2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
3 Free Software Foundation, Inc.
4 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
25 #include "coretypes.h"
29 #include "basic-block.h"
30 #include "diagnostic.h"
31 #include "tree-inline.h"
32 #include "tree-flow.h"
34 #include "tree-dump.h"
38 #include "tree-iterator.h"
40 #include "alloc-pool.h"
42 #include "tree-pass.h"
45 #include "langhooks.h"
47 #include "tree-ssa-sccvn.h"
48 #include "tree-scalar-evolution.h"
54 1. Avail sets can be shared by making an avail_find_leader that
55 walks up the dominator tree and looks in those avail sets.
56 This might affect code optimality, it's unclear right now.
57 2. Strength reduction can be performed by anticipating expressions
58 we can repair later on.
59 3. We can do back-substitution or smarter value numbering to catch
60 commutative expressions split up over multiple statements.
63 /* For ease of terminology, "expression node" in the below refers to
64 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
65 represent the actual statement containing the expressions we care about,
66 and we cache the value number by putting it in the expression. */
70 First we walk the statements to generate the AVAIL sets, the
71 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
72 generation of values/expressions by a given block. We use them
73 when computing the ANTIC sets. The AVAIL sets consist of
74 SSA_NAME's that represent values, so we know what values are
75 available in what blocks. AVAIL is a forward dataflow problem. In
76 SSA, values are never killed, so we don't need a kill set, or a
77 fixpoint iteration, in order to calculate the AVAIL sets. In
78 traditional parlance, AVAIL sets tell us the downsafety of the
81 Next, we generate the ANTIC sets. These sets represent the
82 anticipatable expressions. ANTIC is a backwards dataflow
83 problem. An expression is anticipatable in a given block if it could
84 be generated in that block. This means that if we had to perform
85 an insertion in that block, of the value of that expression, we
86 could. Calculating the ANTIC sets requires phi translation of
87 expressions, because the flow goes backwards through phis. We must
88 iterate to a fixpoint of the ANTIC sets, because we have a kill
89 set. Even in SSA form, values are not live over the entire
90 function, only from their definition point onwards. So we have to
91 remove values from the ANTIC set once we go past the definition
92 point of the leaders that make them up.
93 compute_antic/compute_antic_aux performs this computation.
95 Third, we perform insertions to make partially redundant
96 expressions fully redundant.
98 An expression is partially redundant (excluding partial
101 1. It is AVAIL in some, but not all, of the predecessors of a
103 2. It is ANTIC in all the predecessors.
105 In order to make it fully redundant, we insert the expression into
106 the predecessors where it is not available, but is ANTIC.
108 For the partial anticipation case, we only perform insertion if it
109 is partially anticipated in some block, and fully available in all
112 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
113 performs these steps.
115 Fourth, we eliminate fully redundant expressions.
116 This is a simple statement walk that replaces redundant
117 calculations with the now available values. */
119 /* Representations of value numbers:
121 Value numbers are represented by a representative SSA_NAME. We
122 will create fake SSA_NAME's in situations where we need a
123 representative but do not have one (because it is a complex
124 expression). In order to facilitate storing the value numbers in
125 bitmaps, and keep the number of wasted SSA_NAME's down, we also
126 associate a value_id with each value number, and create full blown
127 ssa_name's only where we actually need them (IE in operands of
128 existing expressions).
130 Theoretically you could replace all the value_id's with
131 SSA_NAME_VERSION, but this would allocate a large number of
132 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
133 It would also require an additional indirection at each point we
136 /* Representation of expressions on value numbers:
138 Expressions consisting of value numbers are represented the same
139 way as our VN internally represents them, with an additional
140 "pre_expr" wrapping around them in order to facilitate storing all
141 of the expressions in the same sets. */
143 /* Representation of sets:
145 The dataflow sets do not need to be sorted in any particular order
146 for the majority of their lifetime, are simply represented as two
147 bitmaps, one that keeps track of values present in the set, and one
148 that keeps track of expressions present in the set.
150 When we need them in topological order, we produce it on demand by
151 transforming the bitmap into an array and sorting it into topo
154 /* Type of expression, used to know which member of the PRE_EXPR union
165 typedef union pre_expr_union_d
170 vn_reference_t reference;
173 typedef struct pre_expr_d
175 enum pre_expr_kind kind;
180 #define PRE_EXPR_NAME(e) (e)->u.name
181 #define PRE_EXPR_NARY(e) (e)->u.nary
182 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
183 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
186 pre_expr_eq (const void *p1, const void *p2)
188 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
189 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
191 if (e1->kind != e2->kind)
197 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
198 PRE_EXPR_CONSTANT (e2));
200 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
202 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
204 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
205 PRE_EXPR_REFERENCE (e2));
212 pre_expr_hash (const void *p1)
214 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
218 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
220 return iterative_hash_hashval_t (SSA_NAME_VERSION (PRE_EXPR_NAME (e)), 0);
222 return PRE_EXPR_NARY (e)->hashcode;
224 return PRE_EXPR_REFERENCE (e)->hashcode;
231 /* Next global expression id number. */
232 static unsigned int next_expression_id;
234 /* Mapping from expression to id number we can use in bitmap sets. */
235 DEF_VEC_P (pre_expr);
236 DEF_VEC_ALLOC_P (pre_expr, heap);
237 static VEC(pre_expr, heap) *expressions;
238 static htab_t expression_to_id;
240 /* Allocate an expression id for EXPR. */
242 static inline unsigned int
243 alloc_expression_id (pre_expr expr)
246 /* Make sure we won't overflow. */
247 gcc_assert (next_expression_id + 1 > next_expression_id);
248 expr->id = next_expression_id++;
249 VEC_safe_push (pre_expr, heap, expressions, expr);
250 slot = htab_find_slot (expression_to_id, expr, INSERT);
253 return next_expression_id - 1;
256 /* Return the expression id for tree EXPR. */
258 static inline unsigned int
259 get_expression_id (const pre_expr expr)
264 static inline unsigned int
265 lookup_expression_id (const pre_expr expr)
269 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
272 return ((pre_expr)*slot)->id;
275 /* Return the existing expression id for EXPR, or create one if one
276 does not exist yet. */
278 static inline unsigned int
279 get_or_alloc_expression_id (pre_expr expr)
281 unsigned int id = lookup_expression_id (expr);
283 return alloc_expression_id (expr);
284 return expr->id = id;
287 /* Return the expression that has expression id ID */
289 static inline pre_expr
290 expression_for_id (unsigned int id)
292 return VEC_index (pre_expr, expressions, id);
295 /* Free the expression id field in all of our expressions,
296 and then destroy the expressions array. */
299 clear_expression_ids (void)
301 VEC_free (pre_expr, heap, expressions);
304 static alloc_pool pre_expr_pool;
306 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
309 get_or_alloc_expr_for_name (tree name)
311 pre_expr result = (pre_expr) pool_alloc (pre_expr_pool);
312 unsigned int result_id;
316 PRE_EXPR_NAME (result) = name;
317 result_id = lookup_expression_id (result);
320 pool_free (pre_expr_pool, result);
321 result = expression_for_id (result_id);
324 get_or_alloc_expression_id (result);
328 static bool in_fre = false;
330 /* An unordered bitmap set. One bitmap tracks values, the other,
332 typedef struct bitmap_set
338 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
339 EXECUTE_IF_SET_IN_BITMAP((set)->expressions, 0, (id), (bi))
341 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
342 EXECUTE_IF_SET_IN_BITMAP((set)->values, 0, (id), (bi))
344 /* Mapping from value id to expressions with that value_id. */
345 DEF_VEC_P (bitmap_set_t);
346 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
347 static VEC(bitmap_set_t, heap) *value_expressions;
349 /* Sets that we need to keep track of. */
350 typedef struct bb_bitmap_sets
352 /* The EXP_GEN set, which represents expressions/values generated in
354 bitmap_set_t exp_gen;
356 /* The PHI_GEN set, which represents PHI results generated in a
358 bitmap_set_t phi_gen;
360 /* The TMP_GEN set, which represents results/temporaries generated
361 in a basic block. IE the LHS of an expression. */
362 bitmap_set_t tmp_gen;
364 /* The AVAIL_OUT set, which represents which values are available in
365 a given basic block. */
366 bitmap_set_t avail_out;
368 /* The ANTIC_IN set, which represents which values are anticipatable
369 in a given basic block. */
370 bitmap_set_t antic_in;
372 /* The PA_IN set, which represents which values are
373 partially anticipatable in a given basic block. */
376 /* The NEW_SETS set, which is used during insertion to augment the
377 AVAIL_OUT set of blocks with the new insertions performed during
378 the current iteration. */
379 bitmap_set_t new_sets;
381 /* A cache for value_dies_in_block_x. */
384 /* True if we have visited this block during ANTIC calculation. */
385 unsigned int visited:1;
387 /* True we have deferred processing this block during ANTIC
388 calculation until its successor is processed. */
389 unsigned int deferred : 1;
392 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
393 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
394 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
395 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
396 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
397 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
398 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
399 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
400 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
401 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
404 /* Basic block list in postorder. */
405 static int *postorder;
407 /* This structure is used to keep track of statistics on what
408 optimization PRE was able to perform. */
411 /* The number of RHS computations eliminated by PRE. */
414 /* The number of new expressions/temporaries generated by PRE. */
417 /* The number of inserts found due to partial anticipation */
420 /* The number of new PHI nodes added by PRE. */
423 /* The number of values found constant. */
428 static bool do_partial_partial;
429 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
430 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
431 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
432 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
433 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
434 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
435 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr, bool);
436 static bitmap_set_t bitmap_set_new (void);
437 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
439 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
441 static unsigned int get_expr_value_id (pre_expr);
443 /* We can add and remove elements and entries to and from sets
444 and hash tables, so we use alloc pools for them. */
446 static alloc_pool bitmap_set_pool;
447 static bitmap_obstack grand_bitmap_obstack;
449 /* To avoid adding 300 temporary variables when we only need one, we
450 only create one temporary variable, on demand, and build ssa names
451 off that. We do have to change the variable if the types don't
452 match the current variable's type. */
454 static tree storetemp;
455 static tree prephitemp;
457 /* Set of blocks with statements that have had its EH information
459 static bitmap need_eh_cleanup;
461 /* The phi_translate_table caches phi translations for a given
462 expression and predecessor. */
464 static htab_t phi_translate_table;
466 /* A three tuple {e, pred, v} used to cache phi translations in the
467 phi_translate_table. */
469 typedef struct expr_pred_trans_d
471 /* The expression. */
474 /* The predecessor block along which we translated the expression. */
477 /* The value that resulted from the translation. */
480 /* The hashcode for the expression, pred pair. This is cached for
483 } *expr_pred_trans_t;
484 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
486 /* Return the hash value for a phi translation table entry. */
489 expr_pred_trans_hash (const void *p)
491 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
495 /* Return true if two phi translation table entries are the same.
496 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
499 expr_pred_trans_eq (const void *p1, const void *p2)
501 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
502 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
503 basic_block b1 = ve1->pred;
504 basic_block b2 = ve2->pred;
506 /* If they are not translations for the same basic block, they can't
510 return pre_expr_eq (ve1->e, ve2->e);
513 /* Search in the phi translation table for the translation of
514 expression E in basic block PRED.
515 Return the translated value, if found, NULL otherwise. */
517 static inline pre_expr
518 phi_trans_lookup (pre_expr e, basic_block pred)
521 struct expr_pred_trans_d ept;
525 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
526 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
531 return ((expr_pred_trans_t) *slot)->v;
535 /* Add the tuple mapping from {expression E, basic block PRED} to
536 value V, to the phi translation table. */
539 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
542 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
544 new_pair->pred = pred;
546 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
549 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
550 new_pair->hashcode, INSERT);
553 *slot = (void *) new_pair;
557 /* Add expression E to the expression set of value id V. */
560 add_to_value (unsigned int v, pre_expr e)
564 gcc_assert (get_expr_value_id (e) == v);
566 if (v >= VEC_length (bitmap_set_t, value_expressions))
568 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
572 set = VEC_index (bitmap_set_t, value_expressions, v);
575 set = bitmap_set_new ();
576 VEC_replace (bitmap_set_t, value_expressions, v, set);
579 bitmap_insert_into_set_1 (set, e, true);
582 /* Create a new bitmap set and return it. */
585 bitmap_set_new (void)
587 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
588 ret->expressions = BITMAP_ALLOC (&grand_bitmap_obstack);
589 ret->values = BITMAP_ALLOC (&grand_bitmap_obstack);
593 /* Return the value id for a PRE expression EXPR. */
596 get_expr_value_id (pre_expr expr)
603 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
606 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
607 add_to_value (id, expr);
612 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
614 return PRE_EXPR_NARY (expr)->value_id;
616 return PRE_EXPR_REFERENCE (expr)->value_id;
622 /* Remove an expression EXPR from a bitmapped set. */
625 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
627 unsigned int val = get_expr_value_id (expr);
628 if (!value_id_constant_p (val))
630 bitmap_clear_bit (set->values, val);
631 bitmap_clear_bit (set->expressions, get_expression_id (expr));
636 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
637 bool allow_constants)
639 unsigned int val = get_expr_value_id (expr);
640 if (allow_constants || !value_id_constant_p (val))
642 /* We specifically expect this and only this function to be able to
643 insert constants into a set. */
644 bitmap_set_bit (set->values, val);
645 bitmap_set_bit (set->expressions, get_or_alloc_expression_id (expr));
649 /* Insert an expression EXPR into a bitmapped set. */
652 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
654 bitmap_insert_into_set_1 (set, expr, false);
657 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
660 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
662 bitmap_copy (dest->expressions, orig->expressions);
663 bitmap_copy (dest->values, orig->values);
667 /* Free memory used up by SET. */
669 bitmap_set_free (bitmap_set_t set)
671 BITMAP_FREE (set->expressions);
672 BITMAP_FREE (set->values);
676 /* Generate an topological-ordered array of bitmap set SET. */
678 static VEC(pre_expr, heap) *
679 sorted_array_from_bitmap_set (bitmap_set_t set)
682 bitmap_iterator bi, bj;
683 VEC(pre_expr, heap) *result = NULL;
685 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
687 /* The number of expressions having a given value is usually
688 relatively small. Thus, rather than making a vector of all
689 the expressions and sorting it by value-id, we walk the values
690 and check in the reverse mapping that tells us what expressions
691 have a given value, to filter those in our set. As a result,
692 the expressions are inserted in value-id order, which means
695 If this is somehow a significant lose for some cases, we can
696 choose which set to walk based on the set size. */
697 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
698 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
700 if (bitmap_bit_p (set->expressions, j))
701 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
708 /* Perform bitmapped set operation DEST &= ORIG. */
711 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
718 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
720 bitmap_and_into (dest->values, orig->values);
721 bitmap_copy (temp, dest->expressions);
722 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
724 pre_expr expr = expression_for_id (i);
725 unsigned int value_id = get_expr_value_id (expr);
726 if (!bitmap_bit_p (dest->values, value_id))
727 bitmap_clear_bit (dest->expressions, i);
733 /* Subtract all values and expressions contained in ORIG from DEST. */
736 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
738 bitmap_set_t result = bitmap_set_new ();
742 bitmap_and_compl (result->expressions, dest->expressions,
745 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
747 pre_expr expr = expression_for_id (i);
748 unsigned int value_id = get_expr_value_id (expr);
749 bitmap_set_bit (result->values, value_id);
755 /* Subtract all the values in bitmap set B from bitmap set A. */
758 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
762 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
764 bitmap_copy (temp, a->expressions);
765 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
767 pre_expr expr = expression_for_id (i);
768 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
769 bitmap_remove_from_set (a, expr);
775 /* Return true if bitmapped set SET contains the value VALUE_ID. */
778 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
780 if (value_id_constant_p (value_id))
783 if (!set || bitmap_empty_p (set->expressions))
786 return bitmap_bit_p (set->values, value_id);
790 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
792 return bitmap_bit_p (set->expressions, get_expression_id (expr));
795 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
798 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
801 bitmap_set_t exprset;
805 if (value_id_constant_p (lookfor))
808 if (!bitmap_set_contains_value (set, lookfor))
811 /* The number of expressions having a given value is usually
812 significantly less than the total number of expressions in SET.
813 Thus, rather than check, for each expression in SET, whether it
814 has the value LOOKFOR, we walk the reverse mapping that tells us
815 what expressions have a given value, and see if any of those
816 expressions are in our set. For large testcases, this is about
817 5-10x faster than walking the bitmap. If this is somehow a
818 significant lose for some cases, we can choose which set to walk
819 based on the set size. */
820 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
821 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
823 if (bitmap_bit_p (set->expressions, i))
825 bitmap_clear_bit (set->expressions, i);
826 bitmap_set_bit (set->expressions, get_expression_id (expr));
832 /* Return true if two bitmap sets are equal. */
835 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
837 return bitmap_equal_p (a->values, b->values);
840 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
841 and add it otherwise. */
844 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
846 unsigned int val = get_expr_value_id (expr);
848 if (bitmap_set_contains_value (set, val))
849 bitmap_set_replace_value (set, val, expr);
851 bitmap_insert_into_set (set, expr);
854 /* Insert EXPR into SET if EXPR's value is not already present in
858 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
860 unsigned int val = get_expr_value_id (expr);
862 if (value_id_constant_p (val))
865 if (!bitmap_set_contains_value (set, val))
866 bitmap_insert_into_set (set, expr);
869 /* Print out EXPR to outfile. */
872 print_pre_expr (FILE *outfile, const pre_expr expr)
877 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
880 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
885 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
886 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
887 for (i = 0; i < nary->length; i++)
889 print_generic_expr (outfile, nary->op[i], 0);
890 if (i != (unsigned) nary->length - 1)
891 fprintf (outfile, ",");
893 fprintf (outfile, "}");
899 vn_reference_op_t vro;
901 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
902 fprintf (outfile, "{");
904 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
907 bool closebrace = false;
908 if (vro->opcode != SSA_NAME
909 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
911 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
914 fprintf (outfile, "<");
920 print_generic_expr (outfile, vro->op0, 0);
923 fprintf (outfile, ",");
924 print_generic_expr (outfile, vro->op1, 0);
928 fprintf (outfile, ",");
929 print_generic_expr (outfile, vro->op2, 0);
933 fprintf (outfile, ">");
934 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
935 fprintf (outfile, ",");
937 fprintf (outfile, "}");
940 fprintf (outfile, "@");
941 print_generic_expr (outfile, ref->vuse, 0);
947 void debug_pre_expr (pre_expr);
949 /* Like print_pre_expr but always prints to stderr. */
951 debug_pre_expr (pre_expr e)
953 print_pre_expr (stderr, e);
954 fprintf (stderr, "\n");
957 /* Print out SET to OUTFILE. */
960 print_bitmap_set (FILE *outfile, bitmap_set_t set,
961 const char *setname, int blockindex)
963 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
970 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
972 const pre_expr expr = expression_for_id (i);
975 fprintf (outfile, ", ");
977 print_pre_expr (outfile, expr);
979 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
982 fprintf (outfile, " }\n");
985 void debug_bitmap_set (bitmap_set_t);
988 debug_bitmap_set (bitmap_set_t set)
990 print_bitmap_set (stderr, set, "debug", 0);
993 /* Print out the expressions that have VAL to OUTFILE. */
996 print_value_expressions (FILE *outfile, unsigned int val)
998 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1002 sprintf (s, "%04d", val);
1003 print_bitmap_set (outfile, set, s, 0);
1009 debug_value_expressions (unsigned int val)
1011 print_value_expressions (stderr, val);
1014 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1018 get_or_alloc_expr_for_constant (tree constant)
1020 unsigned int result_id;
1021 unsigned int value_id;
1022 pre_expr newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1023 newexpr->kind = CONSTANT;
1024 PRE_EXPR_CONSTANT (newexpr) = constant;
1025 result_id = lookup_expression_id (newexpr);
1028 pool_free (pre_expr_pool, newexpr);
1029 newexpr = expression_for_id (result_id);
1032 value_id = get_or_alloc_constant_value_id (constant);
1033 get_or_alloc_expression_id (newexpr);
1034 add_to_value (value_id, newexpr);
1038 /* Given a value id V, find the actual tree representing the constant
1039 value if there is one, and return it. Return NULL if we can't find
1043 get_constant_for_value_id (unsigned int v)
1045 if (value_id_constant_p (v))
1049 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1051 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1053 pre_expr expr = expression_for_id (i);
1054 if (expr->kind == CONSTANT)
1055 return PRE_EXPR_CONSTANT (expr);
1061 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1062 Currently only supports constants and SSA_NAMES. */
1064 get_or_alloc_expr_for (tree t)
1066 if (TREE_CODE (t) == SSA_NAME)
1067 return get_or_alloc_expr_for_name (t);
1068 else if (is_gimple_min_invariant (t))
1069 return get_or_alloc_expr_for_constant (t);
1072 /* More complex expressions can result from SCCVN expression
1073 simplification that inserts values for them. As they all
1074 do not have VOPs the get handled by the nary ops struct. */
1075 vn_nary_op_t result;
1076 unsigned int result_id;
1077 vn_nary_op_lookup (t, &result);
1080 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1082 PRE_EXPR_NARY (e) = result;
1083 result_id = lookup_expression_id (e);
1086 pool_free (pre_expr_pool, e);
1087 e = expression_for_id (result_id);
1090 alloc_expression_id (e);
1097 /* Return the folded version of T if T, when folded, is a gimple
1098 min_invariant. Otherwise, return T. */
1101 fully_constant_expression (pre_expr e)
1109 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1110 switch (TREE_CODE_CLASS (nary->opcode))
1112 case tcc_expression:
1113 if (nary->opcode == TRUTH_NOT_EXPR)
1115 if (nary->opcode != TRUTH_AND_EXPR
1116 && nary->opcode != TRUTH_OR_EXPR
1117 && nary->opcode != TRUTH_XOR_EXPR)
1121 case tcc_comparison:
1123 /* We have to go from trees to pre exprs to value ids to
1125 tree naryop0 = nary->op[0];
1126 tree naryop1 = nary->op[1];
1128 if (!is_gimple_min_invariant (naryop0))
1130 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1131 unsigned int vrep0 = get_expr_value_id (rep0);
1132 tree const0 = get_constant_for_value_id (vrep0);
1134 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1136 if (!is_gimple_min_invariant (naryop1))
1138 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1139 unsigned int vrep1 = get_expr_value_id (rep1);
1140 tree const1 = get_constant_for_value_id (vrep1);
1142 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1144 result = fold_binary (nary->opcode, nary->type,
1146 if (result && is_gimple_min_invariant (result))
1147 return get_or_alloc_expr_for_constant (result);
1148 /* We might have simplified the expression to a
1149 SSA_NAME for example from x_1 * 1. But we cannot
1150 insert a PHI for x_1 unconditionally as x_1 might
1151 not be available readily. */
1155 if (nary->opcode != REALPART_EXPR
1156 && nary->opcode != IMAGPART_EXPR
1157 && nary->opcode != VIEW_CONVERT_EXPR)
1163 /* We have to go from trees to pre exprs to value ids to
1165 tree naryop0 = nary->op[0];
1166 tree const0, result;
1167 if (is_gimple_min_invariant (naryop0))
1171 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1172 unsigned int vrep0 = get_expr_value_id (rep0);
1173 const0 = get_constant_for_value_id (vrep0);
1178 tree type1 = TREE_TYPE (nary->op[0]);
1179 const0 = fold_convert (type1, const0);
1180 result = fold_unary (nary->opcode, nary->type, const0);
1182 if (result && is_gimple_min_invariant (result))
1183 return get_or_alloc_expr_for_constant (result);
1192 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1193 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1194 vn_reference_op_t op;
1196 /* Try to simplify the translated expression if it is
1197 a call to a builtin function with at most two arguments. */
1198 op = VEC_index (vn_reference_op_s, operands, 0);
1199 if (op->opcode == CALL_EXPR
1200 && TREE_CODE (op->op0) == ADDR_EXPR
1201 && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL
1202 && DECL_BUILT_IN (TREE_OPERAND (op->op0, 0))
1203 && VEC_length (vn_reference_op_s, operands) >= 2
1204 && VEC_length (vn_reference_op_s, operands) <= 3)
1206 vn_reference_op_t arg0, arg1 = NULL;
1207 bool anyconst = false;
1208 arg0 = VEC_index (vn_reference_op_s, operands, 1);
1209 if (VEC_length (vn_reference_op_s, operands) > 2)
1210 arg1 = VEC_index (vn_reference_op_s, operands, 2);
1211 if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant
1212 || (arg0->opcode == ADDR_EXPR
1213 && is_gimple_min_invariant (arg0->op0)))
1216 && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant
1217 || (arg1->opcode == ADDR_EXPR
1218 && is_gimple_min_invariant (arg1->op0))))
1222 tree folded = build_call_expr (TREE_OPERAND (op->op0, 0),
1225 arg1 ? arg1->op0 : NULL);
1227 && TREE_CODE (folded) == NOP_EXPR)
1228 folded = TREE_OPERAND (folded, 0);
1230 && is_gimple_min_invariant (folded))
1231 return get_or_alloc_expr_for_constant (folded);
1242 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1243 it has the value it would have in BLOCK. Set *SAME_VALID to true
1244 in case the new vuse doesn't change the value id of the OPERANDS. */
1247 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1248 alias_set_type set, tree type, tree vuse,
1249 basic_block phiblock,
1250 basic_block block, bool *same_valid)
1252 gimple phi = SSA_NAME_DEF_STMT (vuse);
1259 if (gimple_bb (phi) != phiblock)
1262 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1264 /* Use the alias-oracle to find either the PHI node in this block,
1265 the first VUSE used in this block that is equivalent to vuse or
1266 the first VUSE which definition in this block kills the value. */
1267 if (gimple_code (phi) == GIMPLE_PHI)
1268 e = find_edge (block, phiblock);
1269 else if (use_oracle)
1270 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1272 vuse = gimple_vuse (phi);
1273 phi = SSA_NAME_DEF_STMT (vuse);
1274 if (gimple_bb (phi) != phiblock)
1276 if (gimple_code (phi) == GIMPLE_PHI)
1278 e = find_edge (block, phiblock);
1289 bitmap visited = NULL;
1290 /* Try to find a vuse that dominates this phi node by skipping
1291 non-clobbering statements. */
1292 vuse = get_continuation_for_phi (phi, &ref, &visited);
1294 BITMAP_FREE (visited);
1300 /* If we didn't find any, the value ID can't stay the same,
1301 but return the translated vuse. */
1302 *same_valid = false;
1303 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1305 /* ??? We would like to return vuse here as this is the canonical
1306 upmost vdef that this reference is associated with. But during
1307 insertion of the references into the hash tables we only ever
1308 directly insert with their direct gimple_vuse, hence returning
1309 something else would make us not find the other expression. */
1310 return PHI_ARG_DEF (phi, e->dest_idx);
1316 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1317 SET2. This is used to avoid making a set consisting of the union
1318 of PA_IN and ANTIC_IN during insert. */
1320 static inline pre_expr
1321 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1325 result = bitmap_find_leader (set1, val, NULL);
1326 if (!result && set2)
1327 result = bitmap_find_leader (set2, val, NULL);
1331 /* Get the tree type for our PRE expression e. */
1334 get_expr_type (const pre_expr e)
1339 return TREE_TYPE (PRE_EXPR_NAME (e));
1341 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1343 return PRE_EXPR_REFERENCE (e)->type;
1345 return PRE_EXPR_NARY (e)->type;
1350 /* Get a representative SSA_NAME for a given expression.
1351 Since all of our sub-expressions are treated as values, we require
1352 them to be SSA_NAME's for simplicity.
1353 Prior versions of GVNPRE used to use "value handles" here, so that
1354 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1355 either case, the operands are really values (IE we do not expect
1356 them to be usable without finding leaders). */
1359 get_representative_for (const pre_expr e)
1363 unsigned int value_id = get_expr_value_id (e);
1368 return PRE_EXPR_NAME (e);
1370 return PRE_EXPR_CONSTANT (e);
1374 /* Go through all of the expressions representing this value
1375 and pick out an SSA_NAME. */
1378 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1380 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1382 pre_expr rep = expression_for_id (i);
1383 if (rep->kind == NAME)
1384 return PRE_EXPR_NAME (rep);
1389 /* If we reached here we couldn't find an SSA_NAME. This can
1390 happen when we've discovered a value that has never appeared in
1391 the program as set to an SSA_NAME, most likely as the result of
1396 "Could not find SSA_NAME representative for expression:");
1397 print_pre_expr (dump_file, e);
1398 fprintf (dump_file, "\n");
1401 exprtype = get_expr_type (e);
1403 /* Build and insert the assignment of the end result to the temporary
1404 that we will return. */
1405 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1407 pretemp = create_tmp_var (exprtype, "pretmp");
1408 get_var_ann (pretemp);
1411 name = make_ssa_name (pretemp, gimple_build_nop ());
1412 VN_INFO_GET (name)->value_id = value_id;
1413 if (e->kind == CONSTANT)
1414 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1416 VN_INFO (name)->valnum = name;
1418 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1421 fprintf (dump_file, "Created SSA_NAME representative ");
1422 print_generic_expr (dump_file, name, 0);
1423 fprintf (dump_file, " for expression:");
1424 print_pre_expr (dump_file, e);
1425 fprintf (dump_file, "\n");
1434 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1435 the phis in PRED. Return NULL if we can't find a leader for each part
1436 of the translated expression. */
1439 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1440 basic_block pred, basic_block phiblock)
1442 pre_expr oldexpr = expr;
1448 if (value_id_constant_p (get_expr_value_id (expr)))
1451 phitrans = phi_trans_lookup (expr, pred);
1457 /* Constants contain no values that need translation. */
1464 bool changed = false;
1465 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1466 struct vn_nary_op_s newnary;
1467 /* The NARY structure is only guaranteed to have been
1468 allocated to the nary->length operands. */
1469 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1470 - sizeof (tree) * (4 - nary->length)));
1472 for (i = 0; i < newnary.length; i++)
1474 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1478 pre_expr leader, result;
1479 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1480 leader = find_leader_in_sets (op_val_id, set1, set2);
1481 result = phi_translate (leader, set1, set2, pred, phiblock);
1482 if (result && result != leader)
1484 tree name = get_representative_for (result);
1487 newnary.op[i] = name;
1492 changed |= newnary.op[i] != nary->op[i];
1499 tree result = vn_nary_op_lookup_pieces (newnary.length,
1507 unsigned int new_val_id;
1509 expr = (pre_expr) pool_alloc (pre_expr_pool);
1512 if (result && is_gimple_min_invariant (result))
1513 return get_or_alloc_expr_for_constant (result);
1518 PRE_EXPR_NARY (expr) = nary;
1519 constant = fully_constant_expression (expr);
1520 if (constant != expr)
1523 new_val_id = nary->value_id;
1524 get_or_alloc_expression_id (expr);
1528 new_val_id = get_next_value_id ();
1529 VEC_safe_grow_cleared (bitmap_set_t, heap,
1531 get_max_value_id() + 1);
1532 nary = vn_nary_op_insert_pieces (newnary.length,
1539 result, new_val_id);
1540 PRE_EXPR_NARY (expr) = nary;
1541 constant = fully_constant_expression (expr);
1542 if (constant != expr)
1544 get_or_alloc_expression_id (expr);
1546 add_to_value (new_val_id, expr);
1548 phi_trans_add (oldexpr, expr, pred);
1555 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1556 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1557 tree vuse = ref->vuse;
1558 tree newvuse = vuse;
1559 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1560 bool changed = false, same_valid = true;
1562 vn_reference_op_t operand;
1563 vn_reference_t newref;
1566 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1570 tree oldop0 = operand->op0;
1571 tree oldop1 = operand->op1;
1572 tree oldop2 = operand->op2;
1576 tree type = operand->type;
1577 vn_reference_op_s newop = *operand;
1579 if (op0 && TREE_CODE (op0) == SSA_NAME)
1581 unsigned int op_val_id = VN_INFO (op0)->value_id;
1582 leader = find_leader_in_sets (op_val_id, set1, set2);
1583 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1584 if (opresult && opresult != leader)
1586 tree name = get_representative_for (opresult);
1594 changed |= op0 != oldop0;
1596 if (op1 && TREE_CODE (op1) == SSA_NAME)
1598 unsigned int op_val_id = VN_INFO (op1)->value_id;
1599 leader = find_leader_in_sets (op_val_id, set1, set2);
1600 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1601 if (opresult && opresult != leader)
1603 tree name = get_representative_for (opresult);
1611 /* We can't possibly insert these. */
1612 else if (op1 && !is_gimple_min_invariant (op1))
1614 changed |= op1 != oldop1;
1615 if (op2 && TREE_CODE (op2) == SSA_NAME)
1617 unsigned int op_val_id = VN_INFO (op2)->value_id;
1618 leader = find_leader_in_sets (op_val_id, set1, set2);
1619 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1620 if (opresult && opresult != leader)
1622 tree name = get_representative_for (opresult);
1630 /* We can't possibly insert these. */
1631 else if (op2 && !is_gimple_min_invariant (op2))
1633 changed |= op2 != oldop2;
1636 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1637 /* We may have changed from an SSA_NAME to a constant */
1638 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1639 newop.opcode = TREE_CODE (op0);
1644 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1645 /* If it transforms from an SSA_NAME to an address, fold with
1646 a preceding indirect reference. */
1647 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1648 && VEC_index (vn_reference_op_s,
1649 newoperands, j - 1)->opcode == INDIRECT_REF)
1650 vn_reference_fold_indirect (&newoperands, &j);
1652 if (i != VEC_length (vn_reference_op_s, operands))
1655 VEC_free (vn_reference_op_s, heap, newoperands);
1661 newvuse = translate_vuse_through_block (newoperands,
1662 ref->set, ref->type,
1663 vuse, phiblock, pred,
1665 if (newvuse == NULL_TREE)
1667 VEC_free (vn_reference_op_s, heap, newoperands);
1672 if (changed || newvuse != vuse)
1674 unsigned int new_val_id;
1677 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1682 VEC_free (vn_reference_op_s, heap, newoperands);
1684 if (result && is_gimple_min_invariant (result))
1686 gcc_assert (!newoperands);
1687 return get_or_alloc_expr_for_constant (result);
1690 expr = (pre_expr) pool_alloc (pre_expr_pool);
1691 expr->kind = REFERENCE;
1696 PRE_EXPR_REFERENCE (expr) = newref;
1697 constant = fully_constant_expression (expr);
1698 if (constant != expr)
1701 new_val_id = newref->value_id;
1702 get_or_alloc_expression_id (expr);
1706 if (changed || !same_valid)
1708 new_val_id = get_next_value_id ();
1709 VEC_safe_grow_cleared (bitmap_set_t, heap,
1711 get_max_value_id() + 1);
1714 new_val_id = ref->value_id;
1715 newref = vn_reference_insert_pieces (newvuse, ref->set,
1718 result, new_val_id);
1720 PRE_EXPR_REFERENCE (expr) = newref;
1721 constant = fully_constant_expression (expr);
1722 if (constant != expr)
1724 get_or_alloc_expression_id (expr);
1726 add_to_value (new_val_id, expr);
1728 VEC_free (vn_reference_op_s, heap, newoperands);
1729 phi_trans_add (oldexpr, expr, pred);
1739 tree name = PRE_EXPR_NAME (expr);
1741 def_stmt = SSA_NAME_DEF_STMT (name);
1742 if (gimple_code (def_stmt) == GIMPLE_PHI
1743 && gimple_bb (def_stmt) == phiblock)
1748 e = find_edge (pred, gimple_bb (phi));
1751 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1754 if (TREE_CODE (def) == SSA_NAME)
1755 def = VN_INFO (def)->valnum;
1757 /* Handle constant. */
1758 if (is_gimple_min_invariant (def))
1759 return get_or_alloc_expr_for_constant (def);
1761 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1764 newexpr = get_or_alloc_expr_for_name (def);
1775 /* For each expression in SET, translate the values through phi nodes
1776 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1777 expressions in DEST. */
1780 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1781 basic_block phiblock)
1783 VEC (pre_expr, heap) *exprs;
1787 if (!phi_nodes (phiblock))
1789 bitmap_set_copy (dest, set);
1793 exprs = sorted_array_from_bitmap_set (set);
1794 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1796 pre_expr translated;
1797 translated = phi_translate (expr, set, NULL, pred, phiblock);
1799 /* Don't add empty translations to the cache */
1801 phi_trans_add (expr, translated, pred);
1803 if (translated != NULL)
1804 bitmap_value_insert_into_set (dest, translated);
1806 VEC_free (pre_expr, heap, exprs);
1809 /* Find the leader for a value (i.e., the name representing that
1810 value) in a given set, and return it. If STMT is non-NULL it
1811 makes sure the defining statement for the leader dominates it.
1812 Return NULL if no leader is found. */
1815 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1817 if (value_id_constant_p (val))
1821 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1823 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1825 pre_expr expr = expression_for_id (i);
1826 if (expr->kind == CONSTANT)
1830 if (bitmap_set_contains_value (set, val))
1832 /* Rather than walk the entire bitmap of expressions, and see
1833 whether any of them has the value we are looking for, we look
1834 at the reverse mapping, which tells us the set of expressions
1835 that have a given value (IE value->expressions with that
1836 value) and see if any of those expressions are in our set.
1837 The number of expressions per value is usually significantly
1838 less than the number of expressions in the set. In fact, for
1839 large testcases, doing it this way is roughly 5-10x faster
1840 than walking the bitmap.
1841 If this is somehow a significant lose for some cases, we can
1842 choose which set to walk based on which set is smaller. */
1845 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1847 EXECUTE_IF_AND_IN_BITMAP (exprset->expressions,
1848 set->expressions, 0, i, bi)
1850 pre_expr val = expression_for_id (i);
1851 /* At the point where stmt is not null, there should always
1852 be an SSA_NAME first in the list of expressions. */
1855 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1856 if (gimple_code (def_stmt) != GIMPLE_PHI
1857 && gimple_bb (def_stmt) == gimple_bb (stmt)
1858 && gimple_uid (def_stmt) >= gimple_uid (stmt))
1867 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1868 BLOCK by seeing if it is not killed in the block. Note that we are
1869 only determining whether there is a store that kills it. Because
1870 of the order in which clean iterates over values, we are guaranteed
1871 that altered operands will have caused us to be eliminated from the
1872 ANTIC_IN set already. */
1875 value_dies_in_block_x (pre_expr expr, basic_block block)
1877 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1878 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1880 gimple_stmt_iterator gsi;
1881 unsigned id = get_expression_id (expr);
1888 /* Lookup a previously calculated result. */
1889 if (EXPR_DIES (block)
1890 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1891 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1893 /* A memory expression {e, VUSE} dies in the block if there is a
1894 statement that may clobber e. If, starting statement walk from the
1895 top of the basic block, a statement uses VUSE there can be no kill
1896 inbetween that use and the original statement that loaded {e, VUSE},
1897 so we can stop walking. */
1898 ref.base = NULL_TREE;
1899 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1901 tree def_vuse, def_vdef;
1902 def = gsi_stmt (gsi);
1903 def_vuse = gimple_vuse (def);
1904 def_vdef = gimple_vdef (def);
1906 /* Not a memory statement. */
1910 /* Not a may-def. */
1913 /* A load with the same VUSE, we're done. */
1914 if (def_vuse == vuse)
1920 /* Init ref only if we really need it. */
1921 if (ref.base == NULL_TREE
1922 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1928 /* If the statement may clobber expr, it dies. */
1929 if (stmt_may_clobber_ref_p_1 (def, &ref))
1936 /* Remember the result. */
1937 if (!EXPR_DIES (block))
1938 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1939 bitmap_set_bit (EXPR_DIES (block), id * 2);
1941 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1947 #define union_contains_value(SET1, SET2, VAL) \
1948 (bitmap_set_contains_value ((SET1), (VAL)) \
1949 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
1951 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
1954 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
1955 vn_reference_op_t vro)
1957 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
1959 struct pre_expr_d temp;
1962 PRE_EXPR_NAME (&temp) = vro->op0;
1963 temp.id = lookup_expression_id (&temp);
1966 if (!union_contains_value (set1, set2,
1967 get_expr_value_id (&temp)))
1970 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
1972 struct pre_expr_d temp;
1975 PRE_EXPR_NAME (&temp) = vro->op1;
1976 temp.id = lookup_expression_id (&temp);
1979 if (!union_contains_value (set1, set2,
1980 get_expr_value_id (&temp)))
1984 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
1986 struct pre_expr_d temp;
1989 PRE_EXPR_NAME (&temp) = vro->op2;
1990 temp.id = lookup_expression_id (&temp);
1993 if (!union_contains_value (set1, set2,
1994 get_expr_value_id (&temp)))
2001 /* Determine if the expression EXPR is valid in SET1 U SET2.
2002 ONLY SET2 CAN BE NULL.
2003 This means that we have a leader for each part of the expression
2004 (if it consists of values), or the expression is an SSA_NAME.
2005 For loads/calls, we also see if the vuse is killed in this block. */
2008 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2014 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2018 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2019 for (i = 0; i < nary->length; i++)
2021 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2023 struct pre_expr_d temp;
2026 PRE_EXPR_NAME (&temp) = nary->op[i];
2027 temp.id = lookup_expression_id (&temp);
2030 if (!union_contains_value (set1, set2,
2031 get_expr_value_id (&temp)))
2040 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2041 vn_reference_op_t vro;
2044 for (i = 0; VEC_iterate (vn_reference_op_s, ref->operands, i, vro); i++)
2046 if (!vro_valid_in_sets (set1, set2, vro))
2051 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2052 if (!gimple_nop_p (def_stmt)
2053 && gimple_bb (def_stmt) != block
2054 && !dominated_by_p (CDI_DOMINATORS,
2055 block, gimple_bb (def_stmt)))
2058 return !value_dies_in_block_x (expr, block);
2065 /* Clean the set of expressions that are no longer valid in SET1 or
2066 SET2. This means expressions that are made up of values we have no
2067 leaders for in SET1 or SET2. This version is used for partial
2068 anticipation, which means it is not valid in either ANTIC_IN or
2072 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2074 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2078 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2080 if (!valid_in_sets (set1, set2, expr, block))
2081 bitmap_remove_from_set (set1, expr);
2083 VEC_free (pre_expr, heap, exprs);
2086 /* Clean the set of expressions that are no longer valid in SET. This
2087 means expressions that are made up of values we have no leaders for
2091 clean (bitmap_set_t set, basic_block block)
2093 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2097 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2099 if (!valid_in_sets (set, NULL, expr, block))
2100 bitmap_remove_from_set (set, expr);
2102 VEC_free (pre_expr, heap, exprs);
2105 static sbitmap has_abnormal_preds;
2107 /* List of blocks that may have changed during ANTIC computation and
2108 thus need to be iterated over. */
2110 static sbitmap changed_blocks;
2112 /* Decide whether to defer a block for a later iteration, or PHI
2113 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2114 should defer the block, and true if we processed it. */
2117 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2118 basic_block block, basic_block phiblock)
2120 if (!BB_VISITED (phiblock))
2122 SET_BIT (changed_blocks, block->index);
2123 BB_VISITED (block) = 0;
2124 BB_DEFERRED (block) = 1;
2128 phi_translate_set (dest, source, block, phiblock);
2132 /* Compute the ANTIC set for BLOCK.
2134 If succs(BLOCK) > 1 then
2135 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2136 else if succs(BLOCK) == 1 then
2137 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2139 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2143 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2145 bool changed = false;
2146 bitmap_set_t S, old, ANTIC_OUT;
2152 old = ANTIC_OUT = S = NULL;
2153 BB_VISITED (block) = 1;
2155 /* If any edges from predecessors are abnormal, antic_in is empty,
2157 if (block_has_abnormal_pred_edge)
2158 goto maybe_dump_sets;
2160 old = ANTIC_IN (block);
2161 ANTIC_OUT = bitmap_set_new ();
2163 /* If the block has no successors, ANTIC_OUT is empty. */
2164 if (EDGE_COUNT (block->succs) == 0)
2166 /* If we have one successor, we could have some phi nodes to
2167 translate through. */
2168 else if (single_succ_p (block))
2170 basic_block succ_bb = single_succ (block);
2172 /* We trade iterations of the dataflow equations for having to
2173 phi translate the maximal set, which is incredibly slow
2174 (since the maximal set often has 300+ members, even when you
2175 have a small number of blocks).
2176 Basically, we defer the computation of ANTIC for this block
2177 until we have processed it's successor, which will inevitably
2178 have a *much* smaller set of values to phi translate once
2179 clean has been run on it.
2180 The cost of doing this is that we technically perform more
2181 iterations, however, they are lower cost iterations.
2183 Timings for PRE on tramp3d-v4:
2184 without maximal set fix: 11 seconds
2185 with maximal set fix/without deferring: 26 seconds
2186 with maximal set fix/with deferring: 11 seconds
2189 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2193 goto maybe_dump_sets;
2196 /* If we have multiple successors, we take the intersection of all of
2197 them. Note that in the case of loop exit phi nodes, we may have
2198 phis to translate through. */
2201 VEC(basic_block, heap) * worklist;
2203 basic_block bprime, first = NULL;
2205 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2206 FOR_EACH_EDGE (e, ei, block->succs)
2209 && BB_VISITED (e->dest))
2211 else if (BB_VISITED (e->dest))
2212 VEC_quick_push (basic_block, worklist, e->dest);
2215 /* Of multiple successors we have to have visited one already. */
2218 SET_BIT (changed_blocks, block->index);
2219 BB_VISITED (block) = 0;
2220 BB_DEFERRED (block) = 1;
2222 VEC_free (basic_block, heap, worklist);
2223 goto maybe_dump_sets;
2226 if (phi_nodes (first))
2227 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2229 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2231 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2233 if (phi_nodes (bprime))
2235 bitmap_set_t tmp = bitmap_set_new ();
2236 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2237 bitmap_set_and (ANTIC_OUT, tmp);
2238 bitmap_set_free (tmp);
2241 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2243 VEC_free (basic_block, heap, worklist);
2246 /* Generate ANTIC_OUT - TMP_GEN. */
2247 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2249 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2250 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2253 /* Then union in the ANTIC_OUT - TMP_GEN values,
2254 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2255 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2256 bitmap_value_insert_into_set (ANTIC_IN (block),
2257 expression_for_id (bii));
2259 clean (ANTIC_IN (block), block);
2261 /* !old->expressions can happen when we deferred a block. */
2262 if (!old->expressions || !bitmap_set_equal (old, ANTIC_IN (block)))
2265 SET_BIT (changed_blocks, block->index);
2266 FOR_EACH_EDGE (e, ei, block->preds)
2267 SET_BIT (changed_blocks, e->src->index);
2270 RESET_BIT (changed_blocks, block->index);
2273 if (dump_file && (dump_flags & TDF_DETAILS))
2275 if (!BB_DEFERRED (block) || BB_VISITED (block))
2278 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2280 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2284 print_bitmap_set (dump_file, S, "S", block->index);
2289 "Block %d was deferred for a future iteration.\n",
2294 bitmap_set_free (old);
2296 bitmap_set_free (S);
2298 bitmap_set_free (ANTIC_OUT);
2302 /* Compute PARTIAL_ANTIC for BLOCK.
2304 If succs(BLOCK) > 1 then
2305 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2306 in ANTIC_OUT for all succ(BLOCK)
2307 else if succs(BLOCK) == 1 then
2308 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2310 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2315 compute_partial_antic_aux (basic_block block,
2316 bool block_has_abnormal_pred_edge)
2318 bool changed = false;
2319 bitmap_set_t old_PA_IN;
2320 bitmap_set_t PA_OUT;
2323 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2325 old_PA_IN = PA_OUT = NULL;
2327 /* If any edges from predecessors are abnormal, antic_in is empty,
2329 if (block_has_abnormal_pred_edge)
2330 goto maybe_dump_sets;
2332 /* If there are too many partially anticipatable values in the
2333 block, phi_translate_set can take an exponential time: stop
2334 before the translation starts. */
2336 && single_succ_p (block)
2337 && bitmap_count_bits (PA_IN (single_succ (block))->values) > max_pa)
2338 goto maybe_dump_sets;
2340 old_PA_IN = PA_IN (block);
2341 PA_OUT = bitmap_set_new ();
2343 /* If the block has no successors, ANTIC_OUT is empty. */
2344 if (EDGE_COUNT (block->succs) == 0)
2346 /* If we have one successor, we could have some phi nodes to
2347 translate through. Note that we can't phi translate across DFS
2348 back edges in partial antic, because it uses a union operation on
2349 the successors. For recurrences like IV's, we will end up
2350 generating a new value in the set on each go around (i + 3 (VH.1)
2351 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2352 else if (single_succ_p (block))
2354 basic_block succ = single_succ (block);
2355 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2356 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2358 /* If we have multiple successors, we take the union of all of
2362 VEC(basic_block, heap) * worklist;
2366 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2367 FOR_EACH_EDGE (e, ei, block->succs)
2369 if (e->flags & EDGE_DFS_BACK)
2371 VEC_quick_push (basic_block, worklist, e->dest);
2373 if (VEC_length (basic_block, worklist) > 0)
2375 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2380 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2381 bitmap_value_insert_into_set (PA_OUT,
2382 expression_for_id (i));
2383 if (phi_nodes (bprime))
2385 bitmap_set_t pa_in = bitmap_set_new ();
2386 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2387 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2388 bitmap_value_insert_into_set (PA_OUT,
2389 expression_for_id (i));
2390 bitmap_set_free (pa_in);
2393 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2394 bitmap_value_insert_into_set (PA_OUT,
2395 expression_for_id (i));
2398 VEC_free (basic_block, heap, worklist);
2401 /* PA_IN starts with PA_OUT - TMP_GEN.
2402 Then we subtract things from ANTIC_IN. */
2403 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2405 /* For partial antic, we want to put back in the phi results, since
2406 we will properly avoid making them partially antic over backedges. */
2407 bitmap_ior_into (PA_IN (block)->values, PHI_GEN (block)->values);
2408 bitmap_ior_into (PA_IN (block)->expressions, PHI_GEN (block)->expressions);
2410 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2411 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2413 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2415 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2418 SET_BIT (changed_blocks, block->index);
2419 FOR_EACH_EDGE (e, ei, block->preds)
2420 SET_BIT (changed_blocks, e->src->index);
2423 RESET_BIT (changed_blocks, block->index);
2426 if (dump_file && (dump_flags & TDF_DETAILS))
2429 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2431 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2434 bitmap_set_free (old_PA_IN);
2436 bitmap_set_free (PA_OUT);
2440 /* Compute ANTIC and partial ANTIC sets. */
2443 compute_antic (void)
2445 bool changed = true;
2446 int num_iterations = 0;
2450 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2451 We pre-build the map of blocks with incoming abnormal edges here. */
2452 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2453 sbitmap_zero (has_abnormal_preds);
2460 FOR_EACH_EDGE (e, ei, block->preds)
2462 e->flags &= ~EDGE_DFS_BACK;
2463 if (e->flags & EDGE_ABNORMAL)
2465 SET_BIT (has_abnormal_preds, block->index);
2470 BB_VISITED (block) = 0;
2471 BB_DEFERRED (block) = 0;
2472 /* While we are here, give empty ANTIC_IN sets to each block. */
2473 ANTIC_IN (block) = bitmap_set_new ();
2474 PA_IN (block) = bitmap_set_new ();
2477 /* At the exit block we anticipate nothing. */
2478 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2479 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2480 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2482 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2483 sbitmap_ones (changed_blocks);
2486 if (dump_file && (dump_flags & TDF_DETAILS))
2487 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2490 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2492 if (TEST_BIT (changed_blocks, postorder[i]))
2494 basic_block block = BASIC_BLOCK (postorder[i]);
2495 changed |= compute_antic_aux (block,
2496 TEST_BIT (has_abnormal_preds,
2500 #ifdef ENABLE_CHECKING
2501 /* Theoretically possible, but *highly* unlikely. */
2502 gcc_assert (num_iterations < 500);
2506 statistics_histogram_event (cfun, "compute_antic iterations",
2509 if (do_partial_partial)
2511 sbitmap_ones (changed_blocks);
2512 mark_dfs_back_edges ();
2517 if (dump_file && (dump_flags & TDF_DETAILS))
2518 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2521 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2523 if (TEST_BIT (changed_blocks, postorder[i]))
2525 basic_block block = BASIC_BLOCK (postorder[i]);
2527 |= compute_partial_antic_aux (block,
2528 TEST_BIT (has_abnormal_preds,
2532 #ifdef ENABLE_CHECKING
2533 /* Theoretically possible, but *highly* unlikely. */
2534 gcc_assert (num_iterations < 500);
2537 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2540 sbitmap_free (has_abnormal_preds);
2541 sbitmap_free (changed_blocks);
2544 /* Return true if we can value number the call in STMT. This is true
2545 if we have a pure or constant call. */
2548 can_value_number_call (gimple stmt)
2550 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2555 /* Return true if OP is a tree which we can perform PRE on.
2556 This may not match the operations we can value number, but in
2557 a perfect world would. */
2560 can_PRE_operation (tree op)
2562 return UNARY_CLASS_P (op)
2563 || BINARY_CLASS_P (op)
2564 || COMPARISON_CLASS_P (op)
2565 || TREE_CODE (op) == INDIRECT_REF
2566 || TREE_CODE (op) == COMPONENT_REF
2567 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2568 || TREE_CODE (op) == CALL_EXPR
2569 || TREE_CODE (op) == ARRAY_REF;
2573 /* Inserted expressions are placed onto this worklist, which is used
2574 for performing quick dead code elimination of insertions we made
2575 that didn't turn out to be necessary. */
2576 static VEC(gimple,heap) *inserted_exprs;
2577 static bitmap inserted_phi_names;
2579 /* Pool allocated fake store expressions are placed onto this
2580 worklist, which, after performing dead code elimination, is walked
2581 to see which expressions need to be put into GC'able memory */
2582 static VEC(gimple, heap) *need_creation;
2584 /* The actual worker for create_component_ref_by_pieces. */
2587 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2588 unsigned int *operand, gimple_seq *stmts,
2591 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2595 switch (currop->opcode)
2599 tree folded, sc = currop->op1;
2600 unsigned int nargs = 0;
2601 tree *args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2602 ref->operands) - 1);
2603 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2605 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2610 folded = build_call_array (currop->type,
2611 TREE_CODE (currop->op0) == FUNCTION_DECL
2612 ? build_fold_addr_expr (currop->op0)
2618 pre_expr scexpr = get_or_alloc_expr_for (sc);
2619 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2622 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2627 case TARGET_MEM_REF:
2629 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2632 tree genop0 = NULL_TREE;
2633 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2639 op0expr = get_or_alloc_expr_for (currop->op0);
2640 genop0 = find_or_generate_expression (block, op0expr,
2645 if (DECL_P (baseop))
2646 return build6 (TARGET_MEM_REF, currop->type,
2648 genop0, currop->op1, currop->op2,
2649 unshare_expr (nextop->op1));
2651 return build6 (TARGET_MEM_REF, currop->type,
2653 genop0, currop->op1, currop->op2,
2654 unshare_expr (nextop->op1));
2660 gcc_assert (is_gimple_min_invariant (currop->op0));
2666 case VIEW_CONVERT_EXPR:
2669 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2674 folded = fold_build1 (currop->opcode, currop->type,
2679 case ALIGN_INDIRECT_REF:
2680 case MISALIGNED_INDIRECT_REF:
2684 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2689 genop1 = fold_convert (build_pointer_type (currop->type),
2692 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2693 folded = fold_build2 (currop->opcode, currop->type,
2694 genop1, currop->op1);
2696 folded = fold_build1 (currop->opcode, currop->type,
2704 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2706 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2707 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2713 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2716 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2719 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2724 /* For array ref vn_reference_op's, operand 1 of the array ref
2725 is op0 of the reference op and operand 3 of the array ref is
2727 case ARRAY_RANGE_REF:
2731 tree genop1 = currop->op0;
2733 tree genop2 = currop->op1;
2735 tree genop3 = currop->op2;
2737 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2741 op1expr = get_or_alloc_expr_for (genop1);
2742 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2747 op2expr = get_or_alloc_expr_for (genop2);
2748 genop2 = find_or_generate_expression (block, op2expr, stmts,
2755 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2756 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2757 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2758 op3expr = get_or_alloc_expr_for (genop3);
2759 genop3 = find_or_generate_expression (block, op3expr, stmts,
2764 return build4 (currop->opcode, currop->type, genop0, genop1,
2771 tree genop2 = currop->op1;
2773 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2777 /* op1 should be a FIELD_DECL, which are represented by
2782 op2expr = get_or_alloc_expr_for (genop2);
2783 genop2 = find_or_generate_expression (block, op2expr, stmts,
2789 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2795 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2796 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2817 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2818 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2819 trying to rename aggregates into ssa form directly, which is a no no.
2821 Thus, this routine doesn't create temporaries, it just builds a
2822 single access expression for the array, calling
2823 find_or_generate_expression to build the innermost pieces.
2825 This function is a subroutine of create_expression_by_pieces, and
2826 should not be called on it's own unless you really know what you
2830 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2831 gimple_seq *stmts, gimple domstmt)
2833 unsigned int op = 0;
2834 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2837 /* Find a leader for an expression, or generate one using
2838 create_expression_by_pieces if it's ANTIC but
2840 BLOCK is the basic_block we are looking for leaders in.
2841 EXPR is the expression to find a leader or generate for.
2842 STMTS is the statement list to put the inserted expressions on.
2843 Returns the SSA_NAME of the LHS of the generated expression or the
2845 DOMSTMT if non-NULL is a statement that should be dominated by
2846 all uses in the generated expression. If DOMSTMT is non-NULL this
2847 routine can fail and return NULL_TREE. Otherwise it will assert
2851 find_or_generate_expression (basic_block block, pre_expr expr,
2852 gimple_seq *stmts, gimple domstmt)
2854 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2855 get_expr_value_id (expr), domstmt);
2859 if (leader->kind == NAME)
2860 genop = PRE_EXPR_NAME (leader);
2861 else if (leader->kind == CONSTANT)
2862 genop = PRE_EXPR_CONSTANT (leader);
2865 /* If it's still NULL, it must be a complex expression, so generate
2866 it recursively. Not so for FRE though. */
2870 bitmap_set_t exprset;
2871 unsigned int lookfor = get_expr_value_id (expr);
2872 bool handled = false;
2876 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2877 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2879 pre_expr temp = expression_for_id (i);
2880 if (temp->kind != NAME)
2883 genop = create_expression_by_pieces (block, temp, stmts,
2885 get_expr_type (expr));
2889 if (!handled && domstmt)
2892 gcc_assert (handled);
2897 #define NECESSARY GF_PLF_1
2899 /* Create an expression in pieces, so that we can handle very complex
2900 expressions that may be ANTIC, but not necessary GIMPLE.
2901 BLOCK is the basic block the expression will be inserted into,
2902 EXPR is the expression to insert (in value form)
2903 STMTS is a statement list to append the necessary insertions into.
2905 This function will die if we hit some value that shouldn't be
2906 ANTIC but is (IE there is no leader for it, or its components).
2907 This function may also generate expressions that are themselves
2908 partially or fully redundant. Those that are will be either made
2909 fully redundant during the next iteration of insert (for partially
2910 redundant ones), or eliminated by eliminate (for fully redundant
2913 If DOMSTMT is non-NULL then we make sure that all uses in the
2914 expressions dominate that statement. In this case the function
2915 can return NULL_TREE to signal failure. */
2918 create_expression_by_pieces (basic_block block, pre_expr expr,
2919 gimple_seq *stmts, gimple domstmt, tree type)
2923 gimple_seq forced_stmts = NULL;
2924 unsigned int value_id;
2925 gimple_stmt_iterator gsi;
2926 tree exprtype = type ? type : get_expr_type (expr);
2932 /* We may hit the NAME/CONSTANT case if we have to convert types
2933 that value numbering saw through. */
2935 folded = PRE_EXPR_NAME (expr);
2938 folded = PRE_EXPR_CONSTANT (expr);
2942 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2943 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
2948 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2949 switch (nary->length)
2953 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2954 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
2955 tree genop1 = find_or_generate_expression (block, op1,
2957 tree genop2 = find_or_generate_expression (block, op2,
2959 if (!genop1 || !genop2)
2961 genop1 = fold_convert (TREE_TYPE (nary->op[0]),
2963 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
2964 may be a constant with the wrong type. */
2965 if (nary->opcode == POINTER_PLUS_EXPR)
2966 genop2 = fold_convert (sizetype, genop2);
2968 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
2970 folded = fold_build2 (nary->opcode, nary->type,
2976 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2977 tree genop1 = find_or_generate_expression (block, op1,
2981 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
2983 folded = fold_build1 (nary->opcode, nary->type,
2996 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2997 folded = fold_convert (exprtype, folded);
2999 /* Force the generated expression to be a sequence of GIMPLE
3001 We have to call unshare_expr because force_gimple_operand may
3002 modify the tree we pass to it. */
3003 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3006 /* If we have any intermediate expressions to the value sets, add them
3007 to the value sets and chain them in the instruction stream. */
3010 gsi = gsi_start (forced_stmts);
3011 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3013 gimple stmt = gsi_stmt (gsi);
3014 tree forcedname = gimple_get_lhs (stmt);
3017 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3018 if (TREE_CODE (forcedname) == SSA_NAME)
3020 VN_INFO_GET (forcedname)->valnum = forcedname;
3021 VN_INFO (forcedname)->value_id = get_next_value_id ();
3022 nameexpr = get_or_alloc_expr_for_name (forcedname);
3023 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3025 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3026 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3028 mark_symbols_for_renaming (stmt);
3030 gimple_seq_add_seq (stmts, forced_stmts);
3033 /* Build and insert the assignment of the end result to the temporary
3034 that we will return. */
3035 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3037 pretemp = create_tmp_var (exprtype, "pretmp");
3038 get_var_ann (pretemp);
3042 add_referenced_var (temp);
3044 if (TREE_CODE (exprtype) == COMPLEX_TYPE
3045 || TREE_CODE (exprtype) == VECTOR_TYPE)
3046 DECL_GIMPLE_REG_P (temp) = 1;
3048 newstmt = gimple_build_assign (temp, folded);
3049 name = make_ssa_name (temp, newstmt);
3050 gimple_assign_set_lhs (newstmt, name);
3051 gimple_set_plf (newstmt, NECESSARY, false);
3053 gimple_seq_add_stmt (stmts, newstmt);
3054 VEC_safe_push (gimple, heap, inserted_exprs, newstmt);
3056 /* All the symbols in NEWEXPR should be put into SSA form. */
3057 mark_symbols_for_renaming (newstmt);
3059 /* Add a value number to the temporary.
3060 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3061 we are creating the expression by pieces, and this particular piece of
3062 the expression may have been represented. There is no harm in replacing
3064 VN_INFO_GET (name)->valnum = name;
3065 value_id = get_expr_value_id (expr);
3066 VN_INFO (name)->value_id = value_id;
3067 nameexpr = get_or_alloc_expr_for_name (name);
3068 add_to_value (value_id, nameexpr);
3070 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3071 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3073 pre_stats.insertions++;
3074 if (dump_file && (dump_flags & TDF_DETAILS))
3076 fprintf (dump_file, "Inserted ");
3077 print_gimple_stmt (dump_file, newstmt, 0, 0);
3078 fprintf (dump_file, " in predecessor %d\n", block->index);
3085 /* Returns true if we want to inhibit the insertions of PHI nodes
3086 for the given EXPR for basic block BB (a member of a loop).
3087 We want to do this, when we fear that the induction variable we
3088 create might inhibit vectorization. */
3091 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3093 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3094 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3095 vn_reference_op_t op;
3098 /* If we aren't going to vectorize we don't inhibit anything. */
3099 if (!flag_tree_vectorize)
3102 /* Otherwise we inhibit the insertion when the address of the
3103 memory reference is a simple induction variable. In other
3104 cases the vectorizer won't do anything anyway (either it's
3105 loop invariant or a complicated expression). */
3106 for (i = 0; VEC_iterate (vn_reference_op_s, ops, i, op); ++i)
3111 case ARRAY_RANGE_REF:
3112 if (TREE_CODE (op->op0) != SSA_NAME)
3117 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3119 /* Default defs are loop invariant. */
3122 /* Defined outside this loop, also loop invariant. */
3123 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3125 /* If it's a simple induction variable inhibit insertion,
3126 the vectorizer might be interested in this one. */
3127 if (simple_iv (bb->loop_father, bb->loop_father,
3128 op->op0, &iv, true))
3130 /* No simple IV, vectorizer can't do anything, hence no
3131 reason to inhibit the transformation for this operand. */
3141 /* Insert the to-be-made-available values of expression EXPRNUM for each
3142 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3143 merge the result with a phi node, given the same value number as
3144 NODE. Return true if we have inserted new stuff. */
3147 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3150 pre_expr expr = expression_for_id (exprnum);
3152 unsigned int val = get_expr_value_id (expr);
3154 bool insertions = false;
3159 tree type = get_expr_type (expr);
3163 if (dump_file && (dump_flags & TDF_DETAILS))
3165 fprintf (dump_file, "Found partial redundancy for expression ");
3166 print_pre_expr (dump_file, expr);
3167 fprintf (dump_file, " (%04d)\n", val);
3170 /* Make sure we aren't creating an induction variable. */
3171 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3173 bool firstinsideloop = false;
3174 bool secondinsideloop = false;
3175 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3176 EDGE_PRED (block, 0)->src);
3177 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3178 EDGE_PRED (block, 1)->src);
3179 /* Induction variables only have one edge inside the loop. */
3180 if ((firstinsideloop ^ secondinsideloop)
3181 && (expr->kind != REFERENCE
3182 || inhibit_phi_insertion (block, expr)))
3184 if (dump_file && (dump_flags & TDF_DETAILS))
3185 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3190 /* Make sure we are not inserting trapping expressions. */
3191 FOR_EACH_EDGE (pred, ei, block->preds)
3194 eprime = avail[bprime->index];
3195 if (eprime->kind == NARY
3196 && vn_nary_may_trap (PRE_EXPR_NARY (eprime)))
3200 /* Make the necessary insertions. */
3201 FOR_EACH_EDGE (pred, ei, block->preds)
3203 gimple_seq stmts = NULL;
3206 eprime = avail[bprime->index];
3208 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3210 builtexpr = create_expression_by_pieces (bprime,
3214 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3215 gsi_insert_seq_on_edge (pred, stmts);
3216 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3219 else if (eprime->kind == CONSTANT)
3221 /* Constants may not have the right type, fold_convert
3222 should give us back a constant with the right type.
3224 tree constant = PRE_EXPR_CONSTANT (eprime);
3225 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3227 tree builtexpr = fold_convert (type, constant);
3228 if (!is_gimple_min_invariant (builtexpr))
3230 tree forcedexpr = force_gimple_operand (builtexpr,
3233 if (!is_gimple_min_invariant (forcedexpr))
3235 if (forcedexpr != builtexpr)
3237 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3238 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3242 gimple_stmt_iterator gsi;
3243 gsi = gsi_start (stmts);
3244 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3246 gimple stmt = gsi_stmt (gsi);
3247 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3248 gimple_set_plf (stmt, NECESSARY, false);
3250 gsi_insert_seq_on_edge (pred, stmts);
3252 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3257 else if (eprime->kind == NAME)
3259 /* We may have to do a conversion because our value
3260 numbering can look through types in certain cases, but
3261 our IL requires all operands of a phi node have the same
3263 tree name = PRE_EXPR_NAME (eprime);
3264 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3268 builtexpr = fold_convert (type, name);
3269 forcedexpr = force_gimple_operand (builtexpr,
3273 if (forcedexpr != name)
3275 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3276 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3281 gimple_stmt_iterator gsi;
3282 gsi = gsi_start (stmts);
3283 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3285 gimple stmt = gsi_stmt (gsi);
3286 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3287 gimple_set_plf (stmt, NECESSARY, false);
3289 gsi_insert_seq_on_edge (pred, stmts);
3291 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3295 /* If we didn't want a phi node, and we made insertions, we still have
3296 inserted new stuff, and thus return true. If we didn't want a phi node,
3297 and didn't make insertions, we haven't added anything new, so return
3299 if (nophi && insertions)
3301 else if (nophi && !insertions)
3304 /* Now build a phi for the new variable. */
3305 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3307 prephitemp = create_tmp_var (type, "prephitmp");
3308 get_var_ann (prephitemp);
3312 add_referenced_var (temp);
3314 if (TREE_CODE (type) == COMPLEX_TYPE
3315 || TREE_CODE (type) == VECTOR_TYPE)
3316 DECL_GIMPLE_REG_P (temp) = 1;
3317 phi = create_phi_node (temp, block);
3319 gimple_set_plf (phi, NECESSARY, false);
3320 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3321 VN_INFO (gimple_phi_result (phi))->value_id = val;
3322 VEC_safe_push (gimple, heap, inserted_exprs, phi);
3323 bitmap_set_bit (inserted_phi_names,
3324 SSA_NAME_VERSION (gimple_phi_result (phi)));
3325 FOR_EACH_EDGE (pred, ei, block->preds)
3327 pre_expr ae = avail[pred->src->index];
3328 gcc_assert (get_expr_type (ae) == type
3329 || useless_type_conversion_p (type, get_expr_type (ae)));
3330 if (ae->kind == CONSTANT)
3331 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3333 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3337 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3338 add_to_value (val, newphi);
3340 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3341 this insertion, since we test for the existence of this value in PHI_GEN
3342 before proceeding with the partial redundancy checks in insert_aux.
3344 The value may exist in AVAIL_OUT, in particular, it could be represented
3345 by the expression we are trying to eliminate, in which case we want the
3346 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3349 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3350 this block, because if it did, it would have existed in our dominator's
3351 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3354 bitmap_insert_into_set (PHI_GEN (block), newphi);
3355 bitmap_value_replace_in_set (AVAIL_OUT (block),
3357 bitmap_insert_into_set (NEW_SETS (block),
3360 if (dump_file && (dump_flags & TDF_DETAILS))
3362 fprintf (dump_file, "Created phi ");
3363 print_gimple_stmt (dump_file, phi, 0, 0);
3364 fprintf (dump_file, " in block %d\n", block->index);
3372 /* Perform insertion of partially redundant values.
3373 For BLOCK, do the following:
3374 1. Propagate the NEW_SETS of the dominator into the current block.
3375 If the block has multiple predecessors,
3376 2a. Iterate over the ANTIC expressions for the block to see if
3377 any of them are partially redundant.
3378 2b. If so, insert them into the necessary predecessors to make
3379 the expression fully redundant.
3380 2c. Insert a new PHI merging the values of the predecessors.
3381 2d. Insert the new PHI, and the new expressions, into the
3383 3. Recursively call ourselves on the dominator children of BLOCK.
3385 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3386 do_regular_insertion and do_partial_insertion.
3391 do_regular_insertion (basic_block block, basic_block dom)
3393 bool new_stuff = false;
3394 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3398 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3400 if (expr->kind != NAME)
3404 bool by_some = false;
3405 bool cant_insert = false;
3406 bool all_same = true;
3407 pre_expr first_s = NULL;
3410 pre_expr eprime = NULL;
3412 pre_expr edoubleprime = NULL;
3413 bool do_insertion = false;
3415 val = get_expr_value_id (expr);
3416 if (bitmap_set_contains_value (PHI_GEN (block), val))
3418 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3420 if (dump_file && (dump_flags & TDF_DETAILS))
3421 fprintf (dump_file, "Found fully redundant value\n");
3425 avail = XCNEWVEC (pre_expr, last_basic_block);
3426 FOR_EACH_EDGE (pred, ei, block->preds)
3428 unsigned int vprime;
3430 /* We should never run insertion for the exit block
3431 and so not come across fake pred edges. */
3432 gcc_assert (!(pred->flags & EDGE_FAKE));
3434 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3437 /* eprime will generally only be NULL if the
3438 value of the expression, translated
3439 through the PHI for this predecessor, is
3440 undefined. If that is the case, we can't
3441 make the expression fully redundant,
3442 because its value is undefined along a
3443 predecessor path. We can thus break out
3444 early because it doesn't matter what the
3445 rest of the results are. */
3452 eprime = fully_constant_expression (eprime);
3453 vprime = get_expr_value_id (eprime);
3454 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3456 if (edoubleprime == NULL)
3458 avail[bprime->index] = eprime;
3463 avail[bprime->index] = edoubleprime;
3465 /* We want to perform insertions to remove a redundancy on
3466 a path in the CFG we want to optimize for speed. */
3467 if (optimize_edge_for_speed_p (pred))
3468 do_insertion = true;
3469 if (first_s == NULL)
3470 first_s = edoubleprime;
3471 else if (!pre_expr_eq (first_s, edoubleprime))
3475 /* If we can insert it, it's not the same value
3476 already existing along every predecessor, and
3477 it's defined by some predecessor, it is
3478 partially redundant. */
3479 if (!cant_insert && !all_same && by_some && do_insertion
3480 && dbg_cnt (treepre_insert))
3482 if (insert_into_preds_of_block (block, get_expression_id (expr),
3486 /* If all edges produce the same value and that value is
3487 an invariant, then the PHI has the same value on all
3488 edges. Note this. */
3489 else if (!cant_insert && all_same && eprime
3490 && (edoubleprime->kind == CONSTANT
3491 || edoubleprime->kind == NAME)
3492 && !value_id_constant_p (val))
3496 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3497 value_expressions, val);
3499 unsigned int new_val = get_expr_value_id (edoubleprime);
3500 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3502 pre_expr expr = expression_for_id (j);
3504 if (expr->kind == NAME)
3506 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3507 /* Just reset the value id and valnum so it is
3508 the same as the constant we have discovered. */
3509 if (edoubleprime->kind == CONSTANT)
3511 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3512 pre_stats.constified++;
3515 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3516 info->value_id = new_val;
3524 VEC_free (pre_expr, heap, exprs);
3529 /* Perform insertion for partially anticipatable expressions. There
3530 is only one case we will perform insertion for these. This case is
3531 if the expression is partially anticipatable, and fully available.
3532 In this case, we know that putting it earlier will enable us to
3533 remove the later computation. */
3537 do_partial_partial_insertion (basic_block block, basic_block dom)
3539 bool new_stuff = false;
3540 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3544 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3546 if (expr->kind != NAME)
3551 bool cant_insert = false;
3554 pre_expr eprime = NULL;
3557 val = get_expr_value_id (expr);
3558 if (bitmap_set_contains_value (PHI_GEN (block), val))
3560 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3563 avail = XCNEWVEC (pre_expr, last_basic_block);
3564 FOR_EACH_EDGE (pred, ei, block->preds)
3566 unsigned int vprime;
3567 pre_expr edoubleprime;
3569 /* We should never run insertion for the exit block
3570 and so not come across fake pred edges. */
3571 gcc_assert (!(pred->flags & EDGE_FAKE));
3573 eprime = phi_translate (expr, ANTIC_IN (block),
3577 /* eprime will generally only be NULL if the
3578 value of the expression, translated
3579 through the PHI for this predecessor, is
3580 undefined. If that is the case, we can't
3581 make the expression fully redundant,
3582 because its value is undefined along a
3583 predecessor path. We can thus break out
3584 early because it doesn't matter what the
3585 rest of the results are. */
3592 eprime = fully_constant_expression (eprime);
3593 vprime = get_expr_value_id (eprime);
3594 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3596 if (edoubleprime == NULL)
3602 avail[bprime->index] = edoubleprime;
3606 /* If we can insert it, it's not the same value
3607 already existing along every predecessor, and
3608 it's defined by some predecessor, it is
3609 partially redundant. */
3610 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3612 pre_stats.pa_insert++;
3613 if (insert_into_preds_of_block (block, get_expression_id (expr),
3621 VEC_free (pre_expr, heap, exprs);
3626 insert_aux (basic_block block)
3629 bool new_stuff = false;
3634 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3639 bitmap_set_t newset = NEW_SETS (dom);
3642 /* Note that we need to value_replace both NEW_SETS, and
3643 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3644 represented by some non-simple expression here that we want
3645 to replace it with. */
3646 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3648 pre_expr expr = expression_for_id (i);
3649 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3650 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3653 if (!single_pred_p (block))
3655 new_stuff |= do_regular_insertion (block, dom);
3656 if (do_partial_partial)
3657 new_stuff |= do_partial_partial_insertion (block, dom);
3661 for (son = first_dom_son (CDI_DOMINATORS, block);
3663 son = next_dom_son (CDI_DOMINATORS, son))
3665 new_stuff |= insert_aux (son);
3671 /* Perform insertion of partially redundant values. */
3676 bool new_stuff = true;
3678 int num_iterations = 0;
3681 NEW_SETS (bb) = bitmap_set_new ();
3686 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3688 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3692 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3695 add_to_exp_gen (basic_block block, tree op)
3700 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3702 result = get_or_alloc_expr_for_name (op);
3703 bitmap_value_insert_into_set (EXP_GEN (block), result);
3707 /* Create value ids for PHI in BLOCK. */
3710 make_values_for_phi (gimple phi, basic_block block)
3712 tree result = gimple_phi_result (phi);
3714 /* We have no need for virtual phis, as they don't represent
3715 actual computations. */
3716 if (is_gimple_reg (result))
3718 pre_expr e = get_or_alloc_expr_for_name (result);
3719 add_to_value (get_expr_value_id (e), e);
3720 bitmap_insert_into_set (PHI_GEN (block), e);
3721 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3725 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3727 tree arg = gimple_phi_arg_def (phi, i);
3728 if (TREE_CODE (arg) == SSA_NAME)
3730 e = get_or_alloc_expr_for_name (arg);
3731 add_to_value (get_expr_value_id (e), e);
3738 /* Compute the AVAIL set for all basic blocks.
3740 This function performs value numbering of the statements in each basic
3741 block. The AVAIL sets are built from information we glean while doing
3742 this value numbering, since the AVAIL sets contain only one entry per
3745 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3746 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3749 compute_avail (void)
3752 basic_block block, son;
3753 basic_block *worklist;
3757 /* We pretend that default definitions are defined in the entry block.
3758 This includes function arguments and the static chain decl. */
3759 for (i = 1; i < num_ssa_names; ++i)
3761 tree name = ssa_name (i);
3764 || !SSA_NAME_IS_DEFAULT_DEF (name)
3765 || has_zero_uses (name)
3766 || !is_gimple_reg (name))
3769 e = get_or_alloc_expr_for_name (name);
3770 add_to_value (get_expr_value_id (e), e);
3772 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3773 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3776 /* Allocate the worklist. */
3777 worklist = XNEWVEC (basic_block, n_basic_blocks);
3779 /* Seed the algorithm by putting the dominator children of the entry
3780 block on the worklist. */
3781 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3783 son = next_dom_son (CDI_DOMINATORS, son))
3784 worklist[sp++] = son;
3786 /* Loop until the worklist is empty. */
3789 gimple_stmt_iterator gsi;
3792 unsigned int stmt_uid = 1;
3794 /* Pick a block from the worklist. */
3795 block = worklist[--sp];
3797 /* Initially, the set of available values in BLOCK is that of
3798 its immediate dominator. */
3799 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3801 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3803 /* Generate values for PHI nodes. */
3804 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3805 make_values_for_phi (gsi_stmt (gsi), block);
3807 /* Now compute value numbers and populate value sets with all
3808 the expressions computed in BLOCK. */
3809 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3814 stmt = gsi_stmt (gsi);
3815 gimple_set_uid (stmt, stmt_uid++);
3817 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3819 pre_expr e = get_or_alloc_expr_for_name (op);
3821 add_to_value (get_expr_value_id (e), e);
3823 bitmap_insert_into_set (TMP_GEN (block), e);
3824 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3827 if (gimple_has_volatile_ops (stmt)
3828 || stmt_could_throw_p (stmt))
3831 switch (gimple_code (stmt))
3834 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3835 add_to_exp_gen (block, op);
3842 vn_reference_op_t vro;
3843 pre_expr result = NULL;
3844 VEC(vn_reference_op_s, heap) *ops = NULL;
3846 if (!can_value_number_call (stmt))
3849 copy_reference_ops_from_call (stmt, &ops);
3850 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3851 gimple_expr_type (stmt),
3853 VEC_free (vn_reference_op_s, heap, ops);
3857 for (i = 0; VEC_iterate (vn_reference_op_s,
3861 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3862 add_to_exp_gen (block, vro->op0);
3863 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3864 add_to_exp_gen (block, vro->op1);
3865 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3866 add_to_exp_gen (block, vro->op2);
3868 result = (pre_expr) pool_alloc (pre_expr_pool);
3869 result->kind = REFERENCE;
3871 PRE_EXPR_REFERENCE (result) = ref;
3873 get_or_alloc_expression_id (result);
3874 add_to_value (get_expr_value_id (result), result);
3876 bitmap_value_insert_into_set (EXP_GEN (block), result);
3882 pre_expr result = NULL;
3883 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
3887 case tcc_comparison:
3892 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
3893 gimple_assign_rhs_code (stmt),
3894 gimple_expr_type (stmt),
3895 gimple_assign_rhs1 (stmt),
3896 gimple_assign_rhs2 (stmt),
3897 NULL_TREE, NULL_TREE, &nary);
3902 for (i = 0; i < nary->length; i++)
3903 if (TREE_CODE (nary->op[i]) == SSA_NAME)
3904 add_to_exp_gen (block, nary->op[i]);
3906 result = (pre_expr) pool_alloc (pre_expr_pool);
3907 result->kind = NARY;
3909 PRE_EXPR_NARY (result) = nary;
3913 case tcc_declaration:
3918 vn_reference_op_t vro;
3920 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3926 for (i = 0; VEC_iterate (vn_reference_op_s,
3930 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3931 add_to_exp_gen (block, vro->op0);
3932 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3933 add_to_exp_gen (block, vro->op1);
3934 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3935 add_to_exp_gen (block, vro->op2);
3937 result = (pre_expr) pool_alloc (pre_expr_pool);
3938 result->kind = REFERENCE;
3940 PRE_EXPR_REFERENCE (result) = ref;
3945 /* For any other statement that we don't
3946 recognize, simply add all referenced
3947 SSA_NAMEs to EXP_GEN. */
3948 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3949 add_to_exp_gen (block, op);
3953 get_or_alloc_expression_id (result);
3954 add_to_value (get_expr_value_id (result), result);
3956 bitmap_value_insert_into_set (EXP_GEN (block), result);
3965 /* Put the dominator children of BLOCK on the worklist of blocks
3966 to compute available sets for. */
3967 for (son = first_dom_son (CDI_DOMINATORS, block);
3969 son = next_dom_son (CDI_DOMINATORS, son))
3970 worklist[sp++] = son;
3976 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
3977 than the available expressions for it. The insertion point is
3978 right before the first use in STMT. Returns the SSA_NAME that should
3979 be used for replacement. */
3982 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
3984 basic_block bb = gimple_bb (stmt);
3985 gimple_stmt_iterator gsi;
3986 gimple_seq stmts = NULL;
3990 /* First create a value expression from the expression we want
3991 to insert and associate it with the value handle for SSA_VN. */
3992 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
3996 /* Then use create_expression_by_pieces to generate a valid
3997 expression to insert at this point of the IL stream. */
3998 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
3999 if (expr == NULL_TREE)
4001 gsi = gsi_for_stmt (stmt);
4002 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4007 /* Eliminate fully redundant computations. */
4012 VEC (gimple, heap) *to_remove = NULL;
4014 unsigned int todo = 0;
4015 gimple_stmt_iterator gsi;
4021 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4023 stmt = gsi_stmt (gsi);
4025 /* Lookup the RHS of the expression, see if we have an
4026 available computation for it. If so, replace the RHS with
4027 the available computation. */
4028 if (gimple_has_lhs (stmt)
4029 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4030 && !gimple_assign_ssa_name_copy_p (stmt)
4031 && (!gimple_assign_single_p (stmt)
4032 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4033 && !gimple_has_volatile_ops (stmt)
4034 && !has_zero_uses (gimple_get_lhs (stmt)))
4036 tree lhs = gimple_get_lhs (stmt);
4037 tree rhs = NULL_TREE;
4039 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4040 pre_expr sprimeexpr;
4042 if (gimple_assign_single_p (stmt))
4043 rhs = gimple_assign_rhs1 (stmt);
4045 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4046 get_expr_value_id (lhsexpr),
4051 if (sprimeexpr->kind == CONSTANT)
4052 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4053 else if (sprimeexpr->kind == NAME)
4054 sprime = PRE_EXPR_NAME (sprimeexpr);
4059 /* If there is no existing leader but SCCVN knows this
4060 value is constant, use that constant. */
4061 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4063 sprime = VN_INFO (lhs)->valnum;
4064 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4065 TREE_TYPE (sprime)))
4066 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4068 if (dump_file && (dump_flags & TDF_DETAILS))
4070 fprintf (dump_file, "Replaced ");
4071 print_gimple_expr (dump_file, stmt, 0, 0);
4072 fprintf (dump_file, " with ");
4073 print_generic_expr (dump_file, sprime, 0);
4074 fprintf (dump_file, " in ");
4075 print_gimple_stmt (dump_file, stmt, 0, 0);
4077 pre_stats.eliminations++;
4078 propagate_tree_value_into_stmt (&gsi, sprime);
4079 stmt = gsi_stmt (gsi);
4084 /* If there is no existing usable leader but SCCVN thinks
4085 it has an expression it wants to use as replacement,
4087 if (!sprime || sprime == lhs)
4089 tree val = VN_INFO (lhs)->valnum;
4091 && TREE_CODE (val) == SSA_NAME
4092 && VN_INFO (val)->needs_insertion
4093 && can_PRE_operation (vn_get_expr_for (val)))
4094 sprime = do_SCCVN_insertion (stmt, val);
4098 && (rhs == NULL_TREE
4099 || TREE_CODE (rhs) != SSA_NAME
4100 || may_propagate_copy (rhs, sprime)))
4102 gcc_assert (sprime != rhs);
4104 if (dump_file && (dump_flags & TDF_DETAILS))
4106 fprintf (dump_file, "Replaced ");
4107 print_gimple_expr (dump_file, stmt, 0, 0);
4108 fprintf (dump_file, " with ");
4109 print_generic_expr (dump_file, sprime, 0);
4110 fprintf (dump_file, " in ");
4111 print_gimple_stmt (dump_file, stmt, 0, 0);
4114 if (TREE_CODE (sprime) == SSA_NAME)
4115 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4117 /* We need to make sure the new and old types actually match,
4118 which may require adding a simple cast, which fold_convert
4120 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4121 && !useless_type_conversion_p (gimple_expr_type (stmt),
4122 TREE_TYPE (sprime)))
4123 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4125 pre_stats.eliminations++;
4126 propagate_tree_value_into_stmt (&gsi, sprime);
4127 stmt = gsi_stmt (gsi);
4130 /* If we removed EH side effects from the statement, clean
4131 its EH information. */
4132 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4134 bitmap_set_bit (need_eh_cleanup,
4135 gimple_bb (stmt)->index);
4136 if (dump_file && (dump_flags & TDF_DETAILS))
4137 fprintf (dump_file, " Removed EH side effects.\n");
4141 /* If the statement is a scalar store, see if the expression
4142 has the same value number as its rhs. If so, the store is
4144 else if (gimple_assign_single_p (stmt)
4145 && !is_gimple_reg (gimple_assign_lhs (stmt))
4146 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4147 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4149 tree rhs = gimple_assign_rhs1 (stmt);
4151 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4152 gimple_vuse (stmt), true, NULL);
4153 if (TREE_CODE (rhs) == SSA_NAME)
4154 rhs = VN_INFO (rhs)->valnum;
4156 && operand_equal_p (val, rhs, 0))
4158 if (dump_file && (dump_flags & TDF_DETAILS))
4160 fprintf (dump_file, "Deleted redundant store ");
4161 print_gimple_stmt (dump_file, stmt, 0, 0);
4164 /* Queue stmt for removal. */
4165 VEC_safe_push (gimple, heap, to_remove, stmt);
4168 /* Visit COND_EXPRs and fold the comparison with the
4169 available value-numbers. */
4170 else if (gimple_code (stmt) == GIMPLE_COND)
4172 tree op0 = gimple_cond_lhs (stmt);
4173 tree op1 = gimple_cond_rhs (stmt);
4176 if (TREE_CODE (op0) == SSA_NAME)
4177 op0 = VN_INFO (op0)->valnum;
4178 if (TREE_CODE (op1) == SSA_NAME)
4179 op1 = VN_INFO (op1)->valnum;
4180 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4182 if (result && TREE_CODE (result) == INTEGER_CST)
4184 if (integer_zerop (result))
4185 gimple_cond_make_false (stmt);
4187 gimple_cond_make_true (stmt);
4189 todo = TODO_cleanup_cfg;
4192 /* Visit indirect calls and turn them into direct calls if
4194 if (gimple_code (stmt) == GIMPLE_CALL
4195 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4197 tree fn = VN_INFO (gimple_call_fn (stmt))->valnum;
4198 if (TREE_CODE (fn) == ADDR_EXPR
4199 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
4201 if (dump_file && (dump_flags & TDF_DETAILS))
4203 fprintf (dump_file, "Replacing call target with ");
4204 print_generic_expr (dump_file, fn, 0);
4205 fprintf (dump_file, " in ");
4206 print_gimple_stmt (dump_file, stmt, 0, 0);
4209 gimple_call_set_fn (stmt, fn);
4211 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4213 bitmap_set_bit (need_eh_cleanup,
4214 gimple_bb (stmt)->index);
4215 if (dump_file && (dump_flags & TDF_DETAILS))
4216 fprintf (dump_file, " Removed EH side effects.\n");
4219 /* Changing an indirect call to a direct call may
4220 have exposed different semantics. This may
4221 require an SSA update. */
4222 todo |= TODO_update_ssa_only_virtuals;
4227 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4229 gimple stmt, phi = gsi_stmt (gsi);
4230 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4231 pre_expr sprimeexpr, resexpr;
4232 gimple_stmt_iterator gsi2;
4234 /* We want to perform redundant PHI elimination. Do so by
4235 replacing the PHI with a single copy if possible.
4236 Do not touch inserted, single-argument or virtual PHIs. */
4237 if (gimple_phi_num_args (phi) == 1
4238 || !is_gimple_reg (res)
4239 || bitmap_bit_p (inserted_phi_names, SSA_NAME_VERSION (res)))
4245 resexpr = get_or_alloc_expr_for_name (res);
4246 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4247 get_expr_value_id (resexpr), NULL);
4250 if (sprimeexpr->kind == CONSTANT)
4251 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4252 else if (sprimeexpr->kind == NAME)
4253 sprime = PRE_EXPR_NAME (sprimeexpr);
4264 if (dump_file && (dump_flags & TDF_DETAILS))
4266 fprintf (dump_file, "Replaced redundant PHI node defining ");
4267 print_generic_expr (dump_file, res, 0);
4268 fprintf (dump_file, " with ");
4269 print_generic_expr (dump_file, sprime, 0);
4270 fprintf (dump_file, "\n");
4273 remove_phi_node (&gsi, false);
4275 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4276 sprime = fold_convert (TREE_TYPE (res), sprime);
4277 stmt = gimple_build_assign (res, sprime);
4278 SSA_NAME_DEF_STMT (res) = stmt;
4279 if (TREE_CODE (sprime) == SSA_NAME)
4280 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4282 gsi2 = gsi_after_labels (b);
4283 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4284 /* Queue the copy for eventual removal. */
4285 VEC_safe_push (gimple, heap, to_remove, stmt);
4286 pre_stats.eliminations++;
4290 /* We cannot remove stmts during BB walk, especially not release SSA
4291 names there as this confuses the VN machinery. The stmts ending
4292 up in to_remove are either stores or simple copies. */
4293 for (i = 0; VEC_iterate (gimple, to_remove, i, stmt); ++i)
4295 tree lhs = gimple_assign_lhs (stmt);
4296 use_operand_p use_p;
4299 /* If there is a single use only, propagate the equivalency
4300 instead of keeping the copy. */
4301 if (TREE_CODE (lhs) == SSA_NAME
4302 && single_imm_use (lhs, &use_p, &use_stmt)
4303 && may_propagate_copy (USE_FROM_PTR (use_p),
4304 gimple_assign_rhs1 (stmt)))
4306 SET_USE (use_p, gimple_assign_rhs1 (stmt));
4307 update_stmt (use_stmt);
4310 /* If this is a store or a now unused copy, remove it. */
4311 if (TREE_CODE (lhs) != SSA_NAME
4312 || has_zero_uses (lhs))
4314 gsi = gsi_for_stmt (stmt);
4315 unlink_stmt_vdef (stmt);
4316 gsi_remove (&gsi, true);
4317 release_defs (stmt);
4320 VEC_free (gimple, heap, to_remove);
4325 /* Borrow a bit of tree-ssa-dce.c for the moment.
4326 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4327 this may be a bit faster, and we may want critical edges kept split. */
4329 /* If OP's defining statement has not already been determined to be necessary,
4330 mark that statement necessary. Return the stmt, if it is newly
4333 static inline gimple
4334 mark_operand_necessary (tree op)
4340 if (TREE_CODE (op) != SSA_NAME)
4343 stmt = SSA_NAME_DEF_STMT (op);
4346 if (gimple_plf (stmt, NECESSARY)
4347 || gimple_nop_p (stmt))
4350 gimple_set_plf (stmt, NECESSARY, true);
4354 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4355 to insert PHI nodes sometimes, and because value numbering of casts isn't
4356 perfect, we sometimes end up inserting dead code. This simple DCE-like
4357 pass removes any insertions we made that weren't actually used. */
4360 remove_dead_inserted_code (void)
4362 VEC(gimple,heap) *worklist = NULL;
4366 worklist = VEC_alloc (gimple, heap, VEC_length (gimple, inserted_exprs));
4367 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4369 if (gimple_plf (t, NECESSARY))
4370 VEC_quick_push (gimple, worklist, t);
4372 while (VEC_length (gimple, worklist) > 0)
4374 t = VEC_pop (gimple, worklist);
4376 /* PHI nodes are somewhat special in that each PHI alternative has
4377 data and control dependencies. All the statements feeding the
4378 PHI node's arguments are always necessary. */
4379 if (gimple_code (t) == GIMPLE_PHI)
4383 VEC_reserve (gimple, heap, worklist, gimple_phi_num_args (t));
4384 for (k = 0; k < gimple_phi_num_args (t); k++)
4386 tree arg = PHI_ARG_DEF (t, k);
4387 if (TREE_CODE (arg) == SSA_NAME)
4389 gimple n = mark_operand_necessary (arg);
4391 VEC_quick_push (gimple, worklist, n);
4397 /* Propagate through the operands. Examine all the USE, VUSE and
4398 VDEF operands in this statement. Mark all the statements
4399 which feed this statement's uses as necessary. */
4403 /* The operands of VDEF expressions are also needed as they
4404 represent potential definitions that may reach this
4405 statement (VDEF operands allow us to follow def-def
4408 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4410 gimple n = mark_operand_necessary (use);
4412 VEC_safe_push (gimple, heap, worklist, n);
4417 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4419 if (!gimple_plf (t, NECESSARY))
4421 gimple_stmt_iterator gsi;
4423 if (dump_file && (dump_flags & TDF_DETAILS))
4425 fprintf (dump_file, "Removing unnecessary insertion:");
4426 print_gimple_stmt (dump_file, t, 0, 0);
4429 gsi = gsi_for_stmt (t);
4430 if (gimple_code (t) == GIMPLE_PHI)
4431 remove_phi_node (&gsi, true);
4434 gsi_remove (&gsi, true);
4439 VEC_free (gimple, heap, worklist);
4442 /* Initialize data structures used by PRE. */
4445 init_pre (bool do_fre)
4449 next_expression_id = 1;
4451 VEC_safe_push (pre_expr, heap, expressions, NULL);
4452 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4453 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4454 get_max_value_id() + 1);
4458 inserted_exprs = NULL;
4459 need_creation = NULL;
4460 pretemp = NULL_TREE;
4461 storetemp = NULL_TREE;
4462 prephitemp = NULL_TREE;
4464 connect_infinite_loops_to_exit ();
4465 memset (&pre_stats, 0, sizeof (pre_stats));
4468 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4469 post_order_compute (postorder, false, false);
4472 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4474 calculate_dominance_info (CDI_POST_DOMINATORS);
4475 calculate_dominance_info (CDI_DOMINATORS);
4477 bitmap_obstack_initialize (&grand_bitmap_obstack);
4478 inserted_phi_names = BITMAP_ALLOC (&grand_bitmap_obstack);
4479 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4480 expr_pred_trans_eq, free);
4481 expression_to_id = htab_create (num_ssa_names * 3,
4484 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4485 sizeof (struct bitmap_set), 30);
4486 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4487 sizeof (struct pre_expr_d), 30);
4490 EXP_GEN (bb) = bitmap_set_new ();
4491 PHI_GEN (bb) = bitmap_set_new ();
4492 TMP_GEN (bb) = bitmap_set_new ();
4493 AVAIL_OUT (bb) = bitmap_set_new ();
4496 need_eh_cleanup = BITMAP_ALLOC (NULL);
4500 /* Deallocate data structures used by PRE. */
4503 fini_pre (bool do_fre)
4508 VEC_free (bitmap_set_t, heap, value_expressions);
4509 VEC_free (gimple, heap, inserted_exprs);
4510 VEC_free (gimple, heap, need_creation);
4511 bitmap_obstack_release (&grand_bitmap_obstack);
4512 free_alloc_pool (bitmap_set_pool);
4513 free_alloc_pool (pre_expr_pool);
4514 htab_delete (phi_translate_table);
4515 htab_delete (expression_to_id);
4523 free_dominance_info (CDI_POST_DOMINATORS);
4525 if (!bitmap_empty_p (need_eh_cleanup))
4527 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4528 cleanup_tree_cfg ();
4531 BITMAP_FREE (need_eh_cleanup);
4534 loop_optimizer_finalize ();
4537 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4538 only wants to do full redundancy elimination. */
4541 execute_pre (bool do_fre)
4543 unsigned int todo = 0;
4545 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4547 /* This has to happen before SCCVN runs because
4548 loop_optimizer_init may create new phis, etc. */
4550 loop_optimizer_init (LOOPS_NORMAL);
4552 if (!run_scc_vn (do_fre))
4556 remove_dead_inserted_code ();
4557 loop_optimizer_finalize ();
4566 /* Collect and value number expressions computed in each basic block. */
4569 if (dump_file && (dump_flags & TDF_DETAILS))
4575 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4576 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4577 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4578 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4582 /* Insert can get quite slow on an incredibly large number of basic
4583 blocks due to some quadratic behavior. Until this behavior is
4584 fixed, don't run it when he have an incredibly large number of
4585 bb's. If we aren't going to run insert, there is no point in
4586 computing ANTIC, either, even though it's plenty fast. */
4587 if (!do_fre && n_basic_blocks < 4000)
4593 /* Remove all the redundant expressions. */
4594 todo |= eliminate ();
4596 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4597 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4598 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4599 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4600 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4602 /* Make sure to remove fake edges before committing our inserts.
4603 This makes sure we don't end up with extra critical edges that
4604 we would need to split. */
4605 remove_fake_exit_edges ();
4606 gsi_commit_edge_inserts ();
4608 clear_expression_ids ();
4611 remove_dead_inserted_code ();
4619 /* Gate and execute functions for PRE. */
4624 return execute_pre (false);
4630 return flag_tree_pre != 0;
4633 struct gimple_opt_pass pass_pre =
4638 gate_pre, /* gate */
4639 do_pre, /* execute */
4642 0, /* static_pass_number */
4643 TV_TREE_PRE, /* tv_id */
4644 PROP_no_crit_edges | PROP_cfg
4645 | PROP_ssa, /* properties_required */
4646 0, /* properties_provided */
4647 0, /* properties_destroyed */
4648 TODO_rebuild_alias, /* todo_flags_start */
4649 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4650 | TODO_verify_ssa /* todo_flags_finish */
4655 /* Gate and execute functions for FRE. */
4660 return execute_pre (true);
4666 return flag_tree_fre != 0;
4669 struct gimple_opt_pass pass_fre =
4674 gate_fre, /* gate */
4675 execute_fre, /* execute */
4678 0, /* static_pass_number */
4679 TV_TREE_FRE, /* tv_id */
4680 PROP_cfg | PROP_ssa, /* properties_required */
4681 0, /* properties_provided */
4682 0, /* properties_destroyed */
4683 0, /* todo_flags_start */
4684 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */