2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
3 Free Software Foundation, Inc.
4 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
25 #include "coretypes.h"
29 #include "basic-block.h"
30 #include "diagnostic.h"
31 #include "tree-inline.h"
32 #include "tree-flow.h"
34 #include "tree-dump.h"
38 #include "tree-iterator.h"
40 #include "alloc-pool.h"
42 #include "tree-pass.h"
45 #include "langhooks.h"
47 #include "tree-ssa-sccvn.h"
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_hashval_t (SSA_NAME_VERSION (PRE_EXPR_NAME (e)), 0);
221 return PRE_EXPR_NARY (e)->hashcode;
223 return PRE_EXPR_REFERENCE (e)->hashcode;
230 /* Next global expression id number. */
231 static unsigned int next_expression_id;
233 /* Mapping from expression to id number we can use in bitmap sets. */
234 DEF_VEC_P (pre_expr);
235 DEF_VEC_ALLOC_P (pre_expr, heap);
236 static VEC(pre_expr, heap) *expressions;
237 static htab_t expression_to_id;
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 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
341 EXECUTE_IF_SET_IN_BITMAP((set)->values, 0, (id), (bi))
343 /* Mapping from value id to expressions with that value_id. */
344 DEF_VEC_P (bitmap_set_t);
345 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
346 static VEC(bitmap_set_t, heap) *value_expressions;
348 /* Sets that we need to keep track of. */
349 typedef struct bb_bitmap_sets
351 /* The EXP_GEN set, which represents expressions/values generated in
353 bitmap_set_t exp_gen;
355 /* The PHI_GEN set, which represents PHI results generated in a
357 bitmap_set_t phi_gen;
359 /* The TMP_GEN set, which represents results/temporaries generated
360 in a basic block. IE the LHS of an expression. */
361 bitmap_set_t tmp_gen;
363 /* The AVAIL_OUT set, which represents which values are available in
364 a given basic block. */
365 bitmap_set_t avail_out;
367 /* The ANTIC_IN set, which represents which values are anticipatable
368 in a given basic block. */
369 bitmap_set_t antic_in;
371 /* The PA_IN set, which represents which values are
372 partially anticipatable in a given basic block. */
375 /* The NEW_SETS set, which is used during insertion to augment the
376 AVAIL_OUT set of blocks with the new insertions performed during
377 the current iteration. */
378 bitmap_set_t new_sets;
380 /* True if we have visited this block during ANTIC calculation. */
381 unsigned int visited:1;
383 /* True we have deferred processing this block during ANTIC
384 calculation until its successor is processed. */
385 unsigned int deferred : 1;
388 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
389 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
390 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
391 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
392 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
393 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
394 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
395 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
396 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
399 /* Basic block list in postorder. */
400 static int *postorder;
402 /* This structure is used to keep track of statistics on what
403 optimization PRE was able to perform. */
406 /* The number of RHS computations eliminated by PRE. */
409 /* The number of new expressions/temporaries generated by PRE. */
412 /* The number of inserts found due to partial anticipation */
415 /* The number of new PHI nodes added by PRE. */
418 /* The number of values found constant. */
423 static bool do_partial_partial;
424 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
425 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
426 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
427 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
428 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
429 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
430 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr, bool);
431 static bitmap_set_t bitmap_set_new (void);
432 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
434 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
436 static unsigned int get_expr_value_id (pre_expr);
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 gcc_assert (get_expr_value_id (e) == v);
564 if (v >= VEC_length (bitmap_set_t, value_expressions))
566 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
570 set = VEC_index (bitmap_set_t, value_expressions, v);
573 set = bitmap_set_new ();
574 VEC_replace (bitmap_set_t, value_expressions, v, set);
577 bitmap_insert_into_set_1 (set, e, true);
580 /* Create a new bitmap set and return it. */
583 bitmap_set_new (void)
585 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
586 ret->expressions = BITMAP_ALLOC (&grand_bitmap_obstack);
587 ret->values = BITMAP_ALLOC (&grand_bitmap_obstack);
591 /* Return the value id for a PRE expression EXPR. */
594 get_expr_value_id (pre_expr expr)
601 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
604 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
605 add_to_value (id, expr);
610 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
612 return PRE_EXPR_NARY (expr)->value_id;
614 return PRE_EXPR_REFERENCE (expr)->value_id;
620 /* Remove an expression EXPR from a bitmapped set. */
623 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
625 unsigned int val = get_expr_value_id (expr);
626 if (!value_id_constant_p (val))
628 bitmap_clear_bit (set->values, val);
629 bitmap_clear_bit (set->expressions, get_expression_id (expr));
634 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
635 bool allow_constants)
637 unsigned int val = get_expr_value_id (expr);
638 if (allow_constants || !value_id_constant_p (val))
640 /* We specifically expect this and only this function to be able to
641 insert constants into a set. */
642 bitmap_set_bit (set->values, val);
643 bitmap_set_bit (set->expressions, get_or_alloc_expression_id (expr));
647 /* Insert an expression EXPR into a bitmapped set. */
650 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
652 bitmap_insert_into_set_1 (set, expr, false);
655 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
658 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
660 bitmap_copy (dest->expressions, orig->expressions);
661 bitmap_copy (dest->values, orig->values);
665 /* Free memory used up by SET. */
667 bitmap_set_free (bitmap_set_t set)
669 BITMAP_FREE (set->expressions);
670 BITMAP_FREE (set->values);
674 /* Generate an topological-ordered array of bitmap set SET. */
676 static VEC(pre_expr, heap) *
677 sorted_array_from_bitmap_set (bitmap_set_t set)
680 bitmap_iterator bi, bj;
681 VEC(pre_expr, heap) *result = NULL;
683 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
685 /* The number of expressions having a given value is usually
686 relatively small. Thus, rather than making a vector of all
687 the expressions and sorting it by value-id, we walk the values
688 and check in the reverse mapping that tells us what expressions
689 have a given value, to filter those in our set. As a result,
690 the expressions are inserted in value-id order, which means
693 If this is somehow a significant lose for some cases, we can
694 choose which set to walk based on the set size. */
695 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
696 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
698 if (bitmap_bit_p (set->expressions, j))
699 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
706 /* Perform bitmapped set operation DEST &= ORIG. */
709 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
716 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
718 bitmap_and_into (dest->values, orig->values);
719 bitmap_copy (temp, dest->expressions);
720 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
722 pre_expr expr = expression_for_id (i);
723 unsigned int value_id = get_expr_value_id (expr);
724 if (!bitmap_bit_p (dest->values, value_id))
725 bitmap_clear_bit (dest->expressions, i);
731 /* Subtract all values and expressions contained in ORIG from DEST. */
734 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
736 bitmap_set_t result = bitmap_set_new ();
740 bitmap_and_compl (result->expressions, dest->expressions,
743 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
745 pre_expr expr = expression_for_id (i);
746 unsigned int value_id = get_expr_value_id (expr);
747 bitmap_set_bit (result->values, value_id);
753 /* Subtract all the values in bitmap set B from bitmap set A. */
756 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
760 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
762 bitmap_copy (temp, a->expressions);
763 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
765 pre_expr expr = expression_for_id (i);
766 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
767 bitmap_remove_from_set (a, expr);
773 /* Return true if bitmapped set SET contains the value VALUE_ID. */
776 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
778 if (value_id_constant_p (value_id))
781 if (!set || bitmap_empty_p (set->expressions))
784 return bitmap_bit_p (set->values, value_id);
788 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
790 return bitmap_bit_p (set->expressions, get_expression_id (expr));
793 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
796 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
799 bitmap_set_t exprset;
803 if (value_id_constant_p (lookfor))
806 if (!bitmap_set_contains_value (set, lookfor))
809 /* The number of expressions having a given value is usually
810 significantly less than the total number of expressions in SET.
811 Thus, rather than check, for each expression in SET, whether it
812 has the value LOOKFOR, we walk the reverse mapping that tells us
813 what expressions have a given value, and see if any of those
814 expressions are in our set. For large testcases, this is about
815 5-10x faster than walking the bitmap. If this is somehow a
816 significant lose for some cases, we can choose which set to walk
817 based on the set size. */
818 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
819 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
821 if (bitmap_bit_p (set->expressions, i))
823 bitmap_clear_bit (set->expressions, i);
824 bitmap_set_bit (set->expressions, get_expression_id (expr));
830 /* Return true if two bitmap sets are equal. */
833 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
835 return bitmap_equal_p (a->values, b->values);
838 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
839 and add it otherwise. */
842 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
844 unsigned int val = get_expr_value_id (expr);
846 if (bitmap_set_contains_value (set, val))
847 bitmap_set_replace_value (set, val, expr);
849 bitmap_insert_into_set (set, expr);
852 /* Insert EXPR into SET if EXPR's value is not already present in
856 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
858 unsigned int val = get_expr_value_id (expr);
860 if (value_id_constant_p (val))
863 if (!bitmap_set_contains_value (set, val))
864 bitmap_insert_into_set (set, expr);
867 /* Print out EXPR to outfile. */
870 print_pre_expr (FILE *outfile, const pre_expr expr)
875 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
878 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
883 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
884 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
885 for (i = 0; i < nary->length; i++)
887 print_generic_expr (outfile, nary->op[i], 0);
888 if (i != (unsigned) nary->length - 1)
889 fprintf (outfile, ",");
891 fprintf (outfile, "}");
897 vn_reference_op_t vro;
899 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
900 fprintf (outfile, "{");
902 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
905 if (vro->opcode != SSA_NAME
906 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
907 fprintf (outfile, "%s ", tree_code_name [vro->opcode]);
911 fprintf (outfile, "<");
912 print_generic_expr (outfile, vro->op0, 0);
915 fprintf (outfile, ",");
916 print_generic_expr (outfile, vro->op1, 0);
919 fprintf (outfile, ">");
921 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
922 fprintf (outfile, ",");
924 fprintf (outfile, "}");
929 void debug_pre_expr (pre_expr);
931 /* Like print_pre_expr but always prints to stderr. */
933 debug_pre_expr (pre_expr e)
935 print_pre_expr (stderr, e);
936 fprintf (stderr, "\n");
939 /* Print out SET to OUTFILE. */
942 print_bitmap_set (FILE *outfile, bitmap_set_t set,
943 const char *setname, int blockindex)
945 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
952 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
954 const pre_expr expr = expression_for_id (i);
957 fprintf (outfile, ", ");
959 print_pre_expr (outfile, expr);
961 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
964 fprintf (outfile, " }\n");
967 void debug_bitmap_set (bitmap_set_t);
970 debug_bitmap_set (bitmap_set_t set)
972 print_bitmap_set (stderr, set, "debug", 0);
975 /* Print out the expressions that have VAL to OUTFILE. */
978 print_value_expressions (FILE *outfile, unsigned int val)
980 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
984 sprintf (s, "%04d", val);
985 print_bitmap_set (outfile, set, s, 0);
991 debug_value_expressions (unsigned int val)
993 print_value_expressions (stderr, val);
996 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1000 get_or_alloc_expr_for_constant (tree constant)
1002 unsigned int result_id;
1003 unsigned int value_id;
1004 pre_expr newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1005 newexpr->kind = CONSTANT;
1006 PRE_EXPR_CONSTANT (newexpr) = constant;
1007 result_id = lookup_expression_id (newexpr);
1010 pool_free (pre_expr_pool, newexpr);
1011 newexpr = expression_for_id (result_id);
1014 value_id = get_or_alloc_constant_value_id (constant);
1015 get_or_alloc_expression_id (newexpr);
1016 add_to_value (value_id, newexpr);
1020 /* Given a value id V, find the actual tree representing the constant
1021 value if there is one, and return it. Return NULL if we can't find
1025 get_constant_for_value_id (unsigned int v)
1027 if (value_id_constant_p (v))
1031 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1033 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1035 pre_expr expr = expression_for_id (i);
1036 if (expr->kind == CONSTANT)
1037 return PRE_EXPR_CONSTANT (expr);
1043 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1044 Currently only supports constants and SSA_NAMES. */
1046 get_or_alloc_expr_for (tree t)
1048 if (TREE_CODE (t) == SSA_NAME)
1049 return get_or_alloc_expr_for_name (t);
1050 else if (is_gimple_min_invariant (t))
1051 return get_or_alloc_expr_for_constant (t);
1054 /* More complex expressions can result from SCCVN expression
1055 simplification that inserts values for them. As they all
1056 do not have VOPs the get handled by the nary ops struct. */
1057 vn_nary_op_t result;
1058 unsigned int result_id;
1059 vn_nary_op_lookup (t, &result);
1062 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1064 PRE_EXPR_NARY (e) = result;
1065 result_id = lookup_expression_id (e);
1068 pool_free (pre_expr_pool, e);
1069 e = expression_for_id (result_id);
1072 alloc_expression_id (e);
1079 /* Return the folded version of T if T, when folded, is a gimple
1080 min_invariant. Otherwise, return T. */
1083 fully_constant_expression (pre_expr e)
1091 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1092 switch (TREE_CODE_CLASS (nary->opcode))
1094 case tcc_expression:
1095 if (nary->opcode == TRUTH_NOT_EXPR)
1097 if (nary->opcode != TRUTH_AND_EXPR
1098 && nary->opcode != TRUTH_OR_EXPR
1099 && nary->opcode != TRUTH_XOR_EXPR)
1103 case tcc_comparison:
1105 /* We have to go from trees to pre exprs to value ids to
1107 tree naryop0 = nary->op[0];
1108 tree naryop1 = nary->op[1];
1110 if (!is_gimple_min_invariant (naryop0))
1112 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1113 unsigned int vrep0 = get_expr_value_id (rep0);
1114 tree const0 = get_constant_for_value_id (vrep0);
1116 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1118 if (!is_gimple_min_invariant (naryop1))
1120 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1121 unsigned int vrep1 = get_expr_value_id (rep1);
1122 tree const1 = get_constant_for_value_id (vrep1);
1124 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1126 result = fold_binary (nary->opcode, nary->type,
1128 if (result && is_gimple_min_invariant (result))
1129 return get_or_alloc_expr_for_constant (result);
1130 /* We might have simplified the expression to a
1131 SSA_NAME for example from x_1 * 1. But we cannot
1132 insert a PHI for x_1 unconditionally as x_1 might
1133 not be available readily. */
1137 if (nary->opcode != REALPART_EXPR
1138 && nary->opcode != IMAGPART_EXPR
1139 && nary->opcode != VIEW_CONVERT_EXPR)
1145 /* We have to go from trees to pre exprs to value ids to
1147 tree naryop0 = nary->op[0];
1148 tree const0, result;
1149 if (is_gimple_min_invariant (naryop0))
1153 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1154 unsigned int vrep0 = get_expr_value_id (rep0);
1155 const0 = get_constant_for_value_id (vrep0);
1160 tree type1 = TREE_TYPE (nary->op[0]);
1161 const0 = fold_convert (type1, const0);
1162 result = fold_unary (nary->opcode, nary->type, const0);
1164 if (result && is_gimple_min_invariant (result))
1165 return get_or_alloc_expr_for_constant (result);
1174 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1175 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1176 vn_reference_op_t op;
1178 /* Try to simplify the translated expression if it is
1179 a call to a builtin function with at most two arguments. */
1180 op = VEC_index (vn_reference_op_s, operands, 0);
1181 if (op->opcode == CALL_EXPR
1182 && TREE_CODE (op->op0) == ADDR_EXPR
1183 && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL
1184 && DECL_BUILT_IN (TREE_OPERAND (op->op0, 0))
1185 && VEC_length (vn_reference_op_s, operands) >= 2
1186 && VEC_length (vn_reference_op_s, operands) <= 3)
1188 vn_reference_op_t arg0, arg1 = NULL;
1189 bool anyconst = false;
1190 arg0 = VEC_index (vn_reference_op_s, operands, 1);
1191 if (VEC_length (vn_reference_op_s, operands) > 2)
1192 arg1 = VEC_index (vn_reference_op_s, operands, 2);
1193 if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant
1194 || (arg0->opcode == ADDR_EXPR
1195 && is_gimple_min_invariant (arg0->op0)))
1198 && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant
1199 || (arg1->opcode == ADDR_EXPR
1200 && is_gimple_min_invariant (arg1->op0))))
1204 tree folded = build_call_expr (TREE_OPERAND (op->op0, 0),
1207 arg1 ? arg1->op0 : NULL);
1209 && TREE_CODE (folded) == NOP_EXPR)
1210 folded = TREE_OPERAND (folded, 0);
1212 && is_gimple_min_invariant (folded))
1213 return get_or_alloc_expr_for_constant (folded);
1224 /* Translate the vuses in the VUSES vector backwards through phi nodes
1225 in PHIBLOCK, so that they have the value they would have in
1228 static VEC(tree, gc) *
1229 translate_vuses_through_block (VEC (tree, gc) *vuses,
1230 basic_block phiblock,
1234 VEC(tree, gc) *result = NULL;
1237 for (i = 0; VEC_iterate (tree, vuses, i, oldvuse); i++)
1239 gimple phi = SSA_NAME_DEF_STMT (oldvuse);
1240 if (gimple_code (phi) == GIMPLE_PHI
1241 && gimple_bb (phi) == phiblock)
1243 edge e = find_edge (block, gimple_bb (phi));
1246 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1250 result = VEC_copy (tree, gc, vuses);
1251 VEC_replace (tree, result, i, def);
1257 /* We avoid creating a new copy of the vuses unless something
1258 actually changed, so result can be NULL. */
1261 sort_vuses (result);
1268 /* Like find_leader, but checks for the value existing in SET1 *or*
1269 SET2. This is used to avoid making a set consisting of the union
1270 of PA_IN and ANTIC_IN during insert. */
1272 static inline pre_expr
1273 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1277 result = bitmap_find_leader (set1, val, NULL);
1278 if (!result && set2)
1279 result = bitmap_find_leader (set2, val, NULL);
1283 /* Get the tree type for our PRE expression e. */
1286 get_expr_type (const pre_expr e)
1291 return TREE_TYPE (PRE_EXPR_NAME (e));
1293 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1296 vn_reference_op_t vro;
1298 gcc_assert (PRE_EXPR_REFERENCE (e)->operands);
1299 vro = VEC_index (vn_reference_op_s,
1300 PRE_EXPR_REFERENCE (e)->operands,
1302 /* We don't store type along with COMPONENT_REF because it is
1303 always the same as FIELD_DECL's type. */
1306 gcc_assert (vro->opcode == COMPONENT_REF);
1307 return TREE_TYPE (vro->op0);
1313 return PRE_EXPR_NARY (e)->type;
1318 /* Get a representative SSA_NAME for a given expression.
1319 Since all of our sub-expressions are treated as values, we require
1320 them to be SSA_NAME's for simplicity.
1321 Prior versions of GVNPRE used to use "value handles" here, so that
1322 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1323 either case, the operands are really values (IE we do not expect
1324 them to be usable without finding leaders). */
1327 get_representative_for (const pre_expr e)
1331 unsigned int value_id = get_expr_value_id (e);
1336 return PRE_EXPR_NAME (e);
1338 return PRE_EXPR_CONSTANT (e);
1342 /* Go through all of the expressions representing this value
1343 and pick out an SSA_NAME. */
1346 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1348 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1350 pre_expr rep = expression_for_id (i);
1351 if (rep->kind == NAME)
1352 return PRE_EXPR_NAME (rep);
1357 /* If we reached here we couldn't find an SSA_NAME. This can
1358 happen when we've discovered a value that has never appeared in
1359 the program as set to an SSA_NAME, most likely as the result of
1364 "Could not find SSA_NAME representative for expression:");
1365 print_pre_expr (dump_file, e);
1366 fprintf (dump_file, "\n");
1369 exprtype = get_expr_type (e);
1371 /* Build and insert the assignment of the end result to the temporary
1372 that we will return. */
1373 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1375 pretemp = create_tmp_var (exprtype, "pretmp");
1376 get_var_ann (pretemp);
1379 name = make_ssa_name (pretemp, gimple_build_nop ());
1380 VN_INFO_GET (name)->value_id = value_id;
1381 if (e->kind == CONSTANT)
1382 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1384 VN_INFO (name)->valnum = name;
1386 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1389 fprintf (dump_file, "Created SSA_NAME representative ");
1390 print_generic_expr (dump_file, name, 0);
1391 fprintf (dump_file, " for expression:");
1392 print_pre_expr (dump_file, e);
1393 fprintf (dump_file, "\n");
1402 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1403 the phis in PRED. SEEN is a bitmap saying which expression we have
1404 translated since we started translation of the toplevel expression.
1405 Return NULL if we can't find a leader for each part of the
1406 translated expression. */
1409 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1410 basic_block pred, basic_block phiblock, bitmap seen)
1412 pre_expr oldexpr = expr;
1418 if (value_id_constant_p (get_expr_value_id (expr)))
1421 phitrans = phi_trans_lookup (expr, pred);
1425 /* Prevent cycles when we have recursively dependent leaders. This
1426 can only happen when phi translating the maximal set. */
1429 unsigned int expr_id = get_expression_id (expr);
1430 if (bitmap_bit_p (seen, expr_id))
1432 bitmap_set_bit (seen, expr_id);
1437 /* Constants contain no values that need translation. */
1444 bool changed = false;
1445 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1446 struct vn_nary_op_s newnary;
1447 /* The NARY structure is only guaranteed to have been
1448 allocated to the nary->length operands. */
1449 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1450 - sizeof (tree) * (4 - nary->length)));
1452 for (i = 0; i < newnary.length; i++)
1454 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1458 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1459 pre_expr leader = find_leader_in_sets (op_val_id, set1, set2);
1460 pre_expr result = phi_translate_1 (leader, set1, set2,
1461 pred, phiblock, seen);
1462 if (result && result != leader)
1464 tree name = get_representative_for (result);
1467 newnary.op[i] = name;
1472 changed |= newnary.op[i] != nary->op[i];
1479 tree result = vn_nary_op_lookup_pieces (newnary.length,
1487 unsigned int new_val_id;
1489 expr = (pre_expr) pool_alloc (pre_expr_pool);
1492 if (result && is_gimple_min_invariant (result))
1493 return get_or_alloc_expr_for_constant (result);
1498 PRE_EXPR_NARY (expr) = nary;
1499 constant = fully_constant_expression (expr);
1500 if (constant != expr)
1503 new_val_id = nary->value_id;
1504 get_or_alloc_expression_id (expr);
1508 new_val_id = get_next_value_id ();
1509 VEC_safe_grow_cleared (bitmap_set_t, heap,
1511 get_max_value_id() + 1);
1512 nary = vn_nary_op_insert_pieces (newnary.length,
1519 result, new_val_id);
1520 PRE_EXPR_NARY (expr) = nary;
1521 constant = fully_constant_expression (expr);
1522 if (constant != expr)
1524 get_or_alloc_expression_id (expr);
1526 add_to_value (new_val_id, expr);
1528 phi_trans_add (oldexpr, expr, pred);
1535 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1536 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1537 VEC (tree, gc) *vuses = ref->vuses;
1538 VEC (tree, gc) *newvuses = vuses;
1539 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1540 bool changed = false;
1542 vn_reference_op_t operand;
1543 vn_reference_t newref;
1545 for (i = 0; VEC_iterate (vn_reference_op_s, operands, i, operand); i++)
1549 tree oldop0 = operand->op0;
1550 tree oldop1 = operand->op1;
1551 tree oldop2 = operand->op2;
1555 tree type = operand->type;
1556 vn_reference_op_s newop = *operand;
1558 if (op0 && TREE_CODE (op0) == SSA_NAME)
1560 unsigned int op_val_id = VN_INFO (op0)->value_id;
1561 leader = find_leader_in_sets (op_val_id, set1, set2);
1562 opresult = phi_translate_1 (leader, set1, set2,
1563 pred, phiblock, seen);
1564 if (opresult && opresult != leader)
1566 tree name = get_representative_for (opresult);
1574 changed |= op0 != oldop0;
1576 if (op1 && TREE_CODE (op1) == SSA_NAME)
1578 unsigned int op_val_id = VN_INFO (op1)->value_id;
1579 leader = find_leader_in_sets (op_val_id, set1, set2);
1580 opresult = phi_translate_1 (leader, set1, set2,
1581 pred, phiblock, seen);
1582 if (opresult && opresult != leader)
1584 tree name = get_representative_for (opresult);
1592 changed |= op1 != oldop1;
1593 if (op2 && TREE_CODE (op2) == SSA_NAME)
1595 unsigned int op_val_id = VN_INFO (op2)->value_id;
1596 leader = find_leader_in_sets (op_val_id, set1, set2);
1597 opresult = phi_translate_1 (leader, set1, set2,
1598 pred, phiblock, seen);
1599 if (opresult && opresult != leader)
1601 tree name = get_representative_for (opresult);
1609 changed |= op2 != oldop2;
1612 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1613 /* We may have changed from an SSA_NAME to a constant */
1614 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1615 newop.opcode = TREE_CODE (op0);
1620 VEC_replace (vn_reference_op_s, newoperands, i, &newop);
1622 if (i != VEC_length (vn_reference_op_s, operands))
1625 VEC_free (vn_reference_op_s, heap, newoperands);
1629 newvuses = translate_vuses_through_block (vuses, phiblock, pred);
1630 changed |= newvuses != vuses;
1634 unsigned int new_val_id;
1637 tree result = vn_reference_lookup_pieces (newvuses,
1641 VEC_free (vn_reference_op_s, heap, newoperands);
1643 if (result && is_gimple_min_invariant (result))
1645 gcc_assert (!newoperands);
1646 return get_or_alloc_expr_for_constant (result);
1649 expr = (pre_expr) pool_alloc (pre_expr_pool);
1650 expr->kind = REFERENCE;
1655 PRE_EXPR_REFERENCE (expr) = newref;
1656 constant = fully_constant_expression (expr);
1657 if (constant != expr)
1660 new_val_id = newref->value_id;
1661 get_or_alloc_expression_id (expr);
1665 new_val_id = get_next_value_id ();
1666 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
1667 get_max_value_id() + 1);
1668 newref = vn_reference_insert_pieces (newvuses,
1670 result, new_val_id);
1672 PRE_EXPR_REFERENCE (expr) = newref;
1673 constant = fully_constant_expression (expr);
1674 if (constant != expr)
1676 get_or_alloc_expression_id (expr);
1678 add_to_value (new_val_id, expr);
1680 VEC_free (vn_reference_op_s, heap, newoperands);
1681 phi_trans_add (oldexpr, expr, pred);
1691 tree name = PRE_EXPR_NAME (expr);
1693 def_stmt = SSA_NAME_DEF_STMT (name);
1694 if (gimple_code (def_stmt) == GIMPLE_PHI
1695 && gimple_bb (def_stmt) == phiblock)
1700 e = find_edge (pred, gimple_bb (phi));
1703 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1706 if (TREE_CODE (def) == SSA_NAME)
1707 def = VN_INFO (def)->valnum;
1709 /* Handle constant. */
1710 if (is_gimple_min_invariant (def))
1711 return get_or_alloc_expr_for_constant (def);
1713 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1716 newexpr = get_or_alloc_expr_for_name (def);
1727 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1729 Return NULL if we can't find a leader for each part of the
1730 translated expression. */
1733 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1734 basic_block pred, basic_block phiblock)
1736 bitmap_clear (seen_during_translate);
1737 return phi_translate_1 (expr, set1, set2, pred, phiblock,
1738 seen_during_translate);
1741 /* For each expression in SET, translate the values through phi nodes
1742 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1743 expressions in DEST. */
1746 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1747 basic_block phiblock)
1749 VEC (pre_expr, heap) *exprs;
1753 if (!phi_nodes (phiblock))
1755 bitmap_set_copy (dest, set);
1759 exprs = sorted_array_from_bitmap_set (set);
1760 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1762 pre_expr translated;
1763 translated = phi_translate (expr, set, NULL, pred, phiblock);
1765 /* Don't add empty translations to the cache */
1767 phi_trans_add (expr, translated, pred);
1769 if (translated != NULL)
1770 bitmap_value_insert_into_set (dest, translated);
1772 VEC_free (pre_expr, heap, exprs);
1775 /* Find the leader for a value (i.e., the name representing that
1776 value) in a given set, and return it. If STMT is non-NULL it
1777 makes sure the defining statement for the leader dominates it.
1778 Return NULL if no leader is found. */
1781 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1783 if (value_id_constant_p (val))
1787 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1789 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1791 pre_expr expr = expression_for_id (i);
1792 if (expr->kind == CONSTANT)
1796 if (bitmap_set_contains_value (set, val))
1798 /* Rather than walk the entire bitmap of expressions, and see
1799 whether any of them has the value we are looking for, we look
1800 at the reverse mapping, which tells us the set of expressions
1801 that have a given value (IE value->expressions with that
1802 value) and see if any of those expressions are in our set.
1803 The number of expressions per value is usually significantly
1804 less than the number of expressions in the set. In fact, for
1805 large testcases, doing it this way is roughly 5-10x faster
1806 than walking the bitmap.
1807 If this is somehow a significant lose for some cases, we can
1808 choose which set to walk based on which set is smaller. */
1811 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1813 EXECUTE_IF_AND_IN_BITMAP (exprset->expressions,
1814 set->expressions, 0, i, bi)
1816 pre_expr val = expression_for_id (i);
1817 /* At the point where stmt is not null, there should always
1818 be an SSA_NAME first in the list of expressions. */
1821 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1822 if (gimple_code (def_stmt) != GIMPLE_PHI
1823 && gimple_bb (def_stmt) == gimple_bb (stmt)
1824 && gimple_uid (def_stmt) >= gimple_uid (stmt))
1833 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1834 BLOCK by seeing if it is not killed in the block. Note that we are
1835 only determining whether there is a store that kills it. Because
1836 of the order in which clean iterates over values, we are guaranteed
1837 that altered operands will have caused us to be eliminated from the
1838 ANTIC_IN set already. */
1841 value_dies_in_block_x (pre_expr expr, basic_block block)
1845 VEC (tree, gc) *vuses = PRE_EXPR_REFERENCE (expr)->vuses;
1847 /* Conservatively, a value dies if it's vuses are defined in this
1848 block, unless they come from phi nodes (which are merge operations,
1849 rather than stores. */
1850 for (i = 0; VEC_iterate (tree, vuses, i, vuse); i++)
1852 gimple def = SSA_NAME_DEF_STMT (vuse);
1854 if (gimple_bb (def) != block)
1856 if (gimple_code (def) == GIMPLE_PHI)
1864 #define union_contains_value(SET1, SET2, VAL) \
1865 (bitmap_set_contains_value ((SET1), (VAL)) \
1866 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
1868 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
1871 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
1872 vn_reference_op_t vro)
1874 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
1876 struct pre_expr_d temp;
1879 PRE_EXPR_NAME (&temp) = vro->op0;
1880 temp.id = lookup_expression_id (&temp);
1883 if (!union_contains_value (set1, set2,
1884 get_expr_value_id (&temp)))
1887 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
1889 struct pre_expr_d temp;
1892 PRE_EXPR_NAME (&temp) = vro->op1;
1893 temp.id = lookup_expression_id (&temp);
1896 if (!union_contains_value (set1, set2,
1897 get_expr_value_id (&temp)))
1901 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
1903 struct pre_expr_d temp;
1906 PRE_EXPR_NAME (&temp) = vro->op2;
1907 temp.id = lookup_expression_id (&temp);
1910 if (!union_contains_value (set1, set2,
1911 get_expr_value_id (&temp)))
1918 /* Determine if the expression EXPR is valid in SET1 U SET2.
1919 ONLY SET2 CAN BE NULL.
1920 This means that we have a leader for each part of the expression
1921 (if it consists of values), or the expression is an SSA_NAME.
1922 For loads/calls, we also see if the vuses are killed in this block.
1926 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
1932 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
1936 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1937 for (i = 0; i < nary->length; i++)
1939 if (TREE_CODE (nary->op[i]) == SSA_NAME)
1941 struct pre_expr_d temp;
1944 PRE_EXPR_NAME (&temp) = nary->op[i];
1945 temp.id = lookup_expression_id (&temp);
1948 if (!union_contains_value (set1, set2,
1949 get_expr_value_id (&temp)))
1958 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1959 vn_reference_op_t vro;
1962 for (i = 0; VEC_iterate (vn_reference_op_s, ref->operands, i, vro); i++)
1964 if (!vro_valid_in_sets (set1, set2, vro))
1967 return !value_dies_in_block_x (expr, block);
1974 /* Clean the set of expressions that are no longer valid in SET1 or
1975 SET2. This means expressions that are made up of values we have no
1976 leaders for in SET1 or SET2. This version is used for partial
1977 anticipation, which means it is not valid in either ANTIC_IN or
1981 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
1983 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
1987 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1989 if (!valid_in_sets (set1, set2, expr, block))
1990 bitmap_remove_from_set (set1, expr);
1992 VEC_free (pre_expr, heap, exprs);
1995 /* Clean the set of expressions that are no longer valid in SET. This
1996 means expressions that are made up of values we have no leaders for
2000 clean (bitmap_set_t set, basic_block block)
2002 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2006 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2008 if (!valid_in_sets (set, NULL, expr, block))
2009 bitmap_remove_from_set (set, expr);
2011 VEC_free (pre_expr, heap, exprs);
2014 static sbitmap has_abnormal_preds;
2016 /* List of blocks that may have changed during ANTIC computation and
2017 thus need to be iterated over. */
2019 static sbitmap changed_blocks;
2021 /* Decide whether to defer a block for a later iteration, or PHI
2022 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2023 should defer the block, and true if we processed it. */
2026 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2027 basic_block block, basic_block phiblock)
2029 if (!BB_VISITED (phiblock))
2031 SET_BIT (changed_blocks, block->index);
2032 BB_VISITED (block) = 0;
2033 BB_DEFERRED (block) = 1;
2037 phi_translate_set (dest, source, block, phiblock);
2041 /* Compute the ANTIC set for BLOCK.
2043 If succs(BLOCK) > 1 then
2044 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2045 else if succs(BLOCK) == 1 then
2046 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2048 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2052 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2054 bool changed = false;
2055 bitmap_set_t S, old, ANTIC_OUT;
2061 old = ANTIC_OUT = S = NULL;
2062 BB_VISITED (block) = 1;
2064 /* If any edges from predecessors are abnormal, antic_in is empty,
2066 if (block_has_abnormal_pred_edge)
2067 goto maybe_dump_sets;
2069 old = ANTIC_IN (block);
2070 ANTIC_OUT = bitmap_set_new ();
2072 /* If the block has no successors, ANTIC_OUT is empty. */
2073 if (EDGE_COUNT (block->succs) == 0)
2075 /* If we have one successor, we could have some phi nodes to
2076 translate through. */
2077 else if (single_succ_p (block))
2079 basic_block succ_bb = single_succ (block);
2081 /* We trade iterations of the dataflow equations for having to
2082 phi translate the maximal set, which is incredibly slow
2083 (since the maximal set often has 300+ members, even when you
2084 have a small number of blocks).
2085 Basically, we defer the computation of ANTIC for this block
2086 until we have processed it's successor, which will inevitably
2087 have a *much* smaller set of values to phi translate once
2088 clean has been run on it.
2089 The cost of doing this is that we technically perform more
2090 iterations, however, they are lower cost iterations.
2092 Timings for PRE on tramp3d-v4:
2093 without maximal set fix: 11 seconds
2094 with maximal set fix/without deferring: 26 seconds
2095 with maximal set fix/with deferring: 11 seconds
2098 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2102 goto maybe_dump_sets;
2105 /* If we have multiple successors, we take the intersection of all of
2106 them. Note that in the case of loop exit phi nodes, we may have
2107 phis to translate through. */
2110 VEC(basic_block, heap) * worklist;
2112 basic_block bprime, first = NULL;
2114 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2115 FOR_EACH_EDGE (e, ei, block->succs)
2118 && BB_VISITED (e->dest))
2120 else if (BB_VISITED (e->dest))
2121 VEC_quick_push (basic_block, worklist, e->dest);
2124 /* Of multiple successors we have to have visited one already. */
2127 SET_BIT (changed_blocks, block->index);
2128 BB_VISITED (block) = 0;
2129 BB_DEFERRED (block) = 1;
2131 VEC_free (basic_block, heap, worklist);
2132 goto maybe_dump_sets;
2135 if (phi_nodes (first))
2136 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2138 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2140 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2142 if (phi_nodes (bprime))
2144 bitmap_set_t tmp = bitmap_set_new ();
2145 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2146 bitmap_set_and (ANTIC_OUT, tmp);
2147 bitmap_set_free (tmp);
2150 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2152 VEC_free (basic_block, heap, worklist);
2155 /* Generate ANTIC_OUT - TMP_GEN. */
2156 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2158 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2159 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2162 /* Then union in the ANTIC_OUT - TMP_GEN values,
2163 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2164 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2165 bitmap_value_insert_into_set (ANTIC_IN (block),
2166 expression_for_id (bii));
2168 clean (ANTIC_IN (block), block);
2170 /* !old->expressions can happen when we deferred a block. */
2171 if (!old->expressions || !bitmap_set_equal (old, ANTIC_IN (block)))
2174 SET_BIT (changed_blocks, block->index);
2175 FOR_EACH_EDGE (e, ei, block->preds)
2176 SET_BIT (changed_blocks, e->src->index);
2179 RESET_BIT (changed_blocks, block->index);
2182 if (dump_file && (dump_flags & TDF_DETAILS))
2184 if (!BB_DEFERRED (block) || BB_VISITED (block))
2187 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2189 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2193 print_bitmap_set (dump_file, S, "S", block->index);
2198 "Block %d was deferred for a future iteration.\n",
2203 bitmap_set_free (old);
2205 bitmap_set_free (S);
2207 bitmap_set_free (ANTIC_OUT);
2211 /* Compute PARTIAL_ANTIC for BLOCK.
2213 If succs(BLOCK) > 1 then
2214 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2215 in ANTIC_OUT for all succ(BLOCK)
2216 else if succs(BLOCK) == 1 then
2217 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2219 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2224 compute_partial_antic_aux (basic_block block,
2225 bool block_has_abnormal_pred_edge)
2227 bool changed = false;
2228 bitmap_set_t old_PA_IN;
2229 bitmap_set_t PA_OUT;
2232 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2234 old_PA_IN = PA_OUT = NULL;
2236 /* If any edges from predecessors are abnormal, antic_in is empty,
2238 if (block_has_abnormal_pred_edge)
2239 goto maybe_dump_sets;
2241 /* If there are too many partially anticipatable values in the
2242 block, phi_translate_set can take an exponential time: stop
2243 before the translation starts. */
2245 && single_succ_p (block)
2246 && bitmap_count_bits (PA_IN (single_succ (block))->values) > max_pa)
2247 goto maybe_dump_sets;
2249 old_PA_IN = PA_IN (block);
2250 PA_OUT = bitmap_set_new ();
2252 /* If the block has no successors, ANTIC_OUT is empty. */
2253 if (EDGE_COUNT (block->succs) == 0)
2255 /* If we have one successor, we could have some phi nodes to
2256 translate through. Note that we can't phi translate across DFS
2257 back edges in partial antic, because it uses a union operation on
2258 the successors. For recurrences like IV's, we will end up
2259 generating a new value in the set on each go around (i + 3 (VH.1)
2260 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2261 else if (single_succ_p (block))
2263 basic_block succ = single_succ (block);
2264 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2265 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2267 /* If we have multiple successors, we take the union of all of
2271 VEC(basic_block, heap) * worklist;
2275 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2276 FOR_EACH_EDGE (e, ei, block->succs)
2278 if (e->flags & EDGE_DFS_BACK)
2280 VEC_quick_push (basic_block, worklist, e->dest);
2282 if (VEC_length (basic_block, worklist) > 0)
2284 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2289 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2290 bitmap_value_insert_into_set (PA_OUT,
2291 expression_for_id (i));
2292 if (phi_nodes (bprime))
2294 bitmap_set_t pa_in = bitmap_set_new ();
2295 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2296 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2297 bitmap_value_insert_into_set (PA_OUT,
2298 expression_for_id (i));
2299 bitmap_set_free (pa_in);
2302 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2303 bitmap_value_insert_into_set (PA_OUT,
2304 expression_for_id (i));
2307 VEC_free (basic_block, heap, worklist);
2310 /* PA_IN starts with PA_OUT - TMP_GEN.
2311 Then we subtract things from ANTIC_IN. */
2312 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2314 /* For partial antic, we want to put back in the phi results, since
2315 we will properly avoid making them partially antic over backedges. */
2316 bitmap_ior_into (PA_IN (block)->values, PHI_GEN (block)->values);
2317 bitmap_ior_into (PA_IN (block)->expressions, PHI_GEN (block)->expressions);
2319 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2320 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2322 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2324 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2327 SET_BIT (changed_blocks, block->index);
2328 FOR_EACH_EDGE (e, ei, block->preds)
2329 SET_BIT (changed_blocks, e->src->index);
2332 RESET_BIT (changed_blocks, block->index);
2335 if (dump_file && (dump_flags & TDF_DETAILS))
2338 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2340 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2343 bitmap_set_free (old_PA_IN);
2345 bitmap_set_free (PA_OUT);
2349 /* Compute ANTIC and partial ANTIC sets. */
2352 compute_antic (void)
2354 bool changed = true;
2355 int num_iterations = 0;
2359 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2360 We pre-build the map of blocks with incoming abnormal edges here. */
2361 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2362 sbitmap_zero (has_abnormal_preds);
2369 FOR_EACH_EDGE (e, ei, block->preds)
2371 e->flags &= ~EDGE_DFS_BACK;
2372 if (e->flags & EDGE_ABNORMAL)
2374 SET_BIT (has_abnormal_preds, block->index);
2379 BB_VISITED (block) = 0;
2380 BB_DEFERRED (block) = 0;
2381 /* While we are here, give empty ANTIC_IN sets to each block. */
2382 ANTIC_IN (block) = bitmap_set_new ();
2383 PA_IN (block) = bitmap_set_new ();
2386 /* At the exit block we anticipate nothing. */
2387 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2388 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2389 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2391 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2392 sbitmap_ones (changed_blocks);
2395 if (dump_file && (dump_flags & TDF_DETAILS))
2396 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2399 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2401 if (TEST_BIT (changed_blocks, postorder[i]))
2403 basic_block block = BASIC_BLOCK (postorder[i]);
2404 changed |= compute_antic_aux (block,
2405 TEST_BIT (has_abnormal_preds,
2409 #ifdef ENABLE_CHECKING
2410 /* Theoretically possible, but *highly* unlikely. */
2411 gcc_assert (num_iterations < 500);
2415 statistics_histogram_event (cfun, "compute_antic iterations",
2418 if (do_partial_partial)
2420 sbitmap_ones (changed_blocks);
2421 mark_dfs_back_edges ();
2426 if (dump_file && (dump_flags & TDF_DETAILS))
2427 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2430 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2432 if (TEST_BIT (changed_blocks, postorder[i]))
2434 basic_block block = BASIC_BLOCK (postorder[i]);
2436 |= compute_partial_antic_aux (block,
2437 TEST_BIT (has_abnormal_preds,
2441 #ifdef ENABLE_CHECKING
2442 /* Theoretically possible, but *highly* unlikely. */
2443 gcc_assert (num_iterations < 500);
2446 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2449 sbitmap_free (has_abnormal_preds);
2450 sbitmap_free (changed_blocks);
2453 /* Return true if we can value number the call in STMT. This is true
2454 if we have a pure or constant call. */
2457 can_value_number_call (gimple stmt)
2459 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2464 /* Return true if OP is an exception handler related operation, such as
2465 FILTER_EXPR or EXC_PTR_EXPR. */
2468 is_exception_related (gimple stmt)
2470 return (is_gimple_assign (stmt)
2471 && (gimple_assign_rhs_code (stmt) == FILTER_EXPR
2472 || gimple_assign_rhs_code (stmt) == EXC_PTR_EXPR));
2475 /* Return true if OP is a tree which we can perform PRE on
2476 on. This may not match the operations we can value number, but in
2477 a perfect world would. */
2480 can_PRE_operation (tree op)
2482 return UNARY_CLASS_P (op)
2483 || BINARY_CLASS_P (op)
2484 || COMPARISON_CLASS_P (op)
2485 || TREE_CODE (op) == INDIRECT_REF
2486 || TREE_CODE (op) == COMPONENT_REF
2487 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2488 || TREE_CODE (op) == CALL_EXPR
2489 || TREE_CODE (op) == ARRAY_REF;
2493 /* Inserted expressions are placed onto this worklist, which is used
2494 for performing quick dead code elimination of insertions we made
2495 that didn't turn out to be necessary. */
2496 static VEC(gimple,heap) *inserted_exprs;
2498 /* Pool allocated fake store expressions are placed onto this
2499 worklist, which, after performing dead code elimination, is walked
2500 to see which expressions need to be put into GC'able memory */
2501 static VEC(gimple, heap) *need_creation;
2503 /* The actual worker for create_component_ref_by_pieces. */
2506 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2507 unsigned int *operand, gimple_seq *stmts,
2510 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2514 switch (currop->opcode)
2518 tree folded, sc = currop->op1;
2519 unsigned int nargs = 0;
2520 tree *args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2521 ref->operands) - 1);
2522 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2524 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2529 folded = build_call_array (currop->type,
2530 TREE_CODE (currop->op0) == FUNCTION_DECL
2531 ? build_fold_addr_expr (currop->op0)
2537 pre_expr scexpr = get_or_alloc_expr_for (sc);
2538 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2541 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2549 gcc_assert (is_gimple_min_invariant (currop->op0));
2555 case VIEW_CONVERT_EXPR:
2558 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2563 folded = fold_build1 (currop->opcode, currop->type,
2568 case ALIGN_INDIRECT_REF:
2569 case MISALIGNED_INDIRECT_REF:
2573 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2578 genop1 = fold_convert (build_pointer_type (currop->type),
2581 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2582 folded = fold_build2 (currop->opcode, currop->type,
2583 genop1, currop->op1);
2585 folded = fold_build1 (currop->opcode, currop->type,
2593 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2595 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2596 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2602 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2605 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2608 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2613 /* For array ref vn_reference_op's, operand 1 of the array ref
2614 is op0 of the reference op and operand 3 of the array ref is
2616 case ARRAY_RANGE_REF:
2620 tree genop1 = currop->op0;
2622 tree genop2 = currop->op1;
2625 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2629 op1expr = get_or_alloc_expr_for (genop1);
2630 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2635 op2expr = get_or_alloc_expr_for (genop2);
2636 genop2 = find_or_generate_expression (block, op2expr, stmts,
2642 genop3 = currop->op2;
2643 return build4 (currop->opcode, currop->type, genop0, genop1,
2650 tree genop2 = currop->op1;
2652 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2656 /* op1 should be a FIELD_DECL, which are represented by
2661 op2expr = get_or_alloc_expr_for (genop2);
2662 genop2 = find_or_generate_expression (block, op2expr, stmts,
2668 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2674 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2675 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2696 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2697 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2698 trying to rename aggregates into ssa form directly, which is a no no.
2700 Thus, this routine doesn't create temporaries, it just builds a
2701 single access expression for the array, calling
2702 find_or_generate_expression to build the innermost pieces.
2704 This function is a subroutine of create_expression_by_pieces, and
2705 should not be called on it's own unless you really know what you
2709 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2710 gimple_seq *stmts, gimple domstmt)
2712 unsigned int op = 0;
2713 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2716 /* Find a leader for an expression, or generate one using
2717 create_expression_by_pieces if it's ANTIC but
2719 BLOCK is the basic_block we are looking for leaders in.
2720 EXPR is the expression to find a leader or generate for.
2721 STMTS is the statement list to put the inserted expressions on.
2722 Returns the SSA_NAME of the LHS of the generated expression or the
2724 DOMSTMT if non-NULL is a statement that should be dominated by
2725 all uses in the generated expression. If DOMSTMT is non-NULL this
2726 routine can fail and return NULL_TREE. Otherwise it will assert
2730 find_or_generate_expression (basic_block block, pre_expr expr,
2731 gimple_seq *stmts, gimple domstmt)
2733 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2734 get_expr_value_id (expr), domstmt);
2738 if (leader->kind == NAME)
2739 genop = PRE_EXPR_NAME (leader);
2740 else if (leader->kind == CONSTANT)
2741 genop = PRE_EXPR_CONSTANT (leader);
2744 /* If it's still NULL, it must be a complex expression, so generate
2745 it recursively. Not so for FRE though. */
2749 bitmap_set_t exprset;
2750 unsigned int lookfor = get_expr_value_id (expr);
2751 bool handled = false;
2755 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2756 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2758 pre_expr temp = expression_for_id (i);
2759 if (temp->kind != NAME)
2762 genop = create_expression_by_pieces (block, temp, stmts,
2764 get_expr_type (expr));
2768 if (!handled && domstmt)
2771 gcc_assert (handled);
2776 #define NECESSARY GF_PLF_1
2778 /* Create an expression in pieces, so that we can handle very complex
2779 expressions that may be ANTIC, but not necessary GIMPLE.
2780 BLOCK is the basic block the expression will be inserted into,
2781 EXPR is the expression to insert (in value form)
2782 STMTS is a statement list to append the necessary insertions into.
2784 This function will die if we hit some value that shouldn't be
2785 ANTIC but is (IE there is no leader for it, or its components).
2786 This function may also generate expressions that are themselves
2787 partially or fully redundant. Those that are will be either made
2788 fully redundant during the next iteration of insert (for partially
2789 redundant ones), or eliminated by eliminate (for fully redundant
2792 If DOMSTMT is non-NULL then we make sure that all uses in the
2793 expressions dominate that statement. In this case the function
2794 can return NULL_TREE to signal failure. */
2797 create_expression_by_pieces (basic_block block, pre_expr expr,
2798 gimple_seq *stmts, gimple domstmt, tree type)
2801 tree folded, newexpr;
2802 gimple_seq forced_stmts;
2803 unsigned int value_id;
2804 gimple_stmt_iterator gsi;
2805 tree exprtype = type ? type : get_expr_type (expr);
2811 /* We may hit the NAME/CONSTANT case if we have to convert types
2812 that value numbering saw through. */
2814 folded = PRE_EXPR_NAME (expr);
2817 folded = PRE_EXPR_CONSTANT (expr);
2821 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2822 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
2827 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2828 switch (nary->length)
2832 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2833 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
2834 tree genop1 = find_or_generate_expression (block, op1,
2836 tree genop2 = find_or_generate_expression (block, op2,
2838 if (!genop1 || !genop2)
2840 genop1 = fold_convert (TREE_TYPE (nary->op[0]),
2842 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
2843 may be a constant with the wrong type. */
2844 if (nary->opcode == POINTER_PLUS_EXPR)
2845 genop2 = fold_convert (sizetype, genop2);
2847 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
2849 folded = fold_build2 (nary->opcode, nary->type,
2855 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2856 tree genop1 = find_or_generate_expression (block, op1,
2860 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
2862 folded = fold_build1 (nary->opcode, nary->type,
2874 folded = fold_convert (exprtype, folded);
2875 /* Force the generated expression to be a sequence of GIMPLE
2877 We have to call unshare_expr because force_gimple_operand may
2878 modify the tree we pass to it. */
2879 newexpr = force_gimple_operand (unshare_expr (folded), &forced_stmts,
2882 /* If we have any intermediate expressions to the value sets, add them
2883 to the value sets and chain them in the instruction stream. */
2886 gsi = gsi_start (forced_stmts);
2887 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2889 gimple stmt = gsi_stmt (gsi);
2890 tree forcedname = gimple_get_lhs (stmt);
2893 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
2894 if (TREE_CODE (forcedname) == SSA_NAME)
2896 VN_INFO_GET (forcedname)->valnum = forcedname;
2897 VN_INFO (forcedname)->value_id = get_next_value_id ();
2898 nameexpr = get_or_alloc_expr_for_name (forcedname);
2899 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2901 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2902 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2904 mark_symbols_for_renaming (stmt);
2906 gimple_seq_add_seq (stmts, forced_stmts);
2909 /* Build and insert the assignment of the end result to the temporary
2910 that we will return. */
2911 if (!pretemp || exprtype != TREE_TYPE (pretemp))
2913 pretemp = create_tmp_var (exprtype, "pretmp");
2914 get_var_ann (pretemp);
2918 add_referenced_var (temp);
2920 if (TREE_CODE (exprtype) == COMPLEX_TYPE
2921 || TREE_CODE (exprtype) == VECTOR_TYPE)
2922 DECL_GIMPLE_REG_P (temp) = 1;
2924 newstmt = gimple_build_assign (temp, newexpr);
2925 name = make_ssa_name (temp, newstmt);
2926 gimple_assign_set_lhs (newstmt, name);
2927 gimple_set_plf (newstmt, NECESSARY, false);
2929 gimple_seq_add_stmt (stmts, newstmt);
2930 VEC_safe_push (gimple, heap, inserted_exprs, newstmt);
2932 /* All the symbols in NEWEXPR should be put into SSA form. */
2933 mark_symbols_for_renaming (newstmt);
2935 /* Add a value number to the temporary.
2936 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2937 we are creating the expression by pieces, and this particular piece of
2938 the expression may have been represented. There is no harm in replacing
2940 VN_INFO_GET (name)->valnum = name;
2941 value_id = get_expr_value_id (expr);
2942 VN_INFO (name)->value_id = value_id;
2943 nameexpr = get_or_alloc_expr_for_name (name);
2944 add_to_value (value_id, nameexpr);
2946 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2947 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2949 pre_stats.insertions++;
2950 if (dump_file && (dump_flags & TDF_DETAILS))
2952 fprintf (dump_file, "Inserted ");
2953 print_gimple_stmt (dump_file, newstmt, 0, 0);
2954 fprintf (dump_file, " in predecessor %d\n", block->index);
2961 /* Insert the to-be-made-available values of expression EXPRNUM for each
2962 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
2963 merge the result with a phi node, given the same value number as
2964 NODE. Return true if we have inserted new stuff. */
2967 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
2970 pre_expr expr = expression_for_id (exprnum);
2972 unsigned int val = get_expr_value_id (expr);
2974 bool insertions = false;
2979 tree type = get_expr_type (expr);
2983 if (dump_file && (dump_flags & TDF_DETAILS))
2985 fprintf (dump_file, "Found partial redundancy for expression ");
2986 print_pre_expr (dump_file, expr);
2987 fprintf (dump_file, " (%04d)\n", val);
2990 /* Make sure we aren't creating an induction variable. */
2991 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2
2992 && expr->kind != REFERENCE)
2994 bool firstinsideloop = false;
2995 bool secondinsideloop = false;
2996 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
2997 EDGE_PRED (block, 0)->src);
2998 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
2999 EDGE_PRED (block, 1)->src);
3000 /* Induction variables only have one edge inside the loop. */
3001 if (firstinsideloop ^ secondinsideloop)
3003 if (dump_file && (dump_flags & TDF_DETAILS))
3004 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3009 /* Make sure we are not inserting trapping expressions. */
3010 FOR_EACH_EDGE (pred, ei, block->preds)
3013 eprime = avail[bprime->index];
3014 if (eprime->kind == NARY
3015 && vn_nary_may_trap (PRE_EXPR_NARY (eprime)))
3019 /* Make the necessary insertions. */
3020 FOR_EACH_EDGE (pred, ei, block->preds)
3022 gimple_seq stmts = NULL;
3025 eprime = avail[bprime->index];
3027 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3029 builtexpr = create_expression_by_pieces (bprime,
3033 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3034 gsi_insert_seq_on_edge (pred, stmts);
3035 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3038 else if (eprime->kind == CONSTANT)
3040 /* Constants may not have the right type, fold_convert
3041 should give us back a constant with the right type.
3043 tree constant = PRE_EXPR_CONSTANT (eprime);
3044 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3046 tree builtexpr = fold_convert (type, constant);
3047 if (!is_gimple_min_invariant (builtexpr))
3049 tree forcedexpr = force_gimple_operand (builtexpr,
3052 if (!is_gimple_min_invariant (forcedexpr))
3054 if (forcedexpr != builtexpr)
3056 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3057 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3061 gimple_stmt_iterator gsi;
3062 gsi = gsi_start (stmts);
3063 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3065 gimple stmt = gsi_stmt (gsi);
3066 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3067 gimple_set_plf (stmt, NECESSARY, false);
3069 gsi_insert_seq_on_edge (pred, stmts);
3071 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3076 else if (eprime->kind == NAME)
3078 /* We may have to do a conversion because our value
3079 numbering can look through types in certain cases, but
3080 our IL requires all operands of a phi node have the same
3082 tree name = PRE_EXPR_NAME (eprime);
3083 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3087 builtexpr = fold_convert (type, name);
3088 forcedexpr = force_gimple_operand (builtexpr,
3092 if (forcedexpr != name)
3094 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3095 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3100 gimple_stmt_iterator gsi;
3101 gsi = gsi_start (stmts);
3102 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3104 gimple stmt = gsi_stmt (gsi);
3105 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3106 gimple_set_plf (stmt, NECESSARY, false);
3108 gsi_insert_seq_on_edge (pred, stmts);
3110 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3114 /* If we didn't want a phi node, and we made insertions, we still have
3115 inserted new stuff, and thus return true. If we didn't want a phi node,
3116 and didn't make insertions, we haven't added anything new, so return
3118 if (nophi && insertions)
3120 else if (nophi && !insertions)
3123 /* Now build a phi for the new variable. */
3124 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3126 prephitemp = create_tmp_var (type, "prephitmp");
3127 get_var_ann (prephitemp);
3131 add_referenced_var (temp);
3133 if (TREE_CODE (type) == COMPLEX_TYPE
3134 || TREE_CODE (type) == VECTOR_TYPE)
3135 DECL_GIMPLE_REG_P (temp) = 1;
3136 phi = create_phi_node (temp, block);
3138 gimple_set_plf (phi, NECESSARY, false);
3139 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3140 VN_INFO (gimple_phi_result (phi))->value_id = val;
3141 VEC_safe_push (gimple, heap, inserted_exprs, phi);
3142 FOR_EACH_EDGE (pred, ei, block->preds)
3144 pre_expr ae = avail[pred->src->index];
3145 gcc_assert (get_expr_type (ae) == type
3146 || useless_type_conversion_p (type, get_expr_type (ae)));
3147 if (ae->kind == CONSTANT)
3148 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred);
3150 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred);
3153 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3154 add_to_value (val, newphi);
3156 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3157 this insertion, since we test for the existence of this value in PHI_GEN
3158 before proceeding with the partial redundancy checks in insert_aux.
3160 The value may exist in AVAIL_OUT, in particular, it could be represented
3161 by the expression we are trying to eliminate, in which case we want the
3162 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3165 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3166 this block, because if it did, it would have existed in our dominator's
3167 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3170 bitmap_insert_into_set (PHI_GEN (block), newphi);
3171 bitmap_value_replace_in_set (AVAIL_OUT (block),
3173 bitmap_insert_into_set (NEW_SETS (block),
3176 if (dump_file && (dump_flags & TDF_DETAILS))
3178 fprintf (dump_file, "Created phi ");
3179 print_gimple_stmt (dump_file, phi, 0, 0);
3180 fprintf (dump_file, " in block %d\n", block->index);
3188 /* Perform insertion of partially redundant values.
3189 For BLOCK, do the following:
3190 1. Propagate the NEW_SETS of the dominator into the current block.
3191 If the block has multiple predecessors,
3192 2a. Iterate over the ANTIC expressions for the block to see if
3193 any of them are partially redundant.
3194 2b. If so, insert them into the necessary predecessors to make
3195 the expression fully redundant.
3196 2c. Insert a new PHI merging the values of the predecessors.
3197 2d. Insert the new PHI, and the new expressions, into the
3199 3. Recursively call ourselves on the dominator children of BLOCK.
3201 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3202 do_regular_insertion and do_partial_insertion.
3207 do_regular_insertion (basic_block block, basic_block dom)
3209 bool new_stuff = false;
3210 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3214 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3216 if (expr->kind != NAME)
3220 bool by_some = false;
3221 bool cant_insert = false;
3222 bool all_same = true;
3223 pre_expr first_s = NULL;
3226 pre_expr eprime = NULL;
3228 pre_expr edoubleprime = NULL;
3230 val = get_expr_value_id (expr);
3231 if (bitmap_set_contains_value (PHI_GEN (block), val))
3233 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3235 if (dump_file && (dump_flags & TDF_DETAILS))
3236 fprintf (dump_file, "Found fully redundant value\n");
3240 avail = XCNEWVEC (pre_expr, last_basic_block);
3241 FOR_EACH_EDGE (pred, ei, block->preds)
3243 unsigned int vprime;
3245 /* We should never run insertion for the exit block
3246 and so not come across fake pred edges. */
3247 gcc_assert (!(pred->flags & EDGE_FAKE));
3249 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3252 /* eprime will generally only be NULL if the
3253 value of the expression, translated
3254 through the PHI for this predecessor, is
3255 undefined. If that is the case, we can't
3256 make the expression fully redundant,
3257 because its value is undefined along a
3258 predecessor path. We can thus break out
3259 early because it doesn't matter what the
3260 rest of the results are. */
3267 eprime = fully_constant_expression (eprime);
3268 vprime = get_expr_value_id (eprime);
3269 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3271 if (edoubleprime == NULL)
3273 avail[bprime->index] = eprime;
3278 avail[bprime->index] = edoubleprime;
3280 if (first_s == NULL)
3281 first_s = edoubleprime;
3282 else if (!pre_expr_eq (first_s, edoubleprime))
3286 /* If we can insert it, it's not the same value
3287 already existing along every predecessor, and
3288 it's defined by some predecessor, it is
3289 partially redundant. */
3290 if (!cant_insert && !all_same && by_some && dbg_cnt (treepre_insert))
3292 if (insert_into_preds_of_block (block, get_expression_id (expr),
3296 /* If all edges produce the same value and that value is
3297 an invariant, then the PHI has the same value on all
3298 edges. Note this. */
3299 else if (!cant_insert && all_same && eprime
3300 && (edoubleprime->kind == CONSTANT
3301 || edoubleprime->kind == NAME)
3302 && !value_id_constant_p (val))
3306 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3307 value_expressions, val);
3309 unsigned int new_val = get_expr_value_id (edoubleprime);
3310 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3312 pre_expr expr = expression_for_id (j);
3314 if (expr->kind == NAME)
3316 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3317 /* Just reset the value id and valnum so it is
3318 the same as the constant we have discovered. */
3319 if (edoubleprime->kind == CONSTANT)
3321 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3322 pre_stats.constified++;
3325 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3326 info->value_id = new_val;
3334 VEC_free (pre_expr, heap, exprs);
3339 /* Perform insertion for partially anticipatable expressions. There
3340 is only one case we will perform insertion for these. This case is
3341 if the expression is partially anticipatable, and fully available.
3342 In this case, we know that putting it earlier will enable us to
3343 remove the later computation. */
3347 do_partial_partial_insertion (basic_block block, basic_block dom)
3349 bool new_stuff = false;
3350 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3354 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3356 if (expr->kind != NAME)
3361 bool cant_insert = false;
3364 pre_expr eprime = NULL;
3367 val = get_expr_value_id (expr);
3368 if (bitmap_set_contains_value (PHI_GEN (block), val))
3370 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3373 avail = XCNEWVEC (pre_expr, last_basic_block);
3374 FOR_EACH_EDGE (pred, ei, block->preds)
3376 unsigned int vprime;
3377 pre_expr edoubleprime;
3379 /* We should never run insertion for the exit block
3380 and so not come across fake pred edges. */
3381 gcc_assert (!(pred->flags & EDGE_FAKE));
3383 eprime = phi_translate (expr, ANTIC_IN (block),
3387 /* eprime will generally only be NULL if the
3388 value of the expression, translated
3389 through the PHI for this predecessor, is
3390 undefined. If that is the case, we can't
3391 make the expression fully redundant,
3392 because its value is undefined along a
3393 predecessor path. We can thus break out
3394 early because it doesn't matter what the
3395 rest of the results are. */
3402 eprime = fully_constant_expression (eprime);
3403 vprime = get_expr_value_id (eprime);
3404 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3406 if (edoubleprime == NULL)
3412 avail[bprime->index] = edoubleprime;
3416 /* If we can insert it, it's not the same value
3417 already existing along every predecessor, and
3418 it's defined by some predecessor, it is
3419 partially redundant. */
3420 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3422 pre_stats.pa_insert++;
3423 if (insert_into_preds_of_block (block, get_expression_id (expr),
3431 VEC_free (pre_expr, heap, exprs);
3436 insert_aux (basic_block block)
3439 bool new_stuff = false;
3444 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3449 bitmap_set_t newset = NEW_SETS (dom);
3452 /* Note that we need to value_replace both NEW_SETS, and
3453 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3454 represented by some non-simple expression here that we want
3455 to replace it with. */
3456 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3458 pre_expr expr = expression_for_id (i);
3459 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3460 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3463 if (!single_pred_p (block))
3465 new_stuff |= do_regular_insertion (block, dom);
3466 if (do_partial_partial)
3467 new_stuff |= do_partial_partial_insertion (block, dom);
3471 for (son = first_dom_son (CDI_DOMINATORS, block);
3473 son = next_dom_son (CDI_DOMINATORS, son))
3475 new_stuff |= insert_aux (son);
3481 /* Perform insertion of partially redundant values. */
3486 bool new_stuff = true;
3488 int num_iterations = 0;
3491 NEW_SETS (bb) = bitmap_set_new ();
3496 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3498 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3502 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3505 add_to_exp_gen (basic_block block, tree op)
3510 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3512 result = get_or_alloc_expr_for_name (op);
3513 bitmap_value_insert_into_set (EXP_GEN (block), result);
3517 /* Create value ids for PHI in BLOCK. */
3520 make_values_for_phi (gimple phi, basic_block block)
3522 tree result = gimple_phi_result (phi);
3524 /* We have no need for virtual phis, as they don't represent
3525 actual computations. */
3526 if (is_gimple_reg (result))
3528 pre_expr e = get_or_alloc_expr_for_name (result);
3529 add_to_value (get_expr_value_id (e), e);
3530 bitmap_insert_into_set (PHI_GEN (block), e);
3531 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3535 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3537 tree arg = gimple_phi_arg_def (phi, i);
3538 if (TREE_CODE (arg) == SSA_NAME)
3540 e = get_or_alloc_expr_for_name (arg);
3541 add_to_value (get_expr_value_id (e), e);
3548 /* Compute the AVAIL set for all basic blocks.
3550 This function performs value numbering of the statements in each basic
3551 block. The AVAIL sets are built from information we glean while doing
3552 this value numbering, since the AVAIL sets contain only one entry per
3555 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3556 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3559 compute_avail (void)
3562 basic_block block, son;
3563 basic_block *worklist;
3567 /* For arguments with default definitions, we pretend they are
3568 defined in the entry block. */
3569 for (param = DECL_ARGUMENTS (current_function_decl);
3571 param = TREE_CHAIN (param))
3573 if (gimple_default_def (cfun, param) != NULL)
3575 tree def = gimple_default_def (cfun, param);
3576 pre_expr e = get_or_alloc_expr_for_name (def);
3578 add_to_value (get_expr_value_id (e), e);
3580 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3581 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3585 /* Likewise for the static chain decl. */
3586 if (cfun->static_chain_decl)
3588 param = cfun->static_chain_decl;
3589 if (gimple_default_def (cfun, param) != NULL)
3591 tree def = gimple_default_def (cfun, param);
3592 pre_expr e = get_or_alloc_expr_for_name (def);
3594 add_to_value (get_expr_value_id (e), e);
3596 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3597 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3601 /* Allocate the worklist. */
3602 worklist = XNEWVEC (basic_block, n_basic_blocks);
3604 /* Seed the algorithm by putting the dominator children of the entry
3605 block on the worklist. */
3606 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3608 son = next_dom_son (CDI_DOMINATORS, son))
3609 worklist[sp++] = son;
3611 /* Loop until the worklist is empty. */
3614 gimple_stmt_iterator gsi;
3617 unsigned int stmt_uid = 1;
3619 /* Pick a block from the worklist. */
3620 block = worklist[--sp];
3622 /* Initially, the set of available values in BLOCK is that of
3623 its immediate dominator. */
3624 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3626 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3628 /* Generate values for PHI nodes. */
3629 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3630 make_values_for_phi (gsi_stmt (gsi), block);
3632 /* Now compute value numbers and populate value sets with all
3633 the expressions computed in BLOCK. */
3634 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3639 stmt = gsi_stmt (gsi);
3640 gimple_set_uid (stmt, stmt_uid++);
3642 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3644 pre_expr e = get_or_alloc_expr_for_name (op);
3646 add_to_value (get_expr_value_id (e), e);
3648 bitmap_insert_into_set (TMP_GEN (block), e);
3649 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3652 if (gimple_has_volatile_ops (stmt)
3653 || stmt_could_throw_p (stmt))
3656 switch (gimple_code (stmt))
3659 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3660 add_to_exp_gen (block, op);
3667 vn_reference_op_t vro;
3668 pre_expr result = NULL;
3669 VEC(vn_reference_op_s, heap) *ops = NULL;
3671 if (!can_value_number_call (stmt))
3674 copy_reference_ops_from_call (stmt, &ops);
3675 vn_reference_lookup_pieces (shared_vuses_from_stmt (stmt),
3677 VEC_free (vn_reference_op_s, heap, ops);
3681 for (i = 0; VEC_iterate (vn_reference_op_s,
3685 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3686 add_to_exp_gen (block, vro->op0);
3687 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3688 add_to_exp_gen (block, vro->op1);
3689 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3690 add_to_exp_gen (block, vro->op2);
3692 result = (pre_expr) pool_alloc (pre_expr_pool);
3693 result->kind = REFERENCE;
3695 PRE_EXPR_REFERENCE (result) = ref;
3697 get_or_alloc_expression_id (result);
3698 add_to_value (get_expr_value_id (result), result);
3700 bitmap_value_insert_into_set (EXP_GEN (block), result);
3706 pre_expr result = NULL;
3707 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
3710 if (is_exception_related (stmt))
3713 case tcc_comparison:
3718 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
3719 gimple_assign_rhs_code (stmt),
3720 gimple_expr_type (stmt),
3721 gimple_assign_rhs1 (stmt),
3722 gimple_assign_rhs2 (stmt),
3723 NULL_TREE, NULL_TREE, &nary);
3728 for (i = 0; i < nary->length; i++)
3729 if (TREE_CODE (nary->op[i]) == SSA_NAME)
3730 add_to_exp_gen (block, nary->op[i]);
3732 result = (pre_expr) pool_alloc (pre_expr_pool);
3733 result->kind = NARY;
3735 PRE_EXPR_NARY (result) = nary;
3739 case tcc_declaration:
3744 vn_reference_op_t vro;
3746 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3747 shared_vuses_from_stmt (stmt),
3752 for (i = 0; VEC_iterate (vn_reference_op_s,
3756 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3757 add_to_exp_gen (block, vro->op0);
3758 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3759 add_to_exp_gen (block, vro->op1);
3760 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3761 add_to_exp_gen (block, vro->op2);
3763 result = (pre_expr) pool_alloc (pre_expr_pool);
3764 result->kind = REFERENCE;
3766 PRE_EXPR_REFERENCE (result) = ref;
3771 /* For any other statement that we don't
3772 recognize, simply add all referenced
3773 SSA_NAMEs to EXP_GEN. */
3774 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3775 add_to_exp_gen (block, op);
3779 get_or_alloc_expression_id (result);
3780 add_to_value (get_expr_value_id (result), result);
3782 bitmap_value_insert_into_set (EXP_GEN (block), result);
3791 /* Put the dominator children of BLOCK on the worklist of blocks
3792 to compute available sets for. */
3793 for (son = first_dom_son (CDI_DOMINATORS, block);
3795 son = next_dom_son (CDI_DOMINATORS, son))
3796 worklist[sp++] = son;
3802 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
3803 than the available expressions for it. The insertion point is
3804 right before the first use in STMT. Returns the SSA_NAME that should
3805 be used for replacement. */
3808 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
3810 basic_block bb = gimple_bb (stmt);
3811 gimple_stmt_iterator gsi;
3812 gimple_seq stmts = NULL;
3816 /* First create a value expression from the expression we want
3817 to insert and associate it with the value handle for SSA_VN. */
3818 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
3822 /* Then use create_expression_by_pieces to generate a valid
3823 expression to insert at this point of the IL stream. */
3824 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
3825 if (expr == NULL_TREE)
3827 gsi = gsi_for_stmt (stmt);
3828 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
3833 /* Eliminate fully redundant computations. */
3839 unsigned int todo = 0;
3843 gimple_stmt_iterator i;
3845 for (i = gsi_start_bb (b); !gsi_end_p (i); gsi_next (&i))
3847 gimple stmt = gsi_stmt (i);
3849 /* Lookup the RHS of the expression, see if we have an
3850 available computation for it. If so, replace the RHS with
3851 the available computation. */
3852 if (gimple_has_lhs (stmt)
3853 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
3854 && !gimple_assign_ssa_name_copy_p (stmt)
3855 && (!gimple_assign_single_p (stmt)
3856 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
3857 && !gimple_has_volatile_ops (stmt)
3858 && !has_zero_uses (gimple_get_lhs (stmt)))
3860 tree lhs = gimple_get_lhs (stmt);
3861 tree rhs = NULL_TREE;
3863 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
3864 pre_expr sprimeexpr;
3866 if (gimple_assign_single_p (stmt))
3867 rhs = gimple_assign_rhs1 (stmt);
3869 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
3870 get_expr_value_id (lhsexpr),
3875 if (sprimeexpr->kind == CONSTANT)
3876 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
3877 else if (sprimeexpr->kind == NAME)
3878 sprime = PRE_EXPR_NAME (sprimeexpr);
3883 /* If there is no existing leader but SCCVN knows this
3884 value is constant, use that constant. */
3885 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
3887 sprime = fold_convert (TREE_TYPE (lhs),
3888 VN_INFO (lhs)->valnum);
3890 if (dump_file && (dump_flags & TDF_DETAILS))
3892 fprintf (dump_file, "Replaced ");
3893 print_gimple_expr (dump_file, stmt, 0, 0);
3894 fprintf (dump_file, " with ");
3895 print_generic_expr (dump_file, sprime, 0);
3896 fprintf (dump_file, " in ");
3897 print_gimple_stmt (dump_file, stmt, 0, 0);
3899 pre_stats.eliminations++;
3900 propagate_tree_value_into_stmt (&i, sprime);
3901 stmt = gsi_stmt (i);
3906 /* If there is no existing usable leader but SCCVN thinks
3907 it has an expression it wants to use as replacement,
3909 if (!sprime || sprime == lhs)
3911 tree val = VN_INFO (lhs)->valnum;
3913 && TREE_CODE (val) == SSA_NAME
3914 && VN_INFO (val)->needs_insertion
3915 && can_PRE_operation (vn_get_expr_for (val)))
3916 sprime = do_SCCVN_insertion (stmt, val);
3920 && (rhs == NULL_TREE
3921 || TREE_CODE (rhs) != SSA_NAME
3922 || may_propagate_copy (rhs, sprime)))
3924 gcc_assert (sprime != rhs);
3926 if (dump_file && (dump_flags & TDF_DETAILS))
3928 fprintf (dump_file, "Replaced ");
3929 print_gimple_expr (dump_file, stmt, 0, 0);
3930 fprintf (dump_file, " with ");
3931 print_generic_expr (dump_file, sprime, 0);
3932 fprintf (dump_file, " in ");
3933 print_gimple_stmt (dump_file, stmt, 0, 0);
3936 if (TREE_CODE (sprime) == SSA_NAME)
3937 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
3939 /* We need to make sure the new and old types actually match,
3940 which may require adding a simple cast, which fold_convert
3942 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
3943 && !useless_type_conversion_p (gimple_expr_type (stmt),
3944 TREE_TYPE (sprime)))
3945 sprime = fold_convert (gimple_expr_type (stmt), sprime);
3947 pre_stats.eliminations++;
3948 propagate_tree_value_into_stmt (&i, sprime);
3949 stmt = gsi_stmt (i);
3952 /* If we removed EH side effects from the statement, clean
3953 its EH information. */
3954 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
3956 bitmap_set_bit (need_eh_cleanup,
3957 gimple_bb (stmt)->index);
3958 if (dump_file && (dump_flags & TDF_DETAILS))
3959 fprintf (dump_file, " Removed EH side effects.\n");
3963 /* Visit COND_EXPRs and fold the comparison with the
3964 available value-numbers. */
3965 else if (gimple_code (stmt) == GIMPLE_COND)
3967 tree op0 = gimple_cond_lhs (stmt);
3968 tree op1 = gimple_cond_rhs (stmt);
3971 if (TREE_CODE (op0) == SSA_NAME)
3972 op0 = VN_INFO (op0)->valnum;
3973 if (TREE_CODE (op1) == SSA_NAME)
3974 op1 = VN_INFO (op1)->valnum;
3975 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
3977 if (result && TREE_CODE (result) == INTEGER_CST)
3979 if (integer_zerop (result))
3980 gimple_cond_make_false (stmt);
3982 gimple_cond_make_true (stmt);
3984 todo = TODO_cleanup_cfg;
3993 /* Borrow a bit of tree-ssa-dce.c for the moment.
3994 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
3995 this may be a bit faster, and we may want critical edges kept split. */
3997 /* If OP's defining statement has not already been determined to be necessary,
3998 mark that statement necessary. Return the stmt, if it is newly
4001 static inline gimple
4002 mark_operand_necessary (tree op)
4008 if (TREE_CODE (op) != SSA_NAME)
4011 stmt = SSA_NAME_DEF_STMT (op);
4014 if (gimple_plf (stmt, NECESSARY)
4015 || gimple_nop_p (stmt))
4018 gimple_set_plf (stmt, NECESSARY, true);
4022 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4023 to insert PHI nodes sometimes, and because value numbering of casts isn't
4024 perfect, we sometimes end up inserting dead code. This simple DCE-like
4025 pass removes any insertions we made that weren't actually used. */
4028 remove_dead_inserted_code (void)
4030 VEC(gimple,heap) *worklist = NULL;
4034 worklist = VEC_alloc (gimple, heap, VEC_length (gimple, inserted_exprs));
4035 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4037 if (gimple_plf (t, NECESSARY))
4038 VEC_quick_push (gimple, worklist, t);
4040 while (VEC_length (gimple, worklist) > 0)
4042 t = VEC_pop (gimple, worklist);
4044 /* PHI nodes are somewhat special in that each PHI alternative has
4045 data and control dependencies. All the statements feeding the
4046 PHI node's arguments are always necessary. */
4047 if (gimple_code (t) == GIMPLE_PHI)
4051 VEC_reserve (gimple, heap, worklist, gimple_phi_num_args (t));
4052 for (k = 0; k < gimple_phi_num_args (t); k++)
4054 tree arg = PHI_ARG_DEF (t, k);
4055 if (TREE_CODE (arg) == SSA_NAME)
4057 gimple n = mark_operand_necessary (arg);
4059 VEC_quick_push (gimple, worklist, n);
4065 /* Propagate through the operands. Examine all the USE, VUSE and
4066 VDEF operands in this statement. Mark all the statements
4067 which feed this statement's uses as necessary. */
4071 /* The operands of VDEF expressions are also needed as they
4072 represent potential definitions that may reach this
4073 statement (VDEF operands allow us to follow def-def
4076 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4078 gimple n = mark_operand_necessary (use);
4080 VEC_safe_push (gimple, heap, worklist, n);
4085 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4087 if (!gimple_plf (t, NECESSARY))
4089 gimple_stmt_iterator gsi;
4091 if (dump_file && (dump_flags & TDF_DETAILS))
4093 fprintf (dump_file, "Removing unnecessary insertion:");
4094 print_gimple_stmt (dump_file, t, 0, 0);
4097 gsi = gsi_for_stmt (t);
4098 if (gimple_code (t) == GIMPLE_PHI)
4099 remove_phi_node (&gsi, true);
4101 gsi_remove (&gsi, true);
4105 VEC_free (gimple, heap, worklist);
4108 /* Initialize data structures used by PRE. */
4111 init_pre (bool do_fre)
4115 next_expression_id = 1;
4117 VEC_safe_push (pre_expr, heap, expressions, NULL);
4118 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4119 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4120 get_max_value_id() + 1);
4124 inserted_exprs = NULL;
4125 need_creation = NULL;
4126 pretemp = NULL_TREE;
4127 storetemp = NULL_TREE;
4128 prephitemp = NULL_TREE;
4130 connect_infinite_loops_to_exit ();
4131 memset (&pre_stats, 0, sizeof (pre_stats));
4134 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4135 post_order_compute (postorder, false, false);
4138 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4140 calculate_dominance_info (CDI_POST_DOMINATORS);
4141 calculate_dominance_info (CDI_DOMINATORS);
4143 bitmap_obstack_initialize (&grand_bitmap_obstack);
4144 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4145 expr_pred_trans_eq, free);
4146 expression_to_id = htab_create (num_ssa_names * 3,
4149 seen_during_translate = BITMAP_ALLOC (&grand_bitmap_obstack);
4150 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4151 sizeof (struct bitmap_set), 30);
4152 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4153 sizeof (struct pre_expr_d), 30);
4156 EXP_GEN (bb) = bitmap_set_new ();
4157 PHI_GEN (bb) = bitmap_set_new ();
4158 TMP_GEN (bb) = bitmap_set_new ();
4159 AVAIL_OUT (bb) = bitmap_set_new ();
4162 need_eh_cleanup = BITMAP_ALLOC (NULL);
4166 /* Deallocate data structures used by PRE. */
4169 fini_pre (bool do_fre)
4174 VEC_free (bitmap_set_t, heap, value_expressions);
4175 VEC_free (gimple, heap, inserted_exprs);
4176 VEC_free (gimple, heap, need_creation);
4177 bitmap_obstack_release (&grand_bitmap_obstack);
4178 free_alloc_pool (bitmap_set_pool);
4179 free_alloc_pool (pre_expr_pool);
4180 htab_delete (phi_translate_table);
4181 htab_delete (expression_to_id);
4189 free_dominance_info (CDI_POST_DOMINATORS);
4191 if (!bitmap_empty_p (need_eh_cleanup))
4193 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4194 cleanup_tree_cfg ();
4197 BITMAP_FREE (need_eh_cleanup);
4200 loop_optimizer_finalize ();
4203 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4204 only wants to do full redundancy elimination. */
4207 execute_pre (bool do_fre ATTRIBUTE_UNUSED)
4209 unsigned int todo = 0;
4211 do_partial_partial = optimize > 2;
4213 /* This has to happen before SCCVN runs because
4214 loop_optimizer_init may create new phis, etc. */
4216 loop_optimizer_init (LOOPS_NORMAL);
4218 if (!run_scc_vn (do_fre))
4222 remove_dead_inserted_code ();
4223 loop_optimizer_finalize ();
4231 /* Collect and value number expressions computed in each basic block. */
4234 if (dump_file && (dump_flags & TDF_DETAILS))
4240 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4241 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4242 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4243 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4247 /* Insert can get quite slow on an incredibly large number of basic
4248 blocks due to some quadratic behavior. Until this behavior is
4249 fixed, don't run it when he have an incredibly large number of
4250 bb's. If we aren't going to run insert, there is no point in
4251 computing ANTIC, either, even though it's plenty fast. */
4252 if (!do_fre && n_basic_blocks < 4000)
4258 /* Remove all the redundant expressions. */
4259 todo |= eliminate ();
4261 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4262 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4263 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4264 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4265 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4267 /* Make sure to remove fake edges before committing our inserts.
4268 This makes sure we don't end up with extra critical edges that
4269 we would need to split. */
4270 remove_fake_exit_edges ();
4271 gsi_commit_edge_inserts ();
4273 clear_expression_ids ();
4276 remove_dead_inserted_code ();
4283 /* Gate and execute functions for PRE. */
4288 return TODO_rebuild_alias | execute_pre (false);
4294 /* PRE tends to generate bigger code. */
4295 return flag_tree_pre != 0 && optimize_function_for_speed_p (cfun);
4298 struct gimple_opt_pass pass_pre =
4303 gate_pre, /* gate */
4304 do_pre, /* execute */
4307 0, /* static_pass_number */
4308 TV_TREE_PRE, /* tv_id */
4309 PROP_no_crit_edges | PROP_cfg
4310 | PROP_ssa | PROP_alias, /* properties_required */
4311 0, /* properties_provided */
4312 0, /* properties_destroyed */
4313 0, /* todo_flags_start */
4314 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4315 | TODO_verify_ssa /* todo_flags_finish */
4320 /* Gate and execute functions for FRE. */
4325 return execute_pre (true);
4331 return flag_tree_fre != 0;
4334 struct gimple_opt_pass pass_fre =
4339 gate_fre, /* gate */
4340 execute_fre, /* execute */
4343 0, /* static_pass_number */
4344 TV_TREE_FRE, /* tv_id */
4345 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
4346 0, /* properties_provided */
4347 0, /* properties_destroyed */
4348 0, /* todo_flags_start */
4349 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */