= get_value_handle (node
->expr
);
1459 bool insertions
= false;
1464 tree type
= TREE_TYPE (avail
[EDGE_PRED (block
, 0)->src
->index
]);
1467 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1469 fprintf (dump_file
, "Found partial redundancy for expression ");
1470 print_generic_expr (dump_file
, node
->expr
, 0);
1471 fprintf (dump_file
, "\n");
1474 /* Make sure we aren't creating an induction variable. */
1475 if (block
->loop_depth
> 0 && EDGE_COUNT (block
->preds
) == 2)
1477 bool firstinsideloop
= false;
1478 bool secondinsideloop
= false;
1479 firstinsideloop
= flow_bb_inside_loop_p (block
->loop_father
,
1480 EDGE_PRED (block
, 0)->src
);
1481 secondinsideloop
= flow_bb_inside_loop_p (block
->loop_father
,
1482 EDGE_PRED (block
, 1)->src
);
1483 /* Induction variables only have one edge inside the loop. */
1484 if (firstinsideloop
^ secondinsideloop
)
1486 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1487 fprintf (dump_file
, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
1493 /* Make the necessary insertions. */
1494 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
1496 tree stmts
= alloc_stmt_list ();
1499 eprime
= avail
[bprime
->index
];
1500 if (BINARY_CLASS_P (eprime
)
1501 || COMPARISON_CLASS_P (eprime
)
1502 || UNARY_CLASS_P (eprime
))
1504 builtexpr
= create_expression_by_pieces (bprime
,
1507 bsi_insert_on_edge (pred
, stmts
);
1508 avail
[bprime
->index
] = builtexpr
;
1512 /* If we didn't want a phi node, and we made insertions, we still have
1513 inserted new stuff, and thus return true. If we didn't want a phi node,
1514 and didn't make insertions, we haven't added anything new, so return
1516 if (nophi
&& insertions
)
1518 else if (nophi
&& !insertions
)
1521 /* Now build a phi for the new variable. */
1522 temp
= create_tmp_var (type
, tmpname
);
1523 add_referenced_tmp_var (temp
);
1524 temp
= create_phi_node (temp
, block
);
1525 NECESSARY (temp
) = 0;
1526 VEC_safe_push (tree
, heap
, inserted_exprs
, temp
);
1527 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
1528 add_phi_arg (temp
, avail
[pred
->src
->index
], pred
);
1530 vn_add (PHI_RESULT (temp
), val
, NULL
);
1532 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
1533 this insertion, since we test for the existence of this value in PHI_GEN
1534 before proceeding with the partial redundancy checks in insert_aux.
1536 The value may exist in AVAIL_OUT, in particular, it could be represented
1537 by the expression we are trying to eliminate, in which case we want the
1538 replacement to occur. If it's not existing in AVAIL_OUT, we want it
1541 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
1542 this block, because if it did, it would have existed in our dominator's
1543 AVAIL_OUT, and would have been skipped due to the full redundancy check.
1546 bitmap_insert_into_set (PHI_GEN (block
),
1548 bitmap_value_replace_in_set (AVAIL_OUT (block
),
1550 bitmap_insert_into_set (NEW_SETS (block
),
1553 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1555 fprintf (dump_file
, "Created phi ");
1556 print_generic_expr (dump_file
, temp
, 0);
1557 fprintf (dump_file
, " in block %d\n", block
->index
);
1565 /* Perform insertion of partially redundant values.
1566 For BLOCK, do the following:
1567 1. Propagate the NEW_SETS of the dominator into the current block.
1568 If the block has multiple predecessors,
1569 2a. Iterate over the ANTIC expressions for the block to see if
1570 any of them are partially redundant.
1571 2b. If so, insert them into the necessary predecessors to make
1572 the expression fully redundant.
1573 2c. Insert a new PHI merging the values of the predecessors.
1574 2d. Insert the new PHI, and the new expressions, into the
1576 3. Recursively call ourselves on the dominator children of BLOCK.
1581 insert_aux (basic_block block
)
1584 bool new_stuff
= false;
1589 dom
= get_immediate_dominator (CDI_DOMINATORS
, block
);
1594 bitmap_set_t newset
= NEW_SETS (dom
);
1597 /* Note that we need to value_replace both NEW_SETS, and
1598 AVAIL_OUT. For both the case of NEW_SETS, the value may be
1599 represented by some non-simple expression here that we want
1600 to replace it with. */
1601 EXECUTE_IF_SET_IN_BITMAP (newset
->expressions
, 0, i
, bi
)
1603 bitmap_value_replace_in_set (NEW_SETS (block
), ssa_name (i
));
1604 bitmap_value_replace_in_set (AVAIL_OUT (block
), ssa_name (i
));
1607 if (!single_pred_p (block
))
1609 value_set_node_t node
;
1610 for (node
= ANTIC_IN (block
)->head
;
1614 if (BINARY_CLASS_P (node
->expr
)
1615 || COMPARISON_CLASS_P (node
->expr
)
1616 || UNARY_CLASS_P (node
->expr
))
1620 bool by_some
= false;
1621 bool cant_insert
= false;
1622 bool all_same
= true;
1623 tree first_s
= NULL
;
1626 tree eprime
= NULL_TREE
;
1629 val
= get_value_handle (node
->expr
);
1630 if (bitmap_set_contains_value (PHI_GEN (block
), val
))
1632 if (bitmap_set_contains_value (AVAIL_OUT (dom
), val
))
1634 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1635 fprintf (dump_file
, "Found fully redundant value\n");
1639 avail
= xcalloc (last_basic_block
, sizeof (tree
));
1640 FOR_EACH_EDGE (pred
, ei
, block
->preds
)
1645 /* This can happen in the very weird case
1646 that our fake infinite loop edges have caused a
1647 critical edge to appear. */
1648 if (EDGE_CRITICAL_P (pred
))
1654 eprime
= phi_translate (node
->expr
,
1658 /* eprime will generally only be NULL if the
1659 value of the expression, translated
1660 through the PHI for this predecessor, is
1661 undefined. If that is the case, we can't
1662 make the expression fully redundant,
1663 because its value is undefined along a
1664 predecessor path. We can thus break out
1665 early because it doesn't matter what the
1666 rest of the results are. */
1673 eprime
= fully_constant_expression (eprime
);
1674 vprime
= get_value_handle (eprime
);
1675 gcc_assert (vprime
);
1676 edoubleprime
= bitmap_find_leader (AVAIL_OUT (bprime
),
1678 if (edoubleprime
== NULL
)
1680 avail
[bprime
->index
] = eprime
;
1685 avail
[bprime
->index
] = edoubleprime
;
1687 if (first_s
== NULL
)
1688 first_s
= edoubleprime
;
1689 else if (!operand_equal_p (first_s
, edoubleprime
,
1694 /* If we can insert it, it's not the same value
1695 already existing along every predecessor, and
1696 it's defined by some predecessor, it is
1697 partially redundant. */
1698 if (!cant_insert
&& !all_same
&& by_some
)
1700 if (insert_into_preds_of_block (block
, node
, avail
,
1704 /* If all edges produce the same value and that value is
1705 an invariant, then the PHI has the same value on all
1706 edges. Note this. */
1707 else if (!cant_insert
&& all_same
&& eprime
1708 && is_gimple_min_invariant (eprime
)
1709 && !is_gimple_min_invariant (val
))
1711 value_set_t exprset
= VALUE_HANDLE_EXPR_SET (val
);
1712 value_set_node_t node
;
1713 for (node
= exprset
->head
; node
; node
= node
->next
)
1715 if (TREE_CODE (node
->expr
) == SSA_NAME
)
1717 vn_add (node
->expr
, eprime
, NULL
);
1718 pre_stats
.constified
++;
1728 for (son
= first_dom_son (CDI_DOMINATORS
, block
);
1730 son
= next_dom_son (CDI_DOMINATORS
, son
))
1732 new_stuff
|= insert_aux (son
);
1738 /* Perform insertion of partially redundant values. */
1743 bool new_stuff
= true;
1745 int num_iterations
= 0;
1748 NEW_SETS (bb
) = bitmap_set_new ();
1754 new_stuff
= insert_aux (ENTRY_BLOCK_PTR
);
1756 if (num_iterations
> 2 && dump_file
&& (dump_flags
& TDF_STATS
))
1757 fprintf (dump_file
, "insert required %d iterations\n", num_iterations
);
1761 /* Return true if VAR is an SSA variable with no defining statement in
1762 this procedure, *AND* isn't a live-on-entry parameter. */
1765 is_undefined_value (tree expr
)
1767 return (TREE_CODE (expr
) == SSA_NAME
1768 && IS_EMPTY_STMT (SSA_NAME_DEF_STMT (expr
))
1769 /* PARM_DECLs and hard registers are always defined. */
1770 && TREE_CODE (SSA_NAME_VAR (expr
)) != PARM_DECL
);
1774 /* Given an SSA variable VAR and an expression EXPR, compute the value
1775 number for EXPR and create a value handle (VAL) for it. If VAR and
1776 EXPR are not the same, associate VAL with VAR. Finally, add VAR to
1777 S1 and its value handle to S2.
1779 VUSES represent the virtual use operands associated with EXPR (if
1780 any). They are used when computing the hash value for EXPR. */
1783 add_to_sets (tree var
, tree expr
, tree stmt
, bitmap_set_t s1
,
1786 tree val
= vn_lookup_or_add (expr
, stmt
);
1788 /* VAR and EXPR may be the same when processing statements for which
1789 we are not computing value numbers (e.g., non-assignments, or
1790 statements that make aliased stores). In those cases, we are
1791 only interested in making VAR available as its own value. */
1793 vn_add (var
, val
, NULL_TREE
);
1796 bitmap_insert_into_set (s1
, var
);
1797 bitmap_value_insert_into_set (s2
, var
);
1801 /* Given a unary or binary expression EXPR, create and return a new
1802 expression with the same structure as EXPR but with its operands
1803 replaced with the value handles of each of the operands of EXPR.
1805 VUSES represent the virtual use operands associated with EXPR (if
1806 any). They are used when computing the hash value for EXPR.
1807 Insert EXPR's operands into the EXP_GEN set for BLOCK. */
1810 create_value_expr_from (tree expr
, basic_block block
, tree stmt
)
1813 enum tree_code code
= TREE_CODE (expr
);
1817 gcc_assert (TREE_CODE_CLASS (code
) == tcc_unary
1818 || TREE_CODE_CLASS (code
) == tcc_binary
1819 || TREE_CODE_CLASS (code
) == tcc_comparison
1820 || TREE_CODE_CLASS (code
) == tcc_reference
);
1822 if (TREE_CODE_CLASS (code
) == tcc_unary
)
1823 pool
= unary_node_pool
;
1824 else if (TREE_CODE_CLASS (code
) == tcc_reference
)
1825 pool
= reference_node_pool
;
1827 pool
= binary_node_pool
;
1829 vexpr
= pool_alloc (pool
);
1830 memcpy (vexpr
, expr
, tree_size (expr
));
1832 for (i
= 0; i
< TREE_CODE_LENGTH (code
); i
++)
1836 op
= TREE_OPERAND (expr
, i
);
1837 if (op
== NULL_TREE
)
1840 /* If OP is a constant that has overflowed, do not value number
1842 if (CONSTANT_CLASS_P (op
)
1843 && TREE_OVERFLOW (op
))
1845 pool_free (pool
, vexpr
);
1849 /* Recursively value-numberize reference ops */
1850 if (REFERENCE_CLASS_P (op
))
1852 tree tempop
= create_value_expr_from (op
, block
, stmt
);
1853 op
= tempop
? tempop
: op
;
1854 val
= vn_lookup_or_add (op
, stmt
);
1857 /* Create a value handle for OP and add it to VEXPR. */
1858 val
= vn_lookup_or_add (op
, NULL
);
1860 if (!is_undefined_value (op
))
1861 value_insert_into_set (EXP_GEN (block
), op
);
1863 if (TREE_CODE (val
) == VALUE_HANDLE
)
1864 TREE_TYPE (val
) = TREE_TYPE (TREE_OPERAND (vexpr
, i
));
1866 TREE_OPERAND (vexpr
, i
) = val
;
1873 /* Compute the AVAIL set for all basic blocks.
1875 This function performs value numbering of the statements in each basic
1876 block. The AVAIL sets are built from information we glean while doing
1877 this value numbering, since the AVAIL sets contain only one entry per
1880 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
1881 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
1884 compute_avail (void)
1886 basic_block block
, son
;
1887 basic_block
*worklist
;
1891 /* For arguments with default definitions, we pretend they are
1892 defined in the entry block. */
1893 for (param
= DECL_ARGUMENTS (current_function_decl
);
1895 param
= TREE_CHAIN (param
))
1897 if (default_def (param
) != NULL
)
1899 tree def
= default_def (param
);
1900 vn_lookup_or_add (def
, NULL
);
1901 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR
), def
);
1902 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR
), def
);
1906 /* Allocate the worklist. */
1907 worklist
= xmalloc (sizeof (basic_block
) * n_basic_blocks
);
1909 /* Seed the algorithm by putting the dominator children of the entry
1910 block on the worklist. */
1911 for (son
= first_dom_son (CDI_DOMINATORS
, ENTRY_BLOCK_PTR
);
1913 son
= next_dom_son (CDI_DOMINATORS
, son
))
1914 worklist
[sp
++] = son
;
1916 /* Loop until the worklist is empty. */
1919 block_stmt_iterator bsi
;
1923 /* Pick a block from the worklist. */
1924 block
= worklist
[--sp
];
1926 /* Initially, the set of available values in BLOCK is that of
1927 its immediate dominator. */
1928 dom
= get_immediate_dominator (CDI_DOMINATORS
, block
);
1930 bitmap_set_copy (AVAIL_OUT (block
), AVAIL_OUT (dom
));
1932 /* Generate values for PHI nodes. */
1933 for (phi
= phi_nodes (block
); phi
; phi
= PHI_CHAIN (phi
))
1934 /* We have no need for virtual phis, as they don't represent
1935 actual computations. */
1936 if (is_gimple_reg (PHI_RESULT (phi
)))
1937 add_to_sets (PHI_RESULT (phi
), PHI_RESULT (phi
), NULL
,
1938 PHI_GEN (block
), AVAIL_OUT (block
));
1940 /* Now compute value numbers and populate value sets with all
1941 the expressions computed in BLOCK. */
1942 for (bsi
= bsi_start (block
); !bsi_end_p (bsi
); bsi_next (&bsi
))
1948 stmt
= bsi_stmt (bsi
);
1949 ann
= stmt_ann (stmt
);
1951 /* We are only interested in assignments of the form
1952 X_i = EXPR, where EXPR represents an "interesting"
1953 computation, it has no volatile operands and X_i
1954 doesn't flow through an abnormal edge. */
1955 if (TREE_CODE (stmt
) == MODIFY_EXPR
1956 && !ann
->has_volatile_ops
1957 && TREE_CODE (TREE_OPERAND (stmt
, 0)) == SSA_NAME
1958 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (TREE_OPERAND (stmt
, 0)))
1960 tree lhs
= TREE_OPERAND (stmt
, 0);
1961 tree rhs
= TREE_OPERAND (stmt
, 1);
1963 STRIP_USELESS_TYPE_CONVERSION (rhs
);
1964 if (UNARY_CLASS_P (rhs
)
1965 || BINARY_CLASS_P (rhs
)
1966 || COMPARISON_CLASS_P (rhs
)
1967 || REFERENCE_CLASS_P (rhs
))
1969 /* For binary, unary, and reference expressions,
1970 create a duplicate expression with the operands
1971 replaced with the value handles of the original
1973 tree newt
= create_value_expr_from (rhs
, block
, stmt
);
1976 add_to_sets (lhs
, newt
, stmt
, TMP_GEN (block
),
1978 value_insert_into_set (EXP_GEN (block
), newt
);
1982 else if (TREE_CODE (rhs
) == SSA_NAME
1983 || is_gimple_min_invariant (rhs
)
1984 || TREE_CODE (rhs
) == ADDR_EXPR
1985 || TREE_INVARIANT (rhs
)
1988 /* Compute a value number for the RHS of the statement
1989 and add its value to the AVAIL_OUT set for the block.
1990 Add the LHS to TMP_GEN. */
1991 add_to_sets (lhs
, rhs
, stmt
, TMP_GEN (block
),
1994 if (TREE_CODE (rhs
) == SSA_NAME
1995 && !is_undefined_value (rhs
))
1996 value_insert_into_set (EXP_GEN (block
), rhs
);
2001 /* For any other statement that we don't recognize, simply
2002 make the names generated by the statement available in
2003 AVAIL_OUT and TMP_GEN. */
2004 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_DEF
)
2005 add_to_sets (op
, op
, NULL
, TMP_GEN (block
), AVAIL_OUT (block
));
2007 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_USE
)
2008 add_to_sets (op
, op
, NULL
, NULL
, AVAIL_OUT (block
));
2011 /* Put the dominator children of BLOCK on the worklist of blocks
2012 to compute available sets for. */
2013 for (son
= first_dom_son (CDI_DOMINATORS
, block
);
2015 son
= next_dom_son (CDI_DOMINATORS
, son
))
2016 worklist
[sp
++] = son
;
2023 /* Eliminate fully redundant computations. */
2032 block_stmt_iterator i
;
2034 for (i
= bsi_start (b
); !bsi_end_p (i
); bsi_next (&i
))
2036 tree stmt
= bsi_stmt (i
);
2038 /* Lookup the RHS of the expression, see if we have an
2039 available computation for it. If so, replace the RHS with
2040 the available computation. */
2041 if (TREE_CODE (stmt
) == MODIFY_EXPR
2042 && TREE_CODE (TREE_OPERAND (stmt
, 0)) == SSA_NAME
2043 && TREE_CODE (TREE_OPERAND (stmt
,1)) != SSA_NAME
2044 && !is_gimple_min_invariant (TREE_OPERAND (stmt
, 1))
2045 && !stmt_ann (stmt
)->has_volatile_ops
)
2047 tree lhs
= TREE_OPERAND (stmt
, 0);
2048 tree
*rhs_p
= &TREE_OPERAND (stmt
, 1);
2051 sprime
= bitmap_find_leader (AVAIL_OUT (b
),
2052 vn_lookup (lhs
, NULL
));
2055 && (TREE_CODE (*rhs_p
) != SSA_NAME
2056 || may_propagate_copy (*rhs_p
, sprime
)))
2058 gcc_assert (sprime
!= *rhs_p
);
2060 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2062 fprintf (dump_file
, "Replaced ");
2063 print_generic_expr (dump_file
, *rhs_p
, 0);
2064 fprintf (dump_file
, " with ");
2065 print_generic_expr (dump_file
, sprime
, 0);
2066 fprintf (dump_file
, " in ");
2067 print_generic_stmt (dump_file
, stmt
, 0);
2069 if (TREE_CODE (sprime
) == SSA_NAME
)
2070 NECESSARY (SSA_NAME_DEF_STMT (sprime
)) = 1;
2071 pre_stats
.eliminations
++;
2072 propagate_tree_value (rhs_p
, sprime
);
2075 /* If we removed EH side effects from the statement, clean
2076 its EH information. */
2077 if (maybe_clean_eh_stmt (stmt
))
2079 bitmap_set_bit (need_eh_cleanup
,
2080 bb_for_stmt (stmt
)->index
);
2081 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2082 fprintf (dump_file
, " Removed EH side effects.\n");
2090 /* Borrow a bit of tree-ssa-dce.c for the moment.
2091 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
2092 this may be a bit faster, and we may want critical edges kept split. */
2094 /* If OP's defining statement has not already been determined to be necessary,
2095 mark that statement necessary. Return the stmt, if it is newly
2099 mark_operand_necessary (tree op
)
2105 stmt
= SSA_NAME_DEF_STMT (op
);
2108 if (NECESSARY (stmt
)
2109 || IS_EMPTY_STMT (stmt
))
2112 NECESSARY (stmt
) = 1;
2116 /* Because we don't follow exactly the standard PRE algorithm, and decide not
2117 to insert PHI nodes sometimes, and because value numbering of casts isn't
2118 perfect, we sometimes end up inserting dead code. This simple DCE-like
2119 pass removes any insertions we made that weren't actually used. */
2122 remove_dead_inserted_code (void)
2124 VEC(tree
,heap
) *worklist
= NULL
;
2128 worklist
= VEC_alloc (tree
, heap
, VEC_length (tree
, inserted_exprs
));
2129 for (i
= 0; VEC_iterate (tree
, inserted_exprs
, i
, t
); i
++)
2132 VEC_quick_push (tree
, worklist
, t
);
2134 while (VEC_length (tree
, worklist
) > 0)
2136 t
= VEC_pop (tree
, worklist
);
2137 if (TREE_CODE (t
) == PHI_NODE
)
2139 /* PHI nodes are somewhat special in that each PHI alternative has
2140 data and control dependencies. All the statements feeding the
2141 PHI node's arguments are always necessary. In aggressive mode,
2142 we also consider the control dependent edges leading to the
2143 predecessor block associated with each PHI alternative as
2147 VEC_reserve (tree
, heap
, worklist
, PHI_NUM_ARGS (t
));
2148 for (k
= 0; k
< PHI_NUM_ARGS (t
); k
++)
2150 tree arg
= PHI_ARG_DEF (t
, k
);
2151 if (TREE_CODE (arg
) == SSA_NAME
)
2153 arg
= mark_operand_necessary (arg
);
2155 VEC_quick_push (tree
, worklist
, arg
);
2161 /* Propagate through the operands. Examine all the USE, VUSE and
2162 V_MAY_DEF operands in this statement. Mark all the statements
2163 which feed this statement's uses as necessary. */
2167 /* The operands of V_MAY_DEF expressions are also needed as they
2168 represent potential definitions that may reach this
2169 statement (V_MAY_DEF operands allow us to follow def-def
2172 FOR_EACH_SSA_TREE_OPERAND (use
, t
, iter
, SSA_OP_ALL_USES
)
2174 tree n
= mark_operand_necessary (use
);
2176 VEC_safe_push (tree
, heap
, worklist
, n
);
2180 for (i
= 0; VEC_iterate (tree
, inserted_exprs
, i
, t
); i
++)
2184 block_stmt_iterator bsi
;
2185 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2187 fprintf (dump_file
, "Removing unnecessary insertion:");
2188 print_generic_stmt (dump_file
, t
, 0);
2190 if (TREE_CODE (t
) == PHI_NODE
)
2192 remove_phi_node (t
, NULL
);
2196 bsi
= bsi_for_stmt (t
);
2201 VEC_free (tree
, heap
, worklist
);
2203 /* Initialize data structures used by PRE. */
2206 init_pre (bool do_fre
)
2210 inserted_exprs
= NULL
;
2213 current_loops
= loop_optimizer_init (dump_file
);
2214 connect_infinite_loops_to_exit ();
2215 memset (&pre_stats
, 0, sizeof (pre_stats
));
2217 /* If block 0 has more than one predecessor, it means that its PHI
2218 nodes will have arguments coming from block -1. This creates
2219 problems for several places in PRE that keep local arrays indexed
2220 by block number. To prevent this, we split the edge coming from
2221 ENTRY_BLOCK_PTR (FIXME, if ENTRY_BLOCK_PTR had an index number
2222 different than -1 we wouldn't have to hack this. tree-ssa-dce.c
2223 needs a similar change). */
2224 if (!single_pred_p (single_succ (ENTRY_BLOCK_PTR
)))
2225 if (!(single_succ_edge (ENTRY_BLOCK_PTR
)->flags
& EDGE_ABNORMAL
))
2226 split_edge (single_succ_edge (ENTRY_BLOCK_PTR
));
2229 bb
->aux
= xcalloc (1, sizeof (struct bb_value_sets
));
2231 bitmap_obstack_initialize (&grand_bitmap_obstack
);
2232 phi_translate_table
= htab_create (511, expr_pred_trans_hash
,
2233 expr_pred_trans_eq
, free
);
2234 value_set_pool
= create_alloc_pool ("Value sets",
2235 sizeof (struct value_set
), 30);
2236 bitmap_set_pool
= create_alloc_pool ("Bitmap sets",
2237 sizeof (struct bitmap_set
), 30);
2238 value_set_node_pool
= create_alloc_pool ("Value set nodes",
2239 sizeof (struct value_set_node
), 30);
2240 calculate_dominance_info (CDI_POST_DOMINATORS
);
2241 calculate_dominance_info (CDI_DOMINATORS
);
2242 binary_node_pool
= create_alloc_pool ("Binary tree nodes",
2243 tree_code_size (PLUS_EXPR
), 30);
2244 unary_node_pool
= create_alloc_pool ("Unary tree nodes",
2245 tree_code_size (NEGATE_EXPR
), 30);
2246 reference_node_pool
= create_alloc_pool ("Reference tree nodes",
2247 tree_code_size (ARRAY_REF
), 30);
2250 EXP_GEN (bb
) = set_new (true);
2251 PHI_GEN (bb
) = bitmap_set_new ();
2252 TMP_GEN (bb
) = bitmap_set_new ();
2253 AVAIL_OUT (bb
) = bitmap_set_new ();
2256 need_eh_cleanup
= BITMAP_ALLOC (NULL
);
2260 /* Deallocate data structures used by PRE. */
2263 fini_pre (bool do_fre
)
2268 VEC_free (tree
, heap
, inserted_exprs
);
2269 bitmap_obstack_release (&grand_bitmap_obstack
);
2270 free_alloc_pool (value_set_pool
);
2271 free_alloc_pool (bitmap_set_pool
);
2272 free_alloc_pool (value_set_node_pool
);
2273 free_alloc_pool (binary_node_pool
);
2274 free_alloc_pool (reference_node_pool
);
2275 free_alloc_pool (unary_node_pool
);
2276 htab_delete (phi_translate_table
);
2277 remove_fake_exit_edges ();
2285 free_dominance_info (CDI_POST_DOMINATORS
);
2288 if (!bitmap_empty_p (need_eh_cleanup
))
2290 tree_purge_all_dead_eh_edges (need_eh_cleanup
);
2291 cleanup_tree_cfg ();
2294 BITMAP_FREE (need_eh_cleanup
);
2296 /* Wipe out pointers to VALUE_HANDLEs. In the not terribly distant
2297 future we will want them to be persistent though. */
2298 for (i
= 0; i
< num_ssa_names
; i
++)
2300 tree name
= ssa_name (i
);
2305 if (SSA_NAME_VALUE (name
)
2306 && TREE_CODE (SSA_NAME_VALUE (name
)) == VALUE_HANDLE
)
2307 SSA_NAME_VALUE (name
) = NULL
;
2309 if (!do_fre
&& current_loops
)
2311 loop_optimizer_finalize (current_loops
, dump_file
);
2312 current_loops
= NULL
;
2317 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
2318 only wants to do full redundancy elimination. */
2321 execute_pre (bool do_fre
)
2325 /* Collect and value number expressions computed in each basic block. */
2328 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2334 print_value_set (dump_file
, EXP_GEN (bb
), "exp_gen", bb
->index
);
2335 bitmap_print_value_set (dump_file
, TMP_GEN (bb
), "tmp_gen",
2337 bitmap_print_value_set (dump_file
, AVAIL_OUT (bb
), "avail_out",
2342 /* Insert can get quite slow on an incredibly large number of basic
2343 blocks due to some quadratic behavior. Until this behavior is
2344 fixed, don't run it when he have an incredibly large number of
2345 bb's. If we aren't going to run insert, there is no point in
2346 computing ANTIC, either, even though it's plenty fast. */
2347 if (!do_fre
&& n_basic_blocks
< 4000)
2353 /* Remove all the redundant expressions. */
2357 if (dump_file
&& (dump_flags
& TDF_STATS
))
2359 fprintf (dump_file
, "Insertions: %d\n", pre_stats
.insertions
);
2360 fprintf (dump_file
, "New PHIs: %d\n", pre_stats
.phis
);
2361 fprintf (dump_file
, "Eliminated: %d\n", pre_stats
.eliminations
);
2362 fprintf (dump_file
, "Constified: %d\n", pre_stats
.constified
);
2365 bsi_commit_edge_inserts ();
2367 remove_dead_inserted_code ();
2373 /* Gate and execute functions for PRE. */
2378 execute_pre (false);
2384 return flag_tree_pre
!= 0;
2387 struct tree_opt_pass pass_pre
=
2390 gate_pre
, /* gate */
2391 do_pre
, /* execute */
2394 0, /* static_pass_number */
2395 TV_TREE_PRE
, /* tv_id */
2396 PROP_no_crit_edges
| PROP_cfg
2397 | PROP_ssa
| PROP_alias
, /* properties_required */
2398 0, /* properties_provided */
2399 0, /* properties_destroyed */
2400 0, /* todo_flags_start */
2401 TODO_dump_func
| TODO_ggc_collect
| TODO_verify_ssa
, /* todo_flags_finish */
2406 /* Gate and execute functions for FRE. */
2417 return flag_tree_fre
!= 0;
2420 struct tree_opt_pass pass_fre
=
2423 gate_fre
, /* gate */
2424 execute_fre
, /* execute */
2427 0, /* static_pass_number */
2428 TV_TREE_FRE
, /* tv_id */
2429 PROP_cfg
| PROP_ssa
| PROP_alias
, /* properties_required */
2430 0, /* properties_provided */
2431 0, /* properties_destroyed */
2432 0, /* todo_flags_start */
2433 TODO_dump_func
| TODO_ggc_collect
| TODO_verify_ssa
, /* todo_flags_finish */