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 /* The phi_translate_table caches phi translations for a given
457 expression and predecessor. */
459 static htab_t phi_translate_table;
461 /* A three tuple {e, pred, v} used to cache phi translations in the
462 phi_translate_table. */
464 typedef struct expr_pred_trans_d
466 /* The expression. */
469 /* The predecessor block along which we translated the expression. */
472 /* The value that resulted from the translation. */
475 /* The hashcode for the expression, pred pair. This is cached for
478 } *expr_pred_trans_t;
479 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
481 /* Return the hash value for a phi translation table entry. */
484 expr_pred_trans_hash (const void *p)
486 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
490 /* Return true if two phi translation table entries are the same.
491 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
494 expr_pred_trans_eq (const void *p1, const void *p2)
496 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
497 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
498 basic_block b1 = ve1->pred;
499 basic_block b2 = ve2->pred;
501 /* If they are not translations for the same basic block, they can't
505 return pre_expr_eq (ve1->e, ve2->e);
508 /* Search in the phi translation table for the translation of
509 expression E in basic block PRED.
510 Return the translated value, if found, NULL otherwise. */
512 static inline pre_expr
513 phi_trans_lookup (pre_expr e, basic_block pred)
516 struct expr_pred_trans_d ept;
520 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
521 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
526 return ((expr_pred_trans_t) *slot)->v;
530 /* Add the tuple mapping from {expression E, basic block PRED} to
531 value V, to the phi translation table. */
534 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
537 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
539 new_pair->pred = pred;
541 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
544 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
545 new_pair->hashcode, INSERT);
548 *slot = (void *) new_pair;
552 /* Add expression E to the expression set of value id V. */
555 add_to_value (unsigned int v, pre_expr e)
559 gcc_assert (get_expr_value_id (e) == v);
561 if (v >= VEC_length (bitmap_set_t, value_expressions))
563 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
567 set = VEC_index (bitmap_set_t, value_expressions, v);
570 set = bitmap_set_new ();
571 VEC_replace (bitmap_set_t, value_expressions, v, set);
574 bitmap_insert_into_set_1 (set, e, true);
577 /* Create a new bitmap set and return it. */
580 bitmap_set_new (void)
582 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
583 ret->expressions = BITMAP_ALLOC (&grand_bitmap_obstack);
584 ret->values = BITMAP_ALLOC (&grand_bitmap_obstack);
588 /* Return the value id for a PRE expression EXPR. */
591 get_expr_value_id (pre_expr expr)
598 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
601 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
602 add_to_value (id, expr);
607 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
609 return PRE_EXPR_NARY (expr)->value_id;
611 return PRE_EXPR_REFERENCE (expr)->value_id;
617 /* Remove an expression EXPR from a bitmapped set. */
620 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
622 unsigned int val = get_expr_value_id (expr);
623 if (!value_id_constant_p (val))
625 bitmap_clear_bit (set->values, val);
626 bitmap_clear_bit (set->expressions, get_expression_id (expr));
631 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
632 bool allow_constants)
634 unsigned int val = get_expr_value_id (expr);
635 if (allow_constants || !value_id_constant_p (val))
637 /* We specifically expect this and only this function to be able to
638 insert constants into a set. */
639 bitmap_set_bit (set->values, val);
640 bitmap_set_bit (set->expressions, get_or_alloc_expression_id (expr));
644 /* Insert an expression EXPR into a bitmapped set. */
647 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
649 bitmap_insert_into_set_1 (set, expr, false);
652 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
655 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
657 bitmap_copy (dest->expressions, orig->expressions);
658 bitmap_copy (dest->values, orig->values);
662 /* Free memory used up by SET. */
664 bitmap_set_free (bitmap_set_t set)
666 BITMAP_FREE (set->expressions);
667 BITMAP_FREE (set->values);
671 /* Generate an topological-ordered array of bitmap set SET. */
673 static VEC(pre_expr, heap) *
674 sorted_array_from_bitmap_set (bitmap_set_t set)
677 bitmap_iterator bi, bj;
678 VEC(pre_expr, heap) *result = NULL;
680 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
682 /* The number of expressions having a given value is usually
683 relatively small. Thus, rather than making a vector of all
684 the expressions and sorting it by value-id, we walk the values
685 and check in the reverse mapping that tells us what expressions
686 have a given value, to filter those in our set. As a result,
687 the expressions are inserted in value-id order, which means
690 If this is somehow a significant lose for some cases, we can
691 choose which set to walk based on the set size. */
692 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
693 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
695 if (bitmap_bit_p (set->expressions, j))
696 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
703 /* Perform bitmapped set operation DEST &= ORIG. */
706 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
713 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
715 bitmap_and_into (dest->values, orig->values);
716 bitmap_copy (temp, dest->expressions);
717 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
719 pre_expr expr = expression_for_id (i);
720 unsigned int value_id = get_expr_value_id (expr);
721 if (!bitmap_bit_p (dest->values, value_id))
722 bitmap_clear_bit (dest->expressions, i);
728 /* Subtract all values and expressions contained in ORIG from DEST. */
731 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
733 bitmap_set_t result = bitmap_set_new ();
737 bitmap_and_compl (result->expressions, dest->expressions,
740 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
742 pre_expr expr = expression_for_id (i);
743 unsigned int value_id = get_expr_value_id (expr);
744 bitmap_set_bit (result->values, value_id);
750 /* Subtract all the values in bitmap set B from bitmap set A. */
753 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
757 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
759 bitmap_copy (temp, a->expressions);
760 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
762 pre_expr expr = expression_for_id (i);
763 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
764 bitmap_remove_from_set (a, expr);
770 /* Return true if bitmapped set SET contains the value VALUE_ID. */
773 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
775 if (value_id_constant_p (value_id))
778 if (!set || bitmap_empty_p (set->expressions))
781 return bitmap_bit_p (set->values, value_id);
785 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
787 return bitmap_bit_p (set->expressions, get_expression_id (expr));
790 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
793 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
796 bitmap_set_t exprset;
800 if (value_id_constant_p (lookfor))
803 if (!bitmap_set_contains_value (set, lookfor))
806 /* The number of expressions having a given value is usually
807 significantly less than the total number of expressions in SET.
808 Thus, rather than check, for each expression in SET, whether it
809 has the value LOOKFOR, we walk the reverse mapping that tells us
810 what expressions have a given value, and see if any of those
811 expressions are in our set. For large testcases, this is about
812 5-10x faster than walking the bitmap. If this is somehow a
813 significant lose for some cases, we can choose which set to walk
814 based on the set size. */
815 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
816 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
818 if (bitmap_bit_p (set->expressions, i))
820 bitmap_clear_bit (set->expressions, i);
821 bitmap_set_bit (set->expressions, get_expression_id (expr));
827 /* Return true if two bitmap sets are equal. */
830 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
832 return bitmap_equal_p (a->values, b->values);
835 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
836 and add it otherwise. */
839 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
841 unsigned int val = get_expr_value_id (expr);
843 if (bitmap_set_contains_value (set, val))
844 bitmap_set_replace_value (set, val, expr);
846 bitmap_insert_into_set (set, expr);
849 /* Insert EXPR into SET if EXPR's value is not already present in
853 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
855 unsigned int val = get_expr_value_id (expr);
857 if (value_id_constant_p (val))
860 if (!bitmap_set_contains_value (set, val))
861 bitmap_insert_into_set (set, expr);
864 /* Print out EXPR to outfile. */
867 print_pre_expr (FILE *outfile, const pre_expr expr)
872 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
875 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
880 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
881 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
882 for (i = 0; i < nary->length; i++)
884 print_generic_expr (outfile, nary->op[i], 0);
885 if (i != (unsigned) nary->length - 1)
886 fprintf (outfile, ",");
888 fprintf (outfile, "}");
894 vn_reference_op_t vro;
896 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
897 fprintf (outfile, "{");
899 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
902 if (vro->opcode != SSA_NAME
903 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
904 fprintf (outfile, "%s ", tree_code_name [vro->opcode]);
908 fprintf (outfile, "<");
909 print_generic_expr (outfile, vro->op0, 0);
912 fprintf (outfile, ",");
913 print_generic_expr (outfile, vro->op1, 0);
916 fprintf (outfile, ">");
918 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
919 fprintf (outfile, ",");
921 fprintf (outfile, "}");
926 void debug_pre_expr (pre_expr);
928 /* Like print_pre_expr but always prints to stderr. */
930 debug_pre_expr (pre_expr e)
932 print_pre_expr (stderr, e);
933 fprintf (stderr, "\n");
936 /* Print out SET to OUTFILE. */
939 print_bitmap_set (FILE *outfile, bitmap_set_t set,
940 const char *setname, int blockindex)
942 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
949 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
951 const pre_expr expr = expression_for_id (i);
954 fprintf (outfile, ", ");
956 print_pre_expr (outfile, expr);
958 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
961 fprintf (outfile, " }\n");
964 void debug_bitmap_set (bitmap_set_t);
967 debug_bitmap_set (bitmap_set_t set)
969 print_bitmap_set (stderr, set, "debug", 0);
972 /* Print out the expressions that have VAL to OUTFILE. */
975 print_value_expressions (FILE *outfile, unsigned int val)
977 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
981 sprintf (s, "%04d", val);
982 print_bitmap_set (outfile, set, s, 0);
988 debug_value_expressions (unsigned int val)
990 print_value_expressions (stderr, val);
993 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
997 get_or_alloc_expr_for_constant (tree constant)
999 unsigned int result_id;
1000 unsigned int value_id;
1001 pre_expr newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1002 newexpr->kind = CONSTANT;
1003 PRE_EXPR_CONSTANT (newexpr) = constant;
1004 result_id = lookup_expression_id (newexpr);
1007 pool_free (pre_expr_pool, newexpr);
1008 newexpr = expression_for_id (result_id);
1011 value_id = get_or_alloc_constant_value_id (constant);
1012 get_or_alloc_expression_id (newexpr);
1013 add_to_value (value_id, newexpr);
1017 /* Given a value id V, find the actual tree representing the constant
1018 value if there is one, and return it. Return NULL if we can't find
1022 get_constant_for_value_id (unsigned int v)
1024 if (value_id_constant_p (v))
1028 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1030 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1032 pre_expr expr = expression_for_id (i);
1033 if (expr->kind == CONSTANT)
1034 return PRE_EXPR_CONSTANT (expr);
1040 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1041 Currently only supports constants and SSA_NAMES. */
1043 get_or_alloc_expr_for (tree t)
1045 if (TREE_CODE (t) == SSA_NAME)
1046 return get_or_alloc_expr_for_name (t);
1047 else if (is_gimple_min_invariant (t))
1048 return get_or_alloc_expr_for_constant (t);
1051 /* More complex expressions can result from SCCVN expression
1052 simplification that inserts values for them. As they all
1053 do not have VOPs the get handled by the nary ops struct. */
1054 vn_nary_op_t result;
1055 unsigned int result_id;
1056 vn_nary_op_lookup (t, &result);
1059 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1061 PRE_EXPR_NARY (e) = result;
1062 result_id = lookup_expression_id (e);
1065 pool_free (pre_expr_pool, e);
1066 e = expression_for_id (result_id);
1069 alloc_expression_id (e);
1076 /* Return the folded version of T if T, when folded, is a gimple
1077 min_invariant. Otherwise, return T. */
1080 fully_constant_expression (pre_expr e)
1088 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1089 switch (TREE_CODE_CLASS (nary->opcode))
1091 case tcc_expression:
1092 if (nary->opcode == TRUTH_NOT_EXPR)
1094 if (nary->opcode != TRUTH_AND_EXPR
1095 && nary->opcode != TRUTH_OR_EXPR
1096 && nary->opcode != TRUTH_XOR_EXPR)
1100 case tcc_comparison:
1102 /* We have to go from trees to pre exprs to value ids to
1104 tree naryop0 = nary->op[0];
1105 tree naryop1 = nary->op[1];
1107 if (!is_gimple_min_invariant (naryop0))
1109 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1110 unsigned int vrep0 = get_expr_value_id (rep0);
1111 tree const0 = get_constant_for_value_id (vrep0);
1113 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1115 if (!is_gimple_min_invariant (naryop1))
1117 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1118 unsigned int vrep1 = get_expr_value_id (rep1);
1119 tree const1 = get_constant_for_value_id (vrep1);
1121 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1123 result = fold_binary (nary->opcode, nary->type,
1125 if (result && is_gimple_min_invariant (result))
1126 return get_or_alloc_expr_for_constant (result);
1127 /* We might have simplified the expression to a
1128 SSA_NAME for example from x_1 * 1. But we cannot
1129 insert a PHI for x_1 unconditionally as x_1 might
1130 not be available readily. */
1134 if (nary->opcode != REALPART_EXPR
1135 && nary->opcode != IMAGPART_EXPR
1136 && nary->opcode != VIEW_CONVERT_EXPR)
1142 /* We have to go from trees to pre exprs to value ids to
1144 tree naryop0 = nary->op[0];
1145 tree const0, result;
1146 if (is_gimple_min_invariant (naryop0))
1150 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1151 unsigned int vrep0 = get_expr_value_id (rep0);
1152 const0 = get_constant_for_value_id (vrep0);
1157 tree type1 = TREE_TYPE (nary->op[0]);
1158 const0 = fold_convert (type1, const0);
1159 result = fold_unary (nary->opcode, nary->type, const0);
1161 if (result && is_gimple_min_invariant (result))
1162 return get_or_alloc_expr_for_constant (result);
1171 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1172 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1173 vn_reference_op_t op;
1175 /* Try to simplify the translated expression if it is
1176 a call to a builtin function with at most two arguments. */
1177 op = VEC_index (vn_reference_op_s, operands, 0);
1178 if (op->opcode == CALL_EXPR
1179 && TREE_CODE (op->op0) == ADDR_EXPR
1180 && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL
1181 && DECL_BUILT_IN (TREE_OPERAND (op->op0, 0))
1182 && VEC_length (vn_reference_op_s, operands) >= 2
1183 && VEC_length (vn_reference_op_s, operands) <= 3)
1185 vn_reference_op_t arg0, arg1 = NULL;
1186 bool anyconst = false;
1187 arg0 = VEC_index (vn_reference_op_s, operands, 1);
1188 if (VEC_length (vn_reference_op_s, operands) > 2)
1189 arg1 = VEC_index (vn_reference_op_s, operands, 2);
1190 if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant
1191 || (arg0->opcode == ADDR_EXPR
1192 && is_gimple_min_invariant (arg0->op0)))
1195 && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant
1196 || (arg1->opcode == ADDR_EXPR
1197 && is_gimple_min_invariant (arg1->op0))))
1201 tree folded = build_call_expr (TREE_OPERAND (op->op0, 0),
1204 arg1 ? arg1->op0 : NULL);
1206 && TREE_CODE (folded) == NOP_EXPR)
1207 folded = TREE_OPERAND (folded, 0);
1209 && is_gimple_min_invariant (folded))
1210 return get_or_alloc_expr_for_constant (folded);
1221 /* Translate the vuses in the VUSES vector backwards through phi nodes
1222 in PHIBLOCK, so that they have the value they would have in
1225 static VEC(tree, gc) *
1226 translate_vuses_through_block (VEC (tree, gc) *vuses,
1227 basic_block phiblock,
1231 VEC(tree, gc) *result = NULL;
1234 for (i = 0; VEC_iterate (tree, vuses, i, oldvuse); i++)
1236 gimple phi = SSA_NAME_DEF_STMT (oldvuse);
1237 if (gimple_code (phi) == GIMPLE_PHI
1238 && gimple_bb (phi) == phiblock)
1240 edge e = find_edge (block, gimple_bb (phi));
1243 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1247 result = VEC_copy (tree, gc, vuses);
1248 VEC_replace (tree, result, i, def);
1254 /* We avoid creating a new copy of the vuses unless something
1255 actually changed, so result can be NULL. */
1258 sort_vuses (result);
1265 /* Like find_leader, but checks for the value existing in SET1 *or*
1266 SET2. This is used to avoid making a set consisting of the union
1267 of PA_IN and ANTIC_IN during insert. */
1269 static inline pre_expr
1270 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1274 result = bitmap_find_leader (set1, val, NULL);
1275 if (!result && set2)
1276 result = bitmap_find_leader (set2, val, NULL);
1280 /* Get the tree type for our PRE expression e. */
1283 get_expr_type (const pre_expr e)
1288 return TREE_TYPE (PRE_EXPR_NAME (e));
1290 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1293 vn_reference_op_t vro;
1295 gcc_assert (PRE_EXPR_REFERENCE (e)->operands);
1296 vro = VEC_index (vn_reference_op_s,
1297 PRE_EXPR_REFERENCE (e)->operands,
1299 /* We don't store type along with COMPONENT_REF because it is
1300 always the same as FIELD_DECL's type. */
1303 gcc_assert (vro->opcode == COMPONENT_REF);
1304 return TREE_TYPE (vro->op0);
1310 return PRE_EXPR_NARY (e)->type;
1315 /* Get a representative SSA_NAME for a given expression.
1316 Since all of our sub-expressions are treated as values, we require
1317 them to be SSA_NAME's for simplicity.
1318 Prior versions of GVNPRE used to use "value handles" here, so that
1319 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1320 either case, the operands are really values (IE we do not expect
1321 them to be usable without finding leaders). */
1324 get_representative_for (const pre_expr e)
1328 unsigned int value_id = get_expr_value_id (e);
1333 return PRE_EXPR_NAME (e);
1335 return PRE_EXPR_CONSTANT (e);
1339 /* Go through all of the expressions representing this value
1340 and pick out an SSA_NAME. */
1343 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1345 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1347 pre_expr rep = expression_for_id (i);
1348 if (rep->kind == NAME)
1349 return PRE_EXPR_NAME (rep);
1354 /* If we reached here we couldn't find an SSA_NAME. This can
1355 happen when we've discovered a value that has never appeared in
1356 the program as set to an SSA_NAME, most likely as the result of
1361 "Could not find SSA_NAME representative for expression:");
1362 print_pre_expr (dump_file, e);
1363 fprintf (dump_file, "\n");
1366 exprtype = get_expr_type (e);
1368 /* Build and insert the assignment of the end result to the temporary
1369 that we will return. */
1370 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1372 pretemp = create_tmp_var (exprtype, "pretmp");
1373 get_var_ann (pretemp);
1376 name = make_ssa_name (pretemp, gimple_build_nop ());
1377 VN_INFO_GET (name)->value_id = value_id;
1378 if (e->kind == CONSTANT)
1379 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1381 VN_INFO (name)->valnum = name;
1383 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1386 fprintf (dump_file, "Created SSA_NAME representative ");
1387 print_generic_expr (dump_file, name, 0);
1388 fprintf (dump_file, " for expression:");
1389 print_pre_expr (dump_file, e);
1390 fprintf (dump_file, "\n");
1399 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1400 basic_block pred, basic_block phiblock);
1402 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1403 the phis in PRED. Return NULL if we can't find a leader for each part
1404 of the translated expression. */
1407 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1408 basic_block pred, basic_block phiblock)
1415 bool changed = false;
1416 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1417 struct vn_nary_op_s newnary;
1418 /* The NARY structure is only guaranteed to have been
1419 allocated to the nary->length operands. */
1420 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1421 - sizeof (tree) * (4 - nary->length)));
1423 for (i = 0; i < newnary.length; i++)
1425 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1429 pre_expr leader, result;
1430 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1431 leader = find_leader_in_sets (op_val_id, set1, set2);
1432 result = phi_translate (leader, set1, set2, pred, phiblock);
1433 if (result && result != leader)
1435 tree name = get_representative_for (result);
1438 newnary.op[i] = name;
1443 changed |= newnary.op[i] != nary->op[i];
1450 tree result = vn_nary_op_lookup_pieces (newnary.length,
1458 unsigned int new_val_id;
1460 expr = (pre_expr) pool_alloc (pre_expr_pool);
1463 if (result && is_gimple_min_invariant (result))
1464 return get_or_alloc_expr_for_constant (result);
1469 PRE_EXPR_NARY (expr) = nary;
1470 constant = fully_constant_expression (expr);
1471 if (constant != expr)
1474 new_val_id = nary->value_id;
1475 get_or_alloc_expression_id (expr);
1479 new_val_id = get_next_value_id ();
1480 VEC_safe_grow_cleared (bitmap_set_t, heap,
1482 get_max_value_id() + 1);
1483 nary = vn_nary_op_insert_pieces (newnary.length,
1490 result, new_val_id);
1491 PRE_EXPR_NARY (expr) = nary;
1492 constant = fully_constant_expression (expr);
1493 if (constant != expr)
1495 get_or_alloc_expression_id (expr);
1497 add_to_value (new_val_id, expr);
1505 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1506 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1507 VEC (tree, gc) *vuses = ref->vuses;
1508 VEC (tree, gc) *newvuses = vuses;
1509 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1510 bool changed = false;
1512 vn_reference_op_t operand;
1513 vn_reference_t newref;
1515 for (i = 0; VEC_iterate (vn_reference_op_s, operands, i, operand); i++)
1519 tree oldop0 = operand->op0;
1520 tree oldop1 = operand->op1;
1521 tree oldop2 = operand->op2;
1525 tree type = operand->type;
1526 vn_reference_op_s newop = *operand;
1528 if (op0 && TREE_CODE (op0) == SSA_NAME)
1530 unsigned int op_val_id = VN_INFO (op0)->value_id;
1531 leader = find_leader_in_sets (op_val_id, set1, set2);
1532 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1533 if (opresult && opresult != leader)
1535 tree name = get_representative_for (opresult);
1543 changed |= op0 != oldop0;
1545 if (op1 && TREE_CODE (op1) == SSA_NAME)
1547 unsigned int op_val_id = VN_INFO (op1)->value_id;
1548 leader = find_leader_in_sets (op_val_id, set1, set2);
1549 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1550 if (opresult && opresult != leader)
1552 tree name = get_representative_for (opresult);
1560 changed |= op1 != oldop1;
1561 if (op2 && TREE_CODE (op2) == SSA_NAME)
1563 unsigned int op_val_id = VN_INFO (op2)->value_id;
1564 leader = find_leader_in_sets (op_val_id, set1, set2);
1565 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1566 if (opresult && opresult != leader)
1568 tree name = get_representative_for (opresult);
1576 changed |= op2 != oldop2;
1579 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1580 /* We may have changed from an SSA_NAME to a constant */
1581 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1582 newop.opcode = TREE_CODE (op0);
1587 VEC_replace (vn_reference_op_s, newoperands, i, &newop);
1589 if (i != VEC_length (vn_reference_op_s, operands))
1592 VEC_free (vn_reference_op_s, heap, newoperands);
1596 newvuses = translate_vuses_through_block (vuses, phiblock, pred);
1597 changed |= newvuses != vuses;
1601 unsigned int new_val_id;
1604 tree result = vn_reference_lookup_pieces (newvuses,
1608 VEC_free (vn_reference_op_s, heap, newoperands);
1610 if (result && is_gimple_min_invariant (result))
1612 gcc_assert (!newoperands);
1613 return get_or_alloc_expr_for_constant (result);
1616 expr = (pre_expr) pool_alloc (pre_expr_pool);
1617 expr->kind = REFERENCE;
1622 PRE_EXPR_REFERENCE (expr) = newref;
1623 constant = fully_constant_expression (expr);
1624 if (constant != expr)
1627 new_val_id = newref->value_id;
1628 get_or_alloc_expression_id (expr);
1632 new_val_id = get_next_value_id ();
1633 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
1634 get_max_value_id() + 1);
1635 newref = vn_reference_insert_pieces (newvuses,
1637 result, new_val_id);
1639 PRE_EXPR_REFERENCE (expr) = newref;
1640 constant = fully_constant_expression (expr);
1641 if (constant != expr)
1643 get_or_alloc_expression_id (expr);
1645 add_to_value (new_val_id, expr);
1647 VEC_free (vn_reference_op_s, heap, newoperands);
1657 tree name = PRE_EXPR_NAME (expr);
1659 def_stmt = SSA_NAME_DEF_STMT (name);
1660 if (gimple_code (def_stmt) == GIMPLE_PHI
1661 && gimple_bb (def_stmt) == phiblock)
1666 e = find_edge (pred, gimple_bb (phi));
1669 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1672 if (TREE_CODE (def) == SSA_NAME)
1673 def = VN_INFO (def)->valnum;
1675 /* Handle constant. */
1676 if (is_gimple_min_invariant (def))
1677 return get_or_alloc_expr_for_constant (def);
1679 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1682 newexpr = get_or_alloc_expr_for_name (def);
1693 /* Wrapper around phi_translate_1 providing caching functionality. */
1696 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1697 basic_block pred, basic_block phiblock)
1704 /* Constants contain no values that need translation. */
1705 if (expr->kind == CONSTANT)
1708 if (value_id_constant_p (get_expr_value_id (expr)))
1711 if (expr->kind != NAME)
1713 phitrans = phi_trans_lookup (expr, pred);
1719 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1721 /* Don't add empty translations to the cache. Neither add
1722 translations of NAMEs as those are cheap to translate. */
1724 && expr->kind != NAME)
1725 phi_trans_add (expr, phitrans, pred);
1731 /* For each expression in SET, translate the values through phi nodes
1732 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1733 expressions in DEST. */
1736 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1737 basic_block phiblock)
1739 VEC (pre_expr, heap) *exprs;
1743 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1745 bitmap_set_copy (dest, set);
1749 exprs = sorted_array_from_bitmap_set (set);
1750 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1752 pre_expr translated;
1753 translated = phi_translate (expr, set, NULL, pred, phiblock);
1757 /* We might end up with multiple expressions from SET being
1758 translated to the same value. In this case we do not want
1759 to retain the NARY or REFERENCE expression but prefer a NAME
1760 which would be the leader. */
1761 if (translated->kind == NAME)
1762 bitmap_value_replace_in_set (dest, translated);
1764 bitmap_value_insert_into_set (dest, translated);
1766 VEC_free (pre_expr, heap, exprs);
1769 /* Find the leader for a value (i.e., the name representing that
1770 value) in a given set, and return it. If STMT is non-NULL it
1771 makes sure the defining statement for the leader dominates it.
1772 Return NULL if no leader is found. */
1775 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1777 if (value_id_constant_p (val))
1781 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1783 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1785 pre_expr expr = expression_for_id (i);
1786 if (expr->kind == CONSTANT)
1790 if (bitmap_set_contains_value (set, val))
1792 /* Rather than walk the entire bitmap of expressions, and see
1793 whether any of them has the value we are looking for, we look
1794 at the reverse mapping, which tells us the set of expressions
1795 that have a given value (IE value->expressions with that
1796 value) and see if any of those expressions are in our set.
1797 The number of expressions per value is usually significantly
1798 less than the number of expressions in the set. In fact, for
1799 large testcases, doing it this way is roughly 5-10x faster
1800 than walking the bitmap.
1801 If this is somehow a significant lose for some cases, we can
1802 choose which set to walk based on which set is smaller. */
1805 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1807 EXECUTE_IF_AND_IN_BITMAP (exprset->expressions,
1808 set->expressions, 0, i, bi)
1810 pre_expr val = expression_for_id (i);
1811 /* At the point where stmt is not null, there should always
1812 be an SSA_NAME first in the list of expressions. */
1815 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1816 if (gimple_code (def_stmt) != GIMPLE_PHI
1817 && gimple_bb (def_stmt) == gimple_bb (stmt)
1818 && gimple_uid (def_stmt) >= gimple_uid (stmt))
1827 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1828 BLOCK by seeing if it is not killed in the block. Note that we are
1829 only determining whether there is a store that kills it. Because
1830 of the order in which clean iterates over values, we are guaranteed
1831 that altered operands will have caused us to be eliminated from the
1832 ANTIC_IN set already. */
1835 value_dies_in_block_x (pre_expr expr, basic_block block)
1839 VEC (tree, gc) *vuses = PRE_EXPR_REFERENCE (expr)->vuses;
1841 /* Conservatively, a value dies if it's vuses are defined in this
1842 block, unless they come from phi nodes (which are merge operations,
1843 rather than stores. */
1844 for (i = 0; VEC_iterate (tree, vuses, i, vuse); i++)
1846 gimple def = SSA_NAME_DEF_STMT (vuse);
1848 if (gimple_bb (def) != block)
1850 if (gimple_code (def) == GIMPLE_PHI)
1858 #define union_contains_value(SET1, SET2, VAL) \
1859 (bitmap_set_contains_value ((SET1), (VAL)) \
1860 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
1862 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
1865 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
1866 vn_reference_op_t vro)
1868 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
1870 struct pre_expr_d temp;
1873 PRE_EXPR_NAME (&temp) = vro->op0;
1874 temp.id = lookup_expression_id (&temp);
1877 if (!union_contains_value (set1, set2,
1878 get_expr_value_id (&temp)))
1881 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
1883 struct pre_expr_d temp;
1886 PRE_EXPR_NAME (&temp) = vro->op1;
1887 temp.id = lookup_expression_id (&temp);
1890 if (!union_contains_value (set1, set2,
1891 get_expr_value_id (&temp)))
1895 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
1897 struct pre_expr_d temp;
1900 PRE_EXPR_NAME (&temp) = vro->op2;
1901 temp.id = lookup_expression_id (&temp);
1904 if (!union_contains_value (set1, set2,
1905 get_expr_value_id (&temp)))
1912 /* Determine if the expression EXPR is valid in SET1 U SET2.
1913 ONLY SET2 CAN BE NULL.
1914 This means that we have a leader for each part of the expression
1915 (if it consists of values), or the expression is an SSA_NAME.
1916 For loads/calls, we also see if the vuses are killed in this block.
1920 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
1926 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
1930 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1931 for (i = 0; i < nary->length; i++)
1933 if (TREE_CODE (nary->op[i]) == SSA_NAME)
1935 struct pre_expr_d temp;
1938 PRE_EXPR_NAME (&temp) = nary->op[i];
1939 temp.id = lookup_expression_id (&temp);
1942 if (!union_contains_value (set1, set2,
1943 get_expr_value_id (&temp)))
1952 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1953 vn_reference_op_t vro;
1956 for (i = 0; VEC_iterate (vn_reference_op_s, ref->operands, i, vro); i++)
1958 if (!vro_valid_in_sets (set1, set2, vro))
1961 return !value_dies_in_block_x (expr, block);
1968 /* Clean the set of expressions that are no longer valid in SET1 or
1969 SET2. This means expressions that are made up of values we have no
1970 leaders for in SET1 or SET2. This version is used for partial
1971 anticipation, which means it is not valid in either ANTIC_IN or
1975 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
1977 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
1981 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1983 if (!valid_in_sets (set1, set2, expr, block))
1984 bitmap_remove_from_set (set1, expr);
1986 VEC_free (pre_expr, heap, exprs);
1989 /* Clean the set of expressions that are no longer valid in SET. This
1990 means expressions that are made up of values we have no leaders for
1994 clean (bitmap_set_t set, basic_block block)
1996 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2000 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2002 if (!valid_in_sets (set, NULL, expr, block))
2003 bitmap_remove_from_set (set, expr);
2005 VEC_free (pre_expr, heap, exprs);
2008 static sbitmap has_abnormal_preds;
2010 /* List of blocks that may have changed during ANTIC computation and
2011 thus need to be iterated over. */
2013 static sbitmap changed_blocks;
2015 /* Decide whether to defer a block for a later iteration, or PHI
2016 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2017 should defer the block, and true if we processed it. */
2020 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2021 basic_block block, basic_block phiblock)
2023 if (!BB_VISITED (phiblock))
2025 SET_BIT (changed_blocks, block->index);
2026 BB_VISITED (block) = 0;
2027 BB_DEFERRED (block) = 1;
2031 phi_translate_set (dest, source, block, phiblock);
2035 /* Compute the ANTIC set for BLOCK.
2037 If succs(BLOCK) > 1 then
2038 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2039 else if succs(BLOCK) == 1 then
2040 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2042 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2046 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2048 bool changed = false;
2049 bitmap_set_t S, old, ANTIC_OUT;
2055 old = ANTIC_OUT = S = NULL;
2056 BB_VISITED (block) = 1;
2058 /* If any edges from predecessors are abnormal, antic_in is empty,
2060 if (block_has_abnormal_pred_edge)
2061 goto maybe_dump_sets;
2063 old = ANTIC_IN (block);
2064 ANTIC_OUT = bitmap_set_new ();
2066 /* If the block has no successors, ANTIC_OUT is empty. */
2067 if (EDGE_COUNT (block->succs) == 0)
2069 /* If we have one successor, we could have some phi nodes to
2070 translate through. */
2071 else if (single_succ_p (block))
2073 basic_block succ_bb = single_succ (block);
2075 /* We trade iterations of the dataflow equations for having to
2076 phi translate the maximal set, which is incredibly slow
2077 (since the maximal set often has 300+ members, even when you
2078 have a small number of blocks).
2079 Basically, we defer the computation of ANTIC for this block
2080 until we have processed it's successor, which will inevitably
2081 have a *much* smaller set of values to phi translate once
2082 clean has been run on it.
2083 The cost of doing this is that we technically perform more
2084 iterations, however, they are lower cost iterations.
2086 Timings for PRE on tramp3d-v4:
2087 without maximal set fix: 11 seconds
2088 with maximal set fix/without deferring: 26 seconds
2089 with maximal set fix/with deferring: 11 seconds
2092 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2096 goto maybe_dump_sets;
2099 /* If we have multiple successors, we take the intersection of all of
2100 them. Note that in the case of loop exit phi nodes, we may have
2101 phis to translate through. */
2104 VEC(basic_block, heap) * worklist;
2106 basic_block bprime, first = NULL;
2108 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2109 FOR_EACH_EDGE (e, ei, block->succs)
2112 && BB_VISITED (e->dest))
2114 else if (BB_VISITED (e->dest))
2115 VEC_quick_push (basic_block, worklist, e->dest);
2118 /* Of multiple successors we have to have visited one already. */
2121 SET_BIT (changed_blocks, block->index);
2122 BB_VISITED (block) = 0;
2123 BB_DEFERRED (block) = 1;
2125 VEC_free (basic_block, heap, worklist);
2126 goto maybe_dump_sets;
2129 if (!gimple_seq_empty_p (phi_nodes (first)))
2130 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2132 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2134 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2136 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2138 bitmap_set_t tmp = bitmap_set_new ();
2139 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2140 bitmap_set_and (ANTIC_OUT, tmp);
2141 bitmap_set_free (tmp);
2144 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2146 VEC_free (basic_block, heap, worklist);
2149 /* Generate ANTIC_OUT - TMP_GEN. */
2150 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2152 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2153 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2156 /* Then union in the ANTIC_OUT - TMP_GEN values,
2157 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2158 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2159 bitmap_value_insert_into_set (ANTIC_IN (block),
2160 expression_for_id (bii));
2162 clean (ANTIC_IN (block), block);
2164 /* !old->expressions can happen when we deferred a block. */
2165 if (!old->expressions || !bitmap_set_equal (old, ANTIC_IN (block)))
2168 SET_BIT (changed_blocks, block->index);
2169 FOR_EACH_EDGE (e, ei, block->preds)
2170 SET_BIT (changed_blocks, e->src->index);
2173 RESET_BIT (changed_blocks, block->index);
2176 if (dump_file && (dump_flags & TDF_DETAILS))
2178 if (!BB_DEFERRED (block) || BB_VISITED (block))
2181 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2183 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2187 print_bitmap_set (dump_file, S, "S", block->index);
2192 "Block %d was deferred for a future iteration.\n",
2197 bitmap_set_free (old);
2199 bitmap_set_free (S);
2201 bitmap_set_free (ANTIC_OUT);
2205 /* Compute PARTIAL_ANTIC for BLOCK.
2207 If succs(BLOCK) > 1 then
2208 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2209 in ANTIC_OUT for all succ(BLOCK)
2210 else if succs(BLOCK) == 1 then
2211 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2213 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2218 compute_partial_antic_aux (basic_block block,
2219 bool block_has_abnormal_pred_edge)
2221 bool changed = false;
2222 bitmap_set_t old_PA_IN;
2223 bitmap_set_t PA_OUT;
2226 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2228 old_PA_IN = PA_OUT = NULL;
2230 /* If any edges from predecessors are abnormal, antic_in is empty,
2232 if (block_has_abnormal_pred_edge)
2233 goto maybe_dump_sets;
2235 /* If there are too many partially anticipatable values in the
2236 block, phi_translate_set can take an exponential time: stop
2237 before the translation starts. */
2239 && single_succ_p (block)
2240 && bitmap_count_bits (PA_IN (single_succ (block))->values) > max_pa)
2241 goto maybe_dump_sets;
2243 old_PA_IN = PA_IN (block);
2244 PA_OUT = bitmap_set_new ();
2246 /* If the block has no successors, ANTIC_OUT is empty. */
2247 if (EDGE_COUNT (block->succs) == 0)
2249 /* If we have one successor, we could have some phi nodes to
2250 translate through. Note that we can't phi translate across DFS
2251 back edges in partial antic, because it uses a union operation on
2252 the successors. For recurrences like IV's, we will end up
2253 generating a new value in the set on each go around (i + 3 (VH.1)
2254 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2255 else if (single_succ_p (block))
2257 basic_block succ = single_succ (block);
2258 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2259 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2261 /* If we have multiple successors, we take the union of all of
2265 VEC(basic_block, heap) * worklist;
2269 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2270 FOR_EACH_EDGE (e, ei, block->succs)
2272 if (e->flags & EDGE_DFS_BACK)
2274 VEC_quick_push (basic_block, worklist, e->dest);
2276 if (VEC_length (basic_block, worklist) > 0)
2278 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2283 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2284 bitmap_value_insert_into_set (PA_OUT,
2285 expression_for_id (i));
2286 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2288 bitmap_set_t pa_in = bitmap_set_new ();
2289 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2290 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2291 bitmap_value_insert_into_set (PA_OUT,
2292 expression_for_id (i));
2293 bitmap_set_free (pa_in);
2296 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2297 bitmap_value_insert_into_set (PA_OUT,
2298 expression_for_id (i));
2301 VEC_free (basic_block, heap, worklist);
2304 /* PA_IN starts with PA_OUT - TMP_GEN.
2305 Then we subtract things from ANTIC_IN. */
2306 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2308 /* For partial antic, we want to put back in the phi results, since
2309 we will properly avoid making them partially antic over backedges. */
2310 bitmap_ior_into (PA_IN (block)->values, PHI_GEN (block)->values);
2311 bitmap_ior_into (PA_IN (block)->expressions, PHI_GEN (block)->expressions);
2313 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2314 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2316 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2318 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2321 SET_BIT (changed_blocks, block->index);
2322 FOR_EACH_EDGE (e, ei, block->preds)
2323 SET_BIT (changed_blocks, e->src->index);
2326 RESET_BIT (changed_blocks, block->index);
2329 if (dump_file && (dump_flags & TDF_DETAILS))
2332 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2334 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2337 bitmap_set_free (old_PA_IN);
2339 bitmap_set_free (PA_OUT);
2343 /* Compute ANTIC and partial ANTIC sets. */
2346 compute_antic (void)
2348 bool changed = true;
2349 int num_iterations = 0;
2353 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2354 We pre-build the map of blocks with incoming abnormal edges here. */
2355 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2356 sbitmap_zero (has_abnormal_preds);
2363 FOR_EACH_EDGE (e, ei, block->preds)
2365 e->flags &= ~EDGE_DFS_BACK;
2366 if (e->flags & EDGE_ABNORMAL)
2368 SET_BIT (has_abnormal_preds, block->index);
2373 BB_VISITED (block) = 0;
2374 BB_DEFERRED (block) = 0;
2375 /* While we are here, give empty ANTIC_IN sets to each block. */
2376 ANTIC_IN (block) = bitmap_set_new ();
2377 PA_IN (block) = bitmap_set_new ();
2380 /* At the exit block we anticipate nothing. */
2381 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2382 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2383 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2385 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2386 sbitmap_ones (changed_blocks);
2389 if (dump_file && (dump_flags & TDF_DETAILS))
2390 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2393 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2395 if (TEST_BIT (changed_blocks, postorder[i]))
2397 basic_block block = BASIC_BLOCK (postorder[i]);
2398 changed |= compute_antic_aux (block,
2399 TEST_BIT (has_abnormal_preds,
2403 #ifdef ENABLE_CHECKING
2404 /* Theoretically possible, but *highly* unlikely. */
2405 gcc_assert (num_iterations < 500);
2409 statistics_histogram_event (cfun, "compute_antic iterations",
2412 if (do_partial_partial)
2414 sbitmap_ones (changed_blocks);
2415 mark_dfs_back_edges ();
2420 if (dump_file && (dump_flags & TDF_DETAILS))
2421 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2424 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2426 if (TEST_BIT (changed_blocks, postorder[i]))
2428 basic_block block = BASIC_BLOCK (postorder[i]);
2430 |= compute_partial_antic_aux (block,
2431 TEST_BIT (has_abnormal_preds,
2435 #ifdef ENABLE_CHECKING
2436 /* Theoretically possible, but *highly* unlikely. */
2437 gcc_assert (num_iterations < 500);
2440 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2443 sbitmap_free (has_abnormal_preds);
2444 sbitmap_free (changed_blocks);
2447 /* Return true if we can value number the call in STMT. This is true
2448 if we have a pure or constant call. */
2451 can_value_number_call (gimple stmt)
2453 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2458 /* Return true if OP is an exception handler related operation, such as
2459 FILTER_EXPR or EXC_PTR_EXPR. */
2462 is_exception_related (gimple stmt)
2464 return (is_gimple_assign (stmt)
2465 && (gimple_assign_rhs_code (stmt) == FILTER_EXPR
2466 || gimple_assign_rhs_code (stmt) == EXC_PTR_EXPR));
2469 /* Return true if OP is a tree which we can perform PRE on
2470 on. This may not match the operations we can value number, but in
2471 a perfect world would. */
2474 can_PRE_operation (tree op)
2476 return UNARY_CLASS_P (op)
2477 || BINARY_CLASS_P (op)
2478 || COMPARISON_CLASS_P (op)
2479 || TREE_CODE (op) == INDIRECT_REF
2480 || TREE_CODE (op) == COMPONENT_REF
2481 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2482 || TREE_CODE (op) == CALL_EXPR
2483 || TREE_CODE (op) == ARRAY_REF;
2487 /* Inserted expressions are placed onto this worklist, which is used
2488 for performing quick dead code elimination of insertions we made
2489 that didn't turn out to be necessary. */
2490 static VEC(gimple,heap) *inserted_exprs;
2492 /* Pool allocated fake store expressions are placed onto this
2493 worklist, which, after performing dead code elimination, is walked
2494 to see which expressions need to be put into GC'able memory */
2495 static VEC(gimple, heap) *need_creation;
2497 /* The actual worker for create_component_ref_by_pieces. */
2500 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2501 unsigned int *operand, gimple_seq *stmts,
2504 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2508 switch (currop->opcode)
2512 tree folded, sc = NULL_TREE;
2513 unsigned int nargs = 0;
2515 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2519 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2520 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2526 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2527 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2531 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2532 ref->operands) - 1);
2533 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2535 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2545 folded = build_call_array (currop->type,
2546 (TREE_CODE (fn) == FUNCTION_DECL
2547 ? build_fold_addr_expr (fn) : fn),
2551 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2558 gcc_assert (is_gimple_min_invariant (currop->op0));
2564 case VIEW_CONVERT_EXPR:
2567 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2572 folded = fold_build1 (currop->opcode, currop->type,
2577 case ALIGN_INDIRECT_REF:
2578 case MISALIGNED_INDIRECT_REF:
2582 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2587 genop1 = fold_convert (build_pointer_type (currop->type),
2590 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2591 folded = fold_build2 (currop->opcode, currop->type,
2592 genop1, currop->op1);
2594 folded = fold_build1 (currop->opcode, currop->type,
2602 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2604 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2605 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2611 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2614 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2617 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2622 /* For array ref vn_reference_op's, operand 1 of the array ref
2623 is op0 of the reference op and operand 3 of the array ref is
2625 case ARRAY_RANGE_REF:
2629 tree genop1 = currop->op0;
2631 tree genop2 = currop->op1;
2634 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2638 op1expr = get_or_alloc_expr_for (genop1);
2639 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2644 op2expr = get_or_alloc_expr_for (genop2);
2645 genop2 = find_or_generate_expression (block, op2expr, stmts,
2651 genop3 = currop->op2;
2652 return build4 (currop->opcode, currop->type, genop0, genop1,
2659 tree genop2 = currop->op1;
2661 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2665 /* op1 should be a FIELD_DECL, which are represented by
2670 op2expr = get_or_alloc_expr_for (genop2);
2671 genop2 = find_or_generate_expression (block, op2expr, stmts,
2677 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2683 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2684 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2705 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2706 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2707 trying to rename aggregates into ssa form directly, which is a no no.
2709 Thus, this routine doesn't create temporaries, it just builds a
2710 single access expression for the array, calling
2711 find_or_generate_expression to build the innermost pieces.
2713 This function is a subroutine of create_expression_by_pieces, and
2714 should not be called on it's own unless you really know what you
2718 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2719 gimple_seq *stmts, gimple domstmt)
2721 unsigned int op = 0;
2722 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2725 /* Find a leader for an expression, or generate one using
2726 create_expression_by_pieces if it's ANTIC but
2728 BLOCK is the basic_block we are looking for leaders in.
2729 EXPR is the expression to find a leader or generate for.
2730 STMTS is the statement list to put the inserted expressions on.
2731 Returns the SSA_NAME of the LHS of the generated expression or the
2733 DOMSTMT if non-NULL is a statement that should be dominated by
2734 all uses in the generated expression. If DOMSTMT is non-NULL this
2735 routine can fail and return NULL_TREE. Otherwise it will assert
2739 find_or_generate_expression (basic_block block, pre_expr expr,
2740 gimple_seq *stmts, gimple domstmt)
2742 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2743 get_expr_value_id (expr), domstmt);
2747 if (leader->kind == NAME)
2748 genop = PRE_EXPR_NAME (leader);
2749 else if (leader->kind == CONSTANT)
2750 genop = PRE_EXPR_CONSTANT (leader);
2753 /* If it's still NULL, it must be a complex expression, so generate
2754 it recursively. Not so for FRE though. */
2758 bitmap_set_t exprset;
2759 unsigned int lookfor = get_expr_value_id (expr);
2760 bool handled = false;
2764 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2765 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2767 pre_expr temp = expression_for_id (i);
2768 if (temp->kind != NAME)
2771 genop = create_expression_by_pieces (block, temp, stmts,
2773 get_expr_type (expr));
2777 if (!handled && domstmt)
2780 gcc_assert (handled);
2785 #define NECESSARY GF_PLF_1
2787 /* Create an expression in pieces, so that we can handle very complex
2788 expressions that may be ANTIC, but not necessary GIMPLE.
2789 BLOCK is the basic block the expression will be inserted into,
2790 EXPR is the expression to insert (in value form)
2791 STMTS is a statement list to append the necessary insertions into.
2793 This function will die if we hit some value that shouldn't be
2794 ANTIC but is (IE there is no leader for it, or its components).
2795 This function may also generate expressions that are themselves
2796 partially or fully redundant. Those that are will be either made
2797 fully redundant during the next iteration of insert (for partially
2798 redundant ones), or eliminated by eliminate (for fully redundant
2801 If DOMSTMT is non-NULL then we make sure that all uses in the
2802 expressions dominate that statement. In this case the function
2803 can return NULL_TREE to signal failure. */
2806 create_expression_by_pieces (basic_block block, pre_expr expr,
2807 gimple_seq *stmts, gimple domstmt, tree type)
2810 tree folded, newexpr;
2811 gimple_seq forced_stmts;
2812 unsigned int value_id;
2813 gimple_stmt_iterator gsi;
2814 tree exprtype = type ? type : get_expr_type (expr);
2820 /* We may hit the NAME/CONSTANT case if we have to convert types
2821 that value numbering saw through. */
2823 folded = PRE_EXPR_NAME (expr);
2826 folded = PRE_EXPR_CONSTANT (expr);
2830 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2831 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
2836 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2837 switch (nary->length)
2841 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2842 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
2843 tree genop1 = find_or_generate_expression (block, op1,
2845 tree genop2 = find_or_generate_expression (block, op2,
2847 if (!genop1 || !genop2)
2849 genop1 = fold_convert (TREE_TYPE (nary->op[0]),
2851 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
2852 may be a constant with the wrong type. */
2853 if (nary->opcode == POINTER_PLUS_EXPR)
2854 genop2 = fold_convert (sizetype, genop2);
2856 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
2858 folded = fold_build2 (nary->opcode, nary->type,
2864 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
2865 tree genop1 = find_or_generate_expression (block, op1,
2869 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
2871 folded = fold_build1 (nary->opcode, nary->type,
2883 folded = fold_convert (exprtype, folded);
2884 /* Force the generated expression to be a sequence of GIMPLE
2886 We have to call unshare_expr because force_gimple_operand may
2887 modify the tree we pass to it. */
2888 newexpr = force_gimple_operand (unshare_expr (folded), &forced_stmts,
2891 /* If we have any intermediate expressions to the value sets, add them
2892 to the value sets and chain them in the instruction stream. */
2895 gsi = gsi_start (forced_stmts);
2896 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2898 gimple stmt = gsi_stmt (gsi);
2899 tree forcedname = gimple_get_lhs (stmt);
2902 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
2903 if (TREE_CODE (forcedname) == SSA_NAME)
2905 VN_INFO_GET (forcedname)->valnum = forcedname;
2906 VN_INFO (forcedname)->value_id = get_next_value_id ();
2907 nameexpr = get_or_alloc_expr_for_name (forcedname);
2908 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2910 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2911 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2913 mark_symbols_for_renaming (stmt);
2915 gimple_seq_add_seq (stmts, forced_stmts);
2918 /* Build and insert the assignment of the end result to the temporary
2919 that we will return. */
2920 if (!pretemp || exprtype != TREE_TYPE (pretemp))
2922 pretemp = create_tmp_var (exprtype, "pretmp");
2923 get_var_ann (pretemp);
2927 add_referenced_var (temp);
2929 if (TREE_CODE (exprtype) == COMPLEX_TYPE
2930 || TREE_CODE (exprtype) == VECTOR_TYPE)
2931 DECL_GIMPLE_REG_P (temp) = 1;
2933 newstmt = gimple_build_assign (temp, newexpr);
2934 name = make_ssa_name (temp, newstmt);
2935 gimple_assign_set_lhs (newstmt, name);
2936 gimple_set_plf (newstmt, NECESSARY, false);
2938 gimple_seq_add_stmt (stmts, newstmt);
2939 VEC_safe_push (gimple, heap, inserted_exprs, newstmt);
2941 /* All the symbols in NEWEXPR should be put into SSA form. */
2942 mark_symbols_for_renaming (newstmt);
2944 /* Add a value number to the temporary.
2945 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2946 we are creating the expression by pieces, and this particular piece of
2947 the expression may have been represented. There is no harm in replacing
2949 VN_INFO_GET (name)->valnum = name;
2950 value_id = get_expr_value_id (expr);
2951 VN_INFO (name)->value_id = value_id;
2952 nameexpr = get_or_alloc_expr_for_name (name);
2953 add_to_value (value_id, nameexpr);
2955 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2956 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2958 pre_stats.insertions++;
2959 if (dump_file && (dump_flags & TDF_DETAILS))
2961 fprintf (dump_file, "Inserted ");
2962 print_gimple_stmt (dump_file, newstmt, 0, 0);
2963 fprintf (dump_file, " in predecessor %d\n", block->index);
2970 /* Insert the to-be-made-available values of expression EXPRNUM for each
2971 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
2972 merge the result with a phi node, given the same value number as
2973 NODE. Return true if we have inserted new stuff. */
2976 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
2979 pre_expr expr = expression_for_id (exprnum);
2981 unsigned int val = get_expr_value_id (expr);
2983 bool insertions = false;
2988 tree type = get_expr_type (expr);
2992 if (dump_file && (dump_flags & TDF_DETAILS))
2994 fprintf (dump_file, "Found partial redundancy for expression ");
2995 print_pre_expr (dump_file, expr);
2996 fprintf (dump_file, " (%04d)\n", val);
2999 /* Make sure we aren't creating an induction variable. */
3000 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2
3001 && expr->kind != REFERENCE)
3003 bool firstinsideloop = false;
3004 bool secondinsideloop = false;
3005 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3006 EDGE_PRED (block, 0)->src);
3007 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3008 EDGE_PRED (block, 1)->src);
3009 /* Induction variables only have one edge inside the loop. */
3010 if (firstinsideloop ^ secondinsideloop)
3012 if (dump_file && (dump_flags & TDF_DETAILS))
3013 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3018 /* Make sure we are not inserting trapping expressions. */
3019 FOR_EACH_EDGE (pred, ei, block->preds)
3022 eprime = avail[bprime->index];
3023 if (eprime->kind == NARY
3024 && vn_nary_may_trap (PRE_EXPR_NARY (eprime)))
3028 /* Make the necessary insertions. */
3029 FOR_EACH_EDGE (pred, ei, block->preds)
3031 gimple_seq stmts = NULL;
3034 eprime = avail[bprime->index];
3036 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3038 builtexpr = create_expression_by_pieces (bprime,
3042 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3043 gsi_insert_seq_on_edge (pred, stmts);
3044 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3047 else if (eprime->kind == CONSTANT)
3049 /* Constants may not have the right type, fold_convert
3050 should give us back a constant with the right type.
3052 tree constant = PRE_EXPR_CONSTANT (eprime);
3053 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3055 tree builtexpr = fold_convert (type, constant);
3056 if (!is_gimple_min_invariant (builtexpr))
3058 tree forcedexpr = force_gimple_operand (builtexpr,
3061 if (!is_gimple_min_invariant (forcedexpr))
3063 if (forcedexpr != builtexpr)
3065 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3066 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3070 gimple_stmt_iterator gsi;
3071 gsi = gsi_start (stmts);
3072 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3074 gimple stmt = gsi_stmt (gsi);
3075 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3076 gimple_set_plf (stmt, NECESSARY, false);
3078 gsi_insert_seq_on_edge (pred, stmts);
3080 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3085 else if (eprime->kind == NAME)
3087 /* We may have to do a conversion because our value
3088 numbering can look through types in certain cases, but
3089 our IL requires all operands of a phi node have the same
3091 tree name = PRE_EXPR_NAME (eprime);
3092 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3096 builtexpr = fold_convert (type, name);
3097 forcedexpr = force_gimple_operand (builtexpr,
3101 if (forcedexpr != name)
3103 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3104 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3109 gimple_stmt_iterator gsi;
3110 gsi = gsi_start (stmts);
3111 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3113 gimple stmt = gsi_stmt (gsi);
3114 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3115 gimple_set_plf (stmt, NECESSARY, false);
3117 gsi_insert_seq_on_edge (pred, stmts);
3119 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3123 /* If we didn't want a phi node, and we made insertions, we still have
3124 inserted new stuff, and thus return true. If we didn't want a phi node,
3125 and didn't make insertions, we haven't added anything new, so return
3127 if (nophi && insertions)
3129 else if (nophi && !insertions)
3132 /* Now build a phi for the new variable. */
3133 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3135 prephitemp = create_tmp_var (type, "prephitmp");
3136 get_var_ann (prephitemp);
3140 add_referenced_var (temp);
3142 if (TREE_CODE (type) == COMPLEX_TYPE
3143 || TREE_CODE (type) == VECTOR_TYPE)
3144 DECL_GIMPLE_REG_P (temp) = 1;
3145 phi = create_phi_node (temp, block);
3147 gimple_set_plf (phi, NECESSARY, false);
3148 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3149 VN_INFO (gimple_phi_result (phi))->value_id = val;
3150 VEC_safe_push (gimple, heap, inserted_exprs, phi);
3151 FOR_EACH_EDGE (pred, ei, block->preds)
3153 pre_expr ae = avail[pred->src->index];
3154 gcc_assert (get_expr_type (ae) == type
3155 || useless_type_conversion_p (type, get_expr_type (ae)));
3156 if (ae->kind == CONSTANT)
3157 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred);
3159 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred);
3162 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3163 add_to_value (val, newphi);
3165 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3166 this insertion, since we test for the existence of this value in PHI_GEN
3167 before proceeding with the partial redundancy checks in insert_aux.
3169 The value may exist in AVAIL_OUT, in particular, it could be represented
3170 by the expression we are trying to eliminate, in which case we want the
3171 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3174 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3175 this block, because if it did, it would have existed in our dominator's
3176 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3179 bitmap_insert_into_set (PHI_GEN (block), newphi);
3180 bitmap_value_replace_in_set (AVAIL_OUT (block),
3182 bitmap_insert_into_set (NEW_SETS (block),
3185 if (dump_file && (dump_flags & TDF_DETAILS))
3187 fprintf (dump_file, "Created phi ");
3188 print_gimple_stmt (dump_file, phi, 0, 0);
3189 fprintf (dump_file, " in block %d\n", block->index);
3197 /* Perform insertion of partially redundant values.
3198 For BLOCK, do the following:
3199 1. Propagate the NEW_SETS of the dominator into the current block.
3200 If the block has multiple predecessors,
3201 2a. Iterate over the ANTIC expressions for the block to see if
3202 any of them are partially redundant.
3203 2b. If so, insert them into the necessary predecessors to make
3204 the expression fully redundant.
3205 2c. Insert a new PHI merging the values of the predecessors.
3206 2d. Insert the new PHI, and the new expressions, into the
3208 3. Recursively call ourselves on the dominator children of BLOCK.
3210 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3211 do_regular_insertion and do_partial_insertion.
3216 do_regular_insertion (basic_block block, basic_block dom)
3218 bool new_stuff = false;
3219 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3223 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3225 if (expr->kind != NAME)
3229 bool by_some = false;
3230 bool cant_insert = false;
3231 bool all_same = true;
3232 pre_expr first_s = NULL;
3235 pre_expr eprime = NULL;
3237 pre_expr edoubleprime = NULL;
3239 val = get_expr_value_id (expr);
3240 if (bitmap_set_contains_value (PHI_GEN (block), val))
3242 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3244 if (dump_file && (dump_flags & TDF_DETAILS))
3245 fprintf (dump_file, "Found fully redundant value\n");
3249 avail = XCNEWVEC (pre_expr, last_basic_block);
3250 FOR_EACH_EDGE (pred, ei, block->preds)
3252 unsigned int vprime;
3254 /* We should never run insertion for the exit block
3255 and so not come across fake pred edges. */
3256 gcc_assert (!(pred->flags & EDGE_FAKE));
3258 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3261 /* eprime will generally only be NULL if the
3262 value of the expression, translated
3263 through the PHI for this predecessor, is
3264 undefined. If that is the case, we can't
3265 make the expression fully redundant,
3266 because its value is undefined along a
3267 predecessor path. We can thus break out
3268 early because it doesn't matter what the
3269 rest of the results are. */
3276 eprime = fully_constant_expression (eprime);
3277 vprime = get_expr_value_id (eprime);
3278 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3280 if (edoubleprime == NULL)
3282 avail[bprime->index] = eprime;
3287 avail[bprime->index] = edoubleprime;
3289 if (first_s == NULL)
3290 first_s = edoubleprime;
3291 else if (!pre_expr_eq (first_s, edoubleprime))
3295 /* If we can insert it, it's not the same value
3296 already existing along every predecessor, and
3297 it's defined by some predecessor, it is
3298 partially redundant. */
3299 if (!cant_insert && !all_same && by_some && dbg_cnt (treepre_insert))
3301 if (insert_into_preds_of_block (block, get_expression_id (expr),
3305 /* If all edges produce the same value and that value is
3306 an invariant, then the PHI has the same value on all
3307 edges. Note this. */
3308 else if (!cant_insert && all_same && eprime
3309 && (edoubleprime->kind == CONSTANT
3310 || edoubleprime->kind == NAME)
3311 && !value_id_constant_p (val))
3315 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3316 value_expressions, val);
3318 unsigned int new_val = get_expr_value_id (edoubleprime);
3319 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3321 pre_expr expr = expression_for_id (j);
3323 if (expr->kind == NAME)
3325 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3326 /* Just reset the value id and valnum so it is
3327 the same as the constant we have discovered. */
3328 if (edoubleprime->kind == CONSTANT)
3330 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3331 pre_stats.constified++;
3334 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3335 info->value_id = new_val;
3343 VEC_free (pre_expr, heap, exprs);
3348 /* Perform insertion for partially anticipatable expressions. There
3349 is only one case we will perform insertion for these. This case is
3350 if the expression is partially anticipatable, and fully available.
3351 In this case, we know that putting it earlier will enable us to
3352 remove the later computation. */
3356 do_partial_partial_insertion (basic_block block, basic_block dom)
3358 bool new_stuff = false;
3359 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3363 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3365 if (expr->kind != NAME)
3370 bool cant_insert = false;
3373 pre_expr eprime = NULL;
3376 val = get_expr_value_id (expr);
3377 if (bitmap_set_contains_value (PHI_GEN (block), val))
3379 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3382 avail = XCNEWVEC (pre_expr, last_basic_block);
3383 FOR_EACH_EDGE (pred, ei, block->preds)
3385 unsigned int vprime;
3386 pre_expr edoubleprime;
3388 /* We should never run insertion for the exit block
3389 and so not come across fake pred edges. */
3390 gcc_assert (!(pred->flags & EDGE_FAKE));
3392 eprime = phi_translate (expr, ANTIC_IN (block),
3396 /* eprime will generally only be NULL if the
3397 value of the expression, translated
3398 through the PHI for this predecessor, is
3399 undefined. If that is the case, we can't
3400 make the expression fully redundant,
3401 because its value is undefined along a
3402 predecessor path. We can thus break out
3403 early because it doesn't matter what the
3404 rest of the results are. */
3411 eprime = fully_constant_expression (eprime);
3412 vprime = get_expr_value_id (eprime);
3413 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3415 if (edoubleprime == NULL)
3421 avail[bprime->index] = edoubleprime;
3425 /* If we can insert it, it's not the same value
3426 already existing along every predecessor, and
3427 it's defined by some predecessor, it is
3428 partially redundant. */
3429 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3431 pre_stats.pa_insert++;
3432 if (insert_into_preds_of_block (block, get_expression_id (expr),
3440 VEC_free (pre_expr, heap, exprs);
3445 insert_aux (basic_block block)
3448 bool new_stuff = false;
3453 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3458 bitmap_set_t newset = NEW_SETS (dom);
3461 /* Note that we need to value_replace both NEW_SETS, and
3462 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3463 represented by some non-simple expression here that we want
3464 to replace it with. */
3465 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3467 pre_expr expr = expression_for_id (i);
3468 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3469 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3472 if (!single_pred_p (block))
3474 new_stuff |= do_regular_insertion (block, dom);
3475 if (do_partial_partial)
3476 new_stuff |= do_partial_partial_insertion (block, dom);
3480 for (son = first_dom_son (CDI_DOMINATORS, block);
3482 son = next_dom_son (CDI_DOMINATORS, son))
3484 new_stuff |= insert_aux (son);
3490 /* Perform insertion of partially redundant values. */
3495 bool new_stuff = true;
3497 int num_iterations = 0;
3500 NEW_SETS (bb) = bitmap_set_new ();
3505 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3507 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3511 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3514 add_to_exp_gen (basic_block block, tree op)
3519 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3521 result = get_or_alloc_expr_for_name (op);
3522 bitmap_value_insert_into_set (EXP_GEN (block), result);
3526 /* Create value ids for PHI in BLOCK. */
3529 make_values_for_phi (gimple phi, basic_block block)
3531 tree result = gimple_phi_result (phi);
3533 /* We have no need for virtual phis, as they don't represent
3534 actual computations. */
3535 if (is_gimple_reg (result))
3537 pre_expr e = get_or_alloc_expr_for_name (result);
3538 add_to_value (get_expr_value_id (e), e);
3539 bitmap_insert_into_set (PHI_GEN (block), e);
3540 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3544 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3546 tree arg = gimple_phi_arg_def (phi, i);
3547 if (TREE_CODE (arg) == SSA_NAME)
3549 e = get_or_alloc_expr_for_name (arg);
3550 add_to_value (get_expr_value_id (e), e);
3557 /* Compute the AVAIL set for all basic blocks.
3559 This function performs value numbering of the statements in each basic
3560 block. The AVAIL sets are built from information we glean while doing
3561 this value numbering, since the AVAIL sets contain only one entry per
3564 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3565 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3568 compute_avail (void)
3571 basic_block block, son;
3572 basic_block *worklist;
3576 /* For arguments with default definitions, we pretend they are
3577 defined in the entry block. */
3578 for (param = DECL_ARGUMENTS (current_function_decl);
3580 param = TREE_CHAIN (param))
3582 if (gimple_default_def (cfun, param) != NULL)
3584 tree def = gimple_default_def (cfun, param);
3585 pre_expr e = get_or_alloc_expr_for_name (def);
3587 add_to_value (get_expr_value_id (e), e);
3589 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3590 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3594 /* Likewise for the static chain decl. */
3595 if (cfun->static_chain_decl)
3597 param = cfun->static_chain_decl;
3598 if (gimple_default_def (cfun, param) != NULL)
3600 tree def = gimple_default_def (cfun, param);
3601 pre_expr e = get_or_alloc_expr_for_name (def);
3603 add_to_value (get_expr_value_id (e), e);
3605 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3606 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3610 /* Allocate the worklist. */
3611 worklist = XNEWVEC (basic_block, n_basic_blocks);
3613 /* Seed the algorithm by putting the dominator children of the entry
3614 block on the worklist. */
3615 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3617 son = next_dom_son (CDI_DOMINATORS, son))
3618 worklist[sp++] = son;
3620 /* Loop until the worklist is empty. */
3623 gimple_stmt_iterator gsi;
3626 unsigned int stmt_uid = 1;
3628 /* Pick a block from the worklist. */
3629 block = worklist[--sp];
3631 /* Initially, the set of available values in BLOCK is that of
3632 its immediate dominator. */
3633 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3635 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3637 /* Generate values for PHI nodes. */
3638 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3639 make_values_for_phi (gsi_stmt (gsi), block);
3641 /* Now compute value numbers and populate value sets with all
3642 the expressions computed in BLOCK. */
3643 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3648 stmt = gsi_stmt (gsi);
3649 gimple_set_uid (stmt, stmt_uid++);
3651 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3653 pre_expr e = get_or_alloc_expr_for_name (op);
3655 add_to_value (get_expr_value_id (e), e);
3657 bitmap_insert_into_set (TMP_GEN (block), e);
3658 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3661 if (gimple_has_volatile_ops (stmt)
3662 || stmt_could_throw_p (stmt))
3665 switch (gimple_code (stmt))
3668 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3669 add_to_exp_gen (block, op);
3676 vn_reference_op_t vro;
3677 pre_expr result = NULL;
3678 VEC(vn_reference_op_s, heap) *ops = NULL;
3680 if (!can_value_number_call (stmt))
3683 copy_reference_ops_from_call (stmt, &ops);
3684 vn_reference_lookup_pieces (shared_vuses_from_stmt (stmt),
3686 VEC_free (vn_reference_op_s, heap, ops);
3690 for (i = 0; VEC_iterate (vn_reference_op_s,
3694 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3695 add_to_exp_gen (block, vro->op0);
3696 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3697 add_to_exp_gen (block, vro->op1);
3698 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3699 add_to_exp_gen (block, vro->op2);
3701 result = (pre_expr) pool_alloc (pre_expr_pool);
3702 result->kind = REFERENCE;
3704 PRE_EXPR_REFERENCE (result) = ref;
3706 get_or_alloc_expression_id (result);
3707 add_to_value (get_expr_value_id (result), result);
3709 bitmap_value_insert_into_set (EXP_GEN (block), result);
3715 pre_expr result = NULL;
3716 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
3719 if (is_exception_related (stmt))
3722 case tcc_comparison:
3727 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
3728 gimple_assign_rhs_code (stmt),
3729 gimple_expr_type (stmt),
3730 gimple_assign_rhs1 (stmt),
3731 gimple_assign_rhs2 (stmt),
3732 NULL_TREE, NULL_TREE, &nary);
3737 for (i = 0; i < nary->length; i++)
3738 if (TREE_CODE (nary->op[i]) == SSA_NAME)
3739 add_to_exp_gen (block, nary->op[i]);
3741 result = (pre_expr) pool_alloc (pre_expr_pool);
3742 result->kind = NARY;
3744 PRE_EXPR_NARY (result) = nary;
3748 case tcc_declaration:
3753 vn_reference_op_t vro;
3755 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3756 shared_vuses_from_stmt (stmt),
3761 for (i = 0; VEC_iterate (vn_reference_op_s,
3765 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3766 add_to_exp_gen (block, vro->op0);
3767 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3768 add_to_exp_gen (block, vro->op1);
3769 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3770 add_to_exp_gen (block, vro->op2);
3772 result = (pre_expr) pool_alloc (pre_expr_pool);
3773 result->kind = REFERENCE;
3775 PRE_EXPR_REFERENCE (result) = ref;
3780 /* For any other statement that we don't
3781 recognize, simply add all referenced
3782 SSA_NAMEs to EXP_GEN. */
3783 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3784 add_to_exp_gen (block, op);
3788 get_or_alloc_expression_id (result);
3789 add_to_value (get_expr_value_id (result), result);
3791 bitmap_value_insert_into_set (EXP_GEN (block), result);
3800 /* Put the dominator children of BLOCK on the worklist of blocks
3801 to compute available sets for. */
3802 for (son = first_dom_son (CDI_DOMINATORS, block);
3804 son = next_dom_son (CDI_DOMINATORS, son))
3805 worklist[sp++] = son;
3811 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
3812 than the available expressions for it. The insertion point is
3813 right before the first use in STMT. Returns the SSA_NAME that should
3814 be used for replacement. */
3817 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
3819 basic_block bb = gimple_bb (stmt);
3820 gimple_stmt_iterator gsi;
3821 gimple_seq stmts = NULL;
3825 /* First create a value expression from the expression we want
3826 to insert and associate it with the value handle for SSA_VN. */
3827 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
3831 /* Then use create_expression_by_pieces to generate a valid
3832 expression to insert at this point of the IL stream. */
3833 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
3834 if (expr == NULL_TREE)
3836 gsi = gsi_for_stmt (stmt);
3837 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
3842 /* Eliminate fully redundant computations. */
3848 unsigned int todo = 0;
3852 gimple_stmt_iterator i;
3854 for (i = gsi_start_bb (b); !gsi_end_p (i); gsi_next (&i))
3856 gimple stmt = gsi_stmt (i);
3858 /* Lookup the RHS of the expression, see if we have an
3859 available computation for it. If so, replace the RHS with
3860 the available computation. */
3861 if (gimple_has_lhs (stmt)
3862 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
3863 && !gimple_assign_ssa_name_copy_p (stmt)
3864 && (!gimple_assign_single_p (stmt)
3865 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
3866 && !gimple_has_volatile_ops (stmt)
3867 && !has_zero_uses (gimple_get_lhs (stmt)))
3869 tree lhs = gimple_get_lhs (stmt);
3870 tree rhs = NULL_TREE;
3872 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
3873 pre_expr sprimeexpr;
3875 if (gimple_assign_single_p (stmt))
3876 rhs = gimple_assign_rhs1 (stmt);
3878 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
3879 get_expr_value_id (lhsexpr),
3884 if (sprimeexpr->kind == CONSTANT)
3885 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
3886 else if (sprimeexpr->kind == NAME)
3887 sprime = PRE_EXPR_NAME (sprimeexpr);
3892 /* If there is no existing leader but SCCVN knows this
3893 value is constant, use that constant. */
3894 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
3896 sprime = fold_convert (TREE_TYPE (lhs),
3897 VN_INFO (lhs)->valnum);
3899 if (dump_file && (dump_flags & TDF_DETAILS))
3901 fprintf (dump_file, "Replaced ");
3902 print_gimple_expr (dump_file, stmt, 0, 0);
3903 fprintf (dump_file, " with ");
3904 print_generic_expr (dump_file, sprime, 0);
3905 fprintf (dump_file, " in ");
3906 print_gimple_stmt (dump_file, stmt, 0, 0);
3908 pre_stats.eliminations++;
3909 propagate_tree_value_into_stmt (&i, sprime);
3910 stmt = gsi_stmt (i);
3915 /* If there is no existing usable leader but SCCVN thinks
3916 it has an expression it wants to use as replacement,
3918 if (!sprime || sprime == lhs)
3920 tree val = VN_INFO (lhs)->valnum;
3922 && TREE_CODE (val) == SSA_NAME
3923 && VN_INFO (val)->needs_insertion
3924 && can_PRE_operation (vn_get_expr_for (val)))
3925 sprime = do_SCCVN_insertion (stmt, val);
3929 && (rhs == NULL_TREE
3930 || TREE_CODE (rhs) != SSA_NAME
3931 || may_propagate_copy (rhs, sprime)))
3933 gcc_assert (sprime != rhs);
3935 if (dump_file && (dump_flags & TDF_DETAILS))
3937 fprintf (dump_file, "Replaced ");
3938 print_gimple_expr (dump_file, stmt, 0, 0);
3939 fprintf (dump_file, " with ");
3940 print_generic_expr (dump_file, sprime, 0);
3941 fprintf (dump_file, " in ");
3942 print_gimple_stmt (dump_file, stmt, 0, 0);
3945 if (TREE_CODE (sprime) == SSA_NAME)
3946 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
3948 /* We need to make sure the new and old types actually match,
3949 which may require adding a simple cast, which fold_convert
3951 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
3952 && !useless_type_conversion_p (gimple_expr_type (stmt),
3953 TREE_TYPE (sprime)))
3954 sprime = fold_convert (gimple_expr_type (stmt), sprime);
3956 pre_stats.eliminations++;
3957 propagate_tree_value_into_stmt (&i, sprime);
3958 stmt = gsi_stmt (i);
3961 /* If we removed EH side effects from the statement, clean
3962 its EH information. */
3963 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
3965 bitmap_set_bit (need_eh_cleanup,
3966 gimple_bb (stmt)->index);
3967 if (dump_file && (dump_flags & TDF_DETAILS))
3968 fprintf (dump_file, " Removed EH side effects.\n");
3972 /* Visit COND_EXPRs and fold the comparison with the
3973 available value-numbers. */
3974 else if (gimple_code (stmt) == GIMPLE_COND)
3976 tree op0 = gimple_cond_lhs (stmt);
3977 tree op1 = gimple_cond_rhs (stmt);
3980 if (TREE_CODE (op0) == SSA_NAME)
3981 op0 = VN_INFO (op0)->valnum;
3982 if (TREE_CODE (op1) == SSA_NAME)
3983 op1 = VN_INFO (op1)->valnum;
3984 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
3986 if (result && TREE_CODE (result) == INTEGER_CST)
3988 if (integer_zerop (result))
3989 gimple_cond_make_false (stmt);
3991 gimple_cond_make_true (stmt);
3993 todo = TODO_cleanup_cfg;
4002 /* Borrow a bit of tree-ssa-dce.c for the moment.
4003 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4004 this may be a bit faster, and we may want critical edges kept split. */
4006 /* If OP's defining statement has not already been determined to be necessary,
4007 mark that statement necessary. Return the stmt, if it is newly
4010 static inline gimple
4011 mark_operand_necessary (tree op)
4017 if (TREE_CODE (op) != SSA_NAME)
4020 stmt = SSA_NAME_DEF_STMT (op);
4023 if (gimple_plf (stmt, NECESSARY)
4024 || gimple_nop_p (stmt))
4027 gimple_set_plf (stmt, NECESSARY, true);
4031 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4032 to insert PHI nodes sometimes, and because value numbering of casts isn't
4033 perfect, we sometimes end up inserting dead code. This simple DCE-like
4034 pass removes any insertions we made that weren't actually used. */
4037 remove_dead_inserted_code (void)
4039 VEC(gimple,heap) *worklist = NULL;
4043 worklist = VEC_alloc (gimple, heap, VEC_length (gimple, inserted_exprs));
4044 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4046 if (gimple_plf (t, NECESSARY))
4047 VEC_quick_push (gimple, worklist, t);
4049 while (VEC_length (gimple, worklist) > 0)
4051 t = VEC_pop (gimple, worklist);
4053 /* PHI nodes are somewhat special in that each PHI alternative has
4054 data and control dependencies. All the statements feeding the
4055 PHI node's arguments are always necessary. */
4056 if (gimple_code (t) == GIMPLE_PHI)
4060 VEC_reserve (gimple, heap, worklist, gimple_phi_num_args (t));
4061 for (k = 0; k < gimple_phi_num_args (t); k++)
4063 tree arg = PHI_ARG_DEF (t, k);
4064 if (TREE_CODE (arg) == SSA_NAME)
4066 gimple n = mark_operand_necessary (arg);
4068 VEC_quick_push (gimple, worklist, n);
4074 /* Propagate through the operands. Examine all the USE, VUSE and
4075 VDEF operands in this statement. Mark all the statements
4076 which feed this statement's uses as necessary. */
4080 /* The operands of VDEF expressions are also needed as they
4081 represent potential definitions that may reach this
4082 statement (VDEF operands allow us to follow def-def
4085 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4087 gimple n = mark_operand_necessary (use);
4089 VEC_safe_push (gimple, heap, worklist, n);
4094 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4096 if (!gimple_plf (t, NECESSARY))
4098 gimple_stmt_iterator gsi;
4100 if (dump_file && (dump_flags & TDF_DETAILS))
4102 fprintf (dump_file, "Removing unnecessary insertion:");
4103 print_gimple_stmt (dump_file, t, 0, 0);
4106 gsi = gsi_for_stmt (t);
4107 if (gimple_code (t) == GIMPLE_PHI)
4108 remove_phi_node (&gsi, true);
4110 gsi_remove (&gsi, true);
4114 VEC_free (gimple, heap, worklist);
4117 /* Initialize data structures used by PRE. */
4120 init_pre (bool do_fre)
4124 next_expression_id = 1;
4126 VEC_safe_push (pre_expr, heap, expressions, NULL);
4127 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4128 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4129 get_max_value_id() + 1);
4133 inserted_exprs = NULL;
4134 need_creation = NULL;
4135 pretemp = NULL_TREE;
4136 storetemp = NULL_TREE;
4137 prephitemp = NULL_TREE;
4139 connect_infinite_loops_to_exit ();
4140 memset (&pre_stats, 0, sizeof (pre_stats));
4143 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4144 post_order_compute (postorder, false, false);
4147 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4149 calculate_dominance_info (CDI_POST_DOMINATORS);
4150 calculate_dominance_info (CDI_DOMINATORS);
4152 bitmap_obstack_initialize (&grand_bitmap_obstack);
4153 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4154 expr_pred_trans_eq, free);
4155 expression_to_id = htab_create (num_ssa_names * 3,
4158 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4159 sizeof (struct bitmap_set), 30);
4160 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4161 sizeof (struct pre_expr_d), 30);
4164 EXP_GEN (bb) = bitmap_set_new ();
4165 PHI_GEN (bb) = bitmap_set_new ();
4166 TMP_GEN (bb) = bitmap_set_new ();
4167 AVAIL_OUT (bb) = bitmap_set_new ();
4170 need_eh_cleanup = BITMAP_ALLOC (NULL);
4174 /* Deallocate data structures used by PRE. */
4177 fini_pre (bool do_fre)
4182 VEC_free (bitmap_set_t, heap, value_expressions);
4183 VEC_free (gimple, heap, inserted_exprs);
4184 VEC_free (gimple, heap, need_creation);
4185 bitmap_obstack_release (&grand_bitmap_obstack);
4186 free_alloc_pool (bitmap_set_pool);
4187 free_alloc_pool (pre_expr_pool);
4188 htab_delete (phi_translate_table);
4189 htab_delete (expression_to_id);
4197 free_dominance_info (CDI_POST_DOMINATORS);
4199 if (!bitmap_empty_p (need_eh_cleanup))
4201 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4202 cleanup_tree_cfg ();
4205 BITMAP_FREE (need_eh_cleanup);
4208 loop_optimizer_finalize ();
4211 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4212 only wants to do full redundancy elimination. */
4215 execute_pre (bool do_fre ATTRIBUTE_UNUSED)
4217 unsigned int todo = 0;
4219 do_partial_partial = optimize > 2;
4221 /* This has to happen before SCCVN runs because
4222 loop_optimizer_init may create new phis, etc. */
4224 loop_optimizer_init (LOOPS_NORMAL);
4226 if (!run_scc_vn (do_fre))
4230 remove_dead_inserted_code ();
4231 loop_optimizer_finalize ();
4239 /* Collect and value number expressions computed in each basic block. */
4242 if (dump_file && (dump_flags & TDF_DETAILS))
4248 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4249 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4250 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4251 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4255 /* Insert can get quite slow on an incredibly large number of basic
4256 blocks due to some quadratic behavior. Until this behavior is
4257 fixed, don't run it when he have an incredibly large number of
4258 bb's. If we aren't going to run insert, there is no point in
4259 computing ANTIC, either, even though it's plenty fast. */
4260 if (!do_fre && n_basic_blocks < 4000)
4266 /* Remove all the redundant expressions. */
4267 todo |= eliminate ();
4269 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4270 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4271 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4272 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4273 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4275 /* Make sure to remove fake edges before committing our inserts.
4276 This makes sure we don't end up with extra critical edges that
4277 we would need to split. */
4278 remove_fake_exit_edges ();
4279 gsi_commit_edge_inserts ();
4281 clear_expression_ids ();
4284 remove_dead_inserted_code ();
4291 /* Gate and execute functions for PRE. */
4296 return TODO_rebuild_alias | execute_pre (false);
4302 /* PRE tends to generate bigger code. */
4303 return flag_tree_pre != 0 && optimize_function_for_speed_p (cfun);
4306 struct gimple_opt_pass pass_pre =
4311 gate_pre, /* gate */
4312 do_pre, /* execute */
4315 0, /* static_pass_number */
4316 TV_TREE_PRE, /* tv_id */
4317 PROP_no_crit_edges | PROP_cfg
4318 | PROP_ssa | PROP_alias, /* properties_required */
4319 0, /* properties_provided */
4320 0, /* properties_destroyed */
4321 0, /* todo_flags_start */
4322 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4323 | TODO_verify_ssa /* todo_flags_finish */
4328 /* Gate and execute functions for FRE. */
4333 return execute_pre (true);
4339 return flag_tree_fre != 0;
4342 struct gimple_opt_pass pass_fre =
4347 gate_fre, /* gate */
4348 execute_fre, /* execute */
4351 0, /* static_pass_number */
4352 TV_TREE_FRE, /* tv_id */
4353 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
4354 0, /* properties_provided */
4355 0, /* properties_destroyed */
4356 0, /* todo_flags_start */
4357 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */