2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007
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"
53 1. Avail sets can be shared by making an avail_find_leader that
54 walks up the dominator tree and looks in those avail sets.
55 This might affect code optimality, it's unclear right now.
56 2. Strength reduction can be performed by anticipating expressions
57 we can repair later on.
58 3. We can do back-substitution or smarter value numbering to catch
59 commutative expressions split up over multiple statements.
62 /* For ease of terminology, "expression node" in the below refers to
63 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
64 represent the actual statement containing the expressions we care about,
65 and we cache the value number by putting it in the expression. */
69 First we walk the statements to generate the AVAIL sets, the
70 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
71 generation of values/expressions by a given block. We use them
72 when computing the ANTIC sets. The AVAIL sets consist of
73 SSA_NAME's that represent values, so we know what values are
74 available in what blocks. AVAIL is a forward dataflow problem. In
75 SSA, values are never killed, so we don't need a kill set, or a
76 fixpoint iteration, in order to calculate the AVAIL sets. In
77 traditional parlance, AVAIL sets tell us the downsafety of the
80 Next, we generate the ANTIC sets. These sets represent the
81 anticipatable expressions. ANTIC is a backwards dataflow
82 problem. An expression is anticipatable in a given block if it could
83 be generated in that block. This means that if we had to perform
84 an insertion in that block, of the value of that expression, we
85 could. Calculating the ANTIC sets requires phi translation of
86 expressions, because the flow goes backwards through phis. We must
87 iterate to a fixpoint of the ANTIC sets, because we have a kill
88 set. Even in SSA form, values are not live over the entire
89 function, only from their definition point onwards. So we have to
90 remove values from the ANTIC set once we go past the definition
91 point of the leaders that make them up.
92 compute_antic/compute_antic_aux performs this computation.
94 Third, we perform insertions to make partially redundant
95 expressions fully redundant.
97 An expression is partially redundant (excluding partial
100 1. It is AVAIL in some, but not all, of the predecessors of a
102 2. It is ANTIC in all the predecessors.
104 In order to make it fully redundant, we insert the expression into
105 the predecessors where it is not available, but is ANTIC.
107 For the partial anticipation case, we only perform insertion if it
108 is partially anticipated in some block, and fully available in all
111 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
112 performs these steps.
114 Fourth, we eliminate fully redundant expressions.
115 This is a simple statement walk that replaces redundant
116 calculations with the now available values. */
118 /* Representations of value numbers:
120 Value numbers are represented by a representative SSA_NAME. We
121 will create fake SSA_NAME's in situations where we need a
122 representative but do not have one (because it is a complex
123 expression). In order to facilitate storing the value numbers in
124 bitmaps, and keep the number of wasted SSA_NAME's down, we also
125 associate a value_id with each value number, and create full blown
126 ssa_name's only where we actually need them (IE in operands of
127 existing expressions).
129 Theoretically you could replace all the value_id's with
130 SSA_NAME_VERSION, but this would allocate a large number of
131 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
132 It would also require an additional indirection at each point we
135 /* Representation of expressions on value numbers:
137 Expressions consisting of value numbers are represented the same
138 way as our VN internally represents them, with an additional
139 "pre_expr" wrapping around them in order to facilitate storing all
140 of the expressions in the same sets. */
142 /* Representation of sets:
144 The dataflow sets do not need to be sorted in any particular order
145 for the majority of their lifetime, are simply represented as two
146 bitmaps, one that keeps track of values present in the set, and one
147 that keeps track of expressions present in the set.
149 When we need them in topological order, we produce it on demand by
150 transforming the bitmap into an array and sorting it into topo
153 /* Type of expression, used to know which member of the PRE_EXPR union
164 typedef union pre_expr_union_d
169 vn_reference_t reference;
172 typedef struct pre_expr_d
174 enum pre_expr_kind kind;
179 #define PRE_EXPR_NAME(e) (e)->u.name
180 #define PRE_EXPR_NARY(e) (e)->u.nary
181 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
182 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
185 pre_expr_eq (const void *p1, const void *p2)
187 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
188 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
190 if (e1->kind != e2->kind)
196 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
197 PRE_EXPR_CONSTANT (e2));
199 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
201 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
203 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
204 PRE_EXPR_REFERENCE (e2));
211 pre_expr_hash (const void *p1)
213 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
217 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
219 return iterative_hash_expr (PRE_EXPR_NAME (e), 0);
221 return vn_nary_op_compute_hash (PRE_EXPR_NARY (e));
223 return vn_reference_compute_hash (PRE_EXPR_REFERENCE (e));
230 /* Next global expression id number. */
231 static unsigned int next_expression_id;
233 /* Mapping from expression to id number we can use in bitmap sets. */
234 DEF_VEC_P (pre_expr);
235 DEF_VEC_ALLOC_P (pre_expr, heap);
236 static VEC(pre_expr, heap) *expressions;
237 static htab_t expression_to_id;
239 /* Allocate an expression id for EXPR. */
241 static inline unsigned int
242 alloc_expression_id (pre_expr expr)
245 /* Make sure we won't overflow. */
246 gcc_assert (next_expression_id + 1 > next_expression_id);
247 expr->id = next_expression_id++;
248 VEC_safe_push (pre_expr, heap, expressions, expr);
249 slot = htab_find_slot (expression_to_id, expr, INSERT);
252 return next_expression_id - 1;
255 /* Return the expression id for tree EXPR. */
257 static inline unsigned int
258 get_expression_id (const pre_expr expr)
263 static inline unsigned int
264 lookup_expression_id (const pre_expr expr)
268 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
271 return ((pre_expr)*slot)->id;
274 /* Return the existing expression id for EXPR, or create one if one
275 does not exist yet. */
277 static inline unsigned int
278 get_or_alloc_expression_id (pre_expr expr)
280 unsigned int id = lookup_expression_id (expr);
282 return alloc_expression_id (expr);
283 return expr->id = id;
286 /* Return the expression that has expression id ID */
288 static inline pre_expr
289 expression_for_id (unsigned int id)
291 return VEC_index (pre_expr, expressions, id);
294 /* Free the expression id field in all of our expressions,
295 and then destroy the expressions array. */
298 clear_expression_ids (void)
300 VEC_free (pre_expr, heap, expressions);
303 static alloc_pool pre_expr_pool;
305 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
308 get_or_alloc_expr_for_name (tree name)
310 pre_expr result = (pre_expr) pool_alloc (pre_expr_pool);
311 unsigned int result_id;
315 PRE_EXPR_NAME (result) = name;
316 result_id = lookup_expression_id (result);
319 pool_free (pre_expr_pool, result);
320 result = expression_for_id (result_id);
323 get_or_alloc_expression_id (result);
327 static bool in_fre = false;
329 /* An unordered bitmap set. One bitmap tracks values, the other,
331 typedef struct bitmap_set
337 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
338 EXECUTE_IF_SET_IN_BITMAP((set)->expressions, 0, (id), (bi))
340 /* Mapping from value id to expressions with that value_id. */
341 DEF_VEC_P (bitmap_set_t);
342 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
343 static VEC(bitmap_set_t, heap) *value_expressions;
345 /* Sets that we need to keep track of. */
346 typedef struct bb_bitmap_sets
348 /* The EXP_GEN set, which represents expressions/values generated in
350 bitmap_set_t exp_gen;
352 /* The PHI_GEN set, which represents PHI results generated in a
354 bitmap_set_t phi_gen;
356 /* The TMP_GEN set, which represents results/temporaries generated
357 in a basic block. IE the LHS of an expression. */
358 bitmap_set_t tmp_gen;
360 /* The AVAIL_OUT set, which represents which values are available in
361 a given basic block. */
362 bitmap_set_t avail_out;
364 /* The ANTIC_IN set, which represents which values are anticipatable
365 in a given basic block. */
366 bitmap_set_t antic_in;
368 /* The PA_IN set, which represents which values are
369 partially anticipatable in a given basic block. */
372 /* The NEW_SETS set, which is used during insertion to augment the
373 AVAIL_OUT set of blocks with the new insertions performed during
374 the current iteration. */
375 bitmap_set_t new_sets;
377 /* True if we have visited this block during ANTIC calculation. */
378 unsigned int visited:1;
380 /* True we have deferred processing this block during ANTIC
381 calculation until its successor is processed. */
382 unsigned int deferred : 1;
385 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
386 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
387 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
388 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
389 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
390 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
391 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
392 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
393 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
396 /* Maximal set of values, used to initialize the ANTIC problem, which
397 is an intersection problem. */
398 static bitmap_set_t maximal_set;
400 /* Basic block list in postorder. */
401 static int *postorder;
403 /* This structure is used to keep track of statistics on what
404 optimization PRE was able to perform. */
407 /* The number of RHS computations eliminated by PRE. */
410 /* The number of new expressions/temporaries generated by PRE. */
413 /* The number of inserts found due to partial anticipation */
416 /* The number of new PHI nodes added by PRE. */
419 /* The number of values found constant. */
424 static bool do_partial_partial;
425 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
426 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
427 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
428 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
429 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
430 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
431 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr, bool);
432 static bitmap_set_t bitmap_set_new (void);
433 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
435 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
438 /* We can add and remove elements and entries to and from sets
439 and hash tables, so we use alloc pools for them. */
441 static alloc_pool bitmap_set_pool;
442 static bitmap_obstack grand_bitmap_obstack;
444 /* To avoid adding 300 temporary variables when we only need one, we
445 only create one temporary variable, on demand, and build ssa names
446 off that. We do have to change the variable if the types don't
447 match the current variable's type. */
449 static tree storetemp;
450 static tree prephitemp;
452 /* Set of blocks with statements that have had its EH information
454 static bitmap need_eh_cleanup;
456 /* Which expressions have been seen during a given phi translation. */
457 static bitmap seen_during_translate;
459 /* The phi_translate_table caches phi translations for a given
460 expression and predecessor. */
462 static htab_t phi_translate_table;
464 /* A three tuple {e, pred, v} used to cache phi translations in the
465 phi_translate_table. */
467 typedef struct expr_pred_trans_d
469 /* The expression. */
472 /* The predecessor block along which we translated the expression. */
475 /* The value that resulted from the translation. */
478 /* The hashcode for the expression, pred pair. This is cached for
481 } *expr_pred_trans_t;
482 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
484 /* Return the hash value for a phi translation table entry. */
487 expr_pred_trans_hash (const void *p)
489 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
493 /* Return true if two phi translation table entries are the same.
494 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
497 expr_pred_trans_eq (const void *p1, const void *p2)
499 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
500 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
501 basic_block b1 = ve1->pred;
502 basic_block b2 = ve2->pred;
504 /* If they are not translations for the same basic block, they can't
508 return pre_expr_eq (ve1->e, ve2->e);
511 /* Search in the phi translation table for the translation of
512 expression E in basic block PRED.
513 Return the translated value, if found, NULL otherwise. */
515 static inline pre_expr
516 phi_trans_lookup (pre_expr e, basic_block pred)
519 struct expr_pred_trans_d ept;
523 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
524 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
529 return ((expr_pred_trans_t) *slot)->v;
533 /* Add the tuple mapping from {expression E, basic block PRED} to
534 value V, to the phi translation table. */
537 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
540 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
542 new_pair->pred = pred;
544 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
547 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
548 new_pair->hashcode, INSERT);
551 *slot = (void *) new_pair;
555 /* Add expression E to the expression set of value id V. */
558 add_to_value (unsigned int v, pre_expr e)
562 if (v >= VEC_length (bitmap_set_t, value_expressions))
564 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
568 set = VEC_index (bitmap_set_t, value_expressions, v);
571 set = bitmap_set_new ();
572 VEC_replace (bitmap_set_t, value_expressions, v, set);
575 bitmap_insert_into_set_1 (set, e, true);
578 /* Create a new bitmap set and return it. */
581 bitmap_set_new (void)
583 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
584 ret->expressions = BITMAP_ALLOC (&grand_bitmap_obstack);
585 ret->values = BITMAP_ALLOC (&grand_bitmap_obstack);
589 /* Return the value id for a PRE expression EXPR. */
592 get_expr_value_id (pre_expr expr)
599 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
602 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
603 add_to_value (id, expr);
608 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
610 return PRE_EXPR_NARY (expr)->value_id;
612 return PRE_EXPR_REFERENCE (expr)->value_id;
618 /* Remove an expression EXPR from a bitmapped set. */
621 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
623 unsigned int val = get_expr_value_id (expr);
624 if (!value_id_constant_p (val))
626 bitmap_clear_bit (set->values, val);
627 bitmap_clear_bit (set->expressions, get_expression_id (expr));
632 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
633 bool allow_constants)
635 unsigned int val = get_expr_value_id (expr);
636 if (allow_constants || !value_id_constant_p (val))
638 /* We specifically expect this and only this function to be able to
639 insert constants into a set. */
640 bitmap_set_bit (set->values, val);
641 bitmap_set_bit (set->expressions, get_or_alloc_expression_id (expr));
645 /* Insert an expression EXPR into a bitmapped set. */
648 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
650 bitmap_insert_into_set_1 (set, expr, false);
653 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
656 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
658 bitmap_copy (dest->expressions, orig->expressions);
659 bitmap_copy (dest->values, orig->values);
663 /* Free memory used up by SET. */
665 bitmap_set_free (bitmap_set_t set)
667 BITMAP_FREE (set->expressions);
668 BITMAP_FREE (set->values);
672 /* A comparison function for use in qsort to top sort a bitmap set. Simply
673 subtracts value ids, since they are created with leaves before
674 their parent users (IE topological order). */
677 value_id_compare (const void *pa, const void *pb)
679 const unsigned int vha = get_expr_value_id (*((const pre_expr *)pa));
680 const unsigned int vhb = get_expr_value_id (*((const pre_expr *)pb));
685 /* Generate an topological-ordered array of bitmap set SET. */
687 static VEC(pre_expr, heap) *
688 sorted_array_from_bitmap_set (bitmap_set_t set)
692 VEC(pre_expr, heap) *result = NULL;
694 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
695 VEC_safe_push (pre_expr, heap, result, expression_for_id (i));
697 qsort (VEC_address (pre_expr, result), VEC_length (pre_expr, result),
698 sizeof (pre_expr), value_id_compare);
703 /* Perform bitmapped set operation DEST &= ORIG. */
706 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
713 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
715 bitmap_and_into (dest->values, orig->values);
716 bitmap_copy (temp, dest->expressions);
717 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
719 pre_expr expr = expression_for_id (i);
720 unsigned int value_id = get_expr_value_id (expr);
721 if (!bitmap_bit_p (dest->values, value_id))
722 bitmap_clear_bit (dest->expressions, i);
728 /* Subtract all values and expressions contained in ORIG from DEST. */
731 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
733 bitmap_set_t result = bitmap_set_new ();
737 bitmap_and_compl (result->expressions, dest->expressions,
740 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
742 pre_expr expr = expression_for_id (i);
743 unsigned int value_id = get_expr_value_id (expr);
744 bitmap_set_bit (result->values, value_id);
750 /* Subtract all the values in bitmap set B from bitmap set A. */
753 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
757 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
759 bitmap_copy (temp, a->expressions);
760 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
762 pre_expr expr = expression_for_id (i);
763 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
764 bitmap_remove_from_set (a, expr);
770 /* Return true if bitmapped set SET contains the value VALUE_ID. */
773 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
775 if (value_id_constant_p (value_id))
778 if (!set || bitmap_empty_p (set->expressions))
781 return bitmap_bit_p (set->values, value_id);
785 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
787 return bitmap_bit_p (set->expressions, get_expression_id (expr));
790 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
793 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
796 bitmap_set_t exprset;
800 if (value_id_constant_p (lookfor))
803 if (!bitmap_set_contains_value (set, lookfor))
806 /* The number of expressions having a given value is usually
807 significantly less than the total number of expressions in SET.
808 Thus, rather than check, for each expression in SET, whether it
809 has the value LOOKFOR, we walk the reverse mapping that tells us
810 what expressions have a given value, and see if any of those
811 expressions are in our set. For large testcases, this is about
812 5-10x faster than walking the bitmap. If this is somehow a
813 significant lose for some cases, we can choose which set to walk
814 based on the set size. */
815 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
816 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
818 if (bitmap_bit_p (set->expressions, i))
820 bitmap_clear_bit (set->expressions, i);
821 bitmap_set_bit (set->expressions, get_expression_id (expr));
827 /* Return true if two bitmap sets are equal. */
830 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
832 return bitmap_equal_p (a->values, b->values);
835 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
836 and add it otherwise. */
839 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
841 unsigned int val = get_expr_value_id (expr);
843 if (bitmap_set_contains_value (set, val))
844 bitmap_set_replace_value (set, val, expr);
846 bitmap_insert_into_set (set, expr);
849 /* Insert EXPR into SET if EXPR's value is not already present in
853 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
855 unsigned int val = get_expr_value_id (expr);
857 if (value_id_constant_p (val))
860 if (!bitmap_set_contains_value (set, val))
861 bitmap_insert_into_set (set, expr);
864 /* Print out EXPR to outfile. */
867 print_pre_expr (FILE *outfile, const pre_expr expr)
872 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
875 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
880 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
881 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
882 for (i = 0; i < nary->length; i++)
884 print_generic_expr (outfile, nary->op[i], 0);
885 if (i != (unsigned) nary->length - 1)
886 fprintf (outfile, ",");
888 fprintf (outfile, "}");
894 vn_reference_op_t vro;
896 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
897 fprintf (outfile, "{");
899 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
902 if (vro->opcode != SSA_NAME
903 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
904 fprintf (outfile, "%s ", tree_code_name [vro->opcode]);
908 fprintf (outfile, "<");
909 print_generic_expr (outfile, vro->op0, 0);
912 fprintf (outfile, ",");
913 print_generic_expr (outfile, vro->op1, 0);
916 fprintf (outfile, ">");
918 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
919 fprintf (outfile, ",");
921 fprintf (outfile, "}");
926 void debug_pre_expr (pre_expr);
928 /* Like print_pre_expr but always prints to stderr. */
930 debug_pre_expr (pre_expr e)
932 print_pre_expr (stderr, e);
933 fprintf (stderr, "\n");
936 /* Print out SET to OUTFILE. */
939 print_bitmap_set (FILE *outfile, bitmap_set_t set,
940 const char *setname, int blockindex)
942 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
949 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
951 const pre_expr expr = expression_for_id (i);
954 fprintf (outfile, ", ");
956 print_pre_expr (outfile, expr);
958 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
961 fprintf (outfile, " }\n");
964 void debug_bitmap_set (bitmap_set_t);
967 debug_bitmap_set (bitmap_set_t set)
969 print_bitmap_set (stderr, set, "debug", 0);
972 /* Print out the expressions that have VAL to OUTFILE. */
975 print_value_expressions (FILE *outfile, unsigned int val)
977 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
981 sprintf (s, "%04d", val);
982 print_bitmap_set (outfile, set, s, 0);
988 debug_value_expressions (unsigned int val)
990 print_value_expressions (stderr, val);
993 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
997 get_or_alloc_expr_for_constant (tree constant)
999 unsigned int result_id;
1000 unsigned int value_id;
1001 pre_expr newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1002 newexpr->kind = CONSTANT;
1003 PRE_EXPR_CONSTANT (newexpr) = constant;
1004 result_id = lookup_expression_id (newexpr);
1007 pool_free (pre_expr_pool, newexpr);
1008 newexpr = expression_for_id (result_id);
1011 value_id = get_or_alloc_constant_value_id (constant);
1012 get_or_alloc_expression_id (newexpr);
1013 add_to_value (value_id, newexpr);
1017 /* Given a value id V, find the actual tree representing the constant
1018 value if there is one, and return it. Return NULL if we can't find
1022 get_constant_for_value_id (unsigned int v)
1024 if (value_id_constant_p (v))
1028 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1030 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1032 pre_expr expr = expression_for_id (i);
1033 if (expr->kind == CONSTANT)
1034 return PRE_EXPR_CONSTANT (expr);
1040 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1041 Currently only supports constants and SSA_NAMES. */
1043 get_or_alloc_expr_for (tree t)
1045 if (TREE_CODE (t) == SSA_NAME)
1046 return get_or_alloc_expr_for_name (t);
1047 else if (is_gimple_min_invariant (t))
1048 return get_or_alloc_expr_for_constant (t);
1051 /* More complex expressions can result from SCCVN expression
1052 simplification that inserts values for them. As they all
1053 do not have VOPs the get handled by the nary ops struct. */
1054 vn_nary_op_t result;
1055 unsigned int result_id;
1056 vn_nary_op_lookup (t, &result);
1059 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1061 PRE_EXPR_NARY (e) = result;
1062 result_id = lookup_expression_id (e);
1065 pool_free (pre_expr_pool, e);
1066 e = expression_for_id (result_id);
1069 alloc_expression_id (e);
1076 /* Return the folded version of T if T, when folded, is a gimple
1077 min_invariant. Otherwise, return T. */
1080 fully_constant_expression (pre_expr e)
1088 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1089 switch (TREE_CODE_CLASS (nary->opcode))
1093 /* We have to go from trees to pre exprs to value ids to
1095 tree naryop0 = nary->op[0];
1096 tree naryop1 = nary->op[1];
1097 tree const0, const1, result;
1098 if (is_gimple_min_invariant (naryop0))
1102 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1103 unsigned int vrep0 = get_expr_value_id (rep0);
1104 const0 = get_constant_for_value_id (vrep0);
1106 if (is_gimple_min_invariant (naryop1))
1110 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1111 unsigned int vrep1 = get_expr_value_id (rep1);
1112 const1 = get_constant_for_value_id (vrep1);
1115 if (const0 && const1)
1117 tree type1 = TREE_TYPE (nary->op[0]);
1118 tree type2 = TREE_TYPE (nary->op[1]);
1119 const0 = fold_convert (type1, const0);
1120 const1 = fold_convert (type2, const1);
1121 result = fold_binary (nary->opcode, nary->type, const0,
1124 if (result && is_gimple_min_invariant (result))
1125 return get_or_alloc_expr_for_constant (result);
1130 /* We have to go from trees to pre exprs to value ids to
1132 tree naryop0 = nary->op[0];
1133 tree const0, result;
1134 if (is_gimple_min_invariant (naryop0))
1138 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1139 unsigned int vrep0 = get_expr_value_id (rep0);
1140 const0 = get_constant_for_value_id (vrep0);
1145 tree type1 = TREE_TYPE (nary->op[0]);
1146 const0 = fold_convert (type1, const0);
1147 result = fold_unary (nary->opcode, nary->type, const0);
1150 if (result && is_gimple_min_invariant (result))
1151 return get_or_alloc_expr_for_constant (result);
1164 /* Translate the vuses in the VUSES vector backwards through phi nodes
1165 in PHIBLOCK, so that they have the value they would have in
1168 static VEC(tree, gc) *
1169 translate_vuses_through_block (VEC (tree, gc) *vuses,
1170 basic_block phiblock,
1174 VEC(tree, gc) *result = NULL;
1177 for (i = 0; VEC_iterate (tree, vuses, i, oldvuse); i++)
1179 gimple phi = SSA_NAME_DEF_STMT (oldvuse);
1180 if (gimple_code (phi) == GIMPLE_PHI
1181 && gimple_bb (phi) == phiblock)
1183 edge e = find_edge (block, gimple_bb (phi));
1186 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1190 result = VEC_copy (tree, gc, vuses);
1191 VEC_replace (tree, result, i, def);
1197 /* We avoid creating a new copy of the vuses unless something
1198 actually changed, so result can be NULL. */
1201 sort_vuses (result);
1208 /* Like find_leader, but checks for the value existing in SET1 *or*
1209 SET2. This is used to avoid making a set consisting of the union
1210 of PA_IN and ANTIC_IN during insert. */
1212 static inline pre_expr
1213 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1217 result = bitmap_find_leader (set1, val, NULL);
1218 if (!result && set2)
1219 result = bitmap_find_leader (set2, val, NULL);
1223 /* Get the tree type for our PRE expression e. */
1226 get_expr_type (const pre_expr e)
1231 return TREE_TYPE (PRE_EXPR_NAME (e));
1233 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1236 vn_reference_op_t vro;
1238 gcc_assert (PRE_EXPR_REFERENCE (e)->operands);
1239 vro = VEC_index (vn_reference_op_s,
1240 PRE_EXPR_REFERENCE (e)->operands,
1242 /* We don't store type along with COMPONENT_REF because it is
1243 always the same as FIELD_DECL's type. */
1246 gcc_assert (vro->opcode == COMPONENT_REF);
1247 return TREE_TYPE (vro->op0);
1253 return PRE_EXPR_NARY (e)->type;
1258 /* Get a representative SSA_NAME for a given expression.
1259 Since all of our sub-expressions are treated as values, we require
1260 them to be SSA_NAME's for simplicity.
1261 Prior versions of GVNPRE used to use "value handles" here, so that
1262 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1263 either case, the operands are really values (IE we do not expect
1264 them to be usable without finding leaders). */
1267 get_representative_for (const pre_expr e)
1271 unsigned int value_id = get_expr_value_id (e);
1276 return PRE_EXPR_NAME (e);
1278 return PRE_EXPR_CONSTANT (e);
1282 /* Go through all of the expressions representing this value
1283 and pick out an SSA_NAME. */
1286 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1288 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1290 pre_expr rep = expression_for_id (i);
1291 if (rep->kind == NAME)
1292 return PRE_EXPR_NAME (rep);
1297 /* If we reached here we couldn't find an SSA_NAME. This can
1298 happen when we've discovered a value that has never appeared in
1299 the program as set to an SSA_NAME, most likely as the result of
1304 "Could not find SSA_NAME representative for expression:");
1305 print_pre_expr (dump_file, e);
1306 fprintf (dump_file, "\n");
1309 exprtype = get_expr_type (e);
1311 /* Build and insert the assignment of the end result to the temporary
1312 that we will return. */
1313 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1315 pretemp = create_tmp_var (exprtype, "pretmp");
1316 get_var_ann (pretemp);
1319 name = make_ssa_name (pretemp, gimple_build_nop ());
1320 VN_INFO_GET (name)->value_id = value_id;
1321 if (e->kind == CONSTANT)
1322 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1324 VN_INFO (name)->valnum = name;
1326 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1329 fprintf (dump_file, "Created SSA_NAME representative ");
1330 print_generic_expr (dump_file, name, 0);
1331 fprintf (dump_file, " for expression:");
1332 print_pre_expr (dump_file, e);
1333 fprintf (dump_file, "\n");
1342 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1343 the phis in PRED. SEEN is a bitmap saying which expression we have
1344 translated since we started translation of the toplevel expression.
1345 Return NULL if we can't find a leader for each part of the
1346 translated expression. */
1349 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1350 basic_block pred, basic_block phiblock, bitmap seen)
1352 pre_expr oldexpr = expr;
1358 if (value_id_constant_p (get_expr_value_id (expr)))
1361 phitrans = phi_trans_lookup (expr, pred);
1365 /* Prevent cycles when we have recursively dependent leaders. This
1366 can only happen when phi translating the maximal set. */
1369 unsigned int expr_id = get_expression_id (expr);
1370 if (bitmap_bit_p (seen, expr_id))
1372 bitmap_set_bit (seen, expr_id);
1377 /* Constants contain no values that need translation. */
1384 bool changed = false;
1385 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1386 struct vn_nary_op_s newnary;
1387 /* The NARY structure is only guaranteed to have been
1388 allocated to the nary->length operands. */
1389 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1390 - sizeof (tree) * (4 - nary->length)));
1392 for (i = 0; i < newnary.length; i++)
1394 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1398 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1399 pre_expr leader = find_leader_in_sets (op_val_id, set1, set2);
1400 pre_expr result = phi_translate_1 (leader, set1, set2,
1401 pred, phiblock, seen);
1402 if (result && result != leader)
1404 tree name = get_representative_for (result);
1407 newnary.op[i] = name;
1412 changed |= newnary.op[i] != nary->op[i];
1419 tree result = vn_nary_op_lookup_pieces (newnary.length,
1427 unsigned int new_val_id;
1429 expr = (pre_expr) pool_alloc (pre_expr_pool);
1432 if (result && is_gimple_min_invariant (result))
1433 return get_or_alloc_expr_for_constant (result);
1438 PRE_EXPR_NARY (expr) = nary;
1439 constant = fully_constant_expression (expr);
1440 if (constant != expr)
1443 new_val_id = nary->value_id;
1444 get_or_alloc_expression_id (expr);
1448 new_val_id = get_next_value_id ();
1449 VEC_safe_grow_cleared (bitmap_set_t, heap,
1451 get_max_value_id() + 1);
1452 nary = vn_nary_op_insert_pieces (newnary.length,
1459 result, new_val_id);
1460 PRE_EXPR_NARY (expr) = nary;
1461 constant = fully_constant_expression (expr);
1462 if (constant != expr)
1464 get_or_alloc_expression_id (expr);
1466 add_to_value (new_val_id, expr);
1468 phi_trans_add (oldexpr, expr, pred);
1474 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1475 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1476 VEC (tree, gc) *vuses = ref->vuses;
1477 VEC (tree, gc) *newvuses = vuses;
1478 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1479 bool changed = false;
1481 vn_reference_op_t operand;
1482 vn_reference_t newref;
1484 for (i = 0; VEC_iterate (vn_reference_op_s, operands, i, operand); i++)
1488 tree oldop0 = operand->op0;
1489 tree oldop1 = operand->op1;
1490 tree oldop2 = operand->op2;
1494 tree type = operand->type;
1495 vn_reference_op_s newop = *operand;
1497 if (op0 && TREE_CODE (op0) == SSA_NAME)
1499 unsigned int op_val_id = VN_INFO (op0)->value_id;
1500 leader = find_leader_in_sets (op_val_id, set1, set2);
1501 opresult = phi_translate_1 (leader, set1, set2,
1502 pred, phiblock, seen);
1503 if (opresult && opresult != leader)
1505 tree name = get_representative_for (opresult);
1513 changed |= op0 != oldop0;
1515 if (op1 && TREE_CODE (op1) == SSA_NAME)
1517 unsigned int op_val_id = VN_INFO (op1)->value_id;
1518 leader = find_leader_in_sets (op_val_id, set1, set2);
1519 opresult = phi_translate_1 (leader, set1, set2,
1520 pred, phiblock, seen);
1521 if (opresult && opresult != leader)
1523 tree name = get_representative_for (opresult);
1531 changed |= op1 != oldop1;
1532 if (op2 && TREE_CODE (op2) == SSA_NAME)
1534 unsigned int op_val_id = VN_INFO (op2)->value_id;
1535 leader = find_leader_in_sets (op_val_id, set1, set2);
1536 opresult = phi_translate_1 (leader, set1, set2,
1537 pred, phiblock, seen);
1538 if (opresult && opresult != leader)
1540 tree name = get_representative_for (opresult);
1548 changed |= op2 != oldop2;
1551 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1552 /* We may have changed from an SSA_NAME to a constant */
1553 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1554 newop.opcode = TREE_CODE (op0);
1559 VEC_replace (vn_reference_op_s, newoperands, i, &newop);
1562 newvuses = translate_vuses_through_block (vuses, phiblock, pred);
1563 changed |= newvuses != vuses;
1567 tree result = vn_reference_lookup_pieces (newvuses,
1570 unsigned int new_val_id;
1573 VEC_free (vn_reference_op_s, heap, newoperands);
1575 if (result && is_gimple_min_invariant (result))
1576 return get_or_alloc_expr_for_constant (result);
1578 expr = (pre_expr) pool_alloc (pre_expr_pool);
1579 expr->kind = REFERENCE;
1584 PRE_EXPR_REFERENCE (expr) = newref;
1585 new_val_id = newref->value_id;
1586 get_or_alloc_expression_id (expr);
1590 new_val_id = get_next_value_id ();
1591 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
1592 get_max_value_id() + 1);
1593 newref = vn_reference_insert_pieces (newvuses,
1595 result, new_val_id);
1596 PRE_EXPR_REFERENCE (expr) = newref;
1597 get_or_alloc_expression_id (expr);
1599 add_to_value (new_val_id, expr);
1601 phi_trans_add (oldexpr, expr, pred);
1610 tree name = PRE_EXPR_NAME (expr);
1612 def_stmt = SSA_NAME_DEF_STMT (name);
1613 if (gimple_code (def_stmt) == GIMPLE_PHI
1614 && gimple_bb (def_stmt) == phiblock)
1619 e = find_edge (pred, gimple_bb (phi));
1622 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1625 /* Handle constant. */
1626 if (is_gimple_min_invariant (def))
1627 return get_or_alloc_expr_for_constant (def);
1629 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1632 newexpr = get_or_alloc_expr_for_name (def);
1643 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1645 Return NULL if we can't find a leader for each part of the
1646 translated expression. */
1649 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1650 basic_block pred, basic_block phiblock)
1652 bitmap_clear (seen_during_translate);
1653 return phi_translate_1 (expr, set1, set2, pred, phiblock,
1654 seen_during_translate);
1657 /* For each expression in SET, translate the values through phi nodes
1658 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1659 expressions in DEST. */
1662 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1663 basic_block phiblock)
1665 VEC (pre_expr, heap) *exprs;
1669 if (!phi_nodes (phiblock))
1671 bitmap_set_copy (dest, set);
1675 exprs = sorted_array_from_bitmap_set (set);
1676 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1678 pre_expr translated;
1679 translated = phi_translate (expr, set, NULL, pred, phiblock);
1681 /* Don't add constants or empty translations to the cache, since
1682 we won't look them up that way, or use the result, anyway. */
1683 if (translated && !value_id_constant_p (get_expr_value_id (translated)))
1684 phi_trans_add (expr, translated, pred);
1686 if (translated != NULL)
1687 bitmap_value_insert_into_set (dest, translated);
1689 VEC_free (pre_expr, heap, exprs);
1692 /* Find the leader for a value (i.e., the name representing that
1693 value) in a given set, and return it. If STMT is non-NULL it
1694 makes sure the defining statement for the leader dominates it.
1695 Return NULL if no leader is found. */
1698 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1700 if (value_id_constant_p (val))
1704 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1706 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1708 pre_expr expr = expression_for_id (i);
1709 if (expr->kind == CONSTANT)
1713 if (bitmap_set_contains_value (set, val))
1715 /* Rather than walk the entire bitmap of expressions, and see
1716 whether any of them has the value we are looking for, we look
1717 at the reverse mapping, which tells us the set of expressions
1718 that have a given value (IE value->expressions with that
1719 value) and see if any of those expressions are in our set.
1720 The number of expressions per value is usually significantly
1721 less than the number of expressions in the set. In fact, for
1722 large testcases, doing it this way is roughly 5-10x faster
1723 than walking the bitmap.
1724 If this is somehow a significant lose for some cases, we can
1725 choose which set to walk based on which set is smaller. */
1728 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1730 EXECUTE_IF_AND_IN_BITMAP (exprset->expressions,
1731 set->expressions, 0, i, bi)
1733 pre_expr val = expression_for_id (i);
1734 /* At the point where stmt is not null, there should always
1735 be an SSA_NAME first in the list of expressions. */
1738 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1739 if (gimple_code (def_stmt) != GIMPLE_PHI
1740 && gimple_bb (def_stmt) == gimple_bb (stmt)
1741 && gimple_uid (def_stmt) >= gimple_uid (stmt))
1750 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1751 BLOCK by seeing if it is not killed in the block. Note that we are
1752 only determining whether there is a store that kills it. Because
1753 of the order in which clean iterates over values, we are guaranteed
1754 that altered operands will have caused us to be eliminated from the
1755 ANTIC_IN set already. */
1758 value_dies_in_block_x (pre_expr expr, basic_block block)
1762 VEC (tree, gc) *vuses = PRE_EXPR_REFERENCE (expr)->vuses;
1764 /* Conservatively, a value dies if it's vuses are defined in this
1765 block, unless they come from phi nodes (which are merge operations,
1766 rather than stores. */
1767 for (i = 0; VEC_iterate (tree, vuses, i, vuse); i++)
1769 gimple def = SSA_NAME_DEF_STMT (vuse);
1771 if (gimple_bb (def) != block)
1773 if (gimple_code (def) == GIMPLE_PHI)
1781 #define union_contains_value(SET1, SET2, VAL) \
1782 (bitmap_set_contains_value ((SET1), (VAL)) \
1783 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
1785 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
1788 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
1789 vn_reference_op_t vro)
1791 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
1793 struct pre_expr_d temp;
1796 PRE_EXPR_NAME (&temp) = vro->op0;
1797 temp.id = lookup_expression_id (&temp);
1800 if (!union_contains_value (set1, set2,
1801 get_expr_value_id (&temp)))
1804 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
1806 struct pre_expr_d temp;
1809 PRE_EXPR_NAME (&temp) = vro->op1;
1810 temp.id = lookup_expression_id (&temp);
1813 if (!union_contains_value (set1, set2,
1814 get_expr_value_id (&temp)))
1818 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
1820 struct pre_expr_d temp;
1823 PRE_EXPR_NAME (&temp) = vro->op2;
1824 temp.id = lookup_expression_id (&temp);
1827 if (!union_contains_value (set1, set2,
1828 get_expr_value_id (&temp)))
1835 /* Determine if the expression EXPR is valid in SET1 U SET2.
1836 ONLY SET2 CAN BE NULL.
1837 This means that we have a leader for each part of the expression
1838 (if it consists of values), or the expression is an SSA_NAME.
1839 For loads/calls, we also see if the vuses are killed in this block.
1843 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
1849 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
1853 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1854 for (i = 0; i < nary->length; i++)
1856 if (TREE_CODE (nary->op[i]) == SSA_NAME)
1858 struct pre_expr_d temp;
1861 PRE_EXPR_NAME (&temp) = nary->op[i];
1862 temp.id = lookup_expression_id (&temp);
1865 if (!union_contains_value (set1, set2,
1866 get_expr_value_id (&temp)))
1875 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1876 vn_reference_op_t vro;
1879 for (i = 0; VEC_iterate (vn_reference_op_s, ref->operands, i, vro); i++)
1881 if (!vro_valid_in_sets (set1, set2, vro))
1884 return !value_dies_in_block_x (expr, block);
1891 /* Clean the set of expressions that are no longer valid in SET1 or
1892 SET2. This means expressions that are made up of values we have no
1893 leaders for in SET1 or SET2. This version is used for partial
1894 anticipation, which means it is not valid in either ANTIC_IN or
1898 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
1900 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
1904 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1906 if (!valid_in_sets (set1, set2, expr, block))
1907 bitmap_remove_from_set (set1, expr);
1909 VEC_free (pre_expr, heap, exprs);
1912 /* Clean the set of expressions that are no longer valid in SET. This
1913 means expressions that are made up of values we have no leaders for
1917 clean (bitmap_set_t set, basic_block block)
1919 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
1923 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1925 if (!valid_in_sets (set, NULL, expr, block))
1926 bitmap_remove_from_set (set, expr);
1928 VEC_free (pre_expr, heap, exprs);
1931 static sbitmap has_abnormal_preds;
1933 /* List of blocks that may have changed during ANTIC computation and
1934 thus need to be iterated over. */
1936 static sbitmap changed_blocks;
1938 /* Decide whether to defer a block for a later iteration, or PHI
1939 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
1940 should defer the block, and true if we processed it. */
1943 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
1944 basic_block block, basic_block phiblock)
1946 if (!BB_VISITED (phiblock))
1948 SET_BIT (changed_blocks, block->index);
1949 BB_VISITED (block) = 0;
1950 BB_DEFERRED (block) = 1;
1954 phi_translate_set (dest, source, block, phiblock);
1958 /* Compute the ANTIC set for BLOCK.
1960 If succs(BLOCK) > 1 then
1961 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
1962 else if succs(BLOCK) == 1 then
1963 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
1965 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
1969 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
1971 bool changed = false;
1972 bitmap_set_t S, old, ANTIC_OUT;
1978 old = ANTIC_OUT = S = NULL;
1979 BB_VISITED (block) = 1;
1981 /* If any edges from predecessors are abnormal, antic_in is empty,
1983 if (block_has_abnormal_pred_edge)
1984 goto maybe_dump_sets;
1986 old = ANTIC_IN (block);
1987 ANTIC_OUT = bitmap_set_new ();
1989 /* If the block has no successors, ANTIC_OUT is empty. */
1990 if (EDGE_COUNT (block->succs) == 0)
1992 /* If we have one successor, we could have some phi nodes to
1993 translate through. */
1994 else if (single_succ_p (block))
1996 basic_block succ_bb = single_succ (block);
1998 /* We trade iterations of the dataflow equations for having to
1999 phi translate the maximal set, which is incredibly slow
2000 (since the maximal set often has 300+ members, even when you
2001 have a small number of blocks).
2002 Basically, we defer the computation of ANTIC for this block
2003 until we have processed it's successor, which will inevitably
2004 have a *much* smaller set of values to phi translate once
2005 clean has been run on it.
2006 The cost of doing this is that we technically perform more
2007 iterations, however, they are lower cost iterations.
2009 Timings for PRE on tramp3d-v4:
2010 without maximal set fix: 11 seconds
2011 with maximal set fix/without deferring: 26 seconds
2012 with maximal set fix/with deferring: 11 seconds
2015 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2019 goto maybe_dump_sets;
2022 /* If we have multiple successors, we take the intersection of all of
2023 them. Note that in the case of loop exit phi nodes, we may have
2024 phis to translate through. */
2027 VEC(basic_block, heap) * worklist;
2029 basic_block bprime, first;
2031 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2032 FOR_EACH_EDGE (e, ei, block->succs)
2033 VEC_quick_push (basic_block, worklist, e->dest);
2034 first = VEC_index (basic_block, worklist, 0);
2036 if (phi_nodes (first))
2038 bitmap_set_t from = ANTIC_IN (first);
2040 if (!BB_VISITED (first))
2042 phi_translate_set (ANTIC_OUT, from, block, first);
2046 if (!BB_VISITED (first))
2047 bitmap_set_copy (ANTIC_OUT, maximal_set);
2049 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2052 for (i = 1; VEC_iterate (basic_block, worklist, i, bprime); i++)
2054 if (phi_nodes (bprime))
2056 bitmap_set_t tmp = bitmap_set_new ();
2057 bitmap_set_t from = ANTIC_IN (bprime);
2059 if (!BB_VISITED (bprime))
2061 phi_translate_set (tmp, from, block, bprime);
2062 bitmap_set_and (ANTIC_OUT, tmp);
2063 bitmap_set_free (tmp);
2067 if (!BB_VISITED (bprime))
2068 bitmap_set_and (ANTIC_OUT, maximal_set);
2070 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2073 VEC_free (basic_block, heap, worklist);
2076 /* Generate ANTIC_OUT - TMP_GEN. */
2077 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2079 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2080 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2083 /* Then union in the ANTIC_OUT - TMP_GEN values,
2084 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2085 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2086 bitmap_value_insert_into_set (ANTIC_IN (block),
2087 expression_for_id (bii));
2089 clean (ANTIC_IN (block), block);
2091 /* !old->expressions can happen when we deferred a block. */
2092 if (!old->expressions || !bitmap_set_equal (old, ANTIC_IN (block)))
2095 SET_BIT (changed_blocks, block->index);
2096 FOR_EACH_EDGE (e, ei, block->preds)
2097 SET_BIT (changed_blocks, e->src->index);
2100 RESET_BIT (changed_blocks, block->index);
2103 if (dump_file && (dump_flags & TDF_DETAILS))
2105 if (!BB_DEFERRED (block) || BB_VISITED (block))
2108 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2110 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2114 print_bitmap_set (dump_file, S, "S", block->index);
2119 "Block %d was deferred for a future iteration.\n",
2124 bitmap_set_free (old);
2126 bitmap_set_free (S);
2128 bitmap_set_free (ANTIC_OUT);
2132 /* Compute PARTIAL_ANTIC for BLOCK.
2134 If succs(BLOCK) > 1 then
2135 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2136 in ANTIC_OUT for all succ(BLOCK)
2137 else if succs(BLOCK) == 1 then
2138 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2140 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2145 compute_partial_antic_aux (basic_block block,
2146 bool block_has_abnormal_pred_edge)
2148 bool changed = false;
2149 bitmap_set_t old_PA_IN;
2150 bitmap_set_t PA_OUT;
2153 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2155 old_PA_IN = PA_OUT = NULL;
2157 /* If any edges from predecessors are abnormal, antic_in is empty,
2159 if (block_has_abnormal_pred_edge)
2160 goto maybe_dump_sets;
2162 /* If there are too many partially anticipatable values in the
2163 block, phi_translate_set can take an exponential time: stop
2164 before the translation starts. */
2166 && single_succ_p (block)
2167 && bitmap_count_bits (PA_IN (single_succ (block))->values) > max_pa)
2168 goto maybe_dump_sets;
2170 old_PA_IN = PA_IN (block);
2171 PA_OUT = bitmap_set_new ();
2173 /* If the block has no successors, ANTIC_OUT is empty. */
2174 if (EDGE_COUNT (block->succs) == 0)
2176 /* If we have one successor, we could have some phi nodes to
2177 translate through. Note that we can't phi translate across DFS
2178 back edges in partial antic, because it uses a union operation on
2179 the successors. For recurrences like IV's, we will end up
2180 generating a new value in the set on each go around (i + 3 (VH.1)
2181 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2182 else if (single_succ_p (block))
2184 basic_block succ = single_succ (block);
2185 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2186 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2188 /* If we have multiple successors, we take the union of all of
2192 VEC(basic_block, heap) * worklist;
2196 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2197 FOR_EACH_EDGE (e, ei, block->succs)
2199 if (e->flags & EDGE_DFS_BACK)
2201 VEC_quick_push (basic_block, worklist, e->dest);
2203 if (VEC_length (basic_block, worklist) > 0)
2205 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2210 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2211 bitmap_value_insert_into_set (PA_OUT,
2212 expression_for_id (i));
2213 if (phi_nodes (bprime))
2215 bitmap_set_t pa_in = bitmap_set_new ();
2216 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2217 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2218 bitmap_value_insert_into_set (PA_OUT,
2219 expression_for_id (i));
2220 bitmap_set_free (pa_in);
2223 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2224 bitmap_value_insert_into_set (PA_OUT,
2225 expression_for_id (i));
2228 VEC_free (basic_block, heap, worklist);
2231 /* PA_IN starts with PA_OUT - TMP_GEN.
2232 Then we subtract things from ANTIC_IN. */
2233 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2235 /* For partial antic, we want to put back in the phi results, since
2236 we will properly avoid making them partially antic over backedges. */
2237 bitmap_ior_into (PA_IN (block)->values, PHI_GEN (block)->values);
2238 bitmap_ior_into (PA_IN (block)->expressions, PHI_GEN (block)->expressions);
2240 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2241 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2243 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2245 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2248 SET_BIT (changed_blocks, block->index);
2249 FOR_EACH_EDGE (e, ei, block->preds)
2250 SET_BIT (changed_blocks, e->src->index);
2253 RESET_BIT (changed_blocks, block->index);
2256 if (dump_file && (dump_flags & TDF_DETAILS))
2259 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2261 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2264 bitmap_set_free (old_PA_IN);
2266 bitmap_set_free (PA_OUT);
2270 /* Compute ANTIC and partial ANTIC sets. */
2273 compute_antic (void)
2275 bool changed = true;
2276 int num_iterations = 0;
2280 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2281 We pre-build the map of blocks with incoming abnormal edges here. */
2282 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2283 sbitmap_zero (has_abnormal_preds);
2290 FOR_EACH_EDGE (e, ei, block->preds)
2292 e->flags &= ~EDGE_DFS_BACK;
2293 if (e->flags & EDGE_ABNORMAL)
2295 SET_BIT (has_abnormal_preds, block->index);
2300 BB_VISITED (block) = 0;
2301 BB_DEFERRED (block) = 0;
2302 /* While we are here, give empty ANTIC_IN sets to each block. */
2303 ANTIC_IN (block) = bitmap_set_new ();
2304 PA_IN (block) = bitmap_set_new ();
2307 /* At the exit block we anticipate nothing. */
2308 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2309 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2310 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2312 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2313 sbitmap_ones (changed_blocks);
2316 if (dump_file && (dump_flags & TDF_DETAILS))
2317 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2320 for (i = 0; i < last_basic_block - NUM_FIXED_BLOCKS; i++)
2322 if (TEST_BIT (changed_blocks, postorder[i]))
2324 basic_block block = BASIC_BLOCK (postorder[i]);
2325 changed |= compute_antic_aux (block,
2326 TEST_BIT (has_abnormal_preds,
2330 #ifdef ENABLE_CHECKING
2331 /* Theoretically possible, but *highly* unlikely. */
2332 gcc_assert (num_iterations < 500);
2336 statistics_histogram_event (cfun, "compute_antic iterations",
2339 if (do_partial_partial)
2341 sbitmap_ones (changed_blocks);
2342 mark_dfs_back_edges ();
2347 if (dump_file && (dump_flags & TDF_DETAILS))
2348 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2351 for (i = 0; i < last_basic_block - NUM_FIXED_BLOCKS; i++)
2353 if (TEST_BIT (changed_blocks, postorder[i]))
2355 basic_block block = BASIC_BLOCK (postorder[i]);
2357 |= compute_partial_antic_aux (block,
2358 TEST_BIT (has_abnormal_preds,
2362 #ifdef ENABLE_CHECKING
2363 /* Theoretically possible, but *highly* unlikely. */
2364 gcc_assert (num_iterations < 500);
2367 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2370 sbitmap_free (has_abnormal_preds);
2371 sbitmap_free (changed_blocks);
2374 /* Return true if we can value number the call in STMT. This is true
2375 if we have a pure or constant call. */
2378 can_value_number_call (gimple stmt)
2380 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2385 /* Return true if OP is an exception handler related operation, such as
2386 FILTER_EXPR or EXC_PTR_EXPR. */
2389 is_exception_related (gimple stmt)
2391 return (is_gimple_assign (stmt)
2392 && (gimple_assign_rhs_code (stmt) == FILTER_EXPR
2393 || gimple_assign_rhs_code (stmt) == EXC_PTR_EXPR));
2396 /* Return true if OP is a tree which we can perform PRE on
2397 on. This may not match the operations we can value number, but in
2398 a perfect world would. */
2401 can_PRE_operation (tree op)
2403 return UNARY_CLASS_P (op)
2404 || BINARY_CLASS_P (op)
2405 || COMPARISON_CLASS_P (op)
2406 || TREE_CODE (op) == INDIRECT_REF
2407 || TREE_CODE (op) == COMPONENT_REF
2408 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2409 || TREE_CODE (op) == CALL_EXPR
2410 || TREE_CODE (op) == ARRAY_REF;
2414 /* Inserted expressions are placed onto this worklist, which is used
2415 for performing quick dead code elimination of insertions we made
2416 that didn't turn out to be necessary. */
2417 static VEC(gimple,heap) *inserted_exprs;
2419 /* Pool allocated fake store expressions are placed onto this
2420 worklist, which, after performing dead code elimination, is walked
2421 to see which expressions need to be put into GC'able memory */
2422 static VEC(gimple, heap) *need_creation;
2424 /* The actual worker for create_component_ref_by_pieces. */
2427 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2428 unsigned int *operand, gimple_seq *stmts,
2431 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2435 switch (currop->opcode)
2439 tree folded, sc = currop->op1;
2440 unsigned int nargs = 0;
2441 tree *args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2442 ref->operands) - 1);
2443 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2445 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2450 folded = build_call_array (currop->type,
2451 TREE_CODE (currop->op0) == FUNCTION_DECL
2452 ? build_fold_addr_expr (currop->op0)
2458 pre_expr scexpr = get_or_alloc_expr_for (sc);
2459 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2462 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2470 gcc_assert (is_gimple_min_invariant (currop->op0));
2476 case VIEW_CONVERT_EXPR:
2479 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2484 folded = fold_build1 (currop->opcode, currop->type,
2489 case ALIGN_INDIRECT_REF:
2490 case MISALIGNED_INDIRECT_REF:
2494 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2499 genop1 = fold_convert (build_pointer_type (currop->type),
2502 folded = fold_build1 (currop->opcode, currop->type,
2510 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2512 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2513 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2519 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2522 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2525 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2530 /* For array ref vn_reference_op's, operand 1 of the array ref
2531 is op0 of the reference op and operand 3 of the array ref is
2533 case ARRAY_RANGE_REF:
2537 tree genop1 = currop->op0;
2539 tree genop2 = currop->op1;
2542 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2546 op1expr = get_or_alloc_expr_for (genop1);
2547 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2552 op2expr = get_or_alloc_expr_for (genop2);
2553 genop2 = find_or_generate_expression (block, op2expr, stmts,
2559 genop3 = currop->op2;
2560 return build4 (currop->opcode, currop->type, genop0, genop1,
2567 tree genop2 = currop->op1;
2569 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2573 /* op1 should be a FIELD_DECL, which are represented by
2578 op2expr = get_or_alloc_expr_for (genop2);
2579 genop2 = find_or_generate_expression (block, op2expr, stmts,
2585 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2591 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2592 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2613 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2614 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2615 trying to rename aggregates into ssa form directly, which is a no no.
2617 Thus, this routine doesn't create temporaries, it just builds a
2618 single access expression for the array, calling
2619 find_or_generate_expression to build the innermost pieces.
2621 This function is a subroutine of create_expression_by_pieces, and
2622 should not be called on it's own unless you really know what you
2626 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2627 gimple_seq *stmts, gimple domstmt)
2629 unsigned int op = 0;
2630 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2633 /* Find a leader for an expression, or generate one using
2634 create_expression_by_pieces if it's ANTIC but
2636 BLOCK is the basic_block we are looking for leaders in.
2637 EXPR is the expression to find a leader or generate for.
2638 STMTS is the statement list to put the inserted expressions on.
2639 Returns the SSA_NAME of the LHS of the generated expression or the
2641 DOMSTMT if non-NULL is a statement that should be dominated by
2642 all uses in the generated expression. If DOMSTMT is non-NULL this
2643 routine can fail and return NULL_TREE. Otherwise it will assert
2647 find_or_generate_expression (basic_block block, pre_expr expr,
2648 gimple_seq *stmts, gimple domstmt)
2650 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2651 get_expr_value_id (expr), domstmt);
2655 if (leader->kind == NAME)
2656 genop = PRE_EXPR_NAME (leader);
2657 else if (leader->kind == CONSTANT)
2658 genop = PRE_EXPR_CONSTANT (leader);
2661 /* If it's still NULL, it must be a complex expression, so generate
2665 bitmap_set_t exprset;
2666 unsigned int lookfor = get_expr_value_id (expr);
2667 bool handled = false;
2671 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2672 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2674 pre_expr temp = expression_for_id (i);
2675 if (temp->kind != NAME)
2678 genop = create_expression_by_pieces (block, temp, stmts,
2680 get_expr_type (expr));
2684 if (!handled && domstmt)
2687 gcc_assert (handled);
2692 #define NECESSARY GF_PLF_1
2694 /* Create an expression in pieces, so that we can handle very complex
2695 expressions that may be ANTIC, but not necessary GIMPLE.
2696 BLOCK is the basic block the expression will be inserted into,
2697 EXPR is the expression to insert (in value form)
2698 STMTS is a statement list to append the necessary insertions into.
2700 This function will die if we hit some value that shouldn't be
2701 ANTIC but is (IE there is no leader for it, or its components).
2702 This function may also generate expressions that are themselves
2703 partially or fully redundant. Those that are will be either made
2704 fully redundant during the next iteration of insert (for partially
2705 redundant ones), or eliminated by eliminate (for fully redundant
2708 If DOMSTMT is non-NULL then we make sure that all uses in the
2709 expressions dominate that statement. In this case the function
2710 can return NULL_TREE to signal failure. */
2713 create_expression_by_pieces (basic_block block, pre_expr expr,
2714 gimple_seq *stmts, gimple domstmt, tree type)
2717 tree folded, newexpr;
2718 gimple_seq forced_stmts;
2719 unsigned int value_id;
2720 gimple_stmt_iterator gsi;
2721 tree exprtype = type ? type : get_expr_type (expr);
2727 /* We may hit the NAME/CONSTANT case if we have to convert types
2728 that value numbering saw through. */
2730 folded = PRE_EXPR_NAME (expr);
2733 folded = PRE_EXPR_CONSTANT (expr);
2737 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2738 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
2743 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2744 switch (nary->length)
2748 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2749 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
2750 tree genop1 = find_or_generate_expression (block, op1,
2752 tree genop2 = find_or_generate_expression (block, op2,
2754 if (!genop1 || !genop2)
2756 genop1 = fold_convert (TREE_TYPE (nary->op[0]),
2758 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
2759 may be a constant with the wrong type. */
2760 if (nary->opcode == POINTER_PLUS_EXPR)
2761 genop2 = fold_convert (sizetype, genop2);
2763 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
2765 folded = fold_build2 (nary->opcode, nary->type,
2771 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2772 tree genop1 = find_or_generate_expression (block, op1,
2776 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
2778 folded = fold_build1 (nary->opcode, nary->type,
2790 folded = fold_convert (exprtype, folded);
2791 /* Force the generated expression to be a sequence of GIMPLE
2793 We have to call unshare_expr because force_gimple_operand may
2794 modify the tree we pass to it. */
2795 newexpr = force_gimple_operand (unshare_expr (folded), &forced_stmts,
2798 /* If we have any intermediate expressions to the value sets, add them
2799 to the value sets and chain them in the instruction stream. */
2802 gsi = gsi_start (forced_stmts);
2803 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2805 gimple stmt = gsi_stmt (gsi);
2806 tree forcedname = gimple_get_lhs (stmt);
2809 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
2810 if (TREE_CODE (forcedname) == SSA_NAME)
2812 VN_INFO_GET (forcedname)->valnum = forcedname;
2813 VN_INFO (forcedname)->value_id = get_next_value_id ();
2814 nameexpr = get_or_alloc_expr_for_name (forcedname);
2815 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2816 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2817 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2819 mark_symbols_for_renaming (stmt);
2821 gimple_seq_add_seq (stmts, forced_stmts);
2824 /* Build and insert the assignment of the end result to the temporary
2825 that we will return. */
2826 if (!pretemp || exprtype != TREE_TYPE (pretemp))
2828 pretemp = create_tmp_var (exprtype, "pretmp");
2829 get_var_ann (pretemp);
2833 add_referenced_var (temp);
2835 if (TREE_CODE (exprtype) == COMPLEX_TYPE
2836 || TREE_CODE (exprtype) == VECTOR_TYPE)
2837 DECL_GIMPLE_REG_P (temp) = 1;
2839 newstmt = gimple_build_assign (temp, newexpr);
2840 name = make_ssa_name (temp, newstmt);
2841 gimple_assign_set_lhs (newstmt, name);
2842 gimple_set_plf (newstmt, NECESSARY, false);
2844 gimple_seq_add_stmt (stmts, newstmt);
2845 VEC_safe_push (gimple, heap, inserted_exprs, newstmt);
2847 /* All the symbols in NEWEXPR should be put into SSA form. */
2848 mark_symbols_for_renaming (newstmt);
2850 /* Add a value number to the temporary.
2851 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2852 we are creating the expression by pieces, and this particular piece of
2853 the expression may have been represented. There is no harm in replacing
2855 VN_INFO_GET (name)->valnum = name;
2856 value_id = get_expr_value_id (expr);
2857 VN_INFO (name)->value_id = value_id;
2858 nameexpr = get_or_alloc_expr_for_name (name);
2859 add_to_value (value_id, nameexpr);
2861 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2862 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2864 pre_stats.insertions++;
2865 if (dump_file && (dump_flags & TDF_DETAILS))
2867 fprintf (dump_file, "Inserted ");
2868 print_gimple_stmt (dump_file, newstmt, 0, 0);
2869 fprintf (dump_file, " in predecessor %d\n", block->index);
2876 /* Insert the to-be-made-available values of expression EXPRNUM for each
2877 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
2878 merge the result with a phi node, given the same value number as
2879 NODE. Return true if we have inserted new stuff. */
2882 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
2885 pre_expr expr = expression_for_id (exprnum);
2887 unsigned int val = get_expr_value_id (expr);
2889 bool insertions = false;
2894 tree type = get_expr_type (expr);
2898 if (dump_file && (dump_flags & TDF_DETAILS))
2900 fprintf (dump_file, "Found partial redundancy for expression ");
2901 print_pre_expr (dump_file, expr);
2902 fprintf (dump_file, " (%04d)\n", val);
2905 /* Make sure we aren't creating an induction variable. */
2906 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2
2907 && expr->kind != REFERENCE)
2909 bool firstinsideloop = false;
2910 bool secondinsideloop = false;
2911 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
2912 EDGE_PRED (block, 0)->src);
2913 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
2914 EDGE_PRED (block, 1)->src);
2915 /* Induction variables only have one edge inside the loop. */
2916 if (firstinsideloop ^ secondinsideloop)
2918 if (dump_file && (dump_flags & TDF_DETAILS))
2919 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
2925 /* Make the necessary insertions. */
2926 FOR_EACH_EDGE (pred, ei, block->preds)
2928 gimple_seq stmts = NULL;
2931 eprime = avail[bprime->index];
2933 if (eprime->kind != NAME && eprime->kind != CONSTANT)
2935 builtexpr = create_expression_by_pieces (bprime,
2939 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
2940 gsi_insert_seq_on_edge (pred, stmts);
2941 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
2944 else if (eprime->kind == CONSTANT)
2946 /* Constants may not have the right type, fold_convert
2947 should give us back a constant with the right type.
2949 tree constant = PRE_EXPR_CONSTANT (eprime);
2950 if (TREE_TYPE (constant) != type)
2952 tree builtexpr = fold_convert (type, constant);
2953 if (is_gimple_min_invariant (builtexpr))
2955 PRE_EXPR_CONSTANT (eprime) = builtexpr;
2959 tree forcedexpr = force_gimple_operand (builtexpr,
2962 if (is_gimple_min_invariant (forcedexpr))
2964 PRE_EXPR_CONSTANT (eprime) = forcedexpr;
2968 if (forcedexpr != builtexpr)
2970 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
2971 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
2975 gimple_stmt_iterator gsi;
2976 gsi = gsi_start (stmts);
2977 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2979 gimple stmt = gsi_stmt (gsi);
2980 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
2981 gimple_set_plf (stmt, NECESSARY, false);
2983 gsi_insert_seq_on_edge (pred, stmts);
2985 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
2990 else if (eprime->kind == NAME)
2992 /* We may have to do a conversion because our value
2993 numbering can look through types in certain cases, but
2994 our IL requires all operands of a phi node have the same
2996 tree name = PRE_EXPR_NAME (eprime);
2997 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3001 builtexpr = fold_convert (type, name);
3002 forcedexpr = force_gimple_operand (builtexpr,
3006 if (forcedexpr != name)
3008 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3009 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3014 gimple_stmt_iterator gsi;
3015 gsi = gsi_start (stmts);
3016 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3018 gimple stmt = gsi_stmt (gsi);
3019 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3020 gimple_set_plf (stmt, NECESSARY, false);
3022 gsi_insert_seq_on_edge (pred, stmts);
3024 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3028 /* If we didn't want a phi node, and we made insertions, we still have
3029 inserted new stuff, and thus return true. If we didn't want a phi node,
3030 and didn't make insertions, we haven't added anything new, so return
3032 if (nophi && insertions)
3034 else if (nophi && !insertions)
3037 /* Now build a phi for the new variable. */
3038 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3040 prephitemp = create_tmp_var (type, "prephitmp");
3041 get_var_ann (prephitemp);
3045 add_referenced_var (temp);
3047 if (TREE_CODE (type) == COMPLEX_TYPE
3048 || TREE_CODE (type) == VECTOR_TYPE)
3049 DECL_GIMPLE_REG_P (temp) = 1;
3050 phi = create_phi_node (temp, block);
3052 gimple_set_plf (phi, NECESSARY, false);
3053 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3054 VN_INFO (gimple_phi_result (phi))->value_id = val;
3055 VEC_safe_push (gimple, heap, inserted_exprs, phi);
3056 FOR_EACH_EDGE (pred, ei, block->preds)
3058 pre_expr ae = avail[pred->src->index];
3059 gcc_assert (get_expr_type (ae) == type
3060 || useless_type_conversion_p (type, get_expr_type (ae)));
3061 if (ae->kind == CONSTANT)
3062 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred);
3064 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred);
3067 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3068 add_to_value (val, newphi);
3070 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3071 this insertion, since we test for the existence of this value in PHI_GEN
3072 before proceeding with the partial redundancy checks in insert_aux.
3074 The value may exist in AVAIL_OUT, in particular, it could be represented
3075 by the expression we are trying to eliminate, in which case we want the
3076 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3079 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3080 this block, because if it did, it would have existed in our dominator's
3081 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3084 bitmap_insert_into_set (PHI_GEN (block), newphi);
3085 bitmap_value_replace_in_set (AVAIL_OUT (block),
3087 bitmap_insert_into_set (NEW_SETS (block),
3090 if (dump_file && (dump_flags & TDF_DETAILS))
3092 fprintf (dump_file, "Created phi ");
3093 print_gimple_stmt (dump_file, phi, 0, 0);
3094 fprintf (dump_file, " in block %d\n", block->index);
3102 /* Perform insertion of partially redundant values.
3103 For BLOCK, do the following:
3104 1. Propagate the NEW_SETS of the dominator into the current block.
3105 If the block has multiple predecessors,
3106 2a. Iterate over the ANTIC expressions for the block to see if
3107 any of them are partially redundant.
3108 2b. If so, insert them into the necessary predecessors to make
3109 the expression fully redundant.
3110 2c. Insert a new PHI merging the values of the predecessors.
3111 2d. Insert the new PHI, and the new expressions, into the
3113 3. Recursively call ourselves on the dominator children of BLOCK.
3115 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3116 do_regular_insertion and do_partial_insertion.
3121 do_regular_insertion (basic_block block, basic_block dom)
3123 bool new_stuff = false;
3124 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3128 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3130 if (expr->kind != NAME)
3134 bool by_some = false;
3135 bool cant_insert = false;
3136 bool all_same = true;
3137 pre_expr first_s = NULL;
3140 pre_expr eprime = NULL;
3143 val = get_expr_value_id (expr);
3144 if (bitmap_set_contains_value (PHI_GEN (block), val))
3146 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3148 if (dump_file && (dump_flags & TDF_DETAILS))
3149 fprintf (dump_file, "Found fully redundant value\n");
3153 avail = XCNEWVEC (pre_expr, last_basic_block);
3154 FOR_EACH_EDGE (pred, ei, block->preds)
3156 unsigned int vprime;
3157 pre_expr edoubleprime;
3159 /* This can happen in the very weird case
3160 that our fake infinite loop edges have caused a
3161 critical edge to appear. */
3162 if (EDGE_CRITICAL_P (pred))
3168 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3171 /* eprime will generally only be NULL if the
3172 value of the expression, translated
3173 through the PHI for this predecessor, is
3174 undefined. If that is the case, we can't
3175 make the expression fully redundant,
3176 because its value is undefined along a
3177 predecessor path. We can thus break out
3178 early because it doesn't matter what the
3179 rest of the results are. */
3186 eprime = fully_constant_expression (eprime);
3187 vprime = get_expr_value_id (eprime);
3188 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3190 if (edoubleprime == NULL)
3192 avail[bprime->index] = eprime;
3197 avail[bprime->index] = edoubleprime;
3199 if (first_s == NULL)
3200 first_s = edoubleprime;
3201 else if (!pre_expr_eq (first_s, edoubleprime))
3205 /* If we can insert it, it's not the same value
3206 already existing along every predecessor, and
3207 it's defined by some predecessor, it is
3208 partially redundant. */
3209 if (!cant_insert && !all_same && by_some && dbg_cnt (treepre_insert))
3211 if (insert_into_preds_of_block (block, get_expression_id (expr),
3215 /* If all edges produce the same value and that value is
3216 an invariant, then the PHI has the same value on all
3217 edges. Note this. */
3218 else if (!cant_insert && all_same && eprime
3219 && eprime->kind == CONSTANT
3220 && !value_id_constant_p (val))
3224 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3225 value_expressions, val);
3227 unsigned int new_val = get_expr_value_id (eprime);
3228 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3230 pre_expr expr = expression_for_id (j);
3232 if (expr->kind == NAME)
3234 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3235 /* Just reset the value id and valnum so it is
3236 the same as the constant we have discovered. */
3237 info->valnum = PRE_EXPR_CONSTANT (eprime);
3238 info->value_id = new_val;
3239 pre_stats.constified++;
3247 VEC_free (pre_expr, heap, exprs);
3252 /* Perform insertion for partially anticipatable expressions. There
3253 is only one case we will perform insertion for these. This case is
3254 if the expression is partially anticipatable, and fully available.
3255 In this case, we know that putting it earlier will enable us to
3256 remove the later computation. */
3260 do_partial_partial_insertion (basic_block block, basic_block dom)
3262 bool new_stuff = false;
3263 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3267 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3269 if (expr->kind != NAME)
3274 bool cant_insert = false;
3277 pre_expr eprime = NULL;
3280 val = get_expr_value_id (expr);
3281 if (bitmap_set_contains_value (PHI_GEN (block), val))
3283 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3286 avail = XCNEWVEC (pre_expr, last_basic_block);
3287 FOR_EACH_EDGE (pred, ei, block->preds)
3289 unsigned int vprime;
3290 pre_expr edoubleprime;
3292 /* This can happen in the very weird case
3293 that our fake infinite loop edges have caused a
3294 critical edge to appear. */
3295 if (EDGE_CRITICAL_P (pred))
3301 eprime = phi_translate (expr, ANTIC_IN (block),
3305 /* eprime will generally only be NULL if the
3306 value of the expression, translated
3307 through the PHI for this predecessor, is
3308 undefined. If that is the case, we can't
3309 make the expression fully redundant,
3310 because its value is undefined along a
3311 predecessor path. We can thus break out
3312 early because it doesn't matter what the
3313 rest of the results are. */
3320 eprime = fully_constant_expression (eprime);
3321 vprime = get_expr_value_id (eprime);
3322 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3324 if (edoubleprime == NULL)
3330 avail[bprime->index] = edoubleprime;
3334 /* If we can insert it, it's not the same value
3335 already existing along every predecessor, and
3336 it's defined by some predecessor, it is
3337 partially redundant. */
3338 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3340 pre_stats.pa_insert++;
3341 if (insert_into_preds_of_block (block, get_expression_id (expr),
3349 VEC_free (pre_expr, heap, exprs);
3354 insert_aux (basic_block block)
3357 bool new_stuff = false;
3362 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3367 bitmap_set_t newset = NEW_SETS (dom);
3370 /* Note that we need to value_replace both NEW_SETS, and
3371 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3372 represented by some non-simple expression here that we want
3373 to replace it with. */
3374 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3376 pre_expr expr = expression_for_id (i);
3377 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3378 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3381 if (!single_pred_p (block))
3383 new_stuff |= do_regular_insertion (block, dom);
3384 if (do_partial_partial)
3385 new_stuff |= do_partial_partial_insertion (block, dom);
3389 for (son = first_dom_son (CDI_DOMINATORS, block);
3391 son = next_dom_son (CDI_DOMINATORS, son))
3393 new_stuff |= insert_aux (son);
3399 /* Perform insertion of partially redundant values. */
3404 bool new_stuff = true;
3406 int num_iterations = 0;
3409 NEW_SETS (bb) = bitmap_set_new ();
3414 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3416 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3420 /* Add OP to EXP_GEN (block), and possibly to the maximal set if it is
3421 not defined by a phi node.
3422 PHI nodes can't go in the maximal sets because they are not in
3423 TMP_GEN, so it is possible to get into non-monotonic situations
3424 during ANTIC calculation, because it will *add* bits. */
3427 add_to_exp_gen (basic_block block, tree op)
3432 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3434 result = get_or_alloc_expr_for_name (op);
3435 bitmap_value_insert_into_set (EXP_GEN (block), result);
3436 if (TREE_CODE (op) != SSA_NAME
3437 || gimple_code (SSA_NAME_DEF_STMT (op)) != GIMPLE_PHI)
3438 bitmap_value_insert_into_set (maximal_set, result);
3442 /* Create value ids for PHI in BLOCK. */
3445 make_values_for_phi (gimple phi, basic_block block)
3447 tree result = gimple_phi_result (phi);
3449 /* We have no need for virtual phis, as they don't represent
3450 actual computations. */
3451 if (is_gimple_reg (result))
3453 pre_expr e = get_or_alloc_expr_for_name (result);
3454 add_to_value (get_expr_value_id (e), e);
3455 bitmap_insert_into_set (PHI_GEN (block), e);
3456 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3460 /* Compute the AVAIL set for all basic blocks.
3462 This function performs value numbering of the statements in each basic
3463 block. The AVAIL sets are built from information we glean while doing
3464 this value numbering, since the AVAIL sets contain only one entry per
3467 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3468 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3471 compute_avail (void)
3474 basic_block block, son;
3475 basic_block *worklist;
3479 /* For arguments with default definitions, we pretend they are
3480 defined in the entry block. */
3481 for (param = DECL_ARGUMENTS (current_function_decl);
3483 param = TREE_CHAIN (param))
3485 if (gimple_default_def (cfun, param) != NULL)
3487 tree def = gimple_default_def (cfun, param);
3488 pre_expr e = get_or_alloc_expr_for_name (def);
3490 add_to_value (get_expr_value_id (e), e);
3493 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3494 bitmap_value_insert_into_set (maximal_set, e);
3496 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3500 /* Likewise for the static chain decl. */
3501 if (cfun->static_chain_decl)
3503 param = cfun->static_chain_decl;
3504 if (gimple_default_def (cfun, param) != NULL)
3506 tree def = gimple_default_def (cfun, param);
3507 pre_expr e = get_or_alloc_expr_for_name (def);
3509 add_to_value (get_expr_value_id (e), e);
3512 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3513 bitmap_value_insert_into_set (maximal_set, e);
3515 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3519 /* Allocate the worklist. */
3520 worklist = XNEWVEC (basic_block, n_basic_blocks);
3522 /* Seed the algorithm by putting the dominator children of the entry
3523 block on the worklist. */
3524 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3526 son = next_dom_son (CDI_DOMINATORS, son))
3527 worklist[sp++] = son;
3529 /* Loop until the worklist is empty. */
3532 gimple_stmt_iterator gsi;
3535 unsigned int stmt_uid = 1;
3537 /* Pick a block from the worklist. */
3538 block = worklist[--sp];
3540 /* Initially, the set of available values in BLOCK is that of
3541 its immediate dominator. */
3542 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3544 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3546 /* Generate values for PHI nodes. */
3547 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3548 make_values_for_phi (gsi_stmt (gsi), block);
3550 /* Now compute value numbers and populate value sets with all
3551 the expressions computed in BLOCK. */
3552 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3557 stmt = gsi_stmt (gsi);
3558 gimple_set_uid (stmt, stmt_uid++);
3560 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3562 pre_expr e = get_or_alloc_expr_for_name (op);
3564 add_to_value (get_expr_value_id (e), e);
3567 bitmap_insert_into_set (TMP_GEN (block), e);
3568 bitmap_value_insert_into_set (maximal_set, e);
3570 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3573 if (gimple_has_volatile_ops (stmt)
3574 || stmt_could_throw_p (stmt))
3577 switch (gimple_code (stmt))
3580 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3581 add_to_exp_gen (block, op);
3588 vn_reference_op_t vro;
3589 pre_expr result = NULL;
3590 VEC(vn_reference_op_s, heap) *ops = NULL;
3592 if (!can_value_number_call (stmt))
3595 copy_reference_ops_from_call (stmt, &ops);
3596 vn_reference_lookup_pieces (shared_vuses_from_stmt (stmt),
3598 VEC_free (vn_reference_op_s, heap, ops);
3602 for (i = 0; VEC_iterate (vn_reference_op_s,
3606 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3607 add_to_exp_gen (block, vro->op0);
3608 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3609 add_to_exp_gen (block, vro->op1);
3610 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3611 add_to_exp_gen (block, vro->op2);
3613 result = (pre_expr) pool_alloc (pre_expr_pool);
3614 result->kind = REFERENCE;
3616 PRE_EXPR_REFERENCE (result) = ref;
3618 get_or_alloc_expression_id (result);
3619 add_to_value (get_expr_value_id (result), result);
3622 bitmap_value_insert_into_set (EXP_GEN (block),
3624 bitmap_value_insert_into_set (maximal_set, result);
3631 pre_expr result = NULL;
3632 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
3635 if (is_exception_related (stmt))
3642 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
3643 gimple_assign_rhs_code (stmt),
3644 gimple_expr_type (stmt),
3645 gimple_assign_rhs1 (stmt),
3646 gimple_assign_rhs2 (stmt),
3647 NULL_TREE, NULL_TREE, &nary);
3652 for (i = 0; i < nary->length; i++)
3653 if (TREE_CODE (nary->op[i]) == SSA_NAME)
3654 add_to_exp_gen (block, nary->op[i]);
3656 result = (pre_expr) pool_alloc (pre_expr_pool);
3657 result->kind = NARY;
3659 PRE_EXPR_NARY (result) = nary;
3663 case tcc_declaration:
3668 vn_reference_op_t vro;
3670 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3671 shared_vuses_from_stmt (stmt),
3676 for (i = 0; VEC_iterate (vn_reference_op_s,
3680 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3681 add_to_exp_gen (block, vro->op0);
3682 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3683 add_to_exp_gen (block, vro->op1);
3684 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3685 add_to_exp_gen (block, vro->op2);
3687 result = (pre_expr) pool_alloc (pre_expr_pool);
3688 result->kind = REFERENCE;
3690 PRE_EXPR_REFERENCE (result) = ref;
3695 /* For any other statement that we don't
3696 recognize, simply add all referenced
3697 SSA_NAMEs to EXP_GEN. */
3698 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3699 add_to_exp_gen (block, op);
3703 get_or_alloc_expression_id (result);
3704 add_to_value (get_expr_value_id (result), result);
3707 bitmap_value_insert_into_set (EXP_GEN (block), result);
3708 bitmap_value_insert_into_set (maximal_set, result);
3718 /* Put the dominator children of BLOCK on the worklist of blocks
3719 to compute available sets for. */
3720 for (son = first_dom_son (CDI_DOMINATORS, block);
3722 son = next_dom_son (CDI_DOMINATORS, son))
3723 worklist[sp++] = son;
3729 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
3730 than the available expressions for it. The insertion point is
3731 right before the first use in STMT. Returns the SSA_NAME that should
3732 be used for replacement. */
3735 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
3737 basic_block bb = gimple_bb (stmt);
3738 gimple_stmt_iterator gsi;
3739 gimple_seq stmts = NULL;
3743 /* First create a value expression from the expression we want
3744 to insert and associate it with the value handle for SSA_VN. */
3745 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
3749 /* Then use create_expression_by_pieces to generate a valid
3750 expression to insert at this point of the IL stream. */
3751 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
3752 if (expr == NULL_TREE)
3754 gsi = gsi_for_stmt (stmt);
3755 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
3760 /* Eliminate fully redundant computations. */
3766 unsigned int todo = 0;
3770 gimple_stmt_iterator i;
3772 for (i = gsi_start_bb (b); !gsi_end_p (i); gsi_next (&i))
3774 gimple stmt = gsi_stmt (i);
3776 /* Lookup the RHS of the expression, see if we have an
3777 available computation for it. If so, replace the RHS with
3778 the available computation. */
3779 if (gimple_has_lhs (stmt)
3780 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
3781 && !gimple_assign_ssa_name_copy_p (stmt)
3782 && (!gimple_assign_single_p (stmt)
3783 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
3784 && !gimple_has_volatile_ops (stmt)
3785 && !has_zero_uses (gimple_get_lhs (stmt)))
3787 tree lhs = gimple_get_lhs (stmt);
3788 tree rhs = NULL_TREE;
3790 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
3791 pre_expr sprimeexpr;
3793 if (gimple_assign_single_p (stmt))
3794 rhs = gimple_assign_rhs1 (stmt);
3796 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
3797 get_expr_value_id (lhsexpr),
3802 if (sprimeexpr->kind == CONSTANT)
3803 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
3804 else if (sprimeexpr->kind == NAME)
3805 sprime = PRE_EXPR_NAME (sprimeexpr);
3810 /* If there is no existing leader but SCCVN knows this
3811 value is constant, use that constant. */
3812 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
3814 sprime = fold_convert (TREE_TYPE (lhs),
3815 VN_INFO (lhs)->valnum);
3817 if (dump_file && (dump_flags & TDF_DETAILS))
3819 fprintf (dump_file, "Replaced ");
3820 print_gimple_expr (dump_file, stmt, 0, 0);
3821 fprintf (dump_file, " with ");
3822 print_generic_expr (dump_file, sprime, 0);
3823 fprintf (dump_file, " in ");
3824 print_gimple_stmt (dump_file, stmt, 0, 0);
3826 pre_stats.eliminations++;
3827 propagate_tree_value_into_stmt (&i, sprime);
3828 stmt = gsi_stmt (i);
3833 /* If there is no existing usable leader but SCCVN thinks
3834 it has an expression it wants to use as replacement,
3836 if (!sprime || sprime == lhs)
3838 tree val = VN_INFO (lhs)->valnum;
3840 && TREE_CODE (val) == SSA_NAME
3841 && VN_INFO (val)->needs_insertion
3842 && can_PRE_operation (vn_get_expr_for (val)))
3843 sprime = do_SCCVN_insertion (stmt, val);
3847 && (rhs == NULL_TREE
3848 || TREE_CODE (rhs) != SSA_NAME
3849 || may_propagate_copy (rhs, sprime)))
3851 gcc_assert (sprime != rhs);
3853 if (dump_file && (dump_flags & TDF_DETAILS))
3855 fprintf (dump_file, "Replaced ");
3856 print_gimple_expr (dump_file, stmt, 0, 0);
3857 fprintf (dump_file, " with ");
3858 print_generic_expr (dump_file, sprime, 0);
3859 fprintf (dump_file, " in ");
3860 print_gimple_stmt (dump_file, stmt, 0, 0);
3863 if (TREE_CODE (sprime) == SSA_NAME)
3864 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
3866 /* We need to make sure the new and old types actually match,
3867 which may require adding a simple cast, which fold_convert
3869 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
3870 && !useless_type_conversion_p (gimple_expr_type (stmt),
3871 TREE_TYPE (sprime)))
3872 sprime = fold_convert (gimple_expr_type (stmt), sprime);
3874 pre_stats.eliminations++;
3875 propagate_tree_value_into_stmt (&i, sprime);
3876 stmt = gsi_stmt (i);
3879 /* If we removed EH side effects from the statement, clean
3880 its EH information. */
3881 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
3883 bitmap_set_bit (need_eh_cleanup,
3884 gimple_bb (stmt)->index);
3885 if (dump_file && (dump_flags & TDF_DETAILS))
3886 fprintf (dump_file, " Removed EH side effects.\n");
3890 /* Visit COND_EXPRs and fold the comparison with the
3891 available value-numbers. */
3892 else if (gimple_code (stmt) == GIMPLE_COND)
3894 tree op0 = gimple_cond_lhs (stmt);
3895 tree op1 = gimple_cond_rhs (stmt);
3898 if (TREE_CODE (op0) == SSA_NAME)
3899 op0 = VN_INFO (op0)->valnum;
3900 if (TREE_CODE (op1) == SSA_NAME)
3901 op1 = VN_INFO (op1)->valnum;
3902 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
3904 if (result && TREE_CODE (result) == INTEGER_CST)
3906 if (integer_zerop (result))
3907 gimple_cond_make_false (stmt);
3909 gimple_cond_make_true (stmt);
3911 todo = TODO_cleanup_cfg;
3920 /* Borrow a bit of tree-ssa-dce.c for the moment.
3921 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
3922 this may be a bit faster, and we may want critical edges kept split. */
3924 /* If OP's defining statement has not already been determined to be necessary,
3925 mark that statement necessary. Return the stmt, if it is newly
3928 static inline gimple
3929 mark_operand_necessary (tree op)
3935 if (TREE_CODE (op) != SSA_NAME)
3938 stmt = SSA_NAME_DEF_STMT (op);
3941 if (gimple_plf (stmt, NECESSARY)
3942 || gimple_nop_p (stmt))
3945 gimple_set_plf (stmt, NECESSARY, true);
3949 /* Because we don't follow exactly the standard PRE algorithm, and decide not
3950 to insert PHI nodes sometimes, and because value numbering of casts isn't
3951 perfect, we sometimes end up inserting dead code. This simple DCE-like
3952 pass removes any insertions we made that weren't actually used. */
3955 remove_dead_inserted_code (void)
3957 VEC(gimple,heap) *worklist = NULL;
3961 worklist = VEC_alloc (gimple, heap, VEC_length (gimple, inserted_exprs));
3962 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
3964 if (gimple_plf (t, NECESSARY))
3965 VEC_quick_push (gimple, worklist, t);
3967 while (VEC_length (gimple, worklist) > 0)
3969 t = VEC_pop (gimple, worklist);
3971 /* PHI nodes are somewhat special in that each PHI alternative has
3972 data and control dependencies. All the statements feeding the
3973 PHI node's arguments are always necessary. */
3974 if (gimple_code (t) == GIMPLE_PHI)
3978 VEC_reserve (gimple, heap, worklist, gimple_phi_num_args (t));
3979 for (k = 0; k < gimple_phi_num_args (t); k++)
3981 tree arg = PHI_ARG_DEF (t, k);
3982 if (TREE_CODE (arg) == SSA_NAME)
3984 gimple n = mark_operand_necessary (arg);
3986 VEC_quick_push (gimple, worklist, n);
3992 /* Propagate through the operands. Examine all the USE, VUSE and
3993 VDEF operands in this statement. Mark all the statements
3994 which feed this statement's uses as necessary. */
3998 /* The operands of VDEF expressions are also needed as they
3999 represent potential definitions that may reach this
4000 statement (VDEF operands allow us to follow def-def
4003 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4005 gimple n = mark_operand_necessary (use);
4007 VEC_safe_push (gimple, heap, worklist, n);
4012 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4014 if (!gimple_plf (t, NECESSARY))
4016 gimple_stmt_iterator gsi;
4018 if (dump_file && (dump_flags & TDF_DETAILS))
4020 fprintf (dump_file, "Removing unnecessary insertion:");
4021 print_gimple_stmt (dump_file, t, 0, 0);
4024 gsi = gsi_for_stmt (t);
4025 if (gimple_code (t) == GIMPLE_PHI)
4026 remove_phi_node (&gsi, true);
4028 gsi_remove (&gsi, true);
4032 VEC_free (gimple, heap, worklist);
4035 /* Initialize data structures used by PRE. */
4038 init_pre (bool do_fre)
4042 next_expression_id = 1;
4044 VEC_safe_push (pre_expr, heap, expressions, NULL);
4045 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4046 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4047 get_max_value_id() + 1);
4051 inserted_exprs = NULL;
4052 need_creation = NULL;
4053 pretemp = NULL_TREE;
4054 storetemp = NULL_TREE;
4055 prephitemp = NULL_TREE;
4057 connect_infinite_loops_to_exit ();
4058 memset (&pre_stats, 0, sizeof (pre_stats));
4061 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4062 post_order_compute (postorder, false, false);
4065 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4067 calculate_dominance_info (CDI_POST_DOMINATORS);
4068 calculate_dominance_info (CDI_DOMINATORS);
4070 bitmap_obstack_initialize (&grand_bitmap_obstack);
4071 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4072 expr_pred_trans_eq, free);
4073 expression_to_id = htab_create (num_ssa_names * 3,
4076 seen_during_translate = BITMAP_ALLOC (&grand_bitmap_obstack);
4077 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4078 sizeof (struct bitmap_set), 30);
4079 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4080 sizeof (struct pre_expr_d), 30);
4083 EXP_GEN (bb) = bitmap_set_new ();
4084 PHI_GEN (bb) = bitmap_set_new ();
4085 TMP_GEN (bb) = bitmap_set_new ();
4086 AVAIL_OUT (bb) = bitmap_set_new ();
4088 maximal_set = in_fre ? NULL : bitmap_set_new ();
4090 need_eh_cleanup = BITMAP_ALLOC (NULL);
4094 /* Deallocate data structures used by PRE. */
4097 fini_pre (bool do_fre)
4102 VEC_free (bitmap_set_t, heap, value_expressions);
4103 VEC_free (gimple, heap, inserted_exprs);
4104 VEC_free (gimple, heap, need_creation);
4105 bitmap_obstack_release (&grand_bitmap_obstack);
4106 free_alloc_pool (bitmap_set_pool);
4107 free_alloc_pool (pre_expr_pool);
4108 htab_delete (phi_translate_table);
4109 htab_delete (expression_to_id);
4110 remove_fake_exit_edges ();
4118 free_dominance_info (CDI_POST_DOMINATORS);
4120 if (!bitmap_empty_p (need_eh_cleanup))
4122 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4123 cleanup_tree_cfg ();
4126 BITMAP_FREE (need_eh_cleanup);
4129 loop_optimizer_finalize ();
4132 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4133 only wants to do full redundancy elimination. */
4136 execute_pre (bool do_fre ATTRIBUTE_UNUSED)
4138 unsigned int todo = 0;
4140 do_partial_partial = optimize > 2;
4142 /* This has to happen before SCCVN runs because
4143 loop_optimizer_init may create new phis, etc. */
4145 loop_optimizer_init (LOOPS_NORMAL);
4147 if (!run_scc_vn (do_fre))
4151 remove_dead_inserted_code ();
4152 loop_optimizer_finalize ();
4160 /* Collect and value number expressions computed in each basic block. */
4163 if (dump_file && (dump_flags & TDF_DETAILS))
4169 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4170 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen",
4172 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out",
4177 /* Insert can get quite slow on an incredibly large number of basic
4178 blocks due to some quadratic behavior. Until this behavior is
4179 fixed, don't run it when he have an incredibly large number of
4180 bb's. If we aren't going to run insert, there is no point in
4181 computing ANTIC, either, even though it's plenty fast. */
4182 if (!do_fre && n_basic_blocks < 4000)
4188 /* Remove all the redundant expressions. */
4189 todo |= eliminate ();
4191 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4192 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4193 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4194 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4195 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4196 gsi_commit_edge_inserts ();
4198 clear_expression_ids ();
4201 remove_dead_inserted_code ();
4208 /* Gate and execute functions for PRE. */
4213 return TODO_rebuild_alias | execute_pre (false);
4219 return flag_tree_pre != 0;
4222 struct gimple_opt_pass pass_pre =
4227 gate_pre, /* gate */
4228 do_pre, /* execute */
4231 0, /* static_pass_number */
4232 TV_TREE_PRE, /* tv_id */
4233 PROP_no_crit_edges | PROP_cfg
4234 | PROP_ssa | PROP_alias, /* properties_required */
4235 0, /* properties_provided */
4236 0, /* properties_destroyed */
4237 0, /* todo_flags_start */
4238 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4239 | TODO_verify_ssa /* todo_flags_finish */
4244 /* Gate and execute functions for FRE. */
4249 return execute_pre (true);
4255 return flag_tree_fre != 0;
4258 struct gimple_opt_pass pass_fre =
4263 gate_fre, /* gate */
4264 execute_fre, /* execute */
4267 0, /* static_pass_number */
4268 TV_TREE_FRE, /* tv_id */
4269 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
4270 0, /* properties_provided */
4271 0, /* properties_destroyed */
4272 0, /* todo_flags_start */
4273 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */