2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
25 #include "coretypes.h"
28 #include "basic-block.h"
29 #include "tree-pretty-print.h"
30 #include "gimple-pretty-print.h"
31 #include "tree-inline.h"
32 #include "tree-flow.h"
34 #include "tree-dump.h"
38 #include "tree-iterator.h"
39 #include "alloc-pool.h"
41 #include "tree-pass.h"
44 #include "langhooks.h"
46 #include "tree-ssa-sccvn.h"
47 #include "tree-scalar-evolution.h"
53 1. Avail sets can be shared by making an avail_find_leader that
54 walks up the dominator tree and looks in those avail sets.
55 This might affect code optimality, it's unclear right now.
56 2. Strength reduction can be performed by anticipating expressions
57 we can repair later on.
58 3. We can do back-substitution or smarter value numbering to catch
59 commutative expressions split up over multiple statements.
62 /* For ease of terminology, "expression node" in the below refers to
63 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
64 represent the actual statement containing the expressions we care about,
65 and we cache the value number by putting it in the expression. */
69 First we walk the statements to generate the AVAIL sets, the
70 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
71 generation of values/expressions by a given block. We use them
72 when computing the ANTIC sets. The AVAIL sets consist of
73 SSA_NAME's that represent values, so we know what values are
74 available in what blocks. AVAIL is a forward dataflow problem. In
75 SSA, values are never killed, so we don't need a kill set, or a
76 fixpoint iteration, in order to calculate the AVAIL sets. In
77 traditional parlance, AVAIL sets tell us the downsafety of the
80 Next, we generate the ANTIC sets. These sets represent the
81 anticipatable expressions. ANTIC is a backwards dataflow
82 problem. An expression is anticipatable in a given block if it could
83 be generated in that block. This means that if we had to perform
84 an insertion in that block, of the value of that expression, we
85 could. Calculating the ANTIC sets requires phi translation of
86 expressions, because the flow goes backwards through phis. We must
87 iterate to a fixpoint of the ANTIC sets, because we have a kill
88 set. Even in SSA form, values are not live over the entire
89 function, only from their definition point onwards. So we have to
90 remove values from the ANTIC set once we go past the definition
91 point of the leaders that make them up.
92 compute_antic/compute_antic_aux performs this computation.
94 Third, we perform insertions to make partially redundant
95 expressions fully redundant.
97 An expression is partially redundant (excluding partial
100 1. It is AVAIL in some, but not all, of the predecessors of a
102 2. It is ANTIC in all the predecessors.
104 In order to make it fully redundant, we insert the expression into
105 the predecessors where it is not available, but is ANTIC.
107 For the partial anticipation case, we only perform insertion if it
108 is partially anticipated in some block, and fully available in all
111 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
112 performs these steps.
114 Fourth, we eliminate fully redundant expressions.
115 This is a simple statement walk that replaces redundant
116 calculations with the now available values. */
118 /* Representations of value numbers:
120 Value numbers are represented by a representative SSA_NAME. We
121 will create fake SSA_NAME's in situations where we need a
122 representative but do not have one (because it is a complex
123 expression). In order to facilitate storing the value numbers in
124 bitmaps, and keep the number of wasted SSA_NAME's down, we also
125 associate a value_id with each value number, and create full blown
126 ssa_name's only where we actually need them (IE in operands of
127 existing expressions).
129 Theoretically you could replace all the value_id's with
130 SSA_NAME_VERSION, but this would allocate a large number of
131 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
132 It would also require an additional indirection at each point we
135 /* Representation of expressions on value numbers:
137 Expressions consisting of value numbers are represented the same
138 way as our VN internally represents them, with an additional
139 "pre_expr" wrapping around them in order to facilitate storing all
140 of the expressions in the same sets. */
142 /* Representation of sets:
144 The dataflow sets do not need to be sorted in any particular order
145 for the majority of their lifetime, are simply represented as two
146 bitmaps, one that keeps track of values present in the set, and one
147 that keeps track of expressions present in the set.
149 When we need them in topological order, we produce it on demand by
150 transforming the bitmap into an array and sorting it into topo
153 /* Type of expression, used to know which member of the PRE_EXPR union
164 typedef union pre_expr_union_d
169 vn_reference_t reference;
172 typedef struct pre_expr_d
174 enum pre_expr_kind kind;
179 #define PRE_EXPR_NAME(e) (e)->u.name
180 #define PRE_EXPR_NARY(e) (e)->u.nary
181 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
182 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
185 pre_expr_eq (const void *p1, const void *p2)
187 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
188 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
190 if (e1->kind != e2->kind)
196 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
197 PRE_EXPR_CONSTANT (e2));
199 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
201 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
203 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
204 PRE_EXPR_REFERENCE (e2));
211 pre_expr_hash (const void *p1)
213 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
217 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
219 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
221 return PRE_EXPR_NARY (e)->hashcode;
223 return PRE_EXPR_REFERENCE (e)->hashcode;
230 /* Next global expression id number. */
231 static unsigned int next_expression_id;
233 /* Mapping from expression to id number we can use in bitmap sets. */
234 DEF_VEC_P (pre_expr);
235 DEF_VEC_ALLOC_P (pre_expr, heap);
236 static VEC(pre_expr, heap) *expressions;
237 static htab_t expression_to_id;
238 static VEC(unsigned, heap) *name_to_id;
240 /* Allocate an expression id for EXPR. */
242 static inline unsigned int
243 alloc_expression_id (pre_expr expr)
246 /* Make sure we won't overflow. */
247 gcc_assert (next_expression_id + 1 > next_expression_id);
248 expr->id = next_expression_id++;
249 VEC_safe_push (pre_expr, heap, expressions, expr);
250 if (expr->kind == NAME)
252 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
253 /* VEC_safe_grow_cleared allocates no headroom. Avoid frequent
254 re-allocations by using VEC_reserve upfront. There is no
255 VEC_quick_grow_cleared unfortunately. */
256 VEC_reserve (unsigned, heap, name_to_id, num_ssa_names);
257 VEC_safe_grow_cleared (unsigned, heap, name_to_id, num_ssa_names);
258 gcc_assert (VEC_index (unsigned, name_to_id, version) == 0);
259 VEC_replace (unsigned, name_to_id, version, expr->id);
263 slot = htab_find_slot (expression_to_id, expr, INSERT);
267 return next_expression_id - 1;
270 /* Return the expression id for tree EXPR. */
272 static inline unsigned int
273 get_expression_id (const pre_expr expr)
278 static inline unsigned int
279 lookup_expression_id (const pre_expr expr)
283 if (expr->kind == NAME)
285 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
286 if (VEC_length (unsigned, name_to_id) <= version)
288 return VEC_index (unsigned, name_to_id, version);
292 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
295 return ((pre_expr)*slot)->id;
299 /* Return the existing expression id for EXPR, or create one if one
300 does not exist yet. */
302 static inline unsigned int
303 get_or_alloc_expression_id (pre_expr expr)
305 unsigned int id = lookup_expression_id (expr);
307 return alloc_expression_id (expr);
308 return expr->id = id;
311 /* Return the expression that has expression id ID */
313 static inline pre_expr
314 expression_for_id (unsigned int id)
316 return VEC_index (pre_expr, expressions, id);
319 /* Free the expression id field in all of our expressions,
320 and then destroy the expressions array. */
323 clear_expression_ids (void)
325 VEC_free (pre_expr, heap, expressions);
328 static alloc_pool pre_expr_pool;
330 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
333 get_or_alloc_expr_for_name (tree name)
335 struct pre_expr_d expr;
337 unsigned int result_id;
341 PRE_EXPR_NAME (&expr) = name;
342 result_id = lookup_expression_id (&expr);
344 return expression_for_id (result_id);
346 result = (pre_expr) pool_alloc (pre_expr_pool);
348 PRE_EXPR_NAME (result) = name;
349 alloc_expression_id (result);
353 static bool in_fre = false;
355 /* An unordered bitmap set. One bitmap tracks values, the other,
357 typedef struct bitmap_set
359 bitmap_head expressions;
363 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
364 EXECUTE_IF_SET_IN_BITMAP(&(set)->expressions, 0, (id), (bi))
366 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
367 EXECUTE_IF_SET_IN_BITMAP(&(set)->values, 0, (id), (bi))
369 /* Mapping from value id to expressions with that value_id. */
370 DEF_VEC_P (bitmap_set_t);
371 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
372 static VEC(bitmap_set_t, heap) *value_expressions;
374 /* Sets that we need to keep track of. */
375 typedef struct bb_bitmap_sets
377 /* The EXP_GEN set, which represents expressions/values generated in
379 bitmap_set_t exp_gen;
381 /* The PHI_GEN set, which represents PHI results generated in a
383 bitmap_set_t phi_gen;
385 /* The TMP_GEN set, which represents results/temporaries generated
386 in a basic block. IE the LHS of an expression. */
387 bitmap_set_t tmp_gen;
389 /* The AVAIL_OUT set, which represents which values are available in
390 a given basic block. */
391 bitmap_set_t avail_out;
393 /* The ANTIC_IN set, which represents which values are anticipatable
394 in a given basic block. */
395 bitmap_set_t antic_in;
397 /* The PA_IN set, which represents which values are
398 partially anticipatable in a given basic block. */
401 /* The NEW_SETS set, which is used during insertion to augment the
402 AVAIL_OUT set of blocks with the new insertions performed during
403 the current iteration. */
404 bitmap_set_t new_sets;
406 /* A cache for value_dies_in_block_x. */
409 /* True if we have visited this block during ANTIC calculation. */
410 unsigned int visited : 1;
412 /* True we have deferred processing this block during ANTIC
413 calculation until its successor is processed. */
414 unsigned int deferred : 1;
416 /* True when the block contains a call that might not return. */
417 unsigned int contains_may_not_return_call : 1;
420 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
421 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
422 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
423 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
424 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
425 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
426 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
427 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
428 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
429 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
430 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
433 /* Basic block list in postorder. */
434 static int *postorder;
436 /* This structure is used to keep track of statistics on what
437 optimization PRE was able to perform. */
440 /* The number of RHS computations eliminated by PRE. */
443 /* The number of new expressions/temporaries generated by PRE. */
446 /* The number of inserts found due to partial anticipation */
449 /* The number of new PHI nodes added by PRE. */
452 /* The number of values found constant. */
457 static bool do_partial_partial;
458 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
459 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
460 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
461 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
462 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
463 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
464 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
466 static bitmap_set_t bitmap_set_new (void);
467 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
469 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
471 static unsigned int get_expr_value_id (pre_expr);
473 /* We can add and remove elements and entries to and from sets
474 and hash tables, so we use alloc pools for them. */
476 static alloc_pool bitmap_set_pool;
477 static bitmap_obstack grand_bitmap_obstack;
479 /* To avoid adding 300 temporary variables when we only need one, we
480 only create one temporary variable, on demand, and build ssa names
481 off that. We do have to change the variable if the types don't
482 match the current variable's type. */
484 static tree storetemp;
485 static tree prephitemp;
487 /* Set of blocks with statements that have had its EH information
489 static bitmap need_eh_cleanup;
491 /* The phi_translate_table caches phi translations for a given
492 expression and predecessor. */
494 static htab_t phi_translate_table;
496 /* A three tuple {e, pred, v} used to cache phi translations in the
497 phi_translate_table. */
499 typedef struct expr_pred_trans_d
501 /* The expression. */
504 /* The predecessor block along which we translated the expression. */
507 /* The value that resulted from the translation. */
510 /* The hashcode for the expression, pred pair. This is cached for
513 } *expr_pred_trans_t;
514 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
516 /* Return the hash value for a phi translation table entry. */
519 expr_pred_trans_hash (const void *p)
521 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
525 /* Return true if two phi translation table entries are the same.
526 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
529 expr_pred_trans_eq (const void *p1, const void *p2)
531 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
532 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
533 basic_block b1 = ve1->pred;
534 basic_block b2 = ve2->pred;
536 /* If they are not translations for the same basic block, they can't
540 return pre_expr_eq (ve1->e, ve2->e);
543 /* Search in the phi translation table for the translation of
544 expression E in basic block PRED.
545 Return the translated value, if found, NULL otherwise. */
547 static inline pre_expr
548 phi_trans_lookup (pre_expr e, basic_block pred)
551 struct expr_pred_trans_d ept;
555 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
556 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
561 return ((expr_pred_trans_t) *slot)->v;
565 /* Add the tuple mapping from {expression E, basic block PRED} to
566 value V, to the phi translation table. */
569 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
572 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
574 new_pair->pred = pred;
576 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
579 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
580 new_pair->hashcode, INSERT);
583 *slot = (void *) new_pair;
587 /* Add expression E to the expression set of value id V. */
590 add_to_value (unsigned int v, pre_expr e)
594 gcc_assert (get_expr_value_id (e) == v);
596 if (v >= VEC_length (bitmap_set_t, value_expressions))
598 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
602 set = VEC_index (bitmap_set_t, value_expressions, v);
605 set = bitmap_set_new ();
606 VEC_replace (bitmap_set_t, value_expressions, v, set);
609 bitmap_insert_into_set_1 (set, e, v, true);
612 /* Create a new bitmap set and return it. */
615 bitmap_set_new (void)
617 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
618 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
619 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
623 /* Return the value id for a PRE expression EXPR. */
626 get_expr_value_id (pre_expr expr)
633 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
636 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
637 add_to_value (id, expr);
642 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
644 return PRE_EXPR_NARY (expr)->value_id;
646 return PRE_EXPR_REFERENCE (expr)->value_id;
652 /* Remove an expression EXPR from a bitmapped set. */
655 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
657 unsigned int val = get_expr_value_id (expr);
658 if (!value_id_constant_p (val))
660 bitmap_clear_bit (&set->values, val);
661 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
666 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
667 unsigned int val, bool allow_constants)
669 if (allow_constants || !value_id_constant_p (val))
671 /* We specifically expect this and only this function to be able to
672 insert constants into a set. */
673 bitmap_set_bit (&set->values, val);
674 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
678 /* Insert an expression EXPR into a bitmapped set. */
681 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
683 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
686 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
689 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
691 bitmap_copy (&dest->expressions, &orig->expressions);
692 bitmap_copy (&dest->values, &orig->values);
696 /* Free memory used up by SET. */
698 bitmap_set_free (bitmap_set_t set)
700 bitmap_clear (&set->expressions);
701 bitmap_clear (&set->values);
705 /* Generate an topological-ordered array of bitmap set SET. */
707 static VEC(pre_expr, heap) *
708 sorted_array_from_bitmap_set (bitmap_set_t set)
711 bitmap_iterator bi, bj;
712 VEC(pre_expr, heap) *result;
714 /* Pre-allocate roughly enough space for the array. */
715 result = VEC_alloc (pre_expr, heap, bitmap_count_bits (&set->values));
717 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
719 /* The number of expressions having a given value is usually
720 relatively small. Thus, rather than making a vector of all
721 the expressions and sorting it by value-id, we walk the values
722 and check in the reverse mapping that tells us what expressions
723 have a given value, to filter those in our set. As a result,
724 the expressions are inserted in value-id order, which means
727 If this is somehow a significant lose for some cases, we can
728 choose which set to walk based on the set size. */
729 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
730 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
732 if (bitmap_bit_p (&set->expressions, j))
733 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
740 /* Perform bitmapped set operation DEST &= ORIG. */
743 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
751 bitmap_initialize (&temp, &grand_bitmap_obstack);
753 bitmap_and_into (&dest->values, &orig->values);
754 bitmap_copy (&temp, &dest->expressions);
755 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
757 pre_expr expr = expression_for_id (i);
758 unsigned int value_id = get_expr_value_id (expr);
759 if (!bitmap_bit_p (&dest->values, value_id))
760 bitmap_clear_bit (&dest->expressions, i);
762 bitmap_clear (&temp);
766 /* Subtract all values and expressions contained in ORIG from DEST. */
769 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
771 bitmap_set_t result = bitmap_set_new ();
775 bitmap_and_compl (&result->expressions, &dest->expressions,
778 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
780 pre_expr expr = expression_for_id (i);
781 unsigned int value_id = get_expr_value_id (expr);
782 bitmap_set_bit (&result->values, value_id);
788 /* Subtract all the values in bitmap set B from bitmap set A. */
791 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
797 bitmap_initialize (&temp, &grand_bitmap_obstack);
799 bitmap_copy (&temp, &a->expressions);
800 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
802 pre_expr expr = expression_for_id (i);
803 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
804 bitmap_remove_from_set (a, expr);
806 bitmap_clear (&temp);
810 /* Return true if bitmapped set SET contains the value VALUE_ID. */
813 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
815 if (value_id_constant_p (value_id))
818 if (!set || bitmap_empty_p (&set->expressions))
821 return bitmap_bit_p (&set->values, value_id);
825 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
827 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
830 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
833 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
836 bitmap_set_t exprset;
840 if (value_id_constant_p (lookfor))
843 if (!bitmap_set_contains_value (set, lookfor))
846 /* The number of expressions having a given value is usually
847 significantly less than the total number of expressions in SET.
848 Thus, rather than check, for each expression in SET, whether it
849 has the value LOOKFOR, we walk the reverse mapping that tells us
850 what expressions have a given value, and see if any of those
851 expressions are in our set. For large testcases, this is about
852 5-10x faster than walking the bitmap. If this is somehow a
853 significant lose for some cases, we can choose which set to walk
854 based on the set size. */
855 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
856 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
858 if (bitmap_clear_bit (&set->expressions, i))
860 bitmap_set_bit (&set->expressions, get_expression_id (expr));
866 /* Return true if two bitmap sets are equal. */
869 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
871 return bitmap_equal_p (&a->values, &b->values);
874 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
875 and add it otherwise. */
878 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
880 unsigned int val = get_expr_value_id (expr);
882 if (bitmap_set_contains_value (set, val))
883 bitmap_set_replace_value (set, val, expr);
885 bitmap_insert_into_set (set, expr);
888 /* Insert EXPR into SET if EXPR's value is not already present in
892 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
894 unsigned int val = get_expr_value_id (expr);
896 #ifdef ENABLE_CHECKING
897 gcc_assert (expr->id == get_or_alloc_expression_id (expr));
900 /* Constant values are always considered to be part of the set. */
901 if (value_id_constant_p (val))
904 /* If the value membership changed, add the expression. */
905 if (bitmap_set_bit (&set->values, val))
906 bitmap_set_bit (&set->expressions, expr->id);
909 /* Print out EXPR to outfile. */
912 print_pre_expr (FILE *outfile, const pre_expr expr)
917 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
920 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
925 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
926 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
927 for (i = 0; i < nary->length; i++)
929 print_generic_expr (outfile, nary->op[i], 0);
930 if (i != (unsigned) nary->length - 1)
931 fprintf (outfile, ",");
933 fprintf (outfile, "}");
939 vn_reference_op_t vro;
941 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
942 fprintf (outfile, "{");
944 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
947 bool closebrace = false;
948 if (vro->opcode != SSA_NAME
949 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
951 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
954 fprintf (outfile, "<");
960 print_generic_expr (outfile, vro->op0, 0);
963 fprintf (outfile, ",");
964 print_generic_expr (outfile, vro->op1, 0);
968 fprintf (outfile, ",");
969 print_generic_expr (outfile, vro->op2, 0);
973 fprintf (outfile, ">");
974 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
975 fprintf (outfile, ",");
977 fprintf (outfile, "}");
980 fprintf (outfile, "@");
981 print_generic_expr (outfile, ref->vuse, 0);
987 void debug_pre_expr (pre_expr);
989 /* Like print_pre_expr but always prints to stderr. */
991 debug_pre_expr (pre_expr e)
993 print_pre_expr (stderr, e);
994 fprintf (stderr, "\n");
997 /* Print out SET to OUTFILE. */
1000 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1001 const char *setname, int blockindex)
1003 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1010 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1012 const pre_expr expr = expression_for_id (i);
1015 fprintf (outfile, ", ");
1017 print_pre_expr (outfile, expr);
1019 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1022 fprintf (outfile, " }\n");
1025 void debug_bitmap_set (bitmap_set_t);
1028 debug_bitmap_set (bitmap_set_t set)
1030 print_bitmap_set (stderr, set, "debug", 0);
1033 /* Print out the expressions that have VAL to OUTFILE. */
1036 print_value_expressions (FILE *outfile, unsigned int val)
1038 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1042 sprintf (s, "%04d", val);
1043 print_bitmap_set (outfile, set, s, 0);
1049 debug_value_expressions (unsigned int val)
1051 print_value_expressions (stderr, val);
1054 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1058 get_or_alloc_expr_for_constant (tree constant)
1060 unsigned int result_id;
1061 unsigned int value_id;
1062 struct pre_expr_d expr;
1065 expr.kind = CONSTANT;
1066 PRE_EXPR_CONSTANT (&expr) = constant;
1067 result_id = lookup_expression_id (&expr);
1069 return expression_for_id (result_id);
1071 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1072 newexpr->kind = CONSTANT;
1073 PRE_EXPR_CONSTANT (newexpr) = constant;
1074 alloc_expression_id (newexpr);
1075 value_id = get_or_alloc_constant_value_id (constant);
1076 add_to_value (value_id, newexpr);
1080 /* Given a value id V, find the actual tree representing the constant
1081 value if there is one, and return it. Return NULL if we can't find
1085 get_constant_for_value_id (unsigned int v)
1087 if (value_id_constant_p (v))
1091 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1093 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1095 pre_expr expr = expression_for_id (i);
1096 if (expr->kind == CONSTANT)
1097 return PRE_EXPR_CONSTANT (expr);
1103 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1104 Currently only supports constants and SSA_NAMES. */
1106 get_or_alloc_expr_for (tree t)
1108 if (TREE_CODE (t) == SSA_NAME)
1109 return get_or_alloc_expr_for_name (t);
1110 else if (is_gimple_min_invariant (t))
1111 return get_or_alloc_expr_for_constant (t);
1114 /* More complex expressions can result from SCCVN expression
1115 simplification that inserts values for them. As they all
1116 do not have VOPs the get handled by the nary ops struct. */
1117 vn_nary_op_t result;
1118 unsigned int result_id;
1119 vn_nary_op_lookup (t, &result);
1122 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1124 PRE_EXPR_NARY (e) = result;
1125 result_id = lookup_expression_id (e);
1128 pool_free (pre_expr_pool, e);
1129 e = expression_for_id (result_id);
1132 alloc_expression_id (e);
1139 /* Return the folded version of T if T, when folded, is a gimple
1140 min_invariant. Otherwise, return T. */
1143 fully_constant_expression (pre_expr e)
1151 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1152 switch (TREE_CODE_CLASS (nary->opcode))
1154 case tcc_expression:
1155 if (nary->opcode == TRUTH_NOT_EXPR)
1157 if (nary->opcode != TRUTH_AND_EXPR
1158 && nary->opcode != TRUTH_OR_EXPR
1159 && nary->opcode != TRUTH_XOR_EXPR)
1163 case tcc_comparison:
1165 /* We have to go from trees to pre exprs to value ids to
1167 tree naryop0 = nary->op[0];
1168 tree naryop1 = nary->op[1];
1170 if (!is_gimple_min_invariant (naryop0))
1172 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1173 unsigned int vrep0 = get_expr_value_id (rep0);
1174 tree const0 = get_constant_for_value_id (vrep0);
1176 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1178 if (!is_gimple_min_invariant (naryop1))
1180 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1181 unsigned int vrep1 = get_expr_value_id (rep1);
1182 tree const1 = get_constant_for_value_id (vrep1);
1184 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1186 result = fold_binary (nary->opcode, nary->type,
1188 if (result && is_gimple_min_invariant (result))
1189 return get_or_alloc_expr_for_constant (result);
1190 /* We might have simplified the expression to a
1191 SSA_NAME for example from x_1 * 1. But we cannot
1192 insert a PHI for x_1 unconditionally as x_1 might
1193 not be available readily. */
1197 if (nary->opcode != REALPART_EXPR
1198 && nary->opcode != IMAGPART_EXPR
1199 && nary->opcode != VIEW_CONVERT_EXPR)
1205 /* We have to go from trees to pre exprs to value ids to
1207 tree naryop0 = nary->op[0];
1208 tree const0, result;
1209 if (is_gimple_min_invariant (naryop0))
1213 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1214 unsigned int vrep0 = get_expr_value_id (rep0);
1215 const0 = get_constant_for_value_id (vrep0);
1220 tree type1 = TREE_TYPE (nary->op[0]);
1221 const0 = fold_convert (type1, const0);
1222 result = fold_unary (nary->opcode, nary->type, const0);
1224 if (result && is_gimple_min_invariant (result))
1225 return get_or_alloc_expr_for_constant (result);
1234 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1236 if ((folded = fully_constant_vn_reference_p (ref)))
1237 return get_or_alloc_expr_for_constant (folded);
1246 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1247 it has the value it would have in BLOCK. Set *SAME_VALID to true
1248 in case the new vuse doesn't change the value id of the OPERANDS. */
1251 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1252 alias_set_type set, tree type, tree vuse,
1253 basic_block phiblock,
1254 basic_block block, bool *same_valid)
1256 gimple phi = SSA_NAME_DEF_STMT (vuse);
1263 if (gimple_bb (phi) != phiblock)
1266 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1268 /* Use the alias-oracle to find either the PHI node in this block,
1269 the first VUSE used in this block that is equivalent to vuse or
1270 the first VUSE which definition in this block kills the value. */
1271 if (gimple_code (phi) == GIMPLE_PHI)
1272 e = find_edge (block, phiblock);
1273 else if (use_oracle)
1274 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1276 vuse = gimple_vuse (phi);
1277 phi = SSA_NAME_DEF_STMT (vuse);
1278 if (gimple_bb (phi) != phiblock)
1280 if (gimple_code (phi) == GIMPLE_PHI)
1282 e = find_edge (block, phiblock);
1293 bitmap visited = NULL;
1294 /* Try to find a vuse that dominates this phi node by skipping
1295 non-clobbering statements. */
1296 vuse = get_continuation_for_phi (phi, &ref, &visited);
1298 BITMAP_FREE (visited);
1304 /* If we didn't find any, the value ID can't stay the same,
1305 but return the translated vuse. */
1306 *same_valid = false;
1307 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1309 /* ??? We would like to return vuse here as this is the canonical
1310 upmost vdef that this reference is associated with. But during
1311 insertion of the references into the hash tables we only ever
1312 directly insert with their direct gimple_vuse, hence returning
1313 something else would make us not find the other expression. */
1314 return PHI_ARG_DEF (phi, e->dest_idx);
1320 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1321 SET2. This is used to avoid making a set consisting of the union
1322 of PA_IN and ANTIC_IN during insert. */
1324 static inline pre_expr
1325 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1329 result = bitmap_find_leader (set1, val, NULL);
1330 if (!result && set2)
1331 result = bitmap_find_leader (set2, val, NULL);
1335 /* Get the tree type for our PRE expression e. */
1338 get_expr_type (const pre_expr e)
1343 return TREE_TYPE (PRE_EXPR_NAME (e));
1345 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1347 return PRE_EXPR_REFERENCE (e)->type;
1349 return PRE_EXPR_NARY (e)->type;
1354 /* Get a representative SSA_NAME for a given expression.
1355 Since all of our sub-expressions are treated as values, we require
1356 them to be SSA_NAME's for simplicity.
1357 Prior versions of GVNPRE used to use "value handles" here, so that
1358 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1359 either case, the operands are really values (IE we do not expect
1360 them to be usable without finding leaders). */
1363 get_representative_for (const pre_expr e)
1367 unsigned int value_id = get_expr_value_id (e);
1372 return PRE_EXPR_NAME (e);
1374 return PRE_EXPR_CONSTANT (e);
1378 /* Go through all of the expressions representing this value
1379 and pick out an SSA_NAME. */
1382 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1384 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1386 pre_expr rep = expression_for_id (i);
1387 if (rep->kind == NAME)
1388 return PRE_EXPR_NAME (rep);
1393 /* If we reached here we couldn't find an SSA_NAME. This can
1394 happen when we've discovered a value that has never appeared in
1395 the program as set to an SSA_NAME, most likely as the result of
1400 "Could not find SSA_NAME representative for expression:");
1401 print_pre_expr (dump_file, e);
1402 fprintf (dump_file, "\n");
1405 exprtype = get_expr_type (e);
1407 /* Build and insert the assignment of the end result to the temporary
1408 that we will return. */
1409 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1411 pretemp = create_tmp_reg (exprtype, "pretmp");
1412 get_var_ann (pretemp);
1415 name = make_ssa_name (pretemp, gimple_build_nop ());
1416 VN_INFO_GET (name)->value_id = value_id;
1417 if (e->kind == CONSTANT)
1418 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1420 VN_INFO (name)->valnum = name;
1422 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1425 fprintf (dump_file, "Created SSA_NAME representative ");
1426 print_generic_expr (dump_file, name, 0);
1427 fprintf (dump_file, " for expression:");
1428 print_pre_expr (dump_file, e);
1429 fprintf (dump_file, "\n");
1438 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1439 basic_block pred, basic_block phiblock);
1441 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1442 the phis in PRED. Return NULL if we can't find a leader for each part
1443 of the translated expression. */
1446 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1447 basic_block pred, basic_block phiblock)
1454 bool changed = false;
1455 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1456 struct vn_nary_op_s newnary;
1457 /* The NARY structure is only guaranteed to have been
1458 allocated to the nary->length operands. */
1459 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1460 - sizeof (tree) * (4 - nary->length)));
1462 for (i = 0; i < newnary.length; i++)
1464 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1468 pre_expr leader, result;
1469 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1470 leader = find_leader_in_sets (op_val_id, set1, set2);
1471 result = phi_translate (leader, set1, set2, pred, phiblock);
1472 if (result && result != leader)
1474 tree name = get_representative_for (result);
1477 newnary.op[i] = name;
1482 changed |= newnary.op[i] != nary->op[i];
1488 unsigned int new_val_id;
1490 tree result = vn_nary_op_lookup_pieces (newnary.length,
1498 if (result && is_gimple_min_invariant (result))
1499 return get_or_alloc_expr_for_constant (result);
1501 expr = (pre_expr) pool_alloc (pre_expr_pool);
1506 PRE_EXPR_NARY (expr) = nary;
1507 constant = fully_constant_expression (expr);
1508 if (constant != expr)
1511 new_val_id = nary->value_id;
1512 get_or_alloc_expression_id (expr);
1516 new_val_id = get_next_value_id ();
1517 VEC_safe_grow_cleared (bitmap_set_t, heap,
1519 get_max_value_id() + 1);
1520 nary = vn_nary_op_insert_pieces (newnary.length,
1527 result, new_val_id);
1528 PRE_EXPR_NARY (expr) = nary;
1529 constant = fully_constant_expression (expr);
1530 if (constant != expr)
1532 get_or_alloc_expression_id (expr);
1534 add_to_value (new_val_id, expr);
1542 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1543 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1544 tree vuse = ref->vuse;
1545 tree newvuse = vuse;
1546 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1547 bool changed = false, same_valid = true;
1549 vn_reference_op_t operand;
1550 vn_reference_t newref;
1553 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1557 tree oldop0 = operand->op0;
1558 tree oldop1 = operand->op1;
1559 tree oldop2 = operand->op2;
1563 tree type = operand->type;
1564 vn_reference_op_s newop = *operand;
1566 if (op0 && TREE_CODE (op0) == SSA_NAME)
1568 unsigned int op_val_id = VN_INFO (op0)->value_id;
1569 leader = find_leader_in_sets (op_val_id, set1, set2);
1570 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1571 if (opresult && opresult != leader)
1573 tree name = get_representative_for (opresult);
1581 changed |= op0 != oldop0;
1583 if (op1 && TREE_CODE (op1) == SSA_NAME)
1585 unsigned int op_val_id = VN_INFO (op1)->value_id;
1586 leader = find_leader_in_sets (op_val_id, set1, set2);
1587 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1588 if (opresult && opresult != leader)
1590 tree name = get_representative_for (opresult);
1598 /* We can't possibly insert these. */
1599 else if (op1 && !is_gimple_min_invariant (op1))
1601 changed |= op1 != oldop1;
1602 if (op2 && TREE_CODE (op2) == SSA_NAME)
1604 unsigned int op_val_id = VN_INFO (op2)->value_id;
1605 leader = find_leader_in_sets (op_val_id, set1, set2);
1606 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1607 if (opresult && opresult != leader)
1609 tree name = get_representative_for (opresult);
1617 /* We can't possibly insert these. */
1618 else if (op2 && !is_gimple_min_invariant (op2))
1620 changed |= op2 != oldop2;
1623 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1624 /* We may have changed from an SSA_NAME to a constant */
1625 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1626 newop.opcode = TREE_CODE (op0);
1631 /* If it transforms a non-constant ARRAY_REF into a constant
1632 one, adjust the constant offset. */
1633 if (newop.opcode == ARRAY_REF
1635 && TREE_CODE (op0) == INTEGER_CST
1636 && TREE_CODE (op1) == INTEGER_CST
1637 && TREE_CODE (op2) == INTEGER_CST)
1639 double_int off = tree_to_double_int (op0);
1640 off = double_int_add (off,
1642 (tree_to_double_int (op1)));
1643 off = double_int_mul (off, tree_to_double_int (op2));
1644 if (double_int_fits_in_shwi_p (off))
1645 newop.off = off.low;
1647 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1648 /* If it transforms from an SSA_NAME to an address, fold with
1649 a preceding indirect reference. */
1650 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1651 && VEC_index (vn_reference_op_s,
1652 newoperands, j - 1)->opcode == MEM_REF)
1653 vn_reference_fold_indirect (&newoperands, &j);
1655 if (i != VEC_length (vn_reference_op_s, operands))
1658 VEC_free (vn_reference_op_s, heap, newoperands);
1664 newvuse = translate_vuse_through_block (newoperands,
1665 ref->set, ref->type,
1666 vuse, phiblock, pred,
1668 if (newvuse == NULL_TREE)
1670 VEC_free (vn_reference_op_s, heap, newoperands);
1675 if (changed || newvuse != vuse)
1677 unsigned int new_val_id;
1679 bool converted = false;
1681 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1686 VEC_free (vn_reference_op_s, heap, newoperands);
1689 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1691 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1695 if (result && is_gimple_min_invariant (result))
1697 gcc_assert (!newoperands);
1698 return get_or_alloc_expr_for_constant (result);
1701 expr = (pre_expr) pool_alloc (pre_expr_pool);
1702 expr->kind = REFERENCE;
1710 gcc_assert (CONVERT_EXPR_P (result)
1711 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1713 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1715 TREE_OPERAND (result, 0),
1716 NULL_TREE, NULL_TREE,
1719 if (nresult && is_gimple_min_invariant (nresult))
1720 return get_or_alloc_expr_for_constant (nresult);
1725 PRE_EXPR_NARY (expr) = nary;
1726 constant = fully_constant_expression (expr);
1727 if (constant != expr)
1730 new_val_id = nary->value_id;
1731 get_or_alloc_expression_id (expr);
1735 new_val_id = get_next_value_id ();
1736 VEC_safe_grow_cleared (bitmap_set_t, heap,
1738 get_max_value_id() + 1);
1739 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1741 TREE_OPERAND (result, 0),
1742 NULL_TREE, NULL_TREE,
1743 NULL_TREE, NULL_TREE,
1745 PRE_EXPR_NARY (expr) = nary;
1746 constant = fully_constant_expression (expr);
1747 if (constant != expr)
1749 get_or_alloc_expression_id (expr);
1754 PRE_EXPR_REFERENCE (expr) = newref;
1755 constant = fully_constant_expression (expr);
1756 if (constant != expr)
1759 new_val_id = newref->value_id;
1760 get_or_alloc_expression_id (expr);
1764 if (changed || !same_valid)
1766 new_val_id = get_next_value_id ();
1767 VEC_safe_grow_cleared (bitmap_set_t, heap,
1769 get_max_value_id() + 1);
1772 new_val_id = ref->value_id;
1773 newref = vn_reference_insert_pieces (newvuse, ref->set,
1776 result, new_val_id);
1778 PRE_EXPR_REFERENCE (expr) = newref;
1779 constant = fully_constant_expression (expr);
1780 if (constant != expr)
1782 get_or_alloc_expression_id (expr);
1784 add_to_value (new_val_id, expr);
1786 VEC_free (vn_reference_op_s, heap, newoperands);
1796 tree name = PRE_EXPR_NAME (expr);
1798 def_stmt = SSA_NAME_DEF_STMT (name);
1799 if (gimple_code (def_stmt) == GIMPLE_PHI
1800 && gimple_bb (def_stmt) == phiblock)
1805 e = find_edge (pred, gimple_bb (phi));
1808 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1811 if (TREE_CODE (def) == SSA_NAME)
1812 def = VN_INFO (def)->valnum;
1814 /* Handle constant. */
1815 if (is_gimple_min_invariant (def))
1816 return get_or_alloc_expr_for_constant (def);
1818 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1821 newexpr = get_or_alloc_expr_for_name (def);
1832 /* Wrapper around phi_translate_1 providing caching functionality. */
1835 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1836 basic_block pred, basic_block phiblock)
1843 /* Constants contain no values that need translation. */
1844 if (expr->kind == CONSTANT)
1847 if (value_id_constant_p (get_expr_value_id (expr)))
1850 if (expr->kind != NAME)
1852 phitrans = phi_trans_lookup (expr, pred);
1858 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1860 /* Don't add empty translations to the cache. Neither add
1861 translations of NAMEs as those are cheap to translate. */
1863 && expr->kind != NAME)
1864 phi_trans_add (expr, phitrans, pred);
1870 /* For each expression in SET, translate the values through phi nodes
1871 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1872 expressions in DEST. */
1875 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1876 basic_block phiblock)
1878 VEC (pre_expr, heap) *exprs;
1882 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1884 bitmap_set_copy (dest, set);
1888 exprs = sorted_array_from_bitmap_set (set);
1889 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1891 pre_expr translated;
1892 translated = phi_translate (expr, set, NULL, pred, phiblock);
1896 /* We might end up with multiple expressions from SET being
1897 translated to the same value. In this case we do not want
1898 to retain the NARY or REFERENCE expression but prefer a NAME
1899 which would be the leader. */
1900 if (translated->kind == NAME)
1901 bitmap_value_replace_in_set (dest, translated);
1903 bitmap_value_insert_into_set (dest, translated);
1905 VEC_free (pre_expr, heap, exprs);
1908 /* Find the leader for a value (i.e., the name representing that
1909 value) in a given set, and return it. If STMT is non-NULL it
1910 makes sure the defining statement for the leader dominates it.
1911 Return NULL if no leader is found. */
1914 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1916 if (value_id_constant_p (val))
1920 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1922 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1924 pre_expr expr = expression_for_id (i);
1925 if (expr->kind == CONSTANT)
1929 if (bitmap_set_contains_value (set, val))
1931 /* Rather than walk the entire bitmap of expressions, and see
1932 whether any of them has the value we are looking for, we look
1933 at the reverse mapping, which tells us the set of expressions
1934 that have a given value (IE value->expressions with that
1935 value) and see if any of those expressions are in our set.
1936 The number of expressions per value is usually significantly
1937 less than the number of expressions in the set. In fact, for
1938 large testcases, doing it this way is roughly 5-10x faster
1939 than walking the bitmap.
1940 If this is somehow a significant lose for some cases, we can
1941 choose which set to walk based on which set is smaller. */
1944 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1946 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1947 &set->expressions, 0, i, bi)
1949 pre_expr val = expression_for_id (i);
1950 /* At the point where stmt is not null, there should always
1951 be an SSA_NAME first in the list of expressions. */
1954 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1955 if (gimple_code (def_stmt) != GIMPLE_PHI
1956 && gimple_bb (def_stmt) == gimple_bb (stmt)
1957 /* PRE insertions are at the end of the basic-block
1959 && (gimple_uid (def_stmt) == 0
1960 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1969 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1970 BLOCK by seeing if it is not killed in the block. Note that we are
1971 only determining whether there is a store that kills it. Because
1972 of the order in which clean iterates over values, we are guaranteed
1973 that altered operands will have caused us to be eliminated from the
1974 ANTIC_IN set already. */
1977 value_dies_in_block_x (pre_expr expr, basic_block block)
1979 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1980 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1982 gimple_stmt_iterator gsi;
1983 unsigned id = get_expression_id (expr);
1990 /* Lookup a previously calculated result. */
1991 if (EXPR_DIES (block)
1992 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1993 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1995 /* A memory expression {e, VUSE} dies in the block if there is a
1996 statement that may clobber e. If, starting statement walk from the
1997 top of the basic block, a statement uses VUSE there can be no kill
1998 inbetween that use and the original statement that loaded {e, VUSE},
1999 so we can stop walking. */
2000 ref.base = NULL_TREE;
2001 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
2003 tree def_vuse, def_vdef;
2004 def = gsi_stmt (gsi);
2005 def_vuse = gimple_vuse (def);
2006 def_vdef = gimple_vdef (def);
2008 /* Not a memory statement. */
2012 /* Not a may-def. */
2015 /* A load with the same VUSE, we're done. */
2016 if (def_vuse == vuse)
2022 /* Init ref only if we really need it. */
2023 if (ref.base == NULL_TREE
2024 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
2030 /* If the statement may clobber expr, it dies. */
2031 if (stmt_may_clobber_ref_p_1 (def, &ref))
2038 /* Remember the result. */
2039 if (!EXPR_DIES (block))
2040 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2041 bitmap_set_bit (EXPR_DIES (block), id * 2);
2043 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2049 #define union_contains_value(SET1, SET2, VAL) \
2050 (bitmap_set_contains_value ((SET1), (VAL)) \
2051 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
2053 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2056 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2057 vn_reference_op_t vro)
2059 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2061 struct pre_expr_d temp;
2064 PRE_EXPR_NAME (&temp) = vro->op0;
2065 temp.id = lookup_expression_id (&temp);
2068 if (!union_contains_value (set1, set2,
2069 get_expr_value_id (&temp)))
2072 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2074 struct pre_expr_d temp;
2077 PRE_EXPR_NAME (&temp) = vro->op1;
2078 temp.id = lookup_expression_id (&temp);
2081 if (!union_contains_value (set1, set2,
2082 get_expr_value_id (&temp)))
2086 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2088 struct pre_expr_d temp;
2091 PRE_EXPR_NAME (&temp) = vro->op2;
2092 temp.id = lookup_expression_id (&temp);
2095 if (!union_contains_value (set1, set2,
2096 get_expr_value_id (&temp)))
2103 /* Determine if the expression EXPR is valid in SET1 U SET2.
2104 ONLY SET2 CAN BE NULL.
2105 This means that we have a leader for each part of the expression
2106 (if it consists of values), or the expression is an SSA_NAME.
2107 For loads/calls, we also see if the vuse is killed in this block. */
2110 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2116 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2120 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2121 for (i = 0; i < nary->length; i++)
2123 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2125 struct pre_expr_d temp;
2128 PRE_EXPR_NAME (&temp) = nary->op[i];
2129 temp.id = lookup_expression_id (&temp);
2132 if (!union_contains_value (set1, set2,
2133 get_expr_value_id (&temp)))
2137 /* If the NARY may trap make sure the block does not contain
2138 a possible exit point.
2139 ??? This is overly conservative if we translate AVAIL_OUT
2140 as the available expression might be after the exit point. */
2141 if (BB_MAY_NOTRETURN (block)
2142 && vn_nary_may_trap (nary))
2149 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2150 vn_reference_op_t vro;
2153 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2155 if (!vro_valid_in_sets (set1, set2, vro))
2160 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2161 if (!gimple_nop_p (def_stmt)
2162 && gimple_bb (def_stmt) != block
2163 && !dominated_by_p (CDI_DOMINATORS,
2164 block, gimple_bb (def_stmt)))
2167 return !value_dies_in_block_x (expr, block);
2174 /* Clean the set of expressions that are no longer valid in SET1 or
2175 SET2. This means expressions that are made up of values we have no
2176 leaders for in SET1 or SET2. This version is used for partial
2177 anticipation, which means it is not valid in either ANTIC_IN or
2181 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2183 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2187 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2189 if (!valid_in_sets (set1, set2, expr, block))
2190 bitmap_remove_from_set (set1, expr);
2192 VEC_free (pre_expr, heap, exprs);
2195 /* Clean the set of expressions that are no longer valid in SET. This
2196 means expressions that are made up of values we have no leaders for
2200 clean (bitmap_set_t set, basic_block block)
2202 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2206 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2208 if (!valid_in_sets (set, NULL, expr, block))
2209 bitmap_remove_from_set (set, expr);
2211 VEC_free (pre_expr, heap, exprs);
2214 static sbitmap has_abnormal_preds;
2216 /* List of blocks that may have changed during ANTIC computation and
2217 thus need to be iterated over. */
2219 static sbitmap changed_blocks;
2221 /* Decide whether to defer a block for a later iteration, or PHI
2222 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2223 should defer the block, and true if we processed it. */
2226 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2227 basic_block block, basic_block phiblock)
2229 if (!BB_VISITED (phiblock))
2231 SET_BIT (changed_blocks, block->index);
2232 BB_VISITED (block) = 0;
2233 BB_DEFERRED (block) = 1;
2237 phi_translate_set (dest, source, block, phiblock);
2241 /* Compute the ANTIC set for BLOCK.
2243 If succs(BLOCK) > 1 then
2244 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2245 else if succs(BLOCK) == 1 then
2246 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2248 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2252 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2254 bool changed = false;
2255 bitmap_set_t S, old, ANTIC_OUT;
2261 old = ANTIC_OUT = S = NULL;
2262 BB_VISITED (block) = 1;
2264 /* If any edges from predecessors are abnormal, antic_in is empty,
2266 if (block_has_abnormal_pred_edge)
2267 goto maybe_dump_sets;
2269 old = ANTIC_IN (block);
2270 ANTIC_OUT = bitmap_set_new ();
2272 /* If the block has no successors, ANTIC_OUT is empty. */
2273 if (EDGE_COUNT (block->succs) == 0)
2275 /* If we have one successor, we could have some phi nodes to
2276 translate through. */
2277 else if (single_succ_p (block))
2279 basic_block succ_bb = single_succ (block);
2281 /* We trade iterations of the dataflow equations for having to
2282 phi translate the maximal set, which is incredibly slow
2283 (since the maximal set often has 300+ members, even when you
2284 have a small number of blocks).
2285 Basically, we defer the computation of ANTIC for this block
2286 until we have processed it's successor, which will inevitably
2287 have a *much* smaller set of values to phi translate once
2288 clean has been run on it.
2289 The cost of doing this is that we technically perform more
2290 iterations, however, they are lower cost iterations.
2292 Timings for PRE on tramp3d-v4:
2293 without maximal set fix: 11 seconds
2294 with maximal set fix/without deferring: 26 seconds
2295 with maximal set fix/with deferring: 11 seconds
2298 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2302 goto maybe_dump_sets;
2305 /* If we have multiple successors, we take the intersection of all of
2306 them. Note that in the case of loop exit phi nodes, we may have
2307 phis to translate through. */
2310 VEC(basic_block, heap) * worklist;
2312 basic_block bprime, first = NULL;
2314 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2315 FOR_EACH_EDGE (e, ei, block->succs)
2318 && BB_VISITED (e->dest))
2320 else if (BB_VISITED (e->dest))
2321 VEC_quick_push (basic_block, worklist, e->dest);
2324 /* Of multiple successors we have to have visited one already. */
2327 SET_BIT (changed_blocks, block->index);
2328 BB_VISITED (block) = 0;
2329 BB_DEFERRED (block) = 1;
2331 VEC_free (basic_block, heap, worklist);
2332 goto maybe_dump_sets;
2335 if (!gimple_seq_empty_p (phi_nodes (first)))
2336 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2338 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2340 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2342 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2344 bitmap_set_t tmp = bitmap_set_new ();
2345 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2346 bitmap_set_and (ANTIC_OUT, tmp);
2347 bitmap_set_free (tmp);
2350 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2352 VEC_free (basic_block, heap, worklist);
2355 /* Generate ANTIC_OUT - TMP_GEN. */
2356 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2358 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2359 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2362 /* Then union in the ANTIC_OUT - TMP_GEN values,
2363 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2364 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2365 bitmap_value_insert_into_set (ANTIC_IN (block),
2366 expression_for_id (bii));
2368 clean (ANTIC_IN (block), block);
2370 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2373 SET_BIT (changed_blocks, block->index);
2374 FOR_EACH_EDGE (e, ei, block->preds)
2375 SET_BIT (changed_blocks, e->src->index);
2378 RESET_BIT (changed_blocks, block->index);
2381 if (dump_file && (dump_flags & TDF_DETAILS))
2383 if (!BB_DEFERRED (block) || BB_VISITED (block))
2386 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2388 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2392 print_bitmap_set (dump_file, S, "S", block->index);
2397 "Block %d was deferred for a future iteration.\n",
2402 bitmap_set_free (old);
2404 bitmap_set_free (S);
2406 bitmap_set_free (ANTIC_OUT);
2410 /* Compute PARTIAL_ANTIC for BLOCK.
2412 If succs(BLOCK) > 1 then
2413 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2414 in ANTIC_OUT for all succ(BLOCK)
2415 else if succs(BLOCK) == 1 then
2416 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2418 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2423 compute_partial_antic_aux (basic_block block,
2424 bool block_has_abnormal_pred_edge)
2426 bool changed = false;
2427 bitmap_set_t old_PA_IN;
2428 bitmap_set_t PA_OUT;
2431 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2433 old_PA_IN = PA_OUT = NULL;
2435 /* If any edges from predecessors are abnormal, antic_in is empty,
2437 if (block_has_abnormal_pred_edge)
2438 goto maybe_dump_sets;
2440 /* If there are too many partially anticipatable values in the
2441 block, phi_translate_set can take an exponential time: stop
2442 before the translation starts. */
2444 && single_succ_p (block)
2445 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2446 goto maybe_dump_sets;
2448 old_PA_IN = PA_IN (block);
2449 PA_OUT = bitmap_set_new ();
2451 /* If the block has no successors, ANTIC_OUT is empty. */
2452 if (EDGE_COUNT (block->succs) == 0)
2454 /* If we have one successor, we could have some phi nodes to
2455 translate through. Note that we can't phi translate across DFS
2456 back edges in partial antic, because it uses a union operation on
2457 the successors. For recurrences like IV's, we will end up
2458 generating a new value in the set on each go around (i + 3 (VH.1)
2459 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2460 else if (single_succ_p (block))
2462 basic_block succ = single_succ (block);
2463 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2464 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2466 /* If we have multiple successors, we take the union of all of
2470 VEC(basic_block, heap) * worklist;
2474 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2475 FOR_EACH_EDGE (e, ei, block->succs)
2477 if (e->flags & EDGE_DFS_BACK)
2479 VEC_quick_push (basic_block, worklist, e->dest);
2481 if (VEC_length (basic_block, worklist) > 0)
2483 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2488 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2489 bitmap_value_insert_into_set (PA_OUT,
2490 expression_for_id (i));
2491 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2493 bitmap_set_t pa_in = bitmap_set_new ();
2494 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2495 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2496 bitmap_value_insert_into_set (PA_OUT,
2497 expression_for_id (i));
2498 bitmap_set_free (pa_in);
2501 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2502 bitmap_value_insert_into_set (PA_OUT,
2503 expression_for_id (i));
2506 VEC_free (basic_block, heap, worklist);
2509 /* PA_IN starts with PA_OUT - TMP_GEN.
2510 Then we subtract things from ANTIC_IN. */
2511 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2513 /* For partial antic, we want to put back in the phi results, since
2514 we will properly avoid making them partially antic over backedges. */
2515 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2516 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2518 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2519 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2521 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2523 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2526 SET_BIT (changed_blocks, block->index);
2527 FOR_EACH_EDGE (e, ei, block->preds)
2528 SET_BIT (changed_blocks, e->src->index);
2531 RESET_BIT (changed_blocks, block->index);
2534 if (dump_file && (dump_flags & TDF_DETAILS))
2537 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2539 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2542 bitmap_set_free (old_PA_IN);
2544 bitmap_set_free (PA_OUT);
2548 /* Compute ANTIC and partial ANTIC sets. */
2551 compute_antic (void)
2553 bool changed = true;
2554 int num_iterations = 0;
2558 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2559 We pre-build the map of blocks with incoming abnormal edges here. */
2560 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2561 sbitmap_zero (has_abnormal_preds);
2568 FOR_EACH_EDGE (e, ei, block->preds)
2570 e->flags &= ~EDGE_DFS_BACK;
2571 if (e->flags & EDGE_ABNORMAL)
2573 SET_BIT (has_abnormal_preds, block->index);
2578 BB_VISITED (block) = 0;
2579 BB_DEFERRED (block) = 0;
2581 /* While we are here, give empty ANTIC_IN sets to each block. */
2582 ANTIC_IN (block) = bitmap_set_new ();
2583 PA_IN (block) = bitmap_set_new ();
2586 /* At the exit block we anticipate nothing. */
2587 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2588 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2589 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2591 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2592 sbitmap_ones (changed_blocks);
2595 if (dump_file && (dump_flags & TDF_DETAILS))
2596 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2599 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2601 if (TEST_BIT (changed_blocks, postorder[i]))
2603 basic_block block = BASIC_BLOCK (postorder[i]);
2604 changed |= compute_antic_aux (block,
2605 TEST_BIT (has_abnormal_preds,
2609 #ifdef ENABLE_CHECKING
2610 /* Theoretically possible, but *highly* unlikely. */
2611 gcc_assert (num_iterations < 500);
2615 statistics_histogram_event (cfun, "compute_antic iterations",
2618 if (do_partial_partial)
2620 sbitmap_ones (changed_blocks);
2621 mark_dfs_back_edges ();
2626 if (dump_file && (dump_flags & TDF_DETAILS))
2627 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2630 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2632 if (TEST_BIT (changed_blocks, postorder[i]))
2634 basic_block block = BASIC_BLOCK (postorder[i]);
2636 |= compute_partial_antic_aux (block,
2637 TEST_BIT (has_abnormal_preds,
2641 #ifdef ENABLE_CHECKING
2642 /* Theoretically possible, but *highly* unlikely. */
2643 gcc_assert (num_iterations < 500);
2646 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2649 sbitmap_free (has_abnormal_preds);
2650 sbitmap_free (changed_blocks);
2653 /* Return true if we can value number the call in STMT. This is true
2654 if we have a pure or constant call. */
2657 can_value_number_call (gimple stmt)
2659 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2664 /* Return true if OP is a tree which we can perform PRE on.
2665 This may not match the operations we can value number, but in
2666 a perfect world would. */
2669 can_PRE_operation (tree op)
2671 return UNARY_CLASS_P (op)
2672 || BINARY_CLASS_P (op)
2673 || COMPARISON_CLASS_P (op)
2674 || TREE_CODE (op) == MEM_REF
2675 || TREE_CODE (op) == COMPONENT_REF
2676 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2677 || TREE_CODE (op) == CALL_EXPR
2678 || TREE_CODE (op) == ARRAY_REF;
2682 /* Inserted expressions are placed onto this worklist, which is used
2683 for performing quick dead code elimination of insertions we made
2684 that didn't turn out to be necessary. */
2685 static bitmap inserted_exprs;
2687 /* Pool allocated fake store expressions are placed onto this
2688 worklist, which, after performing dead code elimination, is walked
2689 to see which expressions need to be put into GC'able memory */
2690 static VEC(gimple, heap) *need_creation;
2692 /* The actual worker for create_component_ref_by_pieces. */
2695 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2696 unsigned int *operand, gimple_seq *stmts,
2699 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2703 switch (currop->opcode)
2707 tree folded, sc = NULL_TREE;
2708 unsigned int nargs = 0;
2710 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2714 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2715 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2721 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2722 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2726 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2727 ref->operands) - 1);
2728 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2730 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2740 folded = build_call_array (currop->type,
2741 (TREE_CODE (fn) == FUNCTION_DECL
2742 ? build_fold_addr_expr (fn) : fn),
2746 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2752 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2754 tree offset = currop->op0;
2757 if (TREE_CODE (baseop) == ADDR_EXPR
2758 && handled_component_p (TREE_OPERAND (baseop, 0)))
2762 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2765 offset = int_const_binop (PLUS_EXPR, offset,
2766 build_int_cst (TREE_TYPE (offset),
2768 baseop = build_fold_addr_expr (base);
2770 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2773 case TARGET_MEM_REF:
2776 tree genop0 = NULL_TREE;
2777 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2783 op0expr = get_or_alloc_expr_for (currop->op0);
2784 genop0 = find_or_generate_expression (block, op0expr,
2789 if (DECL_P (baseop))
2790 return build5 (TARGET_MEM_REF, currop->type,
2792 genop0, currop->op1, currop->op2);
2794 return build5 (TARGET_MEM_REF, currop->type,
2796 genop0, currop->op1, currop->op2);
2802 gcc_assert (is_gimple_min_invariant (currop->op0));
2808 case VIEW_CONVERT_EXPR:
2811 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2816 folded = fold_build1 (currop->opcode, currop->type,
2821 case MISALIGNED_INDIRECT_REF:
2824 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2829 genop1 = fold_convert (build_pointer_type (currop->type),
2832 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2833 folded = fold_build2 (currop->opcode, currop->type,
2834 genop1, currop->op1);
2836 folded = fold_build1 (currop->opcode, currop->type,
2844 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2846 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2847 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2853 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2856 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2859 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2864 /* For array ref vn_reference_op's, operand 1 of the array ref
2865 is op0 of the reference op and operand 3 of the array ref is
2867 case ARRAY_RANGE_REF:
2871 tree genop1 = currop->op0;
2873 tree genop2 = currop->op1;
2875 tree genop3 = currop->op2;
2877 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2881 op1expr = get_or_alloc_expr_for (genop1);
2882 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2887 /* Drop zero minimum index. */
2888 if (tree_int_cst_equal (genop2, integer_zero_node))
2892 op2expr = get_or_alloc_expr_for (genop2);
2893 genop2 = find_or_generate_expression (block, op2expr, stmts,
2901 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2902 /* We can't always put a size in units of the element alignment
2903 here as the element alignment may be not visible. See
2904 PR43783. Simply drop the element size for constant
2906 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2910 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2911 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2912 op3expr = get_or_alloc_expr_for (genop3);
2913 genop3 = find_or_generate_expression (block, op3expr, stmts,
2919 return build4 (currop->opcode, currop->type, genop0, genop1,
2926 tree genop2 = currop->op1;
2928 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2932 /* op1 should be a FIELD_DECL, which are represented by
2937 op2expr = get_or_alloc_expr_for (genop2);
2938 genop2 = find_or_generate_expression (block, op2expr, stmts,
2944 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2950 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2951 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2972 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2973 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2974 trying to rename aggregates into ssa form directly, which is a no no.
2976 Thus, this routine doesn't create temporaries, it just builds a
2977 single access expression for the array, calling
2978 find_or_generate_expression to build the innermost pieces.
2980 This function is a subroutine of create_expression_by_pieces, and
2981 should not be called on it's own unless you really know what you
2985 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2986 gimple_seq *stmts, gimple domstmt)
2988 unsigned int op = 0;
2989 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2992 /* Find a leader for an expression, or generate one using
2993 create_expression_by_pieces if it's ANTIC but
2995 BLOCK is the basic_block we are looking for leaders in.
2996 EXPR is the expression to find a leader or generate for.
2997 STMTS is the statement list to put the inserted expressions on.
2998 Returns the SSA_NAME of the LHS of the generated expression or the
3000 DOMSTMT if non-NULL is a statement that should be dominated by
3001 all uses in the generated expression. If DOMSTMT is non-NULL this
3002 routine can fail and return NULL_TREE. Otherwise it will assert
3006 find_or_generate_expression (basic_block block, pre_expr expr,
3007 gimple_seq *stmts, gimple domstmt)
3009 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
3010 get_expr_value_id (expr), domstmt);
3014 if (leader->kind == NAME)
3015 genop = PRE_EXPR_NAME (leader);
3016 else if (leader->kind == CONSTANT)
3017 genop = PRE_EXPR_CONSTANT (leader);
3020 /* If it's still NULL, it must be a complex expression, so generate
3021 it recursively. Not so if inserting expressions for values generated
3026 bitmap_set_t exprset;
3027 unsigned int lookfor = get_expr_value_id (expr);
3028 bool handled = false;
3032 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3033 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3035 pre_expr temp = expression_for_id (i);
3036 if (temp->kind != NAME)
3039 genop = create_expression_by_pieces (block, temp, stmts,
3041 get_expr_type (expr));
3045 if (!handled && domstmt)
3048 gcc_assert (handled);
3053 #define NECESSARY GF_PLF_1
3055 /* Create an expression in pieces, so that we can handle very complex
3056 expressions that may be ANTIC, but not necessary GIMPLE.
3057 BLOCK is the basic block the expression will be inserted into,
3058 EXPR is the expression to insert (in value form)
3059 STMTS is a statement list to append the necessary insertions into.
3061 This function will die if we hit some value that shouldn't be
3062 ANTIC but is (IE there is no leader for it, or its components).
3063 This function may also generate expressions that are themselves
3064 partially or fully redundant. Those that are will be either made
3065 fully redundant during the next iteration of insert (for partially
3066 redundant ones), or eliminated by eliminate (for fully redundant
3069 If DOMSTMT is non-NULL then we make sure that all uses in the
3070 expressions dominate that statement. In this case the function
3071 can return NULL_TREE to signal failure. */
3074 create_expression_by_pieces (basic_block block, pre_expr expr,
3075 gimple_seq *stmts, gimple domstmt, tree type)
3079 gimple_seq forced_stmts = NULL;
3080 unsigned int value_id;
3081 gimple_stmt_iterator gsi;
3082 tree exprtype = type ? type : get_expr_type (expr);
3088 /* We may hit the NAME/CONSTANT case if we have to convert types
3089 that value numbering saw through. */
3091 folded = PRE_EXPR_NAME (expr);
3094 folded = PRE_EXPR_CONSTANT (expr);
3098 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3099 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3104 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3105 switch (nary->length)
3109 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3110 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
3111 tree genop1 = find_or_generate_expression (block, op1,
3113 tree genop2 = find_or_generate_expression (block, op2,
3115 if (!genop1 || !genop2)
3117 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
3118 may be a constant with the wrong type. */
3119 if (nary->opcode == POINTER_PLUS_EXPR)
3121 genop1 = fold_convert (nary->type, genop1);
3122 genop2 = fold_convert (sizetype, genop2);
3126 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3127 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
3130 folded = fold_build2 (nary->opcode, nary->type,
3136 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3137 tree genop1 = find_or_generate_expression (block, op1,
3141 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3143 folded = fold_build1 (nary->opcode, nary->type,
3156 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3157 folded = fold_convert (exprtype, folded);
3159 /* Force the generated expression to be a sequence of GIMPLE
3161 We have to call unshare_expr because force_gimple_operand may
3162 modify the tree we pass to it. */
3163 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3166 /* If we have any intermediate expressions to the value sets, add them
3167 to the value sets and chain them in the instruction stream. */
3170 gsi = gsi_start (forced_stmts);
3171 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3173 gimple stmt = gsi_stmt (gsi);
3174 tree forcedname = gimple_get_lhs (stmt);
3177 if (TREE_CODE (forcedname) == SSA_NAME)
3179 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3180 VN_INFO_GET (forcedname)->valnum = forcedname;
3181 VN_INFO (forcedname)->value_id = get_next_value_id ();
3182 nameexpr = get_or_alloc_expr_for_name (forcedname);
3183 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3185 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3186 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3188 mark_symbols_for_renaming (stmt);
3190 gimple_seq_add_seq (stmts, forced_stmts);
3193 /* Build and insert the assignment of the end result to the temporary
3194 that we will return. */
3195 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3197 pretemp = create_tmp_reg (exprtype, "pretmp");
3198 get_var_ann (pretemp);
3202 add_referenced_var (temp);
3204 newstmt = gimple_build_assign (temp, folded);
3205 name = make_ssa_name (temp, newstmt);
3206 gimple_assign_set_lhs (newstmt, name);
3207 gimple_set_plf (newstmt, NECESSARY, false);
3209 gimple_seq_add_stmt (stmts, newstmt);
3210 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3212 /* All the symbols in NEWEXPR should be put into SSA form. */
3213 mark_symbols_for_renaming (newstmt);
3215 /* Add a value number to the temporary.
3216 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3217 we are creating the expression by pieces, and this particular piece of
3218 the expression may have been represented. There is no harm in replacing
3220 VN_INFO_GET (name)->valnum = name;
3221 value_id = get_expr_value_id (expr);
3222 VN_INFO (name)->value_id = value_id;
3223 nameexpr = get_or_alloc_expr_for_name (name);
3224 add_to_value (value_id, nameexpr);
3225 if (NEW_SETS (block))
3226 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3227 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3229 pre_stats.insertions++;
3230 if (dump_file && (dump_flags & TDF_DETAILS))
3232 fprintf (dump_file, "Inserted ");
3233 print_gimple_stmt (dump_file, newstmt, 0, 0);
3234 fprintf (dump_file, " in predecessor %d\n", block->index);
3241 /* Returns true if we want to inhibit the insertions of PHI nodes
3242 for the given EXPR for basic block BB (a member of a loop).
3243 We want to do this, when we fear that the induction variable we
3244 create might inhibit vectorization. */
3247 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3249 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3250 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3251 vn_reference_op_t op;
3254 /* If we aren't going to vectorize we don't inhibit anything. */
3255 if (!flag_tree_vectorize)
3258 /* Otherwise we inhibit the insertion when the address of the
3259 memory reference is a simple induction variable. In other
3260 cases the vectorizer won't do anything anyway (either it's
3261 loop invariant or a complicated expression). */
3262 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3267 case ARRAY_RANGE_REF:
3268 if (TREE_CODE (op->op0) != SSA_NAME)
3273 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3275 /* Default defs are loop invariant. */
3278 /* Defined outside this loop, also loop invariant. */
3279 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3281 /* If it's a simple induction variable inhibit insertion,
3282 the vectorizer might be interested in this one. */
3283 if (simple_iv (bb->loop_father, bb->loop_father,
3284 op->op0, &iv, true))
3286 /* No simple IV, vectorizer can't do anything, hence no
3287 reason to inhibit the transformation for this operand. */
3297 /* Insert the to-be-made-available values of expression EXPRNUM for each
3298 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3299 merge the result with a phi node, given the same value number as
3300 NODE. Return true if we have inserted new stuff. */
3303 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3306 pre_expr expr = expression_for_id (exprnum);
3308 unsigned int val = get_expr_value_id (expr);
3310 bool insertions = false;
3315 tree type = get_expr_type (expr);
3319 if (dump_file && (dump_flags & TDF_DETAILS))
3321 fprintf (dump_file, "Found partial redundancy for expression ");
3322 print_pre_expr (dump_file, expr);
3323 fprintf (dump_file, " (%04d)\n", val);
3326 /* Make sure we aren't creating an induction variable. */
3327 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3329 bool firstinsideloop = false;
3330 bool secondinsideloop = false;
3331 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3332 EDGE_PRED (block, 0)->src);
3333 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3334 EDGE_PRED (block, 1)->src);
3335 /* Induction variables only have one edge inside the loop. */
3336 if ((firstinsideloop ^ secondinsideloop)
3337 && (expr->kind != REFERENCE
3338 || inhibit_phi_insertion (block, expr)))
3340 if (dump_file && (dump_flags & TDF_DETAILS))
3341 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3346 /* Make the necessary insertions. */
3347 FOR_EACH_EDGE (pred, ei, block->preds)
3349 gimple_seq stmts = NULL;
3352 eprime = avail[bprime->index];
3354 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3356 builtexpr = create_expression_by_pieces (bprime,
3360 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3361 gsi_insert_seq_on_edge (pred, stmts);
3362 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3365 else if (eprime->kind == CONSTANT)
3367 /* Constants may not have the right type, fold_convert
3368 should give us back a constant with the right type.
3370 tree constant = PRE_EXPR_CONSTANT (eprime);
3371 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3373 tree builtexpr = fold_convert (type, constant);
3374 if (!is_gimple_min_invariant (builtexpr))
3376 tree forcedexpr = force_gimple_operand (builtexpr,
3379 if (!is_gimple_min_invariant (forcedexpr))
3381 if (forcedexpr != builtexpr)
3383 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3384 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3388 gimple_stmt_iterator gsi;
3389 gsi = gsi_start (stmts);
3390 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3392 gimple stmt = gsi_stmt (gsi);
3393 tree lhs = gimple_get_lhs (stmt);
3394 if (TREE_CODE (lhs) == SSA_NAME)
3395 bitmap_set_bit (inserted_exprs,
3396 SSA_NAME_VERSION (lhs));
3397 gimple_set_plf (stmt, NECESSARY, false);
3399 gsi_insert_seq_on_edge (pred, stmts);
3401 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3405 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3408 else if (eprime->kind == NAME)
3410 /* We may have to do a conversion because our value
3411 numbering can look through types in certain cases, but
3412 our IL requires all operands of a phi node have the same
3414 tree name = PRE_EXPR_NAME (eprime);
3415 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3419 builtexpr = fold_convert (type, name);
3420 forcedexpr = force_gimple_operand (builtexpr,
3424 if (forcedexpr != name)
3426 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3427 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3432 gimple_stmt_iterator gsi;
3433 gsi = gsi_start (stmts);
3434 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3436 gimple stmt = gsi_stmt (gsi);
3437 tree lhs = gimple_get_lhs (stmt);
3438 if (TREE_CODE (lhs) == SSA_NAME)
3439 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3440 gimple_set_plf (stmt, NECESSARY, false);
3442 gsi_insert_seq_on_edge (pred, stmts);
3444 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3448 /* If we didn't want a phi node, and we made insertions, we still have
3449 inserted new stuff, and thus return true. If we didn't want a phi node,
3450 and didn't make insertions, we haven't added anything new, so return
3452 if (nophi && insertions)
3454 else if (nophi && !insertions)
3457 /* Now build a phi for the new variable. */
3458 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3460 prephitemp = create_tmp_var (type, "prephitmp");
3461 get_var_ann (prephitemp);
3465 add_referenced_var (temp);
3467 if (TREE_CODE (type) == COMPLEX_TYPE
3468 || TREE_CODE (type) == VECTOR_TYPE)
3469 DECL_GIMPLE_REG_P (temp) = 1;
3470 phi = create_phi_node (temp, block);
3472 gimple_set_plf (phi, NECESSARY, false);
3473 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3474 VN_INFO (gimple_phi_result (phi))->value_id = val;
3475 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3476 FOR_EACH_EDGE (pred, ei, block->preds)
3478 pre_expr ae = avail[pred->src->index];
3479 gcc_assert (get_expr_type (ae) == type
3480 || useless_type_conversion_p (type, get_expr_type (ae)));
3481 if (ae->kind == CONSTANT)
3482 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3484 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3488 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3489 add_to_value (val, newphi);
3491 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3492 this insertion, since we test for the existence of this value in PHI_GEN
3493 before proceeding with the partial redundancy checks in insert_aux.
3495 The value may exist in AVAIL_OUT, in particular, it could be represented
3496 by the expression we are trying to eliminate, in which case we want the
3497 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3500 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3501 this block, because if it did, it would have existed in our dominator's
3502 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3505 bitmap_insert_into_set (PHI_GEN (block), newphi);
3506 bitmap_value_replace_in_set (AVAIL_OUT (block),
3508 bitmap_insert_into_set (NEW_SETS (block),
3511 if (dump_file && (dump_flags & TDF_DETAILS))
3513 fprintf (dump_file, "Created phi ");
3514 print_gimple_stmt (dump_file, phi, 0, 0);
3515 fprintf (dump_file, " in block %d\n", block->index);
3523 /* Perform insertion of partially redundant values.
3524 For BLOCK, do the following:
3525 1. Propagate the NEW_SETS of the dominator into the current block.
3526 If the block has multiple predecessors,
3527 2a. Iterate over the ANTIC expressions for the block to see if
3528 any of them are partially redundant.
3529 2b. If so, insert them into the necessary predecessors to make
3530 the expression fully redundant.
3531 2c. Insert a new PHI merging the values of the predecessors.
3532 2d. Insert the new PHI, and the new expressions, into the
3534 3. Recursively call ourselves on the dominator children of BLOCK.
3536 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3537 do_regular_insertion and do_partial_insertion.
3542 do_regular_insertion (basic_block block, basic_block dom)
3544 bool new_stuff = false;
3545 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3549 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3551 if (expr->kind != NAME)
3555 bool by_some = false;
3556 bool cant_insert = false;
3557 bool all_same = true;
3558 pre_expr first_s = NULL;
3561 pre_expr eprime = NULL;
3563 pre_expr edoubleprime = NULL;
3564 bool do_insertion = false;
3566 val = get_expr_value_id (expr);
3567 if (bitmap_set_contains_value (PHI_GEN (block), val))
3569 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3571 if (dump_file && (dump_flags & TDF_DETAILS))
3572 fprintf (dump_file, "Found fully redundant value\n");
3576 avail = XCNEWVEC (pre_expr, last_basic_block);
3577 FOR_EACH_EDGE (pred, ei, block->preds)
3579 unsigned int vprime;
3581 /* We should never run insertion for the exit block
3582 and so not come across fake pred edges. */
3583 gcc_assert (!(pred->flags & EDGE_FAKE));
3585 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3588 /* eprime will generally only be NULL if the
3589 value of the expression, translated
3590 through the PHI for this predecessor, is
3591 undefined. If that is the case, we can't
3592 make the expression fully redundant,
3593 because its value is undefined along a
3594 predecessor path. We can thus break out
3595 early because it doesn't matter what the
3596 rest of the results are. */
3603 eprime = fully_constant_expression (eprime);
3604 vprime = get_expr_value_id (eprime);
3605 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3607 if (edoubleprime == NULL)
3609 avail[bprime->index] = eprime;
3614 avail[bprime->index] = edoubleprime;
3616 /* We want to perform insertions to remove a redundancy on
3617 a path in the CFG we want to optimize for speed. */
3618 if (optimize_edge_for_speed_p (pred))
3619 do_insertion = true;
3620 if (first_s == NULL)
3621 first_s = edoubleprime;
3622 else if (!pre_expr_eq (first_s, edoubleprime))
3626 /* If we can insert it, it's not the same value
3627 already existing along every predecessor, and
3628 it's defined by some predecessor, it is
3629 partially redundant. */
3630 if (!cant_insert && !all_same && by_some && do_insertion
3631 && dbg_cnt (treepre_insert))
3633 if (insert_into_preds_of_block (block, get_expression_id (expr),
3637 /* If all edges produce the same value and that value is
3638 an invariant, then the PHI has the same value on all
3639 edges. Note this. */
3640 else if (!cant_insert && all_same && eprime
3641 && (edoubleprime->kind == CONSTANT
3642 || edoubleprime->kind == NAME)
3643 && !value_id_constant_p (val))
3647 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3648 value_expressions, val);
3650 unsigned int new_val = get_expr_value_id (edoubleprime);
3651 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3653 pre_expr expr = expression_for_id (j);
3655 if (expr->kind == NAME)
3657 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3658 /* Just reset the value id and valnum so it is
3659 the same as the constant we have discovered. */
3660 if (edoubleprime->kind == CONSTANT)
3662 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3663 pre_stats.constified++;
3666 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3667 info->value_id = new_val;
3675 VEC_free (pre_expr, heap, exprs);
3680 /* Perform insertion for partially anticipatable expressions. There
3681 is only one case we will perform insertion for these. This case is
3682 if the expression is partially anticipatable, and fully available.
3683 In this case, we know that putting it earlier will enable us to
3684 remove the later computation. */
3688 do_partial_partial_insertion (basic_block block, basic_block dom)
3690 bool new_stuff = false;
3691 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3695 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3697 if (expr->kind != NAME)
3702 bool cant_insert = false;
3705 pre_expr eprime = NULL;
3708 val = get_expr_value_id (expr);
3709 if (bitmap_set_contains_value (PHI_GEN (block), val))
3711 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3714 avail = XCNEWVEC (pre_expr, last_basic_block);
3715 FOR_EACH_EDGE (pred, ei, block->preds)
3717 unsigned int vprime;
3718 pre_expr edoubleprime;
3720 /* We should never run insertion for the exit block
3721 and so not come across fake pred edges. */
3722 gcc_assert (!(pred->flags & EDGE_FAKE));
3724 eprime = phi_translate (expr, ANTIC_IN (block),
3728 /* eprime will generally only be NULL if the
3729 value of the expression, translated
3730 through the PHI for this predecessor, is
3731 undefined. If that is the case, we can't
3732 make the expression fully redundant,
3733 because its value is undefined along a
3734 predecessor path. We can thus break out
3735 early because it doesn't matter what the
3736 rest of the results are. */
3743 eprime = fully_constant_expression (eprime);
3744 vprime = get_expr_value_id (eprime);
3745 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3747 if (edoubleprime == NULL)
3753 avail[bprime->index] = edoubleprime;
3757 /* If we can insert it, it's not the same value
3758 already existing along every predecessor, and
3759 it's defined by some predecessor, it is
3760 partially redundant. */
3761 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3763 pre_stats.pa_insert++;
3764 if (insert_into_preds_of_block (block, get_expression_id (expr),
3772 VEC_free (pre_expr, heap, exprs);
3777 insert_aux (basic_block block)
3780 bool new_stuff = false;
3785 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3790 bitmap_set_t newset = NEW_SETS (dom);
3793 /* Note that we need to value_replace both NEW_SETS, and
3794 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3795 represented by some non-simple expression here that we want
3796 to replace it with. */
3797 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3799 pre_expr expr = expression_for_id (i);
3800 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3801 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3804 if (!single_pred_p (block))
3806 new_stuff |= do_regular_insertion (block, dom);
3807 if (do_partial_partial)
3808 new_stuff |= do_partial_partial_insertion (block, dom);
3812 for (son = first_dom_son (CDI_DOMINATORS, block);
3814 son = next_dom_son (CDI_DOMINATORS, son))
3816 new_stuff |= insert_aux (son);
3822 /* Perform insertion of partially redundant values. */
3827 bool new_stuff = true;
3829 int num_iterations = 0;
3832 NEW_SETS (bb) = bitmap_set_new ();
3837 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3839 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3843 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3846 add_to_exp_gen (basic_block block, tree op)
3851 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3853 result = get_or_alloc_expr_for_name (op);
3854 bitmap_value_insert_into_set (EXP_GEN (block), result);
3858 /* Create value ids for PHI in BLOCK. */
3861 make_values_for_phi (gimple phi, basic_block block)
3863 tree result = gimple_phi_result (phi);
3865 /* We have no need for virtual phis, as they don't represent
3866 actual computations. */
3867 if (is_gimple_reg (result))
3869 pre_expr e = get_or_alloc_expr_for_name (result);
3870 add_to_value (get_expr_value_id (e), e);
3871 bitmap_insert_into_set (PHI_GEN (block), e);
3872 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3876 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3878 tree arg = gimple_phi_arg_def (phi, i);
3879 if (TREE_CODE (arg) == SSA_NAME)
3881 e = get_or_alloc_expr_for_name (arg);
3882 add_to_value (get_expr_value_id (e), e);
3889 /* Compute the AVAIL set for all basic blocks.
3891 This function performs value numbering of the statements in each basic
3892 block. The AVAIL sets are built from information we glean while doing
3893 this value numbering, since the AVAIL sets contain only one entry per
3896 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3897 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3900 compute_avail (void)
3903 basic_block block, son;
3904 basic_block *worklist;
3908 /* We pretend that default definitions are defined in the entry block.
3909 This includes function arguments and the static chain decl. */
3910 for (i = 1; i < num_ssa_names; ++i)
3912 tree name = ssa_name (i);
3915 || !SSA_NAME_IS_DEFAULT_DEF (name)
3916 || has_zero_uses (name)
3917 || !is_gimple_reg (name))
3920 e = get_or_alloc_expr_for_name (name);
3921 add_to_value (get_expr_value_id (e), e);
3923 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3924 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3927 /* Allocate the worklist. */
3928 worklist = XNEWVEC (basic_block, n_basic_blocks);
3930 /* Seed the algorithm by putting the dominator children of the entry
3931 block on the worklist. */
3932 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3934 son = next_dom_son (CDI_DOMINATORS, son))
3935 worklist[sp++] = son;
3937 /* Loop until the worklist is empty. */
3940 gimple_stmt_iterator gsi;
3943 unsigned int stmt_uid = 1;
3945 /* Pick a block from the worklist. */
3946 block = worklist[--sp];
3948 /* Initially, the set of available values in BLOCK is that of
3949 its immediate dominator. */
3950 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3952 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3954 /* Generate values for PHI nodes. */
3955 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3956 make_values_for_phi (gsi_stmt (gsi), block);
3958 BB_MAY_NOTRETURN (block) = 0;
3960 /* Now compute value numbers and populate value sets with all
3961 the expressions computed in BLOCK. */
3962 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3967 stmt = gsi_stmt (gsi);
3968 gimple_set_uid (stmt, stmt_uid++);
3970 /* Cache whether the basic-block has any non-visible side-effect
3972 If this isn't a call or it is the last stmt in the
3973 basic-block then the CFG represents things correctly. */
3974 if (is_gimple_call (stmt)
3975 && !stmt_ends_bb_p (stmt))
3977 /* Non-looping const functions always return normally.
3978 Otherwise the call might not return or have side-effects
3979 that forbids hoisting possibly trapping expressions
3981 int flags = gimple_call_flags (stmt);
3982 if (!(flags & ECF_CONST)
3983 || (flags & ECF_LOOPING_CONST_OR_PURE))
3984 BB_MAY_NOTRETURN (block) = 1;
3987 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3989 pre_expr e = get_or_alloc_expr_for_name (op);
3991 add_to_value (get_expr_value_id (e), e);
3993 bitmap_insert_into_set (TMP_GEN (block), e);
3994 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3997 if (gimple_has_volatile_ops (stmt)
3998 || stmt_could_throw_p (stmt))
4001 switch (gimple_code (stmt))
4004 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4005 add_to_exp_gen (block, op);
4012 vn_reference_op_t vro;
4013 pre_expr result = NULL;
4014 VEC(vn_reference_op_s, heap) *ops = NULL;
4016 if (!can_value_number_call (stmt))
4019 copy_reference_ops_from_call (stmt, &ops);
4020 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4021 gimple_expr_type (stmt),
4023 VEC_free (vn_reference_op_s, heap, ops);
4027 for (i = 0; VEC_iterate (vn_reference_op_s,
4031 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4032 add_to_exp_gen (block, vro->op0);
4033 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4034 add_to_exp_gen (block, vro->op1);
4035 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4036 add_to_exp_gen (block, vro->op2);
4038 result = (pre_expr) pool_alloc (pre_expr_pool);
4039 result->kind = REFERENCE;
4041 PRE_EXPR_REFERENCE (result) = ref;
4043 get_or_alloc_expression_id (result);
4044 add_to_value (get_expr_value_id (result), result);
4046 bitmap_value_insert_into_set (EXP_GEN (block), result);
4052 pre_expr result = NULL;
4053 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4057 case tcc_comparison:
4062 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4063 gimple_assign_rhs_code (stmt),
4064 gimple_expr_type (stmt),
4065 gimple_assign_rhs1 (stmt),
4066 gimple_assign_rhs2 (stmt),
4067 NULL_TREE, NULL_TREE, &nary);
4072 for (i = 0; i < nary->length; i++)
4073 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4074 add_to_exp_gen (block, nary->op[i]);
4076 result = (pre_expr) pool_alloc (pre_expr_pool);
4077 result->kind = NARY;
4079 PRE_EXPR_NARY (result) = nary;
4083 case tcc_declaration:
4088 vn_reference_op_t vro;
4090 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4096 for (i = 0; VEC_iterate (vn_reference_op_s,
4100 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4101 add_to_exp_gen (block, vro->op0);
4102 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4103 add_to_exp_gen (block, vro->op1);
4104 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4105 add_to_exp_gen (block, vro->op2);
4107 result = (pre_expr) pool_alloc (pre_expr_pool);
4108 result->kind = REFERENCE;
4110 PRE_EXPR_REFERENCE (result) = ref;
4115 /* For any other statement that we don't
4116 recognize, simply add all referenced
4117 SSA_NAMEs to EXP_GEN. */
4118 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4119 add_to_exp_gen (block, op);
4123 get_or_alloc_expression_id (result);
4124 add_to_value (get_expr_value_id (result), result);
4126 bitmap_value_insert_into_set (EXP_GEN (block), result);
4135 /* Put the dominator children of BLOCK on the worklist of blocks
4136 to compute available sets for. */
4137 for (son = first_dom_son (CDI_DOMINATORS, block);
4139 son = next_dom_son (CDI_DOMINATORS, son))
4140 worklist[sp++] = son;
4146 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4147 than the available expressions for it. The insertion point is
4148 right before the first use in STMT. Returns the SSA_NAME that should
4149 be used for replacement. */
4152 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4154 basic_block bb = gimple_bb (stmt);
4155 gimple_stmt_iterator gsi;
4156 gimple_seq stmts = NULL;
4160 /* First create a value expression from the expression we want
4161 to insert and associate it with the value handle for SSA_VN. */
4162 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4166 /* Then use create_expression_by_pieces to generate a valid
4167 expression to insert at this point of the IL stream. */
4168 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4169 if (expr == NULL_TREE)
4171 gsi = gsi_for_stmt (stmt);
4172 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4177 /* Eliminate fully redundant computations. */
4182 VEC (gimple, heap) *to_remove = NULL;
4184 unsigned int todo = 0;
4185 gimple_stmt_iterator gsi;
4191 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4193 stmt = gsi_stmt (gsi);
4195 /* Lookup the RHS of the expression, see if we have an
4196 available computation for it. If so, replace the RHS with
4197 the available computation. */
4198 if (gimple_has_lhs (stmt)
4199 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4200 && !gimple_assign_ssa_name_copy_p (stmt)
4201 && (!gimple_assign_single_p (stmt)
4202 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4203 && !gimple_has_volatile_ops (stmt)
4204 && !has_zero_uses (gimple_get_lhs (stmt)))
4206 tree lhs = gimple_get_lhs (stmt);
4207 tree rhs = NULL_TREE;
4209 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4210 pre_expr sprimeexpr;
4212 if (gimple_assign_single_p (stmt))
4213 rhs = gimple_assign_rhs1 (stmt);
4215 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4216 get_expr_value_id (lhsexpr),
4221 if (sprimeexpr->kind == CONSTANT)
4222 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4223 else if (sprimeexpr->kind == NAME)
4224 sprime = PRE_EXPR_NAME (sprimeexpr);
4229 /* If there is no existing leader but SCCVN knows this
4230 value is constant, use that constant. */
4231 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4233 sprime = VN_INFO (lhs)->valnum;
4234 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4235 TREE_TYPE (sprime)))
4236 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4238 if (dump_file && (dump_flags & TDF_DETAILS))
4240 fprintf (dump_file, "Replaced ");
4241 print_gimple_expr (dump_file, stmt, 0, 0);
4242 fprintf (dump_file, " with ");
4243 print_generic_expr (dump_file, sprime, 0);
4244 fprintf (dump_file, " in ");
4245 print_gimple_stmt (dump_file, stmt, 0, 0);
4247 pre_stats.eliminations++;
4248 propagate_tree_value_into_stmt (&gsi, sprime);
4249 stmt = gsi_stmt (gsi);
4254 /* If there is no existing usable leader but SCCVN thinks
4255 it has an expression it wants to use as replacement,
4257 if (!sprime || sprime == lhs)
4259 tree val = VN_INFO (lhs)->valnum;
4261 && TREE_CODE (val) == SSA_NAME
4262 && VN_INFO (val)->needs_insertion
4263 && can_PRE_operation (vn_get_expr_for (val)))
4264 sprime = do_SCCVN_insertion (stmt, val);
4268 && (rhs == NULL_TREE
4269 || TREE_CODE (rhs) != SSA_NAME
4270 || may_propagate_copy (rhs, sprime)))
4272 gcc_assert (sprime != rhs);
4274 if (dump_file && (dump_flags & TDF_DETAILS))
4276 fprintf (dump_file, "Replaced ");
4277 print_gimple_expr (dump_file, stmt, 0, 0);
4278 fprintf (dump_file, " with ");
4279 print_generic_expr (dump_file, sprime, 0);
4280 fprintf (dump_file, " in ");
4281 print_gimple_stmt (dump_file, stmt, 0, 0);
4284 if (TREE_CODE (sprime) == SSA_NAME)
4285 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4287 /* We need to make sure the new and old types actually match,
4288 which may require adding a simple cast, which fold_convert
4290 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4291 && !useless_type_conversion_p (gimple_expr_type (stmt),
4292 TREE_TYPE (sprime)))
4293 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4295 pre_stats.eliminations++;
4296 propagate_tree_value_into_stmt (&gsi, sprime);
4297 stmt = gsi_stmt (gsi);
4300 /* If we removed EH side effects from the statement, clean
4301 its EH information. */
4302 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4304 bitmap_set_bit (need_eh_cleanup,
4305 gimple_bb (stmt)->index);
4306 if (dump_file && (dump_flags & TDF_DETAILS))
4307 fprintf (dump_file, " Removed EH side effects.\n");
4311 /* If the statement is a scalar store, see if the expression
4312 has the same value number as its rhs. If so, the store is
4314 else if (gimple_assign_single_p (stmt)
4315 && !is_gimple_reg (gimple_assign_lhs (stmt))
4316 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4317 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4319 tree rhs = gimple_assign_rhs1 (stmt);
4321 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4322 gimple_vuse (stmt), true, NULL);
4323 if (TREE_CODE (rhs) == SSA_NAME)
4324 rhs = VN_INFO (rhs)->valnum;
4326 && operand_equal_p (val, rhs, 0))
4328 if (dump_file && (dump_flags & TDF_DETAILS))
4330 fprintf (dump_file, "Deleted redundant store ");
4331 print_gimple_stmt (dump_file, stmt, 0, 0);
4334 /* Queue stmt for removal. */
4335 VEC_safe_push (gimple, heap, to_remove, stmt);
4338 /* Visit COND_EXPRs and fold the comparison with the
4339 available value-numbers. */
4340 else if (gimple_code (stmt) == GIMPLE_COND)
4342 tree op0 = gimple_cond_lhs (stmt);
4343 tree op1 = gimple_cond_rhs (stmt);
4346 if (TREE_CODE (op0) == SSA_NAME)
4347 op0 = VN_INFO (op0)->valnum;
4348 if (TREE_CODE (op1) == SSA_NAME)
4349 op1 = VN_INFO (op1)->valnum;
4350 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4352 if (result && TREE_CODE (result) == INTEGER_CST)
4354 if (integer_zerop (result))
4355 gimple_cond_make_false (stmt);
4357 gimple_cond_make_true (stmt);
4359 todo = TODO_cleanup_cfg;
4362 /* Visit indirect calls and turn them into direct calls if
4364 if (gimple_code (stmt) == GIMPLE_CALL
4365 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4367 tree fn = VN_INFO (gimple_call_fn (stmt))->valnum;
4368 if (TREE_CODE (fn) == ADDR_EXPR
4369 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
4371 if (dump_file && (dump_flags & TDF_DETAILS))
4373 fprintf (dump_file, "Replacing call target with ");
4374 print_generic_expr (dump_file, fn, 0);
4375 fprintf (dump_file, " in ");
4376 print_gimple_stmt (dump_file, stmt, 0, 0);
4379 gimple_call_set_fn (stmt, fn);
4381 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4383 bitmap_set_bit (need_eh_cleanup,
4384 gimple_bb (stmt)->index);
4385 if (dump_file && (dump_flags & TDF_DETAILS))
4386 fprintf (dump_file, " Removed EH side effects.\n");
4389 /* Changing an indirect call to a direct call may
4390 have exposed different semantics. This may
4391 require an SSA update. */
4392 todo |= TODO_update_ssa_only_virtuals;
4397 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4399 gimple stmt, phi = gsi_stmt (gsi);
4400 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4401 pre_expr sprimeexpr, resexpr;
4402 gimple_stmt_iterator gsi2;
4404 /* We want to perform redundant PHI elimination. Do so by
4405 replacing the PHI with a single copy if possible.
4406 Do not touch inserted, single-argument or virtual PHIs. */
4407 if (gimple_phi_num_args (phi) == 1
4408 || !is_gimple_reg (res))
4414 resexpr = get_or_alloc_expr_for_name (res);
4415 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4416 get_expr_value_id (resexpr), NULL);
4419 if (sprimeexpr->kind == CONSTANT)
4420 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4421 else if (sprimeexpr->kind == NAME)
4422 sprime = PRE_EXPR_NAME (sprimeexpr);
4426 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4428 sprime = VN_INFO (res)->valnum;
4429 if (!useless_type_conversion_p (TREE_TYPE (res),
4430 TREE_TYPE (sprime)))
4431 sprime = fold_convert (TREE_TYPE (res), sprime);
4440 if (dump_file && (dump_flags & TDF_DETAILS))
4442 fprintf (dump_file, "Replaced redundant PHI node defining ");
4443 print_generic_expr (dump_file, res, 0);
4444 fprintf (dump_file, " with ");
4445 print_generic_expr (dump_file, sprime, 0);
4446 fprintf (dump_file, "\n");
4449 remove_phi_node (&gsi, false);
4451 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4452 && TREE_CODE (sprime) == SSA_NAME)
4453 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4455 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4456 sprime = fold_convert (TREE_TYPE (res), sprime);
4457 stmt = gimple_build_assign (res, sprime);
4458 SSA_NAME_DEF_STMT (res) = stmt;
4459 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4461 gsi2 = gsi_after_labels (b);
4462 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4463 /* Queue the copy for eventual removal. */
4464 VEC_safe_push (gimple, heap, to_remove, stmt);
4465 /* If we inserted this PHI node ourself, it's not an elimination. */
4466 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4469 pre_stats.eliminations++;
4473 /* We cannot remove stmts during BB walk, especially not release SSA
4474 names there as this confuses the VN machinery. The stmts ending
4475 up in to_remove are either stores or simple copies. */
4476 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4478 tree lhs = gimple_assign_lhs (stmt);
4479 tree rhs = gimple_assign_rhs1 (stmt);
4480 use_operand_p use_p;
4483 /* If there is a single use only, propagate the equivalency
4484 instead of keeping the copy. */
4485 if (TREE_CODE (lhs) == SSA_NAME
4486 && TREE_CODE (rhs) == SSA_NAME
4487 && single_imm_use (lhs, &use_p, &use_stmt)
4488 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4490 SET_USE (use_p, rhs);
4491 update_stmt (use_stmt);
4492 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4493 && TREE_CODE (rhs) == SSA_NAME)
4494 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4497 /* If this is a store or a now unused copy, remove it. */
4498 if (TREE_CODE (lhs) != SSA_NAME
4499 || has_zero_uses (lhs))
4501 gsi = gsi_for_stmt (stmt);
4502 unlink_stmt_vdef (stmt);
4503 gsi_remove (&gsi, true);
4504 if (TREE_CODE (lhs) == SSA_NAME)
4505 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4506 release_defs (stmt);
4509 VEC_free (gimple, heap, to_remove);
4514 /* Borrow a bit of tree-ssa-dce.c for the moment.
4515 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4516 this may be a bit faster, and we may want critical edges kept split. */
4518 /* If OP's defining statement has not already been determined to be necessary,
4519 mark that statement necessary. Return the stmt, if it is newly
4522 static inline gimple
4523 mark_operand_necessary (tree op)
4529 if (TREE_CODE (op) != SSA_NAME)
4532 stmt = SSA_NAME_DEF_STMT (op);
4535 if (gimple_plf (stmt, NECESSARY)
4536 || gimple_nop_p (stmt))
4539 gimple_set_plf (stmt, NECESSARY, true);
4543 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4544 to insert PHI nodes sometimes, and because value numbering of casts isn't
4545 perfect, we sometimes end up inserting dead code. This simple DCE-like
4546 pass removes any insertions we made that weren't actually used. */
4549 remove_dead_inserted_code (void)
4556 worklist = BITMAP_ALLOC (NULL);
4557 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4559 t = SSA_NAME_DEF_STMT (ssa_name (i));
4560 if (gimple_plf (t, NECESSARY))
4561 bitmap_set_bit (worklist, i);
4563 while (!bitmap_empty_p (worklist))
4565 i = bitmap_first_set_bit (worklist);
4566 bitmap_clear_bit (worklist, i);
4567 t = SSA_NAME_DEF_STMT (ssa_name (i));
4569 /* PHI nodes are somewhat special in that each PHI alternative has
4570 data and control dependencies. All the statements feeding the
4571 PHI node's arguments are always necessary. */
4572 if (gimple_code (t) == GIMPLE_PHI)
4576 for (k = 0; k < gimple_phi_num_args (t); k++)
4578 tree arg = PHI_ARG_DEF (t, k);
4579 if (TREE_CODE (arg) == SSA_NAME)
4581 gimple n = mark_operand_necessary (arg);
4583 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4589 /* Propagate through the operands. Examine all the USE, VUSE and
4590 VDEF operands in this statement. Mark all the statements
4591 which feed this statement's uses as necessary. */
4595 /* The operands of VDEF expressions are also needed as they
4596 represent potential definitions that may reach this
4597 statement (VDEF operands allow us to follow def-def
4600 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4602 gimple n = mark_operand_necessary (use);
4604 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4609 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4611 t = SSA_NAME_DEF_STMT (ssa_name (i));
4612 if (!gimple_plf (t, NECESSARY))
4614 gimple_stmt_iterator gsi;
4616 if (dump_file && (dump_flags & TDF_DETAILS))
4618 fprintf (dump_file, "Removing unnecessary insertion:");
4619 print_gimple_stmt (dump_file, t, 0, 0);
4622 gsi = gsi_for_stmt (t);
4623 if (gimple_code (t) == GIMPLE_PHI)
4624 remove_phi_node (&gsi, true);
4627 gsi_remove (&gsi, true);
4632 BITMAP_FREE (worklist);
4635 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4636 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4637 the number of visited blocks. */
4640 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4642 edge_iterator *stack;
4644 int post_order_num = 0;
4647 if (include_entry_exit)
4648 post_order[post_order_num++] = EXIT_BLOCK;
4650 /* Allocate stack for back-tracking up CFG. */
4651 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4654 /* Allocate bitmap to track nodes that have been visited. */
4655 visited = sbitmap_alloc (last_basic_block);
4657 /* None of the nodes in the CFG have been visited yet. */
4658 sbitmap_zero (visited);
4660 /* Push the last edge on to the stack. */
4661 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4669 /* Look at the edge on the top of the stack. */
4671 src = ei_edge (ei)->src;
4672 dest = ei_edge (ei)->dest;
4674 /* Check if the edge destination has been visited yet. */
4675 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4677 /* Mark that we have visited the destination. */
4678 SET_BIT (visited, src->index);
4680 if (EDGE_COUNT (src->preds) > 0)
4681 /* Since the DEST node has been visited for the first
4682 time, check its successors. */
4683 stack[sp++] = ei_start (src->preds);
4685 post_order[post_order_num++] = src->index;
4689 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4690 post_order[post_order_num++] = dest->index;
4692 if (!ei_one_before_end_p (ei))
4693 ei_next (&stack[sp - 1]);
4699 if (include_entry_exit)
4700 post_order[post_order_num++] = ENTRY_BLOCK;
4703 sbitmap_free (visited);
4704 return post_order_num;
4708 /* Initialize data structures used by PRE. */
4711 init_pre (bool do_fre)
4715 next_expression_id = 1;
4717 VEC_safe_push (pre_expr, heap, expressions, NULL);
4718 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4719 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4720 get_max_value_id() + 1);
4725 inserted_exprs = BITMAP_ALLOC (NULL);
4726 need_creation = NULL;
4727 pretemp = NULL_TREE;
4728 storetemp = NULL_TREE;
4729 prephitemp = NULL_TREE;
4731 connect_infinite_loops_to_exit ();
4732 memset (&pre_stats, 0, sizeof (pre_stats));
4735 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4736 my_rev_post_order_compute (postorder, false);
4738 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4740 calculate_dominance_info (CDI_POST_DOMINATORS);
4741 calculate_dominance_info (CDI_DOMINATORS);
4743 bitmap_obstack_initialize (&grand_bitmap_obstack);
4744 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4745 expr_pred_trans_eq, free);
4746 expression_to_id = htab_create (num_ssa_names * 3,
4749 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4750 sizeof (struct bitmap_set), 30);
4751 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4752 sizeof (struct pre_expr_d), 30);
4755 EXP_GEN (bb) = bitmap_set_new ();
4756 PHI_GEN (bb) = bitmap_set_new ();
4757 TMP_GEN (bb) = bitmap_set_new ();
4758 AVAIL_OUT (bb) = bitmap_set_new ();
4761 need_eh_cleanup = BITMAP_ALLOC (NULL);
4765 /* Deallocate data structures used by PRE. */
4768 fini_pre (bool do_fre)
4771 VEC_free (bitmap_set_t, heap, value_expressions);
4772 BITMAP_FREE (inserted_exprs);
4773 VEC_free (gimple, heap, need_creation);
4774 bitmap_obstack_release (&grand_bitmap_obstack);
4775 free_alloc_pool (bitmap_set_pool);
4776 free_alloc_pool (pre_expr_pool);
4777 htab_delete (phi_translate_table);
4778 htab_delete (expression_to_id);
4779 VEC_free (unsigned, heap, name_to_id);
4781 free_aux_for_blocks ();
4783 free_dominance_info (CDI_POST_DOMINATORS);
4785 if (!bitmap_empty_p (need_eh_cleanup))
4787 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4788 cleanup_tree_cfg ();
4791 BITMAP_FREE (need_eh_cleanup);
4794 loop_optimizer_finalize ();
4797 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4798 only wants to do full redundancy elimination. */
4801 execute_pre (bool do_fre)
4803 unsigned int todo = 0;
4805 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4807 /* This has to happen before SCCVN runs because
4808 loop_optimizer_init may create new phis, etc. */
4810 loop_optimizer_init (LOOPS_NORMAL);
4815 loop_optimizer_finalize ();
4823 /* Collect and value number expressions computed in each basic block. */
4826 if (dump_file && (dump_flags & TDF_DETAILS))
4832 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4833 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4834 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4835 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4839 /* Insert can get quite slow on an incredibly large number of basic
4840 blocks due to some quadratic behavior. Until this behavior is
4841 fixed, don't run it when he have an incredibly large number of
4842 bb's. If we aren't going to run insert, there is no point in
4843 computing ANTIC, either, even though it's plenty fast. */
4844 if (!do_fre && n_basic_blocks < 4000)
4850 /* Make sure to remove fake edges before committing our inserts.
4851 This makes sure we don't end up with extra critical edges that
4852 we would need to split. */
4853 remove_fake_exit_edges ();
4854 gsi_commit_edge_inserts ();
4856 /* Remove all the redundant expressions. */
4857 todo |= eliminate ();
4859 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4860 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4861 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4862 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4863 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4865 clear_expression_ids ();
4868 remove_dead_inserted_code ();
4876 /* Gate and execute functions for PRE. */
4881 return execute_pre (false);
4887 return flag_tree_pre != 0;
4890 struct gimple_opt_pass pass_pre =
4895 gate_pre, /* gate */
4896 do_pre, /* execute */
4899 0, /* static_pass_number */
4900 TV_TREE_PRE, /* tv_id */
4901 PROP_no_crit_edges | PROP_cfg
4902 | PROP_ssa, /* properties_required */
4903 0, /* properties_provided */
4904 0, /* properties_destroyed */
4905 TODO_rebuild_alias, /* todo_flags_start */
4906 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4907 | TODO_verify_ssa /* todo_flags_finish */
4912 /* Gate and execute functions for FRE. */
4917 return execute_pre (true);
4923 return flag_tree_fre != 0;
4926 struct gimple_opt_pass pass_fre =
4931 gate_fre, /* gate */
4932 execute_fre, /* execute */
4935 0, /* static_pass_number */
4936 TV_TREE_FRE, /* tv_id */
4937 PROP_cfg | PROP_ssa, /* properties_required */
4938 0, /* properties_provided */
4939 0, /* properties_destroyed */
4940 0, /* todo_flags_start */
4941 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */