2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
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"
28 #include "basic-block.h"
29 #include "tree-pretty-print.h"
30 #include "gimple-pretty-print.h"
31 #include "tree-inline.h"
32 #include "tree-flow.h"
34 #include "tree-dump.h"
38 #include "tree-iterator.h"
39 #include "alloc-pool.h"
41 #include "tree-pass.h"
44 #include "langhooks.h"
46 #include "tree-ssa-sccvn.h"
47 #include "tree-scalar-evolution.h"
53 1. Avail sets can be shared by making an avail_find_leader that
54 walks up the dominator tree and looks in those avail sets.
55 This might affect code optimality, it's unclear right now.
56 2. Strength reduction can be performed by anticipating expressions
57 we can repair later on.
58 3. We can do back-substitution or smarter value numbering to catch
59 commutative expressions split up over multiple statements.
62 /* For ease of terminology, "expression node" in the below refers to
63 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
64 represent the actual statement containing the expressions we care about,
65 and we cache the value number by putting it in the expression. */
69 First we walk the statements to generate the AVAIL sets, the
70 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
71 generation of values/expressions by a given block. We use them
72 when computing the ANTIC sets. The AVAIL sets consist of
73 SSA_NAME's that represent values, so we know what values are
74 available in what blocks. AVAIL is a forward dataflow problem. In
75 SSA, values are never killed, so we don't need a kill set, or a
76 fixpoint iteration, in order to calculate the AVAIL sets. In
77 traditional parlance, AVAIL sets tell us the downsafety of the
80 Next, we generate the ANTIC sets. These sets represent the
81 anticipatable expressions. ANTIC is a backwards dataflow
82 problem. An expression is anticipatable in a given block if it could
83 be generated in that block. This means that if we had to perform
84 an insertion in that block, of the value of that expression, we
85 could. Calculating the ANTIC sets requires phi translation of
86 expressions, because the flow goes backwards through phis. We must
87 iterate to a fixpoint of the ANTIC sets, because we have a kill
88 set. Even in SSA form, values are not live over the entire
89 function, only from their definition point onwards. So we have to
90 remove values from the ANTIC set once we go past the definition
91 point of the leaders that make them up.
92 compute_antic/compute_antic_aux performs this computation.
94 Third, we perform insertions to make partially redundant
95 expressions fully redundant.
97 An expression is partially redundant (excluding partial
100 1. It is AVAIL in some, but not all, of the predecessors of a
102 2. It is ANTIC in all the predecessors.
104 In order to make it fully redundant, we insert the expression into
105 the predecessors where it is not available, but is ANTIC.
107 For the partial anticipation case, we only perform insertion if it
108 is partially anticipated in some block, and fully available in all
111 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
112 performs these steps.
114 Fourth, we eliminate fully redundant expressions.
115 This is a simple statement walk that replaces redundant
116 calculations with the now available values. */
118 /* Representations of value numbers:
120 Value numbers are represented by a representative SSA_NAME. We
121 will create fake SSA_NAME's in situations where we need a
122 representative but do not have one (because it is a complex
123 expression). In order to facilitate storing the value numbers in
124 bitmaps, and keep the number of wasted SSA_NAME's down, we also
125 associate a value_id with each value number, and create full blown
126 ssa_name's only where we actually need them (IE in operands of
127 existing expressions).
129 Theoretically you could replace all the value_id's with
130 SSA_NAME_VERSION, but this would allocate a large number of
131 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
132 It would also require an additional indirection at each point we
135 /* Representation of expressions on value numbers:
137 Expressions consisting of value numbers are represented the same
138 way as our VN internally represents them, with an additional
139 "pre_expr" wrapping around them in order to facilitate storing all
140 of the expressions in the same sets. */
142 /* Representation of sets:
144 The dataflow sets do not need to be sorted in any particular order
145 for the majority of their lifetime, are simply represented as two
146 bitmaps, one that keeps track of values present in the set, and one
147 that keeps track of expressions present in the set.
149 When we need them in topological order, we produce it on demand by
150 transforming the bitmap into an array and sorting it into topo
153 /* Type of expression, used to know which member of the PRE_EXPR union
164 typedef union pre_expr_union_d
169 vn_reference_t reference;
172 typedef struct pre_expr_d
174 enum pre_expr_kind kind;
179 #define PRE_EXPR_NAME(e) (e)->u.name
180 #define PRE_EXPR_NARY(e) (e)->u.nary
181 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
182 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
185 pre_expr_eq (const void *p1, const void *p2)
187 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
188 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
190 if (e1->kind != e2->kind)
196 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
197 PRE_EXPR_CONSTANT (e2));
199 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
201 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
203 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
204 PRE_EXPR_REFERENCE (e2));
211 pre_expr_hash (const void *p1)
213 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
217 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
219 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
221 return PRE_EXPR_NARY (e)->hashcode;
223 return PRE_EXPR_REFERENCE (e)->hashcode;
230 /* Next global expression id number. */
231 static unsigned int next_expression_id;
233 /* Mapping from expression to id number we can use in bitmap sets. */
234 DEF_VEC_P (pre_expr);
235 DEF_VEC_ALLOC_P (pre_expr, heap);
236 static VEC(pre_expr, heap) *expressions;
237 static htab_t expression_to_id;
238 static VEC(unsigned, heap) *name_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 if (expr->kind == NAME)
252 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
253 /* VEC_safe_grow_cleared allocates no headroom. Avoid frequent
254 re-allocations by using VEC_reserve upfront. There is no
255 VEC_quick_grow_cleared unfortunately. */
256 VEC_reserve (unsigned, heap, name_to_id, num_ssa_names);
257 VEC_safe_grow_cleared (unsigned, heap, name_to_id, num_ssa_names);
258 gcc_assert (VEC_index (unsigned, name_to_id, version) == 0);
259 VEC_replace (unsigned, name_to_id, version, expr->id);
263 slot = htab_find_slot (expression_to_id, expr, INSERT);
267 return next_expression_id - 1;
270 /* Return the expression id for tree EXPR. */
272 static inline unsigned int
273 get_expression_id (const pre_expr expr)
278 static inline unsigned int
279 lookup_expression_id (const pre_expr expr)
283 if (expr->kind == NAME)
285 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
286 if (VEC_length (unsigned, name_to_id) <= version)
288 return VEC_index (unsigned, name_to_id, version);
292 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
295 return ((pre_expr)*slot)->id;
299 /* Return the existing expression id for EXPR, or create one if one
300 does not exist yet. */
302 static inline unsigned int
303 get_or_alloc_expression_id (pre_expr expr)
305 unsigned int id = lookup_expression_id (expr);
307 return alloc_expression_id (expr);
308 return expr->id = id;
311 /* Return the expression that has expression id ID */
313 static inline pre_expr
314 expression_for_id (unsigned int id)
316 return VEC_index (pre_expr, expressions, id);
319 /* Free the expression id field in all of our expressions,
320 and then destroy the expressions array. */
323 clear_expression_ids (void)
325 VEC_free (pre_expr, heap, expressions);
328 static alloc_pool pre_expr_pool;
330 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
333 get_or_alloc_expr_for_name (tree name)
335 struct pre_expr_d expr;
337 unsigned int result_id;
341 PRE_EXPR_NAME (&expr) = name;
342 result_id = lookup_expression_id (&expr);
344 return expression_for_id (result_id);
346 result = (pre_expr) pool_alloc (pre_expr_pool);
348 PRE_EXPR_NAME (result) = name;
349 alloc_expression_id (result);
353 static bool in_fre = false;
355 /* An unordered bitmap set. One bitmap tracks values, the other,
357 typedef struct bitmap_set
363 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
364 EXECUTE_IF_SET_IN_BITMAP((set)->expressions, 0, (id), (bi))
366 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
367 EXECUTE_IF_SET_IN_BITMAP((set)->values, 0, (id), (bi))
369 /* Mapping from value id to expressions with that value_id. */
370 DEF_VEC_P (bitmap_set_t);
371 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
372 static VEC(bitmap_set_t, heap) *value_expressions;
374 /* Sets that we need to keep track of. */
375 typedef struct bb_bitmap_sets
377 /* The EXP_GEN set, which represents expressions/values generated in
379 bitmap_set_t exp_gen;
381 /* The PHI_GEN set, which represents PHI results generated in a
383 bitmap_set_t phi_gen;
385 /* The TMP_GEN set, which represents results/temporaries generated
386 in a basic block. IE the LHS of an expression. */
387 bitmap_set_t tmp_gen;
389 /* The AVAIL_OUT set, which represents which values are available in
390 a given basic block. */
391 bitmap_set_t avail_out;
393 /* The ANTIC_IN set, which represents which values are anticipatable
394 in a given basic block. */
395 bitmap_set_t antic_in;
397 /* The PA_IN set, which represents which values are
398 partially anticipatable in a given basic block. */
401 /* The NEW_SETS set, which is used during insertion to augment the
402 AVAIL_OUT set of blocks with the new insertions performed during
403 the current iteration. */
404 bitmap_set_t new_sets;
406 /* A cache for value_dies_in_block_x. */
409 /* True if we have visited this block during ANTIC calculation. */
410 unsigned int visited : 1;
412 /* True we have deferred processing this block during ANTIC
413 calculation until its successor is processed. */
414 unsigned int deferred : 1;
416 /* True when the block contains a call that might not return. */
417 unsigned int contains_may_not_return_call : 1;
420 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
421 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
422 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
423 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
424 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
425 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
426 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
427 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
428 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
429 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
430 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
433 /* Basic block list in postorder. */
434 static int *postorder;
436 /* This structure is used to keep track of statistics on what
437 optimization PRE was able to perform. */
440 /* The number of RHS computations eliminated by PRE. */
443 /* The number of new expressions/temporaries generated by PRE. */
446 /* The number of inserts found due to partial anticipation */
449 /* The number of new PHI nodes added by PRE. */
452 /* The number of values found constant. */
457 static bool do_partial_partial;
458 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
459 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
460 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
461 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
462 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
463 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
464 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
466 static bitmap_set_t bitmap_set_new (void);
467 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
469 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
471 static unsigned int get_expr_value_id (pre_expr);
473 /* We can add and remove elements and entries to and from sets
474 and hash tables, so we use alloc pools for them. */
476 static alloc_pool bitmap_set_pool;
477 static bitmap_obstack grand_bitmap_obstack;
479 /* To avoid adding 300 temporary variables when we only need one, we
480 only create one temporary variable, on demand, and build ssa names
481 off that. We do have to change the variable if the types don't
482 match the current variable's type. */
484 static tree storetemp;
485 static tree prephitemp;
487 /* Set of blocks with statements that have had its EH information
489 static bitmap need_eh_cleanup;
491 /* The phi_translate_table caches phi translations for a given
492 expression and predecessor. */
494 static htab_t phi_translate_table;
496 /* A three tuple {e, pred, v} used to cache phi translations in the
497 phi_translate_table. */
499 typedef struct expr_pred_trans_d
501 /* The expression. */
504 /* The predecessor block along which we translated the expression. */
507 /* The value that resulted from the translation. */
510 /* The hashcode for the expression, pred pair. This is cached for
513 } *expr_pred_trans_t;
514 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
516 /* Return the hash value for a phi translation table entry. */
519 expr_pred_trans_hash (const void *p)
521 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
525 /* Return true if two phi translation table entries are the same.
526 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
529 expr_pred_trans_eq (const void *p1, const void *p2)
531 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
532 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
533 basic_block b1 = ve1->pred;
534 basic_block b2 = ve2->pred;
536 /* If they are not translations for the same basic block, they can't
540 return pre_expr_eq (ve1->e, ve2->e);
543 /* Search in the phi translation table for the translation of
544 expression E in basic block PRED.
545 Return the translated value, if found, NULL otherwise. */
547 static inline pre_expr
548 phi_trans_lookup (pre_expr e, basic_block pred)
551 struct expr_pred_trans_d ept;
555 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
556 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
561 return ((expr_pred_trans_t) *slot)->v;
565 /* Add the tuple mapping from {expression E, basic block PRED} to
566 value V, to the phi translation table. */
569 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
572 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
574 new_pair->pred = pred;
576 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
579 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
580 new_pair->hashcode, INSERT);
583 *slot = (void *) new_pair;
587 /* Add expression E to the expression set of value id V. */
590 add_to_value (unsigned int v, pre_expr e)
594 gcc_assert (get_expr_value_id (e) == v);
596 if (v >= VEC_length (bitmap_set_t, value_expressions))
598 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
602 set = VEC_index (bitmap_set_t, value_expressions, v);
605 set = bitmap_set_new ();
606 VEC_replace (bitmap_set_t, value_expressions, v, set);
609 bitmap_insert_into_set_1 (set, e, v, true);
612 /* Create a new bitmap set and return it. */
615 bitmap_set_new (void)
617 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
618 ret->expressions = BITMAP_ALLOC (&grand_bitmap_obstack);
619 ret->values = BITMAP_ALLOC (&grand_bitmap_obstack);
623 /* Return the value id for a PRE expression EXPR. */
626 get_expr_value_id (pre_expr expr)
633 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
636 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
637 add_to_value (id, expr);
642 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
644 return PRE_EXPR_NARY (expr)->value_id;
646 return PRE_EXPR_REFERENCE (expr)->value_id;
652 /* Remove an expression EXPR from a bitmapped set. */
655 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
657 unsigned int val = get_expr_value_id (expr);
658 if (!value_id_constant_p (val))
660 bitmap_clear_bit (set->values, val);
661 bitmap_clear_bit (set->expressions, get_expression_id (expr));
666 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
667 unsigned int val, bool allow_constants)
669 if (allow_constants || !value_id_constant_p (val))
671 /* We specifically expect this and only this function to be able to
672 insert constants into a set. */
673 bitmap_set_bit (set->values, val);
674 bitmap_set_bit (set->expressions, get_or_alloc_expression_id (expr));
678 /* Insert an expression EXPR into a bitmapped set. */
681 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
683 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
686 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
689 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
691 bitmap_copy (dest->expressions, orig->expressions);
692 bitmap_copy (dest->values, orig->values);
696 /* Free memory used up by SET. */
698 bitmap_set_free (bitmap_set_t set)
700 BITMAP_FREE (set->expressions);
701 BITMAP_FREE (set->values);
705 /* Generate an topological-ordered array of bitmap set SET. */
707 static VEC(pre_expr, heap) *
708 sorted_array_from_bitmap_set (bitmap_set_t set)
711 bitmap_iterator bi, bj;
712 VEC(pre_expr, heap) *result;
714 /* Pre-allocate roughly enough space for the array. */
715 result = VEC_alloc (pre_expr, heap, bitmap_count_bits (set->values));
717 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
719 /* The number of expressions having a given value is usually
720 relatively small. Thus, rather than making a vector of all
721 the expressions and sorting it by value-id, we walk the values
722 and check in the reverse mapping that tells us what expressions
723 have a given value, to filter those in our set. As a result,
724 the expressions are inserted in value-id order, which means
727 If this is somehow a significant lose for some cases, we can
728 choose which set to walk based on the set size. */
729 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
730 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
732 if (bitmap_bit_p (set->expressions, j))
733 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
740 /* Perform bitmapped set operation DEST &= ORIG. */
743 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
750 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
752 bitmap_and_into (dest->values, orig->values);
753 bitmap_copy (temp, dest->expressions);
754 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
756 pre_expr expr = expression_for_id (i);
757 unsigned int value_id = get_expr_value_id (expr);
758 if (!bitmap_bit_p (dest->values, value_id))
759 bitmap_clear_bit (dest->expressions, i);
765 /* Subtract all values and expressions contained in ORIG from DEST. */
768 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
770 bitmap_set_t result = bitmap_set_new ();
774 bitmap_and_compl (result->expressions, dest->expressions,
777 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
779 pre_expr expr = expression_for_id (i);
780 unsigned int value_id = get_expr_value_id (expr);
781 bitmap_set_bit (result->values, value_id);
787 /* Subtract all the values in bitmap set B from bitmap set A. */
790 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
794 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
796 bitmap_copy (temp, a->expressions);
797 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
799 pre_expr expr = expression_for_id (i);
800 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
801 bitmap_remove_from_set (a, expr);
807 /* Return true if bitmapped set SET contains the value VALUE_ID. */
810 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
812 if (value_id_constant_p (value_id))
815 if (!set || bitmap_empty_p (set->expressions))
818 return bitmap_bit_p (set->values, value_id);
822 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
824 return bitmap_bit_p (set->expressions, get_expression_id (expr));
827 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
830 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
833 bitmap_set_t exprset;
837 if (value_id_constant_p (lookfor))
840 if (!bitmap_set_contains_value (set, lookfor))
843 /* The number of expressions having a given value is usually
844 significantly less than the total number of expressions in SET.
845 Thus, rather than check, for each expression in SET, whether it
846 has the value LOOKFOR, we walk the reverse mapping that tells us
847 what expressions have a given value, and see if any of those
848 expressions are in our set. For large testcases, this is about
849 5-10x faster than walking the bitmap. If this is somehow a
850 significant lose for some cases, we can choose which set to walk
851 based on the set size. */
852 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
853 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
855 if (bitmap_bit_p (set->expressions, i))
857 bitmap_clear_bit (set->expressions, i);
858 bitmap_set_bit (set->expressions, get_expression_id (expr));
864 /* Return true if two bitmap sets are equal. */
867 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
869 return bitmap_equal_p (a->values, b->values);
872 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
873 and add it otherwise. */
876 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
878 unsigned int val = get_expr_value_id (expr);
880 if (bitmap_set_contains_value (set, val))
881 bitmap_set_replace_value (set, val, expr);
883 bitmap_insert_into_set (set, expr);
886 /* Insert EXPR into SET if EXPR's value is not already present in
890 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
892 unsigned int val = get_expr_value_id (expr);
894 #ifdef ENABLE_CHECKING
895 gcc_assert (expr->id == get_or_alloc_expression_id (expr));
898 /* Constant values are always considered to be part of the set. */
899 if (value_id_constant_p (val))
902 /* If the value membership changed, add the expression. */
903 if (bitmap_set_bit (set->values, val))
904 bitmap_set_bit (set->expressions, expr->id);
907 /* Print out EXPR to outfile. */
910 print_pre_expr (FILE *outfile, const pre_expr expr)
915 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
918 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
923 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
924 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
925 for (i = 0; i < nary->length; i++)
927 print_generic_expr (outfile, nary->op[i], 0);
928 if (i != (unsigned) nary->length - 1)
929 fprintf (outfile, ",");
931 fprintf (outfile, "}");
937 vn_reference_op_t vro;
939 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
940 fprintf (outfile, "{");
942 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
945 bool closebrace = false;
946 if (vro->opcode != SSA_NAME
947 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
949 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
952 fprintf (outfile, "<");
958 print_generic_expr (outfile, vro->op0, 0);
961 fprintf (outfile, ",");
962 print_generic_expr (outfile, vro->op1, 0);
966 fprintf (outfile, ",");
967 print_generic_expr (outfile, vro->op2, 0);
971 fprintf (outfile, ">");
972 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
973 fprintf (outfile, ",");
975 fprintf (outfile, "}");
978 fprintf (outfile, "@");
979 print_generic_expr (outfile, ref->vuse, 0);
985 void debug_pre_expr (pre_expr);
987 /* Like print_pre_expr but always prints to stderr. */
989 debug_pre_expr (pre_expr e)
991 print_pre_expr (stderr, e);
992 fprintf (stderr, "\n");
995 /* Print out SET to OUTFILE. */
998 print_bitmap_set (FILE *outfile, bitmap_set_t set,
999 const char *setname, int blockindex)
1001 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1008 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1010 const pre_expr expr = expression_for_id (i);
1013 fprintf (outfile, ", ");
1015 print_pre_expr (outfile, expr);
1017 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1020 fprintf (outfile, " }\n");
1023 void debug_bitmap_set (bitmap_set_t);
1026 debug_bitmap_set (bitmap_set_t set)
1028 print_bitmap_set (stderr, set, "debug", 0);
1031 /* Print out the expressions that have VAL to OUTFILE. */
1034 print_value_expressions (FILE *outfile, unsigned int val)
1036 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1040 sprintf (s, "%04d", val);
1041 print_bitmap_set (outfile, set, s, 0);
1047 debug_value_expressions (unsigned int val)
1049 print_value_expressions (stderr, val);
1052 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1056 get_or_alloc_expr_for_constant (tree constant)
1058 unsigned int result_id;
1059 unsigned int value_id;
1060 struct pre_expr_d expr;
1063 expr.kind = CONSTANT;
1064 PRE_EXPR_CONSTANT (&expr) = constant;
1065 result_id = lookup_expression_id (&expr);
1067 return expression_for_id (result_id);
1069 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1070 newexpr->kind = CONSTANT;
1071 PRE_EXPR_CONSTANT (newexpr) = constant;
1072 alloc_expression_id (newexpr);
1073 value_id = get_or_alloc_constant_value_id (constant);
1074 add_to_value (value_id, newexpr);
1078 /* Given a value id V, find the actual tree representing the constant
1079 value if there is one, and return it. Return NULL if we can't find
1083 get_constant_for_value_id (unsigned int v)
1085 if (value_id_constant_p (v))
1089 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1091 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1093 pre_expr expr = expression_for_id (i);
1094 if (expr->kind == CONSTANT)
1095 return PRE_EXPR_CONSTANT (expr);
1101 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1102 Currently only supports constants and SSA_NAMES. */
1104 get_or_alloc_expr_for (tree t)
1106 if (TREE_CODE (t) == SSA_NAME)
1107 return get_or_alloc_expr_for_name (t);
1108 else if (is_gimple_min_invariant (t))
1109 return get_or_alloc_expr_for_constant (t);
1112 /* More complex expressions can result from SCCVN expression
1113 simplification that inserts values for them. As they all
1114 do not have VOPs the get handled by the nary ops struct. */
1115 vn_nary_op_t result;
1116 unsigned int result_id;
1117 vn_nary_op_lookup (t, &result);
1120 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1122 PRE_EXPR_NARY (e) = result;
1123 result_id = lookup_expression_id (e);
1126 pool_free (pre_expr_pool, e);
1127 e = expression_for_id (result_id);
1130 alloc_expression_id (e);
1137 /* Return the folded version of T if T, when folded, is a gimple
1138 min_invariant. Otherwise, return T. */
1141 fully_constant_expression (pre_expr e)
1149 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1150 switch (TREE_CODE_CLASS (nary->opcode))
1152 case tcc_expression:
1153 if (nary->opcode == TRUTH_NOT_EXPR)
1155 if (nary->opcode != TRUTH_AND_EXPR
1156 && nary->opcode != TRUTH_OR_EXPR
1157 && nary->opcode != TRUTH_XOR_EXPR)
1161 case tcc_comparison:
1163 /* We have to go from trees to pre exprs to value ids to
1165 tree naryop0 = nary->op[0];
1166 tree naryop1 = nary->op[1];
1168 if (!is_gimple_min_invariant (naryop0))
1170 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1171 unsigned int vrep0 = get_expr_value_id (rep0);
1172 tree const0 = get_constant_for_value_id (vrep0);
1174 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1176 if (!is_gimple_min_invariant (naryop1))
1178 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1179 unsigned int vrep1 = get_expr_value_id (rep1);
1180 tree const1 = get_constant_for_value_id (vrep1);
1182 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1184 result = fold_binary (nary->opcode, nary->type,
1186 if (result && is_gimple_min_invariant (result))
1187 return get_or_alloc_expr_for_constant (result);
1188 /* We might have simplified the expression to a
1189 SSA_NAME for example from x_1 * 1. But we cannot
1190 insert a PHI for x_1 unconditionally as x_1 might
1191 not be available readily. */
1195 if (nary->opcode != REALPART_EXPR
1196 && nary->opcode != IMAGPART_EXPR
1197 && nary->opcode != VIEW_CONVERT_EXPR)
1203 /* We have to go from trees to pre exprs to value ids to
1205 tree naryop0 = nary->op[0];
1206 tree const0, result;
1207 if (is_gimple_min_invariant (naryop0))
1211 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1212 unsigned int vrep0 = get_expr_value_id (rep0);
1213 const0 = get_constant_for_value_id (vrep0);
1218 tree type1 = TREE_TYPE (nary->op[0]);
1219 const0 = fold_convert (type1, const0);
1220 result = fold_unary (nary->opcode, nary->type, const0);
1222 if (result && is_gimple_min_invariant (result))
1223 return get_or_alloc_expr_for_constant (result);
1232 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1234 if ((folded = fully_constant_vn_reference_p (ref)))
1235 return get_or_alloc_expr_for_constant (folded);
1244 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1245 it has the value it would have in BLOCK. Set *SAME_VALID to true
1246 in case the new vuse doesn't change the value id of the OPERANDS. */
1249 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1250 alias_set_type set, tree type, tree vuse,
1251 basic_block phiblock,
1252 basic_block block, bool *same_valid)
1254 gimple phi = SSA_NAME_DEF_STMT (vuse);
1261 if (gimple_bb (phi) != phiblock)
1264 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1266 /* Use the alias-oracle to find either the PHI node in this block,
1267 the first VUSE used in this block that is equivalent to vuse or
1268 the first VUSE which definition in this block kills the value. */
1269 if (gimple_code (phi) == GIMPLE_PHI)
1270 e = find_edge (block, phiblock);
1271 else if (use_oracle)
1272 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1274 vuse = gimple_vuse (phi);
1275 phi = SSA_NAME_DEF_STMT (vuse);
1276 if (gimple_bb (phi) != phiblock)
1278 if (gimple_code (phi) == GIMPLE_PHI)
1280 e = find_edge (block, phiblock);
1291 bitmap visited = NULL;
1292 /* Try to find a vuse that dominates this phi node by skipping
1293 non-clobbering statements. */
1294 vuse = get_continuation_for_phi (phi, &ref, &visited);
1296 BITMAP_FREE (visited);
1302 /* If we didn't find any, the value ID can't stay the same,
1303 but return the translated vuse. */
1304 *same_valid = false;
1305 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1307 /* ??? We would like to return vuse here as this is the canonical
1308 upmost vdef that this reference is associated with. But during
1309 insertion of the references into the hash tables we only ever
1310 directly insert with their direct gimple_vuse, hence returning
1311 something else would make us not find the other expression. */
1312 return PHI_ARG_DEF (phi, e->dest_idx);
1318 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1319 SET2. This is used to avoid making a set consisting of the union
1320 of PA_IN and ANTIC_IN during insert. */
1322 static inline pre_expr
1323 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1327 result = bitmap_find_leader (set1, val, NULL);
1328 if (!result && set2)
1329 result = bitmap_find_leader (set2, val, NULL);
1333 /* Get the tree type for our PRE expression e. */
1336 get_expr_type (const pre_expr e)
1341 return TREE_TYPE (PRE_EXPR_NAME (e));
1343 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1345 return PRE_EXPR_REFERENCE (e)->type;
1347 return PRE_EXPR_NARY (e)->type;
1352 /* Get a representative SSA_NAME for a given expression.
1353 Since all of our sub-expressions are treated as values, we require
1354 them to be SSA_NAME's for simplicity.
1355 Prior versions of GVNPRE used to use "value handles" here, so that
1356 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1357 either case, the operands are really values (IE we do not expect
1358 them to be usable without finding leaders). */
1361 get_representative_for (const pre_expr e)
1365 unsigned int value_id = get_expr_value_id (e);
1370 return PRE_EXPR_NAME (e);
1372 return PRE_EXPR_CONSTANT (e);
1376 /* Go through all of the expressions representing this value
1377 and pick out an SSA_NAME. */
1380 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1382 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1384 pre_expr rep = expression_for_id (i);
1385 if (rep->kind == NAME)
1386 return PRE_EXPR_NAME (rep);
1391 /* If we reached here we couldn't find an SSA_NAME. This can
1392 happen when we've discovered a value that has never appeared in
1393 the program as set to an SSA_NAME, most likely as the result of
1398 "Could not find SSA_NAME representative for expression:");
1399 print_pre_expr (dump_file, e);
1400 fprintf (dump_file, "\n");
1403 exprtype = get_expr_type (e);
1405 /* Build and insert the assignment of the end result to the temporary
1406 that we will return. */
1407 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1409 pretemp = create_tmp_reg (exprtype, "pretmp");
1410 get_var_ann (pretemp);
1413 name = make_ssa_name (pretemp, gimple_build_nop ());
1414 VN_INFO_GET (name)->value_id = value_id;
1415 if (e->kind == CONSTANT)
1416 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1418 VN_INFO (name)->valnum = name;
1420 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1423 fprintf (dump_file, "Created SSA_NAME representative ");
1424 print_generic_expr (dump_file, name, 0);
1425 fprintf (dump_file, " for expression:");
1426 print_pre_expr (dump_file, e);
1427 fprintf (dump_file, "\n");
1436 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1437 basic_block pred, basic_block phiblock);
1439 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1440 the phis in PRED. Return NULL if we can't find a leader for each part
1441 of the translated expression. */
1444 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1445 basic_block pred, basic_block phiblock)
1452 bool changed = false;
1453 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1454 struct vn_nary_op_s newnary;
1455 /* The NARY structure is only guaranteed to have been
1456 allocated to the nary->length operands. */
1457 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1458 - sizeof (tree) * (4 - nary->length)));
1460 for (i = 0; i < newnary.length; i++)
1462 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1466 pre_expr leader, result;
1467 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1468 leader = find_leader_in_sets (op_val_id, set1, set2);
1469 result = phi_translate (leader, set1, set2, pred, phiblock);
1470 if (result && result != leader)
1472 tree name = get_representative_for (result);
1475 newnary.op[i] = name;
1480 changed |= newnary.op[i] != nary->op[i];
1486 unsigned int new_val_id;
1488 tree result = vn_nary_op_lookup_pieces (newnary.length,
1496 if (result && is_gimple_min_invariant (result))
1497 return get_or_alloc_expr_for_constant (result);
1499 expr = (pre_expr) pool_alloc (pre_expr_pool);
1504 PRE_EXPR_NARY (expr) = nary;
1505 constant = fully_constant_expression (expr);
1506 if (constant != expr)
1509 new_val_id = nary->value_id;
1510 get_or_alloc_expression_id (expr);
1514 new_val_id = get_next_value_id ();
1515 VEC_safe_grow_cleared (bitmap_set_t, heap,
1517 get_max_value_id() + 1);
1518 nary = vn_nary_op_insert_pieces (newnary.length,
1525 result, new_val_id);
1526 PRE_EXPR_NARY (expr) = nary;
1527 constant = fully_constant_expression (expr);
1528 if (constant != expr)
1530 get_or_alloc_expression_id (expr);
1532 add_to_value (new_val_id, expr);
1540 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1541 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1542 tree vuse = ref->vuse;
1543 tree newvuse = vuse;
1544 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1545 bool changed = false, same_valid = true;
1547 vn_reference_op_t operand;
1548 vn_reference_t newref;
1551 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1555 tree oldop0 = operand->op0;
1556 tree oldop1 = operand->op1;
1557 tree oldop2 = operand->op2;
1561 tree type = operand->type;
1562 vn_reference_op_s newop = *operand;
1564 if (op0 && TREE_CODE (op0) == SSA_NAME)
1566 unsigned int op_val_id = VN_INFO (op0)->value_id;
1567 leader = find_leader_in_sets (op_val_id, set1, set2);
1568 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1569 if (opresult && opresult != leader)
1571 tree name = get_representative_for (opresult);
1579 changed |= op0 != oldop0;
1581 if (op1 && TREE_CODE (op1) == SSA_NAME)
1583 unsigned int op_val_id = VN_INFO (op1)->value_id;
1584 leader = find_leader_in_sets (op_val_id, set1, set2);
1585 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1586 if (opresult && opresult != leader)
1588 tree name = get_representative_for (opresult);
1596 /* We can't possibly insert these. */
1597 else if (op1 && !is_gimple_min_invariant (op1))
1599 changed |= op1 != oldop1;
1600 if (op2 && TREE_CODE (op2) == SSA_NAME)
1602 unsigned int op_val_id = VN_INFO (op2)->value_id;
1603 leader = find_leader_in_sets (op_val_id, set1, set2);
1604 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1605 if (opresult && opresult != leader)
1607 tree name = get_representative_for (opresult);
1615 /* We can't possibly insert these. */
1616 else if (op2 && !is_gimple_min_invariant (op2))
1618 changed |= op2 != oldop2;
1621 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1622 /* We may have changed from an SSA_NAME to a constant */
1623 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1624 newop.opcode = TREE_CODE (op0);
1629 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1630 /* If it transforms from an SSA_NAME to an address, fold with
1631 a preceding indirect reference. */
1632 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1633 && VEC_index (vn_reference_op_s,
1634 newoperands, j - 1)->opcode == INDIRECT_REF)
1635 vn_reference_fold_indirect (&newoperands, &j);
1637 if (i != VEC_length (vn_reference_op_s, operands))
1640 VEC_free (vn_reference_op_s, heap, newoperands);
1646 newvuse = translate_vuse_through_block (newoperands,
1647 ref->set, ref->type,
1648 vuse, phiblock, pred,
1650 if (newvuse == NULL_TREE)
1652 VEC_free (vn_reference_op_s, heap, newoperands);
1657 if (changed || newvuse != vuse)
1659 unsigned int new_val_id;
1662 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1667 VEC_free (vn_reference_op_s, heap, newoperands);
1669 if (result && is_gimple_min_invariant (result))
1671 gcc_assert (!newoperands);
1672 return get_or_alloc_expr_for_constant (result);
1675 expr = (pre_expr) pool_alloc (pre_expr_pool);
1676 expr->kind = REFERENCE;
1681 PRE_EXPR_REFERENCE (expr) = newref;
1682 constant = fully_constant_expression (expr);
1683 if (constant != expr)
1686 new_val_id = newref->value_id;
1687 get_or_alloc_expression_id (expr);
1691 if (changed || !same_valid)
1693 new_val_id = get_next_value_id ();
1694 VEC_safe_grow_cleared (bitmap_set_t, heap,
1696 get_max_value_id() + 1);
1699 new_val_id = ref->value_id;
1700 newref = vn_reference_insert_pieces (newvuse, ref->set,
1703 result, new_val_id);
1705 PRE_EXPR_REFERENCE (expr) = newref;
1706 constant = fully_constant_expression (expr);
1707 if (constant != expr)
1709 get_or_alloc_expression_id (expr);
1711 add_to_value (new_val_id, expr);
1713 VEC_free (vn_reference_op_s, heap, newoperands);
1723 tree name = PRE_EXPR_NAME (expr);
1725 def_stmt = SSA_NAME_DEF_STMT (name);
1726 if (gimple_code (def_stmt) == GIMPLE_PHI
1727 && gimple_bb (def_stmt) == phiblock)
1732 e = find_edge (pred, gimple_bb (phi));
1735 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1738 if (TREE_CODE (def) == SSA_NAME)
1739 def = VN_INFO (def)->valnum;
1741 /* Handle constant. */
1742 if (is_gimple_min_invariant (def))
1743 return get_or_alloc_expr_for_constant (def);
1745 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1748 newexpr = get_or_alloc_expr_for_name (def);
1759 /* Wrapper around phi_translate_1 providing caching functionality. */
1762 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1763 basic_block pred, basic_block phiblock)
1770 /* Constants contain no values that need translation. */
1771 if (expr->kind == CONSTANT)
1774 if (value_id_constant_p (get_expr_value_id (expr)))
1777 if (expr->kind != NAME)
1779 phitrans = phi_trans_lookup (expr, pred);
1785 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1787 /* Don't add empty translations to the cache. Neither add
1788 translations of NAMEs as those are cheap to translate. */
1790 && expr->kind != NAME)
1791 phi_trans_add (expr, phitrans, pred);
1797 /* For each expression in SET, translate the values through phi nodes
1798 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1799 expressions in DEST. */
1802 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1803 basic_block phiblock)
1805 VEC (pre_expr, heap) *exprs;
1809 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1811 bitmap_set_copy (dest, set);
1815 exprs = sorted_array_from_bitmap_set (set);
1816 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1818 pre_expr translated;
1819 translated = phi_translate (expr, set, NULL, pred, phiblock);
1823 /* We might end up with multiple expressions from SET being
1824 translated to the same value. In this case we do not want
1825 to retain the NARY or REFERENCE expression but prefer a NAME
1826 which would be the leader. */
1827 if (translated->kind == NAME)
1828 bitmap_value_replace_in_set (dest, translated);
1830 bitmap_value_insert_into_set (dest, translated);
1832 VEC_free (pre_expr, heap, exprs);
1835 /* Find the leader for a value (i.e., the name representing that
1836 value) in a given set, and return it. If STMT is non-NULL it
1837 makes sure the defining statement for the leader dominates it.
1838 Return NULL if no leader is found. */
1841 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1843 if (value_id_constant_p (val))
1847 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1849 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1851 pre_expr expr = expression_for_id (i);
1852 if (expr->kind == CONSTANT)
1856 if (bitmap_set_contains_value (set, val))
1858 /* Rather than walk the entire bitmap of expressions, and see
1859 whether any of them has the value we are looking for, we look
1860 at the reverse mapping, which tells us the set of expressions
1861 that have a given value (IE value->expressions with that
1862 value) and see if any of those expressions are in our set.
1863 The number of expressions per value is usually significantly
1864 less than the number of expressions in the set. In fact, for
1865 large testcases, doing it this way is roughly 5-10x faster
1866 than walking the bitmap.
1867 If this is somehow a significant lose for some cases, we can
1868 choose which set to walk based on which set is smaller. */
1871 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1873 EXECUTE_IF_AND_IN_BITMAP (exprset->expressions,
1874 set->expressions, 0, i, bi)
1876 pre_expr val = expression_for_id (i);
1877 /* At the point where stmt is not null, there should always
1878 be an SSA_NAME first in the list of expressions. */
1881 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1882 if (gimple_code (def_stmt) != GIMPLE_PHI
1883 && gimple_bb (def_stmt) == gimple_bb (stmt)
1884 && gimple_uid (def_stmt) >= gimple_uid (stmt))
1893 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1894 BLOCK by seeing if it is not killed in the block. Note that we are
1895 only determining whether there is a store that kills it. Because
1896 of the order in which clean iterates over values, we are guaranteed
1897 that altered operands will have caused us to be eliminated from the
1898 ANTIC_IN set already. */
1901 value_dies_in_block_x (pre_expr expr, basic_block block)
1903 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1904 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1906 gimple_stmt_iterator gsi;
1907 unsigned id = get_expression_id (expr);
1914 /* Lookup a previously calculated result. */
1915 if (EXPR_DIES (block)
1916 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1917 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1919 /* A memory expression {e, VUSE} dies in the block if there is a
1920 statement that may clobber e. If, starting statement walk from the
1921 top of the basic block, a statement uses VUSE there can be no kill
1922 inbetween that use and the original statement that loaded {e, VUSE},
1923 so we can stop walking. */
1924 ref.base = NULL_TREE;
1925 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1927 tree def_vuse, def_vdef;
1928 def = gsi_stmt (gsi);
1929 def_vuse = gimple_vuse (def);
1930 def_vdef = gimple_vdef (def);
1932 /* Not a memory statement. */
1936 /* Not a may-def. */
1939 /* A load with the same VUSE, we're done. */
1940 if (def_vuse == vuse)
1946 /* Init ref only if we really need it. */
1947 if (ref.base == NULL_TREE
1948 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1954 /* If the statement may clobber expr, it dies. */
1955 if (stmt_may_clobber_ref_p_1 (def, &ref))
1962 /* Remember the result. */
1963 if (!EXPR_DIES (block))
1964 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1965 bitmap_set_bit (EXPR_DIES (block), id * 2);
1967 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1973 #define union_contains_value(SET1, SET2, VAL) \
1974 (bitmap_set_contains_value ((SET1), (VAL)) \
1975 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
1977 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
1980 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
1981 vn_reference_op_t vro)
1983 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
1985 struct pre_expr_d temp;
1988 PRE_EXPR_NAME (&temp) = vro->op0;
1989 temp.id = lookup_expression_id (&temp);
1992 if (!union_contains_value (set1, set2,
1993 get_expr_value_id (&temp)))
1996 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
1998 struct pre_expr_d temp;
2001 PRE_EXPR_NAME (&temp) = vro->op1;
2002 temp.id = lookup_expression_id (&temp);
2005 if (!union_contains_value (set1, set2,
2006 get_expr_value_id (&temp)))
2010 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2012 struct pre_expr_d temp;
2015 PRE_EXPR_NAME (&temp) = vro->op2;
2016 temp.id = lookup_expression_id (&temp);
2019 if (!union_contains_value (set1, set2,
2020 get_expr_value_id (&temp)))
2027 /* Determine if the expression EXPR is valid in SET1 U SET2.
2028 ONLY SET2 CAN BE NULL.
2029 This means that we have a leader for each part of the expression
2030 (if it consists of values), or the expression is an SSA_NAME.
2031 For loads/calls, we also see if the vuse is killed in this block. */
2034 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2040 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2044 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2045 for (i = 0; i < nary->length; i++)
2047 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2049 struct pre_expr_d temp;
2052 PRE_EXPR_NAME (&temp) = nary->op[i];
2053 temp.id = lookup_expression_id (&temp);
2056 if (!union_contains_value (set1, set2,
2057 get_expr_value_id (&temp)))
2061 /* If the NARY may trap make sure the block does not contain
2062 a possible exit point.
2063 ??? This is overly conservative if we translate AVAIL_OUT
2064 as the available expression might be after the exit point. */
2065 if (BB_MAY_NOTRETURN (block)
2066 && vn_nary_may_trap (nary))
2073 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2074 vn_reference_op_t vro;
2077 for (i = 0; VEC_iterate (vn_reference_op_s, ref->operands, i, vro); i++)
2079 if (!vro_valid_in_sets (set1, set2, vro))
2084 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2085 if (!gimple_nop_p (def_stmt)
2086 && gimple_bb (def_stmt) != block
2087 && !dominated_by_p (CDI_DOMINATORS,
2088 block, gimple_bb (def_stmt)))
2091 return !value_dies_in_block_x (expr, block);
2098 /* Clean the set of expressions that are no longer valid in SET1 or
2099 SET2. This means expressions that are made up of values we have no
2100 leaders for in SET1 or SET2. This version is used for partial
2101 anticipation, which means it is not valid in either ANTIC_IN or
2105 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2107 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2111 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2113 if (!valid_in_sets (set1, set2, expr, block))
2114 bitmap_remove_from_set (set1, expr);
2116 VEC_free (pre_expr, heap, exprs);
2119 /* Clean the set of expressions that are no longer valid in SET. This
2120 means expressions that are made up of values we have no leaders for
2124 clean (bitmap_set_t set, basic_block block)
2126 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2130 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2132 if (!valid_in_sets (set, NULL, expr, block))
2133 bitmap_remove_from_set (set, expr);
2135 VEC_free (pre_expr, heap, exprs);
2138 static sbitmap has_abnormal_preds;
2140 /* List of blocks that may have changed during ANTIC computation and
2141 thus need to be iterated over. */
2143 static sbitmap changed_blocks;
2145 /* Decide whether to defer a block for a later iteration, or PHI
2146 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2147 should defer the block, and true if we processed it. */
2150 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2151 basic_block block, basic_block phiblock)
2153 if (!BB_VISITED (phiblock))
2155 SET_BIT (changed_blocks, block->index);
2156 BB_VISITED (block) = 0;
2157 BB_DEFERRED (block) = 1;
2161 phi_translate_set (dest, source, block, phiblock);
2165 /* Compute the ANTIC set for BLOCK.
2167 If succs(BLOCK) > 1 then
2168 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2169 else if succs(BLOCK) == 1 then
2170 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2172 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2176 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2178 bool changed = false;
2179 bitmap_set_t S, old, ANTIC_OUT;
2185 old = ANTIC_OUT = S = NULL;
2186 BB_VISITED (block) = 1;
2188 /* If any edges from predecessors are abnormal, antic_in is empty,
2190 if (block_has_abnormal_pred_edge)
2191 goto maybe_dump_sets;
2193 old = ANTIC_IN (block);
2194 ANTIC_OUT = bitmap_set_new ();
2196 /* If the block has no successors, ANTIC_OUT is empty. */
2197 if (EDGE_COUNT (block->succs) == 0)
2199 /* If we have one successor, we could have some phi nodes to
2200 translate through. */
2201 else if (single_succ_p (block))
2203 basic_block succ_bb = single_succ (block);
2205 /* We trade iterations of the dataflow equations for having to
2206 phi translate the maximal set, which is incredibly slow
2207 (since the maximal set often has 300+ members, even when you
2208 have a small number of blocks).
2209 Basically, we defer the computation of ANTIC for this block
2210 until we have processed it's successor, which will inevitably
2211 have a *much* smaller set of values to phi translate once
2212 clean has been run on it.
2213 The cost of doing this is that we technically perform more
2214 iterations, however, they are lower cost iterations.
2216 Timings for PRE on tramp3d-v4:
2217 without maximal set fix: 11 seconds
2218 with maximal set fix/without deferring: 26 seconds
2219 with maximal set fix/with deferring: 11 seconds
2222 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2226 goto maybe_dump_sets;
2229 /* If we have multiple successors, we take the intersection of all of
2230 them. Note that in the case of loop exit phi nodes, we may have
2231 phis to translate through. */
2234 VEC(basic_block, heap) * worklist;
2236 basic_block bprime, first = NULL;
2238 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2239 FOR_EACH_EDGE (e, ei, block->succs)
2242 && BB_VISITED (e->dest))
2244 else if (BB_VISITED (e->dest))
2245 VEC_quick_push (basic_block, worklist, e->dest);
2248 /* Of multiple successors we have to have visited one already. */
2251 SET_BIT (changed_blocks, block->index);
2252 BB_VISITED (block) = 0;
2253 BB_DEFERRED (block) = 1;
2255 VEC_free (basic_block, heap, worklist);
2256 goto maybe_dump_sets;
2259 if (!gimple_seq_empty_p (phi_nodes (first)))
2260 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2262 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2264 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2266 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2268 bitmap_set_t tmp = bitmap_set_new ();
2269 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2270 bitmap_set_and (ANTIC_OUT, tmp);
2271 bitmap_set_free (tmp);
2274 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2276 VEC_free (basic_block, heap, worklist);
2279 /* Generate ANTIC_OUT - TMP_GEN. */
2280 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2282 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2283 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2286 /* Then union in the ANTIC_OUT - TMP_GEN values,
2287 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2288 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2289 bitmap_value_insert_into_set (ANTIC_IN (block),
2290 expression_for_id (bii));
2292 clean (ANTIC_IN (block), block);
2294 /* !old->expressions can happen when we deferred a block. */
2295 if (!old->expressions || !bitmap_set_equal (old, ANTIC_IN (block)))
2298 SET_BIT (changed_blocks, block->index);
2299 FOR_EACH_EDGE (e, ei, block->preds)
2300 SET_BIT (changed_blocks, e->src->index);
2303 RESET_BIT (changed_blocks, block->index);
2306 if (dump_file && (dump_flags & TDF_DETAILS))
2308 if (!BB_DEFERRED (block) || BB_VISITED (block))
2311 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2313 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2317 print_bitmap_set (dump_file, S, "S", block->index);
2322 "Block %d was deferred for a future iteration.\n",
2327 bitmap_set_free (old);
2329 bitmap_set_free (S);
2331 bitmap_set_free (ANTIC_OUT);
2335 /* Compute PARTIAL_ANTIC for BLOCK.
2337 If succs(BLOCK) > 1 then
2338 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2339 in ANTIC_OUT for all succ(BLOCK)
2340 else if succs(BLOCK) == 1 then
2341 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2343 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2348 compute_partial_antic_aux (basic_block block,
2349 bool block_has_abnormal_pred_edge)
2351 bool changed = false;
2352 bitmap_set_t old_PA_IN;
2353 bitmap_set_t PA_OUT;
2356 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2358 old_PA_IN = PA_OUT = NULL;
2360 /* If any edges from predecessors are abnormal, antic_in is empty,
2362 if (block_has_abnormal_pred_edge)
2363 goto maybe_dump_sets;
2365 /* If there are too many partially anticipatable values in the
2366 block, phi_translate_set can take an exponential time: stop
2367 before the translation starts. */
2369 && single_succ_p (block)
2370 && bitmap_count_bits (PA_IN (single_succ (block))->values) > max_pa)
2371 goto maybe_dump_sets;
2373 old_PA_IN = PA_IN (block);
2374 PA_OUT = bitmap_set_new ();
2376 /* If the block has no successors, ANTIC_OUT is empty. */
2377 if (EDGE_COUNT (block->succs) == 0)
2379 /* If we have one successor, we could have some phi nodes to
2380 translate through. Note that we can't phi translate across DFS
2381 back edges in partial antic, because it uses a union operation on
2382 the successors. For recurrences like IV's, we will end up
2383 generating a new value in the set on each go around (i + 3 (VH.1)
2384 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2385 else if (single_succ_p (block))
2387 basic_block succ = single_succ (block);
2388 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2389 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2391 /* If we have multiple successors, we take the union of all of
2395 VEC(basic_block, heap) * worklist;
2399 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2400 FOR_EACH_EDGE (e, ei, block->succs)
2402 if (e->flags & EDGE_DFS_BACK)
2404 VEC_quick_push (basic_block, worklist, e->dest);
2406 if (VEC_length (basic_block, worklist) > 0)
2408 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2413 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2414 bitmap_value_insert_into_set (PA_OUT,
2415 expression_for_id (i));
2416 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2418 bitmap_set_t pa_in = bitmap_set_new ();
2419 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2420 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2421 bitmap_value_insert_into_set (PA_OUT,
2422 expression_for_id (i));
2423 bitmap_set_free (pa_in);
2426 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2427 bitmap_value_insert_into_set (PA_OUT,
2428 expression_for_id (i));
2431 VEC_free (basic_block, heap, worklist);
2434 /* PA_IN starts with PA_OUT - TMP_GEN.
2435 Then we subtract things from ANTIC_IN. */
2436 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2438 /* For partial antic, we want to put back in the phi results, since
2439 we will properly avoid making them partially antic over backedges. */
2440 bitmap_ior_into (PA_IN (block)->values, PHI_GEN (block)->values);
2441 bitmap_ior_into (PA_IN (block)->expressions, PHI_GEN (block)->expressions);
2443 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2444 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2446 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2448 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2451 SET_BIT (changed_blocks, block->index);
2452 FOR_EACH_EDGE (e, ei, block->preds)
2453 SET_BIT (changed_blocks, e->src->index);
2456 RESET_BIT (changed_blocks, block->index);
2459 if (dump_file && (dump_flags & TDF_DETAILS))
2462 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2464 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2467 bitmap_set_free (old_PA_IN);
2469 bitmap_set_free (PA_OUT);
2473 /* Compute ANTIC and partial ANTIC sets. */
2476 compute_antic (void)
2478 bool changed = true;
2479 int num_iterations = 0;
2483 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2484 We pre-build the map of blocks with incoming abnormal edges here. */
2485 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2486 sbitmap_zero (has_abnormal_preds);
2493 FOR_EACH_EDGE (e, ei, block->preds)
2495 e->flags &= ~EDGE_DFS_BACK;
2496 if (e->flags & EDGE_ABNORMAL)
2498 SET_BIT (has_abnormal_preds, block->index);
2503 BB_VISITED (block) = 0;
2504 BB_DEFERRED (block) = 0;
2506 /* While we are here, give empty ANTIC_IN sets to each block. */
2507 ANTIC_IN (block) = bitmap_set_new ();
2508 PA_IN (block) = bitmap_set_new ();
2511 /* At the exit block we anticipate nothing. */
2512 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2513 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2514 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2516 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2517 sbitmap_ones (changed_blocks);
2520 if (dump_file && (dump_flags & TDF_DETAILS))
2521 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2524 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2526 if (TEST_BIT (changed_blocks, postorder[i]))
2528 basic_block block = BASIC_BLOCK (postorder[i]);
2529 changed |= compute_antic_aux (block,
2530 TEST_BIT (has_abnormal_preds,
2534 #ifdef ENABLE_CHECKING
2535 /* Theoretically possible, but *highly* unlikely. */
2536 gcc_assert (num_iterations < 500);
2540 statistics_histogram_event (cfun, "compute_antic iterations",
2543 if (do_partial_partial)
2545 sbitmap_ones (changed_blocks);
2546 mark_dfs_back_edges ();
2551 if (dump_file && (dump_flags & TDF_DETAILS))
2552 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2555 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2557 if (TEST_BIT (changed_blocks, postorder[i]))
2559 basic_block block = BASIC_BLOCK (postorder[i]);
2561 |= compute_partial_antic_aux (block,
2562 TEST_BIT (has_abnormal_preds,
2566 #ifdef ENABLE_CHECKING
2567 /* Theoretically possible, but *highly* unlikely. */
2568 gcc_assert (num_iterations < 500);
2571 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2574 sbitmap_free (has_abnormal_preds);
2575 sbitmap_free (changed_blocks);
2578 /* Return true if we can value number the call in STMT. This is true
2579 if we have a pure or constant call. */
2582 can_value_number_call (gimple stmt)
2584 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2589 /* Return true if OP is a tree which we can perform PRE on.
2590 This may not match the operations we can value number, but in
2591 a perfect world would. */
2594 can_PRE_operation (tree op)
2596 return UNARY_CLASS_P (op)
2597 || BINARY_CLASS_P (op)
2598 || COMPARISON_CLASS_P (op)
2599 || TREE_CODE (op) == INDIRECT_REF
2600 || TREE_CODE (op) == COMPONENT_REF
2601 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2602 || TREE_CODE (op) == CALL_EXPR
2603 || TREE_CODE (op) == ARRAY_REF;
2607 /* Inserted expressions are placed onto this worklist, which is used
2608 for performing quick dead code elimination of insertions we made
2609 that didn't turn out to be necessary. */
2610 static bitmap inserted_exprs;
2612 /* Pool allocated fake store expressions are placed onto this
2613 worklist, which, after performing dead code elimination, is walked
2614 to see which expressions need to be put into GC'able memory */
2615 static VEC(gimple, heap) *need_creation;
2617 /* The actual worker for create_component_ref_by_pieces. */
2620 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2621 unsigned int *operand, gimple_seq *stmts,
2624 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2628 switch (currop->opcode)
2632 tree folded, sc = NULL_TREE;
2633 unsigned int nargs = 0;
2635 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2639 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2640 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2646 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2647 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2651 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2652 ref->operands) - 1);
2653 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2655 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2665 folded = build_call_array (currop->type,
2666 (TREE_CODE (fn) == FUNCTION_DECL
2667 ? build_fold_addr_expr (fn) : fn),
2671 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2675 case TARGET_MEM_REF:
2677 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2680 tree genop0 = NULL_TREE;
2681 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2687 op0expr = get_or_alloc_expr_for (currop->op0);
2688 genop0 = find_or_generate_expression (block, op0expr,
2693 if (DECL_P (baseop))
2694 return build6 (TARGET_MEM_REF, currop->type,
2696 genop0, currop->op1, currop->op2,
2697 unshare_expr (nextop->op1));
2699 return build6 (TARGET_MEM_REF, currop->type,
2701 genop0, currop->op1, currop->op2,
2702 unshare_expr (nextop->op1));
2708 gcc_assert (is_gimple_min_invariant (currop->op0));
2714 case VIEW_CONVERT_EXPR:
2717 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2722 folded = fold_build1 (currop->opcode, currop->type,
2727 case ALIGN_INDIRECT_REF:
2728 case MISALIGNED_INDIRECT_REF:
2732 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2737 genop1 = fold_convert (build_pointer_type (currop->type),
2740 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2741 folded = fold_build2 (currop->opcode, currop->type,
2742 genop1, currop->op1);
2744 folded = fold_build1 (currop->opcode, currop->type,
2752 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2754 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2755 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2761 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2764 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2767 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2772 /* For array ref vn_reference_op's, operand 1 of the array ref
2773 is op0 of the reference op and operand 3 of the array ref is
2775 case ARRAY_RANGE_REF:
2779 tree genop1 = currop->op0;
2781 tree genop2 = currop->op1;
2783 tree genop3 = currop->op2;
2785 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2789 op1expr = get_or_alloc_expr_for (genop1);
2790 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2795 /* Drop zero minimum index. */
2796 if (tree_int_cst_equal (genop2, integer_zero_node))
2800 op2expr = get_or_alloc_expr_for (genop2);
2801 genop2 = find_or_generate_expression (block, op2expr, stmts,
2809 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2810 /* We can't always put a size in units of the element alignment
2811 here as the element alignment may be not visible. See
2812 PR43783. Simply drop the element size for constant
2814 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2818 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2819 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2820 op3expr = get_or_alloc_expr_for (genop3);
2821 genop3 = find_or_generate_expression (block, op3expr, stmts,
2827 return build4 (currop->opcode, currop->type, genop0, genop1,
2834 tree genop2 = currop->op1;
2836 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2840 /* op1 should be a FIELD_DECL, which are represented by
2845 op2expr = get_or_alloc_expr_for (genop2);
2846 genop2 = find_or_generate_expression (block, op2expr, stmts,
2852 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2858 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2859 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2880 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2881 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2882 trying to rename aggregates into ssa form directly, which is a no no.
2884 Thus, this routine doesn't create temporaries, it just builds a
2885 single access expression for the array, calling
2886 find_or_generate_expression to build the innermost pieces.
2888 This function is a subroutine of create_expression_by_pieces, and
2889 should not be called on it's own unless you really know what you
2893 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2894 gimple_seq *stmts, gimple domstmt)
2896 unsigned int op = 0;
2897 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2900 /* Find a leader for an expression, or generate one using
2901 create_expression_by_pieces if it's ANTIC but
2903 BLOCK is the basic_block we are looking for leaders in.
2904 EXPR is the expression to find a leader or generate for.
2905 STMTS is the statement list to put the inserted expressions on.
2906 Returns the SSA_NAME of the LHS of the generated expression or the
2908 DOMSTMT if non-NULL is a statement that should be dominated by
2909 all uses in the generated expression. If DOMSTMT is non-NULL this
2910 routine can fail and return NULL_TREE. Otherwise it will assert
2914 find_or_generate_expression (basic_block block, pre_expr expr,
2915 gimple_seq *stmts, gimple domstmt)
2917 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2918 get_expr_value_id (expr), domstmt);
2922 if (leader->kind == NAME)
2923 genop = PRE_EXPR_NAME (leader);
2924 else if (leader->kind == CONSTANT)
2925 genop = PRE_EXPR_CONSTANT (leader);
2928 /* If it's still NULL, it must be a complex expression, so generate
2929 it recursively. Not so for FRE though. */
2933 bitmap_set_t exprset;
2934 unsigned int lookfor = get_expr_value_id (expr);
2935 bool handled = false;
2939 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2940 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2942 pre_expr temp = expression_for_id (i);
2943 if (temp->kind != NAME)
2946 genop = create_expression_by_pieces (block, temp, stmts,
2948 get_expr_type (expr));
2952 if (!handled && domstmt)
2955 gcc_assert (handled);
2960 #define NECESSARY GF_PLF_1
2962 /* Create an expression in pieces, so that we can handle very complex
2963 expressions that may be ANTIC, but not necessary GIMPLE.
2964 BLOCK is the basic block the expression will be inserted into,
2965 EXPR is the expression to insert (in value form)
2966 STMTS is a statement list to append the necessary insertions into.
2968 This function will die if we hit some value that shouldn't be
2969 ANTIC but is (IE there is no leader for it, or its components).
2970 This function may also generate expressions that are themselves
2971 partially or fully redundant. Those that are will be either made
2972 fully redundant during the next iteration of insert (for partially
2973 redundant ones), or eliminated by eliminate (for fully redundant
2976 If DOMSTMT is non-NULL then we make sure that all uses in the
2977 expressions dominate that statement. In this case the function
2978 can return NULL_TREE to signal failure. */
2981 create_expression_by_pieces (basic_block block, pre_expr expr,
2982 gimple_seq *stmts, gimple domstmt, tree type)
2986 gimple_seq forced_stmts = NULL;
2987 unsigned int value_id;
2988 gimple_stmt_iterator gsi;
2989 tree exprtype = type ? type : get_expr_type (expr);
2995 /* We may hit the NAME/CONSTANT case if we have to convert types
2996 that value numbering saw through. */
2998 folded = PRE_EXPR_NAME (expr);
3001 folded = PRE_EXPR_CONSTANT (expr);
3005 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3006 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3011 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3012 switch (nary->length)
3016 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3017 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
3018 tree genop1 = find_or_generate_expression (block, op1,
3020 tree genop2 = find_or_generate_expression (block, op2,
3022 if (!genop1 || !genop2)
3024 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
3025 may be a constant with the wrong type. */
3026 if (nary->opcode == POINTER_PLUS_EXPR)
3028 genop1 = fold_convert (nary->type, genop1);
3029 genop2 = fold_convert (sizetype, genop2);
3033 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3034 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
3037 folded = fold_build2 (nary->opcode, nary->type,
3043 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3044 tree genop1 = find_or_generate_expression (block, op1,
3048 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3050 folded = fold_build1 (nary->opcode, nary->type,
3063 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3064 folded = fold_convert (exprtype, folded);
3066 /* Force the generated expression to be a sequence of GIMPLE
3068 We have to call unshare_expr because force_gimple_operand may
3069 modify the tree we pass to it. */
3070 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3073 /* If we have any intermediate expressions to the value sets, add them
3074 to the value sets and chain them in the instruction stream. */
3077 gsi = gsi_start (forced_stmts);
3078 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3080 gimple stmt = gsi_stmt (gsi);
3081 tree forcedname = gimple_get_lhs (stmt);
3084 if (TREE_CODE (forcedname) == SSA_NAME)
3086 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3087 VN_INFO_GET (forcedname)->valnum = forcedname;
3088 VN_INFO (forcedname)->value_id = get_next_value_id ();
3089 nameexpr = get_or_alloc_expr_for_name (forcedname);
3090 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3092 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3093 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3095 mark_symbols_for_renaming (stmt);
3097 gimple_seq_add_seq (stmts, forced_stmts);
3100 /* Build and insert the assignment of the end result to the temporary
3101 that we will return. */
3102 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3104 pretemp = create_tmp_reg (exprtype, "pretmp");
3105 get_var_ann (pretemp);
3109 add_referenced_var (temp);
3111 newstmt = gimple_build_assign (temp, folded);
3112 name = make_ssa_name (temp, newstmt);
3113 gimple_assign_set_lhs (newstmt, name);
3114 gimple_set_plf (newstmt, NECESSARY, false);
3116 gimple_seq_add_stmt (stmts, newstmt);
3117 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3119 /* All the symbols in NEWEXPR should be put into SSA form. */
3120 mark_symbols_for_renaming (newstmt);
3122 /* Add a value number to the temporary.
3123 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3124 we are creating the expression by pieces, and this particular piece of
3125 the expression may have been represented. There is no harm in replacing
3127 VN_INFO_GET (name)->valnum = name;
3128 value_id = get_expr_value_id (expr);
3129 VN_INFO (name)->value_id = value_id;
3130 nameexpr = get_or_alloc_expr_for_name (name);
3131 add_to_value (value_id, nameexpr);
3133 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3134 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3136 pre_stats.insertions++;
3137 if (dump_file && (dump_flags & TDF_DETAILS))
3139 fprintf (dump_file, "Inserted ");
3140 print_gimple_stmt (dump_file, newstmt, 0, 0);
3141 fprintf (dump_file, " in predecessor %d\n", block->index);
3148 /* Returns true if we want to inhibit the insertions of PHI nodes
3149 for the given EXPR for basic block BB (a member of a loop).
3150 We want to do this, when we fear that the induction variable we
3151 create might inhibit vectorization. */
3154 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3156 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3157 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3158 vn_reference_op_t op;
3161 /* If we aren't going to vectorize we don't inhibit anything. */
3162 if (!flag_tree_vectorize)
3165 /* Otherwise we inhibit the insertion when the address of the
3166 memory reference is a simple induction variable. In other
3167 cases the vectorizer won't do anything anyway (either it's
3168 loop invariant or a complicated expression). */
3169 for (i = 0; VEC_iterate (vn_reference_op_s, ops, i, op); ++i)
3174 case ARRAY_RANGE_REF:
3175 if (TREE_CODE (op->op0) != SSA_NAME)
3180 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3182 /* Default defs are loop invariant. */
3185 /* Defined outside this loop, also loop invariant. */
3186 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3188 /* If it's a simple induction variable inhibit insertion,
3189 the vectorizer might be interested in this one. */
3190 if (simple_iv (bb->loop_father, bb->loop_father,
3191 op->op0, &iv, true))
3193 /* No simple IV, vectorizer can't do anything, hence no
3194 reason to inhibit the transformation for this operand. */
3204 /* Insert the to-be-made-available values of expression EXPRNUM for each
3205 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3206 merge the result with a phi node, given the same value number as
3207 NODE. Return true if we have inserted new stuff. */
3210 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3213 pre_expr expr = expression_for_id (exprnum);
3215 unsigned int val = get_expr_value_id (expr);
3217 bool insertions = false;
3222 tree type = get_expr_type (expr);
3226 if (dump_file && (dump_flags & TDF_DETAILS))
3228 fprintf (dump_file, "Found partial redundancy for expression ");
3229 print_pre_expr (dump_file, expr);
3230 fprintf (dump_file, " (%04d)\n", val);
3233 /* Make sure we aren't creating an induction variable. */
3234 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3236 bool firstinsideloop = false;
3237 bool secondinsideloop = false;
3238 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3239 EDGE_PRED (block, 0)->src);
3240 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3241 EDGE_PRED (block, 1)->src);
3242 /* Induction variables only have one edge inside the loop. */
3243 if ((firstinsideloop ^ secondinsideloop)
3244 && (expr->kind != REFERENCE
3245 || inhibit_phi_insertion (block, expr)))
3247 if (dump_file && (dump_flags & TDF_DETAILS))
3248 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3253 /* Make the necessary insertions. */
3254 FOR_EACH_EDGE (pred, ei, block->preds)
3256 gimple_seq stmts = NULL;
3259 eprime = avail[bprime->index];
3261 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3263 builtexpr = create_expression_by_pieces (bprime,
3267 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3268 gsi_insert_seq_on_edge (pred, stmts);
3269 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3272 else if (eprime->kind == CONSTANT)
3274 /* Constants may not have the right type, fold_convert
3275 should give us back a constant with the right type.
3277 tree constant = PRE_EXPR_CONSTANT (eprime);
3278 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3280 tree builtexpr = fold_convert (type, constant);
3281 if (!is_gimple_min_invariant (builtexpr))
3283 tree forcedexpr = force_gimple_operand (builtexpr,
3286 if (!is_gimple_min_invariant (forcedexpr))
3288 if (forcedexpr != builtexpr)
3290 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3291 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3295 gimple_stmt_iterator gsi;
3296 gsi = gsi_start (stmts);
3297 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3299 gimple stmt = gsi_stmt (gsi);
3300 tree lhs = gimple_get_lhs (stmt);
3301 if (TREE_CODE (lhs) == SSA_NAME)
3302 bitmap_set_bit (inserted_exprs,
3303 SSA_NAME_VERSION (lhs));
3304 gimple_set_plf (stmt, NECESSARY, false);
3306 gsi_insert_seq_on_edge (pred, stmts);
3308 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3313 else if (eprime->kind == NAME)
3315 /* We may have to do a conversion because our value
3316 numbering can look through types in certain cases, but
3317 our IL requires all operands of a phi node have the same
3319 tree name = PRE_EXPR_NAME (eprime);
3320 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3324 builtexpr = fold_convert (type, name);
3325 forcedexpr = force_gimple_operand (builtexpr,
3329 if (forcedexpr != name)
3331 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3332 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3337 gimple_stmt_iterator gsi;
3338 gsi = gsi_start (stmts);
3339 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3341 gimple stmt = gsi_stmt (gsi);
3342 tree lhs = gimple_get_lhs (stmt);
3343 if (TREE_CODE (lhs) == SSA_NAME)
3344 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3345 gimple_set_plf (stmt, NECESSARY, false);
3347 gsi_insert_seq_on_edge (pred, stmts);
3349 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3353 /* If we didn't want a phi node, and we made insertions, we still have
3354 inserted new stuff, and thus return true. If we didn't want a phi node,
3355 and didn't make insertions, we haven't added anything new, so return
3357 if (nophi && insertions)
3359 else if (nophi && !insertions)
3362 /* Now build a phi for the new variable. */
3363 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3365 prephitemp = create_tmp_var (type, "prephitmp");
3366 get_var_ann (prephitemp);
3370 add_referenced_var (temp);
3372 if (TREE_CODE (type) == COMPLEX_TYPE
3373 || TREE_CODE (type) == VECTOR_TYPE)
3374 DECL_GIMPLE_REG_P (temp) = 1;
3375 phi = create_phi_node (temp, block);
3377 gimple_set_plf (phi, NECESSARY, false);
3378 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3379 VN_INFO (gimple_phi_result (phi))->value_id = val;
3380 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3381 FOR_EACH_EDGE (pred, ei, block->preds)
3383 pre_expr ae = avail[pred->src->index];
3384 gcc_assert (get_expr_type (ae) == type
3385 || useless_type_conversion_p (type, get_expr_type (ae)));
3386 if (ae->kind == CONSTANT)
3387 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3389 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3393 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3394 add_to_value (val, newphi);
3396 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3397 this insertion, since we test for the existence of this value in PHI_GEN
3398 before proceeding with the partial redundancy checks in insert_aux.
3400 The value may exist in AVAIL_OUT, in particular, it could be represented
3401 by the expression we are trying to eliminate, in which case we want the
3402 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3405 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3406 this block, because if it did, it would have existed in our dominator's
3407 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3410 bitmap_insert_into_set (PHI_GEN (block), newphi);
3411 bitmap_value_replace_in_set (AVAIL_OUT (block),
3413 bitmap_insert_into_set (NEW_SETS (block),
3416 if (dump_file && (dump_flags & TDF_DETAILS))
3418 fprintf (dump_file, "Created phi ");
3419 print_gimple_stmt (dump_file, phi, 0, 0);
3420 fprintf (dump_file, " in block %d\n", block->index);
3428 /* Perform insertion of partially redundant values.
3429 For BLOCK, do the following:
3430 1. Propagate the NEW_SETS of the dominator into the current block.
3431 If the block has multiple predecessors,
3432 2a. Iterate over the ANTIC expressions for the block to see if
3433 any of them are partially redundant.
3434 2b. If so, insert them into the necessary predecessors to make
3435 the expression fully redundant.
3436 2c. Insert a new PHI merging the values of the predecessors.
3437 2d. Insert the new PHI, and the new expressions, into the
3439 3. Recursively call ourselves on the dominator children of BLOCK.
3441 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3442 do_regular_insertion and do_partial_insertion.
3447 do_regular_insertion (basic_block block, basic_block dom)
3449 bool new_stuff = false;
3450 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3454 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3456 if (expr->kind != NAME)
3460 bool by_some = false;
3461 bool cant_insert = false;
3462 bool all_same = true;
3463 pre_expr first_s = NULL;
3466 pre_expr eprime = NULL;
3468 pre_expr edoubleprime = NULL;
3469 bool do_insertion = false;
3471 val = get_expr_value_id (expr);
3472 if (bitmap_set_contains_value (PHI_GEN (block), val))
3474 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3476 if (dump_file && (dump_flags & TDF_DETAILS))
3477 fprintf (dump_file, "Found fully redundant value\n");
3481 avail = XCNEWVEC (pre_expr, last_basic_block);
3482 FOR_EACH_EDGE (pred, ei, block->preds)
3484 unsigned int vprime;
3486 /* We should never run insertion for the exit block
3487 and so not come across fake pred edges. */
3488 gcc_assert (!(pred->flags & EDGE_FAKE));
3490 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3493 /* eprime will generally only be NULL if the
3494 value of the expression, translated
3495 through the PHI for this predecessor, is
3496 undefined. If that is the case, we can't
3497 make the expression fully redundant,
3498 because its value is undefined along a
3499 predecessor path. We can thus break out
3500 early because it doesn't matter what the
3501 rest of the results are. */
3508 eprime = fully_constant_expression (eprime);
3509 vprime = get_expr_value_id (eprime);
3510 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3512 if (edoubleprime == NULL)
3514 avail[bprime->index] = eprime;
3519 avail[bprime->index] = edoubleprime;
3521 /* We want to perform insertions to remove a redundancy on
3522 a path in the CFG we want to optimize for speed. */
3523 if (optimize_edge_for_speed_p (pred))
3524 do_insertion = true;
3525 if (first_s == NULL)
3526 first_s = edoubleprime;
3527 else if (!pre_expr_eq (first_s, edoubleprime))
3531 /* If we can insert it, it's not the same value
3532 already existing along every predecessor, and
3533 it's defined by some predecessor, it is
3534 partially redundant. */
3535 if (!cant_insert && !all_same && by_some && do_insertion
3536 && dbg_cnt (treepre_insert))
3538 if (insert_into_preds_of_block (block, get_expression_id (expr),
3542 /* If all edges produce the same value and that value is
3543 an invariant, then the PHI has the same value on all
3544 edges. Note this. */
3545 else if (!cant_insert && all_same && eprime
3546 && (edoubleprime->kind == CONSTANT
3547 || edoubleprime->kind == NAME)
3548 && !value_id_constant_p (val))
3552 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3553 value_expressions, val);
3555 unsigned int new_val = get_expr_value_id (edoubleprime);
3556 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3558 pre_expr expr = expression_for_id (j);
3560 if (expr->kind == NAME)
3562 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3563 /* Just reset the value id and valnum so it is
3564 the same as the constant we have discovered. */
3565 if (edoubleprime->kind == CONSTANT)
3567 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3568 pre_stats.constified++;
3571 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3572 info->value_id = new_val;
3580 VEC_free (pre_expr, heap, exprs);
3585 /* Perform insertion for partially anticipatable expressions. There
3586 is only one case we will perform insertion for these. This case is
3587 if the expression is partially anticipatable, and fully available.
3588 In this case, we know that putting it earlier will enable us to
3589 remove the later computation. */
3593 do_partial_partial_insertion (basic_block block, basic_block dom)
3595 bool new_stuff = false;
3596 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3600 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3602 if (expr->kind != NAME)
3607 bool cant_insert = false;
3610 pre_expr eprime = NULL;
3613 val = get_expr_value_id (expr);
3614 if (bitmap_set_contains_value (PHI_GEN (block), val))
3616 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3619 avail = XCNEWVEC (pre_expr, last_basic_block);
3620 FOR_EACH_EDGE (pred, ei, block->preds)
3622 unsigned int vprime;
3623 pre_expr edoubleprime;
3625 /* We should never run insertion for the exit block
3626 and so not come across fake pred edges. */
3627 gcc_assert (!(pred->flags & EDGE_FAKE));
3629 eprime = phi_translate (expr, ANTIC_IN (block),
3633 /* eprime will generally only be NULL if the
3634 value of the expression, translated
3635 through the PHI for this predecessor, is
3636 undefined. If that is the case, we can't
3637 make the expression fully redundant,
3638 because its value is undefined along a
3639 predecessor path. We can thus break out
3640 early because it doesn't matter what the
3641 rest of the results are. */
3648 eprime = fully_constant_expression (eprime);
3649 vprime = get_expr_value_id (eprime);
3650 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3652 if (edoubleprime == NULL)
3658 avail[bprime->index] = edoubleprime;
3662 /* If we can insert it, it's not the same value
3663 already existing along every predecessor, and
3664 it's defined by some predecessor, it is
3665 partially redundant. */
3666 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3668 pre_stats.pa_insert++;
3669 if (insert_into_preds_of_block (block, get_expression_id (expr),
3677 VEC_free (pre_expr, heap, exprs);
3682 insert_aux (basic_block block)
3685 bool new_stuff = false;
3690 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3695 bitmap_set_t newset = NEW_SETS (dom);
3698 /* Note that we need to value_replace both NEW_SETS, and
3699 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3700 represented by some non-simple expression here that we want
3701 to replace it with. */
3702 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3704 pre_expr expr = expression_for_id (i);
3705 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3706 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3709 if (!single_pred_p (block))
3711 new_stuff |= do_regular_insertion (block, dom);
3712 if (do_partial_partial)
3713 new_stuff |= do_partial_partial_insertion (block, dom);
3717 for (son = first_dom_son (CDI_DOMINATORS, block);
3719 son = next_dom_son (CDI_DOMINATORS, son))
3721 new_stuff |= insert_aux (son);
3727 /* Perform insertion of partially redundant values. */
3732 bool new_stuff = true;
3734 int num_iterations = 0;
3737 NEW_SETS (bb) = bitmap_set_new ();
3742 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3744 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3748 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3751 add_to_exp_gen (basic_block block, tree op)
3756 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3758 result = get_or_alloc_expr_for_name (op);
3759 bitmap_value_insert_into_set (EXP_GEN (block), result);
3763 /* Create value ids for PHI in BLOCK. */
3766 make_values_for_phi (gimple phi, basic_block block)
3768 tree result = gimple_phi_result (phi);
3770 /* We have no need for virtual phis, as they don't represent
3771 actual computations. */
3772 if (is_gimple_reg (result))
3774 pre_expr e = get_or_alloc_expr_for_name (result);
3775 add_to_value (get_expr_value_id (e), e);
3776 bitmap_insert_into_set (PHI_GEN (block), e);
3777 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3781 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3783 tree arg = gimple_phi_arg_def (phi, i);
3784 if (TREE_CODE (arg) == SSA_NAME)
3786 e = get_or_alloc_expr_for_name (arg);
3787 add_to_value (get_expr_value_id (e), e);
3794 /* Compute the AVAIL set for all basic blocks.
3796 This function performs value numbering of the statements in each basic
3797 block. The AVAIL sets are built from information we glean while doing
3798 this value numbering, since the AVAIL sets contain only one entry per
3801 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3802 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3805 compute_avail (void)
3808 basic_block block, son;
3809 basic_block *worklist;
3813 /* We pretend that default definitions are defined in the entry block.
3814 This includes function arguments and the static chain decl. */
3815 for (i = 1; i < num_ssa_names; ++i)
3817 tree name = ssa_name (i);
3820 || !SSA_NAME_IS_DEFAULT_DEF (name)
3821 || has_zero_uses (name)
3822 || !is_gimple_reg (name))
3825 e = get_or_alloc_expr_for_name (name);
3826 add_to_value (get_expr_value_id (e), e);
3828 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3829 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3832 /* Allocate the worklist. */
3833 worklist = XNEWVEC (basic_block, n_basic_blocks);
3835 /* Seed the algorithm by putting the dominator children of the entry
3836 block on the worklist. */
3837 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3839 son = next_dom_son (CDI_DOMINATORS, son))
3840 worklist[sp++] = son;
3842 /* Loop until the worklist is empty. */
3845 gimple_stmt_iterator gsi;
3848 unsigned int stmt_uid = 1;
3850 /* Pick a block from the worklist. */
3851 block = worklist[--sp];
3853 /* Initially, the set of available values in BLOCK is that of
3854 its immediate dominator. */
3855 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3857 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3859 /* Generate values for PHI nodes. */
3860 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3861 make_values_for_phi (gsi_stmt (gsi), block);
3863 BB_MAY_NOTRETURN (block) = 0;
3865 /* Now compute value numbers and populate value sets with all
3866 the expressions computed in BLOCK. */
3867 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3872 stmt = gsi_stmt (gsi);
3873 gimple_set_uid (stmt, stmt_uid++);
3875 /* Cache whether the basic-block has any non-visible side-effect
3877 If this isn't a call or it is the last stmt in the
3878 basic-block then the CFG represents things correctly. */
3879 if (is_gimple_call (stmt)
3880 && !stmt_ends_bb_p (stmt))
3882 /* Non-looping const functions always return normally.
3883 Otherwise the call might not return or have side-effects
3884 that forbids hoisting possibly trapping expressions
3886 int flags = gimple_call_flags (stmt);
3887 if (!(flags & ECF_CONST)
3888 || (flags & ECF_LOOPING_CONST_OR_PURE))
3889 BB_MAY_NOTRETURN (block) = 1;
3892 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3894 pre_expr e = get_or_alloc_expr_for_name (op);
3896 add_to_value (get_expr_value_id (e), e);
3898 bitmap_insert_into_set (TMP_GEN (block), e);
3899 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3902 if (gimple_has_volatile_ops (stmt)
3903 || stmt_could_throw_p (stmt))
3906 switch (gimple_code (stmt))
3909 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3910 add_to_exp_gen (block, op);
3917 vn_reference_op_t vro;
3918 pre_expr result = NULL;
3919 VEC(vn_reference_op_s, heap) *ops = NULL;
3921 if (!can_value_number_call (stmt))
3924 copy_reference_ops_from_call (stmt, &ops);
3925 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3926 gimple_expr_type (stmt),
3928 VEC_free (vn_reference_op_s, heap, ops);
3932 for (i = 0; VEC_iterate (vn_reference_op_s,
3936 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3937 add_to_exp_gen (block, vro->op0);
3938 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3939 add_to_exp_gen (block, vro->op1);
3940 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3941 add_to_exp_gen (block, vro->op2);
3943 result = (pre_expr) pool_alloc (pre_expr_pool);
3944 result->kind = REFERENCE;
3946 PRE_EXPR_REFERENCE (result) = ref;
3948 get_or_alloc_expression_id (result);
3949 add_to_value (get_expr_value_id (result), result);
3951 bitmap_value_insert_into_set (EXP_GEN (block), result);
3957 pre_expr result = NULL;
3958 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
3962 case tcc_comparison:
3967 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
3968 gimple_assign_rhs_code (stmt),
3969 gimple_expr_type (stmt),
3970 gimple_assign_rhs1 (stmt),
3971 gimple_assign_rhs2 (stmt),
3972 NULL_TREE, NULL_TREE, &nary);
3977 for (i = 0; i < nary->length; i++)
3978 if (TREE_CODE (nary->op[i]) == SSA_NAME)
3979 add_to_exp_gen (block, nary->op[i]);
3981 result = (pre_expr) pool_alloc (pre_expr_pool);
3982 result->kind = NARY;
3984 PRE_EXPR_NARY (result) = nary;
3988 case tcc_declaration:
3993 vn_reference_op_t vro;
3995 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4001 for (i = 0; VEC_iterate (vn_reference_op_s,
4005 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4006 add_to_exp_gen (block, vro->op0);
4007 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4008 add_to_exp_gen (block, vro->op1);
4009 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4010 add_to_exp_gen (block, vro->op2);
4012 result = (pre_expr) pool_alloc (pre_expr_pool);
4013 result->kind = REFERENCE;
4015 PRE_EXPR_REFERENCE (result) = ref;
4020 /* For any other statement that we don't
4021 recognize, simply add all referenced
4022 SSA_NAMEs to EXP_GEN. */
4023 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4024 add_to_exp_gen (block, op);
4028 get_or_alloc_expression_id (result);
4029 add_to_value (get_expr_value_id (result), result);
4031 bitmap_value_insert_into_set (EXP_GEN (block), result);
4040 /* Put the dominator children of BLOCK on the worklist of blocks
4041 to compute available sets for. */
4042 for (son = first_dom_son (CDI_DOMINATORS, block);
4044 son = next_dom_son (CDI_DOMINATORS, son))
4045 worklist[sp++] = son;
4051 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4052 than the available expressions for it. The insertion point is
4053 right before the first use in STMT. Returns the SSA_NAME that should
4054 be used for replacement. */
4057 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4059 basic_block bb = gimple_bb (stmt);
4060 gimple_stmt_iterator gsi;
4061 gimple_seq stmts = NULL;
4065 /* First create a value expression from the expression we want
4066 to insert and associate it with the value handle for SSA_VN. */
4067 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4071 /* Then use create_expression_by_pieces to generate a valid
4072 expression to insert at this point of the IL stream. */
4073 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4074 if (expr == NULL_TREE)
4076 gsi = gsi_for_stmt (stmt);
4077 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4082 /* Eliminate fully redundant computations. */
4087 VEC (gimple, heap) *to_remove = NULL;
4089 unsigned int todo = 0;
4090 gimple_stmt_iterator gsi;
4096 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4098 stmt = gsi_stmt (gsi);
4100 /* Lookup the RHS of the expression, see if we have an
4101 available computation for it. If so, replace the RHS with
4102 the available computation. */
4103 if (gimple_has_lhs (stmt)
4104 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4105 && !gimple_assign_ssa_name_copy_p (stmt)
4106 && (!gimple_assign_single_p (stmt)
4107 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4108 && !gimple_has_volatile_ops (stmt)
4109 && !has_zero_uses (gimple_get_lhs (stmt)))
4111 tree lhs = gimple_get_lhs (stmt);
4112 tree rhs = NULL_TREE;
4114 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4115 pre_expr sprimeexpr;
4117 if (gimple_assign_single_p (stmt))
4118 rhs = gimple_assign_rhs1 (stmt);
4120 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4121 get_expr_value_id (lhsexpr),
4126 if (sprimeexpr->kind == CONSTANT)
4127 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4128 else if (sprimeexpr->kind == NAME)
4129 sprime = PRE_EXPR_NAME (sprimeexpr);
4134 /* If there is no existing leader but SCCVN knows this
4135 value is constant, use that constant. */
4136 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4138 sprime = VN_INFO (lhs)->valnum;
4139 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4140 TREE_TYPE (sprime)))
4141 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4143 if (dump_file && (dump_flags & TDF_DETAILS))
4145 fprintf (dump_file, "Replaced ");
4146 print_gimple_expr (dump_file, stmt, 0, 0);
4147 fprintf (dump_file, " with ");
4148 print_generic_expr (dump_file, sprime, 0);
4149 fprintf (dump_file, " in ");
4150 print_gimple_stmt (dump_file, stmt, 0, 0);
4152 pre_stats.eliminations++;
4153 propagate_tree_value_into_stmt (&gsi, sprime);
4154 stmt = gsi_stmt (gsi);
4159 /* If there is no existing usable leader but SCCVN thinks
4160 it has an expression it wants to use as replacement,
4162 if (!sprime || sprime == lhs)
4164 tree val = VN_INFO (lhs)->valnum;
4166 && TREE_CODE (val) == SSA_NAME
4167 && VN_INFO (val)->needs_insertion
4168 && can_PRE_operation (vn_get_expr_for (val)))
4169 sprime = do_SCCVN_insertion (stmt, val);
4173 && (rhs == NULL_TREE
4174 || TREE_CODE (rhs) != SSA_NAME
4175 || may_propagate_copy (rhs, sprime)))
4177 gcc_assert (sprime != rhs);
4179 if (dump_file && (dump_flags & TDF_DETAILS))
4181 fprintf (dump_file, "Replaced ");
4182 print_gimple_expr (dump_file, stmt, 0, 0);
4183 fprintf (dump_file, " with ");
4184 print_generic_expr (dump_file, sprime, 0);
4185 fprintf (dump_file, " in ");
4186 print_gimple_stmt (dump_file, stmt, 0, 0);
4189 if (TREE_CODE (sprime) == SSA_NAME)
4190 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4192 /* We need to make sure the new and old types actually match,
4193 which may require adding a simple cast, which fold_convert
4195 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4196 && !useless_type_conversion_p (gimple_expr_type (stmt),
4197 TREE_TYPE (sprime)))
4198 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4200 pre_stats.eliminations++;
4201 propagate_tree_value_into_stmt (&gsi, sprime);
4202 stmt = gsi_stmt (gsi);
4205 /* If we removed EH side effects from the statement, clean
4206 its EH information. */
4207 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4209 bitmap_set_bit (need_eh_cleanup,
4210 gimple_bb (stmt)->index);
4211 if (dump_file && (dump_flags & TDF_DETAILS))
4212 fprintf (dump_file, " Removed EH side effects.\n");
4216 /* If the statement is a scalar store, see if the expression
4217 has the same value number as its rhs. If so, the store is
4219 else if (gimple_assign_single_p (stmt)
4220 && !is_gimple_reg (gimple_assign_lhs (stmt))
4221 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4222 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4224 tree rhs = gimple_assign_rhs1 (stmt);
4226 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4227 gimple_vuse (stmt), true, NULL);
4228 if (TREE_CODE (rhs) == SSA_NAME)
4229 rhs = VN_INFO (rhs)->valnum;
4231 && operand_equal_p (val, rhs, 0))
4233 if (dump_file && (dump_flags & TDF_DETAILS))
4235 fprintf (dump_file, "Deleted redundant store ");
4236 print_gimple_stmt (dump_file, stmt, 0, 0);
4239 /* Queue stmt for removal. */
4240 VEC_safe_push (gimple, heap, to_remove, stmt);
4243 /* Visit COND_EXPRs and fold the comparison with the
4244 available value-numbers. */
4245 else if (gimple_code (stmt) == GIMPLE_COND)
4247 tree op0 = gimple_cond_lhs (stmt);
4248 tree op1 = gimple_cond_rhs (stmt);
4251 if (TREE_CODE (op0) == SSA_NAME)
4252 op0 = VN_INFO (op0)->valnum;
4253 if (TREE_CODE (op1) == SSA_NAME)
4254 op1 = VN_INFO (op1)->valnum;
4255 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4257 if (result && TREE_CODE (result) == INTEGER_CST)
4259 if (integer_zerop (result))
4260 gimple_cond_make_false (stmt);
4262 gimple_cond_make_true (stmt);
4264 todo = TODO_cleanup_cfg;
4267 /* Visit indirect calls and turn them into direct calls if
4269 if (gimple_code (stmt) == GIMPLE_CALL
4270 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4272 tree fn = VN_INFO (gimple_call_fn (stmt))->valnum;
4273 if (TREE_CODE (fn) == ADDR_EXPR
4274 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
4276 if (dump_file && (dump_flags & TDF_DETAILS))
4278 fprintf (dump_file, "Replacing call target with ");
4279 print_generic_expr (dump_file, fn, 0);
4280 fprintf (dump_file, " in ");
4281 print_gimple_stmt (dump_file, stmt, 0, 0);
4284 gimple_call_set_fn (stmt, fn);
4286 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4288 bitmap_set_bit (need_eh_cleanup,
4289 gimple_bb (stmt)->index);
4290 if (dump_file && (dump_flags & TDF_DETAILS))
4291 fprintf (dump_file, " Removed EH side effects.\n");
4294 /* Changing an indirect call to a direct call may
4295 have exposed different semantics. This may
4296 require an SSA update. */
4297 todo |= TODO_update_ssa_only_virtuals;
4302 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4304 gimple stmt, phi = gsi_stmt (gsi);
4305 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4306 pre_expr sprimeexpr, resexpr;
4307 gimple_stmt_iterator gsi2;
4309 /* We want to perform redundant PHI elimination. Do so by
4310 replacing the PHI with a single copy if possible.
4311 Do not touch inserted, single-argument or virtual PHIs. */
4312 if (gimple_phi_num_args (phi) == 1
4313 || !is_gimple_reg (res))
4319 resexpr = get_or_alloc_expr_for_name (res);
4320 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4321 get_expr_value_id (resexpr), NULL);
4324 if (sprimeexpr->kind == CONSTANT)
4325 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4326 else if (sprimeexpr->kind == NAME)
4327 sprime = PRE_EXPR_NAME (sprimeexpr);
4338 if (dump_file && (dump_flags & TDF_DETAILS))
4340 fprintf (dump_file, "Replaced redundant PHI node defining ");
4341 print_generic_expr (dump_file, res, 0);
4342 fprintf (dump_file, " with ");
4343 print_generic_expr (dump_file, sprime, 0);
4344 fprintf (dump_file, "\n");
4347 remove_phi_node (&gsi, false);
4349 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4350 && TREE_CODE (sprime) == SSA_NAME)
4351 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4353 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4354 sprime = fold_convert (TREE_TYPE (res), sprime);
4355 stmt = gimple_build_assign (res, sprime);
4356 SSA_NAME_DEF_STMT (res) = stmt;
4357 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4359 gsi2 = gsi_after_labels (b);
4360 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4361 /* Queue the copy for eventual removal. */
4362 VEC_safe_push (gimple, heap, to_remove, stmt);
4363 /* If we inserted this PHI node ourself, it's not an elimination. */
4364 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4367 pre_stats.eliminations++;
4371 /* We cannot remove stmts during BB walk, especially not release SSA
4372 names there as this confuses the VN machinery. The stmts ending
4373 up in to_remove are either stores or simple copies. */
4374 for (i = 0; VEC_iterate (gimple, to_remove, i, stmt); ++i)
4376 tree lhs = gimple_assign_lhs (stmt);
4377 tree rhs = gimple_assign_rhs1 (stmt);
4378 use_operand_p use_p;
4381 /* If there is a single use only, propagate the equivalency
4382 instead of keeping the copy. */
4383 if (TREE_CODE (lhs) == SSA_NAME
4384 && TREE_CODE (rhs) == SSA_NAME
4385 && single_imm_use (lhs, &use_p, &use_stmt)
4386 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4388 SET_USE (use_p, rhs);
4389 update_stmt (use_stmt);
4390 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4391 && TREE_CODE (rhs) == SSA_NAME)
4392 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4395 /* If this is a store or a now unused copy, remove it. */
4396 if (TREE_CODE (lhs) != SSA_NAME
4397 || has_zero_uses (lhs))
4399 gsi = gsi_for_stmt (stmt);
4400 unlink_stmt_vdef (stmt);
4401 gsi_remove (&gsi, true);
4402 if (TREE_CODE (lhs) == SSA_NAME)
4403 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4404 release_defs (stmt);
4407 VEC_free (gimple, heap, to_remove);
4412 /* Borrow a bit of tree-ssa-dce.c for the moment.
4413 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4414 this may be a bit faster, and we may want critical edges kept split. */
4416 /* If OP's defining statement has not already been determined to be necessary,
4417 mark that statement necessary. Return the stmt, if it is newly
4420 static inline gimple
4421 mark_operand_necessary (tree op)
4427 if (TREE_CODE (op) != SSA_NAME)
4430 stmt = SSA_NAME_DEF_STMT (op);
4433 if (gimple_plf (stmt, NECESSARY)
4434 || gimple_nop_p (stmt))
4437 gimple_set_plf (stmt, NECESSARY, true);
4441 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4442 to insert PHI nodes sometimes, and because value numbering of casts isn't
4443 perfect, we sometimes end up inserting dead code. This simple DCE-like
4444 pass removes any insertions we made that weren't actually used. */
4447 remove_dead_inserted_code (void)
4454 worklist = BITMAP_ALLOC (NULL);
4455 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4457 t = SSA_NAME_DEF_STMT (ssa_name (i));
4458 if (gimple_plf (t, NECESSARY))
4459 bitmap_set_bit (worklist, i);
4461 while (!bitmap_empty_p (worklist))
4463 i = bitmap_first_set_bit (worklist);
4464 bitmap_clear_bit (worklist, i);
4465 t = SSA_NAME_DEF_STMT (ssa_name (i));
4467 /* PHI nodes are somewhat special in that each PHI alternative has
4468 data and control dependencies. All the statements feeding the
4469 PHI node's arguments are always necessary. */
4470 if (gimple_code (t) == GIMPLE_PHI)
4474 for (k = 0; k < gimple_phi_num_args (t); k++)
4476 tree arg = PHI_ARG_DEF (t, k);
4477 if (TREE_CODE (arg) == SSA_NAME)
4479 gimple n = mark_operand_necessary (arg);
4481 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4487 /* Propagate through the operands. Examine all the USE, VUSE and
4488 VDEF operands in this statement. Mark all the statements
4489 which feed this statement's uses as necessary. */
4493 /* The operands of VDEF expressions are also needed as they
4494 represent potential definitions that may reach this
4495 statement (VDEF operands allow us to follow def-def
4498 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4500 gimple n = mark_operand_necessary (use);
4502 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4507 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4509 t = SSA_NAME_DEF_STMT (ssa_name (i));
4510 if (!gimple_plf (t, NECESSARY))
4512 gimple_stmt_iterator gsi;
4514 if (dump_file && (dump_flags & TDF_DETAILS))
4516 fprintf (dump_file, "Removing unnecessary insertion:");
4517 print_gimple_stmt (dump_file, t, 0, 0);
4520 gsi = gsi_for_stmt (t);
4521 if (gimple_code (t) == GIMPLE_PHI)
4522 remove_phi_node (&gsi, true);
4525 gsi_remove (&gsi, true);
4530 BITMAP_FREE (worklist);
4533 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4534 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4535 the number of visited blocks. */
4538 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4540 edge_iterator *stack;
4542 int post_order_num = 0;
4545 if (include_entry_exit)
4546 post_order[post_order_num++] = EXIT_BLOCK;
4548 /* Allocate stack for back-tracking up CFG. */
4549 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4552 /* Allocate bitmap to track nodes that have been visited. */
4553 visited = sbitmap_alloc (last_basic_block);
4555 /* None of the nodes in the CFG have been visited yet. */
4556 sbitmap_zero (visited);
4558 /* Push the last edge on to the stack. */
4559 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4567 /* Look at the edge on the top of the stack. */
4569 src = ei_edge (ei)->src;
4570 dest = ei_edge (ei)->dest;
4572 /* Check if the edge destination has been visited yet. */
4573 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4575 /* Mark that we have visited the destination. */
4576 SET_BIT (visited, src->index);
4578 if (EDGE_COUNT (src->preds) > 0)
4579 /* Since the DEST node has been visited for the first
4580 time, check its successors. */
4581 stack[sp++] = ei_start (src->preds);
4583 post_order[post_order_num++] = src->index;
4587 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4588 post_order[post_order_num++] = dest->index;
4590 if (!ei_one_before_end_p (ei))
4591 ei_next (&stack[sp - 1]);
4597 if (include_entry_exit)
4598 post_order[post_order_num++] = ENTRY_BLOCK;
4601 sbitmap_free (visited);
4602 return post_order_num;
4606 /* Initialize data structures used by PRE. */
4609 init_pre (bool do_fre)
4613 next_expression_id = 1;
4615 VEC_safe_push (pre_expr, heap, expressions, NULL);
4616 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4617 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4618 get_max_value_id() + 1);
4623 inserted_exprs = BITMAP_ALLOC (NULL);
4624 need_creation = NULL;
4625 pretemp = NULL_TREE;
4626 storetemp = NULL_TREE;
4627 prephitemp = NULL_TREE;
4629 connect_infinite_loops_to_exit ();
4630 memset (&pre_stats, 0, sizeof (pre_stats));
4633 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4634 my_rev_post_order_compute (postorder, false);
4637 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4639 calculate_dominance_info (CDI_POST_DOMINATORS);
4640 calculate_dominance_info (CDI_DOMINATORS);
4642 bitmap_obstack_initialize (&grand_bitmap_obstack);
4643 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4644 expr_pred_trans_eq, free);
4645 expression_to_id = htab_create (num_ssa_names * 3,
4648 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4649 sizeof (struct bitmap_set), 30);
4650 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4651 sizeof (struct pre_expr_d), 30);
4654 EXP_GEN (bb) = bitmap_set_new ();
4655 PHI_GEN (bb) = bitmap_set_new ();
4656 TMP_GEN (bb) = bitmap_set_new ();
4657 AVAIL_OUT (bb) = bitmap_set_new ();
4660 need_eh_cleanup = BITMAP_ALLOC (NULL);
4664 /* Deallocate data structures used by PRE. */
4667 fini_pre (bool do_fre)
4672 VEC_free (bitmap_set_t, heap, value_expressions);
4673 BITMAP_FREE (inserted_exprs);
4674 VEC_free (gimple, heap, need_creation);
4675 bitmap_obstack_release (&grand_bitmap_obstack);
4676 free_alloc_pool (bitmap_set_pool);
4677 free_alloc_pool (pre_expr_pool);
4678 htab_delete (phi_translate_table);
4679 htab_delete (expression_to_id);
4680 VEC_free (unsigned, heap, name_to_id);
4688 free_dominance_info (CDI_POST_DOMINATORS);
4690 if (!bitmap_empty_p (need_eh_cleanup))
4692 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4693 cleanup_tree_cfg ();
4696 BITMAP_FREE (need_eh_cleanup);
4699 loop_optimizer_finalize ();
4702 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4703 only wants to do full redundancy elimination. */
4706 execute_pre (bool do_fre)
4708 unsigned int todo = 0;
4710 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4712 /* This has to happen before SCCVN runs because
4713 loop_optimizer_init may create new phis, etc. */
4715 loop_optimizer_init (LOOPS_NORMAL);
4717 if (!run_scc_vn (do_fre))
4720 loop_optimizer_finalize ();
4728 /* Collect and value number expressions computed in each basic block. */
4731 if (dump_file && (dump_flags & TDF_DETAILS))
4737 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4738 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4739 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4740 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4744 /* Insert can get quite slow on an incredibly large number of basic
4745 blocks due to some quadratic behavior. Until this behavior is
4746 fixed, don't run it when he have an incredibly large number of
4747 bb's. If we aren't going to run insert, there is no point in
4748 computing ANTIC, either, even though it's plenty fast. */
4749 if (!do_fre && n_basic_blocks < 4000)
4755 /* Make sure to remove fake edges before committing our inserts.
4756 This makes sure we don't end up with extra critical edges that
4757 we would need to split. */
4758 remove_fake_exit_edges ();
4759 gsi_commit_edge_inserts ();
4761 /* Remove all the redundant expressions. */
4762 todo |= eliminate ();
4764 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4765 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4766 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4767 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4768 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4770 clear_expression_ids ();
4773 remove_dead_inserted_code ();
4781 /* Gate and execute functions for PRE. */
4786 return execute_pre (false);
4792 return flag_tree_pre != 0;
4795 struct gimple_opt_pass pass_pre =
4800 gate_pre, /* gate */
4801 do_pre, /* execute */
4804 0, /* static_pass_number */
4805 TV_TREE_PRE, /* tv_id */
4806 PROP_no_crit_edges | PROP_cfg
4807 | PROP_ssa, /* properties_required */
4808 0, /* properties_provided */
4809 0, /* properties_destroyed */
4810 TODO_rebuild_alias, /* todo_flags_start */
4811 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4812 | TODO_verify_ssa /* todo_flags_finish */
4817 /* Gate and execute functions for FRE. */
4822 return execute_pre (true);
4828 return flag_tree_fre != 0;
4831 struct gimple_opt_pass pass_fre =
4836 gate_fre, /* gate */
4837 execute_fre, /* execute */
4840 0, /* static_pass_number */
4841 TV_TREE_FRE, /* tv_id */
4842 PROP_cfg | PROP_ssa, /* properties_required */
4843 0, /* properties_provided */
4844 0, /* properties_destroyed */
4845 0, /* todo_flags_start */
4846 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */