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 their EH properties changed. */
488 static bitmap need_eh_cleanup;
490 /* Set of blocks with statements that have had their AB properties changed. */
491 static bitmap need_ab_cleanup;
493 /* The phi_translate_table caches phi translations for a given
494 expression and predecessor. */
496 static htab_t phi_translate_table;
498 /* A three tuple {e, pred, v} used to cache phi translations in the
499 phi_translate_table. */
501 typedef struct expr_pred_trans_d
503 /* The expression. */
506 /* The predecessor block along which we translated the expression. */
509 /* The value that resulted from the translation. */
512 /* The hashcode for the expression, pred pair. This is cached for
515 } *expr_pred_trans_t;
516 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
518 /* Return the hash value for a phi translation table entry. */
521 expr_pred_trans_hash (const void *p)
523 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
527 /* Return true if two phi translation table entries are the same.
528 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
531 expr_pred_trans_eq (const void *p1, const void *p2)
533 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
534 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
535 basic_block b1 = ve1->pred;
536 basic_block b2 = ve2->pred;
538 /* If they are not translations for the same basic block, they can't
542 return pre_expr_eq (ve1->e, ve2->e);
545 /* Search in the phi translation table for the translation of
546 expression E in basic block PRED.
547 Return the translated value, if found, NULL otherwise. */
549 static inline pre_expr
550 phi_trans_lookup (pre_expr e, basic_block pred)
553 struct expr_pred_trans_d ept;
557 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
558 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
563 return ((expr_pred_trans_t) *slot)->v;
567 /* Add the tuple mapping from {expression E, basic block PRED} to
568 value V, to the phi translation table. */
571 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
574 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
576 new_pair->pred = pred;
578 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
581 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
582 new_pair->hashcode, INSERT);
585 *slot = (void *) new_pair;
589 /* Add expression E to the expression set of value id V. */
592 add_to_value (unsigned int v, pre_expr e)
596 gcc_assert (get_expr_value_id (e) == v);
598 if (v >= VEC_length (bitmap_set_t, value_expressions))
600 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
604 set = VEC_index (bitmap_set_t, value_expressions, v);
607 set = bitmap_set_new ();
608 VEC_replace (bitmap_set_t, value_expressions, v, set);
611 bitmap_insert_into_set_1 (set, e, v, true);
614 /* Create a new bitmap set and return it. */
617 bitmap_set_new (void)
619 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
620 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
621 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
625 /* Return the value id for a PRE expression EXPR. */
628 get_expr_value_id (pre_expr expr)
635 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
638 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
639 add_to_value (id, expr);
644 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
646 return PRE_EXPR_NARY (expr)->value_id;
648 return PRE_EXPR_REFERENCE (expr)->value_id;
654 /* Remove an expression EXPR from a bitmapped set. */
657 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
659 unsigned int val = get_expr_value_id (expr);
660 if (!value_id_constant_p (val))
662 bitmap_clear_bit (&set->values, val);
663 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
668 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
669 unsigned int val, bool allow_constants)
671 if (allow_constants || !value_id_constant_p (val))
673 /* We specifically expect this and only this function to be able to
674 insert constants into a set. */
675 bitmap_set_bit (&set->values, val);
676 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
680 /* Insert an expression EXPR into a bitmapped set. */
683 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
685 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
688 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
691 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
693 bitmap_copy (&dest->expressions, &orig->expressions);
694 bitmap_copy (&dest->values, &orig->values);
698 /* Free memory used up by SET. */
700 bitmap_set_free (bitmap_set_t set)
702 bitmap_clear (&set->expressions);
703 bitmap_clear (&set->values);
707 /* Generate an topological-ordered array of bitmap set SET. */
709 static VEC(pre_expr, heap) *
710 sorted_array_from_bitmap_set (bitmap_set_t set)
713 bitmap_iterator bi, bj;
714 VEC(pre_expr, heap) *result;
716 /* Pre-allocate roughly enough space for the array. */
717 result = VEC_alloc (pre_expr, heap, bitmap_count_bits (&set->values));
719 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
721 /* The number of expressions having a given value is usually
722 relatively small. Thus, rather than making a vector of all
723 the expressions and sorting it by value-id, we walk the values
724 and check in the reverse mapping that tells us what expressions
725 have a given value, to filter those in our set. As a result,
726 the expressions are inserted in value-id order, which means
729 If this is somehow a significant lose for some cases, we can
730 choose which set to walk based on the set size. */
731 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
732 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
734 if (bitmap_bit_p (&set->expressions, j))
735 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
742 /* Perform bitmapped set operation DEST &= ORIG. */
745 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
753 bitmap_initialize (&temp, &grand_bitmap_obstack);
755 bitmap_and_into (&dest->values, &orig->values);
756 bitmap_copy (&temp, &dest->expressions);
757 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
759 pre_expr expr = expression_for_id (i);
760 unsigned int value_id = get_expr_value_id (expr);
761 if (!bitmap_bit_p (&dest->values, value_id))
762 bitmap_clear_bit (&dest->expressions, i);
764 bitmap_clear (&temp);
768 /* Subtract all values and expressions contained in ORIG from DEST. */
771 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
773 bitmap_set_t result = bitmap_set_new ();
777 bitmap_and_compl (&result->expressions, &dest->expressions,
780 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
782 pre_expr expr = expression_for_id (i);
783 unsigned int value_id = get_expr_value_id (expr);
784 bitmap_set_bit (&result->values, value_id);
790 /* Subtract all the values in bitmap set B from bitmap set A. */
793 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
799 bitmap_initialize (&temp, &grand_bitmap_obstack);
801 bitmap_copy (&temp, &a->expressions);
802 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
804 pre_expr expr = expression_for_id (i);
805 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
806 bitmap_remove_from_set (a, expr);
808 bitmap_clear (&temp);
812 /* Return true if bitmapped set SET contains the value VALUE_ID. */
815 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
817 if (value_id_constant_p (value_id))
820 if (!set || bitmap_empty_p (&set->expressions))
823 return bitmap_bit_p (&set->values, value_id);
827 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
829 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
832 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
835 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
838 bitmap_set_t exprset;
842 if (value_id_constant_p (lookfor))
845 if (!bitmap_set_contains_value (set, lookfor))
848 /* The number of expressions having a given value is usually
849 significantly less than the total number of expressions in SET.
850 Thus, rather than check, for each expression in SET, whether it
851 has the value LOOKFOR, we walk the reverse mapping that tells us
852 what expressions have a given value, and see if any of those
853 expressions are in our set. For large testcases, this is about
854 5-10x faster than walking the bitmap. If this is somehow a
855 significant lose for some cases, we can choose which set to walk
856 based on the set size. */
857 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
858 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
860 if (bitmap_clear_bit (&set->expressions, i))
862 bitmap_set_bit (&set->expressions, get_expression_id (expr));
868 /* Return true if two bitmap sets are equal. */
871 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
873 return bitmap_equal_p (&a->values, &b->values);
876 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
877 and add it otherwise. */
880 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
882 unsigned int val = get_expr_value_id (expr);
884 if (bitmap_set_contains_value (set, val))
885 bitmap_set_replace_value (set, val, expr);
887 bitmap_insert_into_set (set, expr);
890 /* Insert EXPR into SET if EXPR's value is not already present in
894 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
896 unsigned int val = get_expr_value_id (expr);
898 #ifdef ENABLE_CHECKING
899 gcc_assert (expr->id == get_or_alloc_expression_id (expr));
902 /* Constant values are always considered to be part of the set. */
903 if (value_id_constant_p (val))
906 /* If the value membership changed, add the expression. */
907 if (bitmap_set_bit (&set->values, val))
908 bitmap_set_bit (&set->expressions, expr->id);
911 /* Print out EXPR to outfile. */
914 print_pre_expr (FILE *outfile, const pre_expr expr)
919 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
922 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
927 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
928 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
929 for (i = 0; i < nary->length; i++)
931 print_generic_expr (outfile, nary->op[i], 0);
932 if (i != (unsigned) nary->length - 1)
933 fprintf (outfile, ",");
935 fprintf (outfile, "}");
941 vn_reference_op_t vro;
943 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
944 fprintf (outfile, "{");
946 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
949 bool closebrace = false;
950 if (vro->opcode != SSA_NAME
951 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
953 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
956 fprintf (outfile, "<");
962 print_generic_expr (outfile, vro->op0, 0);
965 fprintf (outfile, ",");
966 print_generic_expr (outfile, vro->op1, 0);
970 fprintf (outfile, ",");
971 print_generic_expr (outfile, vro->op2, 0);
975 fprintf (outfile, ">");
976 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
977 fprintf (outfile, ",");
979 fprintf (outfile, "}");
982 fprintf (outfile, "@");
983 print_generic_expr (outfile, ref->vuse, 0);
989 void debug_pre_expr (pre_expr);
991 /* Like print_pre_expr but always prints to stderr. */
993 debug_pre_expr (pre_expr e)
995 print_pre_expr (stderr, e);
996 fprintf (stderr, "\n");
999 /* Print out SET to OUTFILE. */
1002 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1003 const char *setname, int blockindex)
1005 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1012 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1014 const pre_expr expr = expression_for_id (i);
1017 fprintf (outfile, ", ");
1019 print_pre_expr (outfile, expr);
1021 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1024 fprintf (outfile, " }\n");
1027 void debug_bitmap_set (bitmap_set_t);
1030 debug_bitmap_set (bitmap_set_t set)
1032 print_bitmap_set (stderr, set, "debug", 0);
1035 /* Print out the expressions that have VAL to OUTFILE. */
1038 print_value_expressions (FILE *outfile, unsigned int val)
1040 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1044 sprintf (s, "%04d", val);
1045 print_bitmap_set (outfile, set, s, 0);
1051 debug_value_expressions (unsigned int val)
1053 print_value_expressions (stderr, val);
1056 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1060 get_or_alloc_expr_for_constant (tree constant)
1062 unsigned int result_id;
1063 unsigned int value_id;
1064 struct pre_expr_d expr;
1067 expr.kind = CONSTANT;
1068 PRE_EXPR_CONSTANT (&expr) = constant;
1069 result_id = lookup_expression_id (&expr);
1071 return expression_for_id (result_id);
1073 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1074 newexpr->kind = CONSTANT;
1075 PRE_EXPR_CONSTANT (newexpr) = constant;
1076 alloc_expression_id (newexpr);
1077 value_id = get_or_alloc_constant_value_id (constant);
1078 add_to_value (value_id, newexpr);
1082 /* Given a value id V, find the actual tree representing the constant
1083 value if there is one, and return it. Return NULL if we can't find
1087 get_constant_for_value_id (unsigned int v)
1089 if (value_id_constant_p (v))
1093 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1095 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1097 pre_expr expr = expression_for_id (i);
1098 if (expr->kind == CONSTANT)
1099 return PRE_EXPR_CONSTANT (expr);
1105 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1106 Currently only supports constants and SSA_NAMES. */
1108 get_or_alloc_expr_for (tree t)
1110 if (TREE_CODE (t) == SSA_NAME)
1111 return get_or_alloc_expr_for_name (t);
1112 else if (is_gimple_min_invariant (t))
1113 return get_or_alloc_expr_for_constant (t);
1116 /* More complex expressions can result from SCCVN expression
1117 simplification that inserts values for them. As they all
1118 do not have VOPs the get handled by the nary ops struct. */
1119 vn_nary_op_t result;
1120 unsigned int result_id;
1121 vn_nary_op_lookup (t, &result);
1124 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1126 PRE_EXPR_NARY (e) = result;
1127 result_id = lookup_expression_id (e);
1130 pool_free (pre_expr_pool, e);
1131 e = expression_for_id (result_id);
1134 alloc_expression_id (e);
1141 /* Return the folded version of T if T, when folded, is a gimple
1142 min_invariant. Otherwise, return T. */
1145 fully_constant_expression (pre_expr e)
1153 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1154 switch (TREE_CODE_CLASS (nary->opcode))
1156 case tcc_expression:
1157 if (nary->opcode == TRUTH_NOT_EXPR)
1159 if (nary->opcode != TRUTH_AND_EXPR
1160 && nary->opcode != TRUTH_OR_EXPR
1161 && nary->opcode != TRUTH_XOR_EXPR)
1165 case tcc_comparison:
1167 /* We have to go from trees to pre exprs to value ids to
1169 tree naryop0 = nary->op[0];
1170 tree naryop1 = nary->op[1];
1172 if (!is_gimple_min_invariant (naryop0))
1174 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1175 unsigned int vrep0 = get_expr_value_id (rep0);
1176 tree const0 = get_constant_for_value_id (vrep0);
1178 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1180 if (!is_gimple_min_invariant (naryop1))
1182 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1183 unsigned int vrep1 = get_expr_value_id (rep1);
1184 tree const1 = get_constant_for_value_id (vrep1);
1186 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1188 result = fold_binary (nary->opcode, nary->type,
1190 if (result && is_gimple_min_invariant (result))
1191 return get_or_alloc_expr_for_constant (result);
1192 /* We might have simplified the expression to a
1193 SSA_NAME for example from x_1 * 1. But we cannot
1194 insert a PHI for x_1 unconditionally as x_1 might
1195 not be available readily. */
1199 if (nary->opcode != REALPART_EXPR
1200 && nary->opcode != IMAGPART_EXPR
1201 && nary->opcode != VIEW_CONVERT_EXPR)
1207 /* We have to go from trees to pre exprs to value ids to
1209 tree naryop0 = nary->op[0];
1210 tree const0, result;
1211 if (is_gimple_min_invariant (naryop0))
1215 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1216 unsigned int vrep0 = get_expr_value_id (rep0);
1217 const0 = get_constant_for_value_id (vrep0);
1222 tree type1 = TREE_TYPE (nary->op[0]);
1223 const0 = fold_convert (type1, const0);
1224 result = fold_unary (nary->opcode, nary->type, const0);
1226 if (result && is_gimple_min_invariant (result))
1227 return get_or_alloc_expr_for_constant (result);
1236 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1238 if ((folded = fully_constant_vn_reference_p (ref)))
1239 return get_or_alloc_expr_for_constant (folded);
1248 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1249 it has the value it would have in BLOCK. Set *SAME_VALID to true
1250 in case the new vuse doesn't change the value id of the OPERANDS. */
1253 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1254 alias_set_type set, tree type, tree vuse,
1255 basic_block phiblock,
1256 basic_block block, bool *same_valid)
1258 gimple phi = SSA_NAME_DEF_STMT (vuse);
1265 if (gimple_bb (phi) != phiblock)
1268 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1270 /* Use the alias-oracle to find either the PHI node in this block,
1271 the first VUSE used in this block that is equivalent to vuse or
1272 the first VUSE which definition in this block kills the value. */
1273 if (gimple_code (phi) == GIMPLE_PHI)
1274 e = find_edge (block, phiblock);
1275 else if (use_oracle)
1276 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1278 vuse = gimple_vuse (phi);
1279 phi = SSA_NAME_DEF_STMT (vuse);
1280 if (gimple_bb (phi) != phiblock)
1282 if (gimple_code (phi) == GIMPLE_PHI)
1284 e = find_edge (block, phiblock);
1295 bitmap visited = NULL;
1296 /* Try to find a vuse that dominates this phi node by skipping
1297 non-clobbering statements. */
1298 vuse = get_continuation_for_phi (phi, &ref, &visited);
1300 BITMAP_FREE (visited);
1306 /* If we didn't find any, the value ID can't stay the same,
1307 but return the translated vuse. */
1308 *same_valid = false;
1309 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1311 /* ??? We would like to return vuse here as this is the canonical
1312 upmost vdef that this reference is associated with. But during
1313 insertion of the references into the hash tables we only ever
1314 directly insert with their direct gimple_vuse, hence returning
1315 something else would make us not find the other expression. */
1316 return PHI_ARG_DEF (phi, e->dest_idx);
1322 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1323 SET2. This is used to avoid making a set consisting of the union
1324 of PA_IN and ANTIC_IN during insert. */
1326 static inline pre_expr
1327 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1331 result = bitmap_find_leader (set1, val, NULL);
1332 if (!result && set2)
1333 result = bitmap_find_leader (set2, val, NULL);
1337 /* Get the tree type for our PRE expression e. */
1340 get_expr_type (const pre_expr e)
1345 return TREE_TYPE (PRE_EXPR_NAME (e));
1347 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1349 return PRE_EXPR_REFERENCE (e)->type;
1351 return PRE_EXPR_NARY (e)->type;
1356 /* Get a representative SSA_NAME for a given expression.
1357 Since all of our sub-expressions are treated as values, we require
1358 them to be SSA_NAME's for simplicity.
1359 Prior versions of GVNPRE used to use "value handles" here, so that
1360 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1361 either case, the operands are really values (IE we do not expect
1362 them to be usable without finding leaders). */
1365 get_representative_for (const pre_expr e)
1369 unsigned int value_id = get_expr_value_id (e);
1374 return PRE_EXPR_NAME (e);
1376 return PRE_EXPR_CONSTANT (e);
1380 /* Go through all of the expressions representing this value
1381 and pick out an SSA_NAME. */
1384 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1386 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1388 pre_expr rep = expression_for_id (i);
1389 if (rep->kind == NAME)
1390 return PRE_EXPR_NAME (rep);
1395 /* If we reached here we couldn't find an SSA_NAME. This can
1396 happen when we've discovered a value that has never appeared in
1397 the program as set to an SSA_NAME, most likely as the result of
1402 "Could not find SSA_NAME representative for expression:");
1403 print_pre_expr (dump_file, e);
1404 fprintf (dump_file, "\n");
1407 exprtype = get_expr_type (e);
1409 /* Build and insert the assignment of the end result to the temporary
1410 that we will return. */
1411 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1413 pretemp = create_tmp_reg (exprtype, "pretmp");
1414 get_var_ann (pretemp);
1417 name = make_ssa_name (pretemp, gimple_build_nop ());
1418 VN_INFO_GET (name)->value_id = value_id;
1419 if (e->kind == CONSTANT)
1420 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1422 VN_INFO (name)->valnum = name;
1424 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1427 fprintf (dump_file, "Created SSA_NAME representative ");
1428 print_generic_expr (dump_file, name, 0);
1429 fprintf (dump_file, " for expression:");
1430 print_pre_expr (dump_file, e);
1431 fprintf (dump_file, "\n");
1440 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1441 basic_block pred, basic_block phiblock);
1443 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1444 the phis in PRED. Return NULL if we can't find a leader for each part
1445 of the translated expression. */
1448 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1449 basic_block pred, basic_block phiblock)
1456 bool changed = false;
1457 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1458 struct vn_nary_op_s newnary;
1459 /* The NARY structure is only guaranteed to have been
1460 allocated to the nary->length operands. */
1461 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1462 - sizeof (tree) * (4 - nary->length)));
1464 for (i = 0; i < newnary.length; i++)
1466 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1470 pre_expr leader, result;
1471 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1472 leader = find_leader_in_sets (op_val_id, set1, set2);
1473 result = phi_translate (leader, set1, set2, pred, phiblock);
1474 if (result && result != leader)
1476 tree name = get_representative_for (result);
1479 newnary.op[i] = name;
1484 changed |= newnary.op[i] != nary->op[i];
1490 unsigned int new_val_id;
1492 tree result = vn_nary_op_lookup_pieces (newnary.length,
1500 if (result && is_gimple_min_invariant (result))
1501 return get_or_alloc_expr_for_constant (result);
1503 expr = (pre_expr) pool_alloc (pre_expr_pool);
1508 PRE_EXPR_NARY (expr) = nary;
1509 constant = fully_constant_expression (expr);
1510 if (constant != expr)
1513 new_val_id = nary->value_id;
1514 get_or_alloc_expression_id (expr);
1518 new_val_id = get_next_value_id ();
1519 VEC_safe_grow_cleared (bitmap_set_t, heap,
1521 get_max_value_id() + 1);
1522 nary = vn_nary_op_insert_pieces (newnary.length,
1529 result, new_val_id);
1530 PRE_EXPR_NARY (expr) = nary;
1531 constant = fully_constant_expression (expr);
1532 if (constant != expr)
1534 get_or_alloc_expression_id (expr);
1536 add_to_value (new_val_id, expr);
1544 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1545 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1546 tree vuse = ref->vuse;
1547 tree newvuse = vuse;
1548 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1549 bool changed = false, same_valid = true;
1551 vn_reference_op_t operand;
1552 vn_reference_t newref;
1555 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1559 tree oldop0 = operand->op0;
1560 tree oldop1 = operand->op1;
1561 tree oldop2 = operand->op2;
1565 tree type = operand->type;
1566 vn_reference_op_s newop = *operand;
1568 if (op0 && TREE_CODE (op0) == SSA_NAME)
1570 unsigned int op_val_id = VN_INFO (op0)->value_id;
1571 leader = find_leader_in_sets (op_val_id, set1, set2);
1572 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1573 if (opresult && opresult != leader)
1575 tree name = get_representative_for (opresult);
1583 changed |= op0 != oldop0;
1585 if (op1 && TREE_CODE (op1) == SSA_NAME)
1587 unsigned int op_val_id = VN_INFO (op1)->value_id;
1588 leader = find_leader_in_sets (op_val_id, set1, set2);
1589 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1590 if (opresult && opresult != leader)
1592 tree name = get_representative_for (opresult);
1600 /* We can't possibly insert these. */
1601 else if (op1 && !is_gimple_min_invariant (op1))
1603 changed |= op1 != oldop1;
1604 if (op2 && TREE_CODE (op2) == SSA_NAME)
1606 unsigned int op_val_id = VN_INFO (op2)->value_id;
1607 leader = find_leader_in_sets (op_val_id, set1, set2);
1608 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1609 if (opresult && opresult != leader)
1611 tree name = get_representative_for (opresult);
1619 /* We can't possibly insert these. */
1620 else if (op2 && !is_gimple_min_invariant (op2))
1622 changed |= op2 != oldop2;
1625 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1626 /* We may have changed from an SSA_NAME to a constant */
1627 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1628 newop.opcode = TREE_CODE (op0);
1633 /* If it transforms a non-constant ARRAY_REF into a constant
1634 one, adjust the constant offset. */
1635 if (newop.opcode == ARRAY_REF
1637 && TREE_CODE (op0) == INTEGER_CST
1638 && TREE_CODE (op1) == INTEGER_CST
1639 && TREE_CODE (op2) == INTEGER_CST)
1641 double_int off = tree_to_double_int (op0);
1642 off = double_int_add (off,
1644 (tree_to_double_int (op1)));
1645 off = double_int_mul (off, tree_to_double_int (op2));
1646 if (double_int_fits_in_shwi_p (off))
1647 newop.off = off.low;
1649 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1650 /* If it transforms from an SSA_NAME to an address, fold with
1651 a preceding indirect reference. */
1652 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1653 && VEC_index (vn_reference_op_s,
1654 newoperands, j - 1)->opcode == MEM_REF)
1655 vn_reference_fold_indirect (&newoperands, &j);
1657 if (i != VEC_length (vn_reference_op_s, operands))
1660 VEC_free (vn_reference_op_s, heap, newoperands);
1666 newvuse = translate_vuse_through_block (newoperands,
1667 ref->set, ref->type,
1668 vuse, phiblock, pred,
1670 if (newvuse == NULL_TREE)
1672 VEC_free (vn_reference_op_s, heap, newoperands);
1677 if (changed || newvuse != vuse)
1679 unsigned int new_val_id;
1681 bool converted = false;
1683 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1688 VEC_free (vn_reference_op_s, heap, newoperands);
1691 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1693 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1697 if (result && is_gimple_min_invariant (result))
1699 gcc_assert (!newoperands);
1700 return get_or_alloc_expr_for_constant (result);
1703 expr = (pre_expr) pool_alloc (pre_expr_pool);
1704 expr->kind = REFERENCE;
1712 gcc_assert (CONVERT_EXPR_P (result)
1713 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1715 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1717 TREE_OPERAND (result, 0),
1718 NULL_TREE, NULL_TREE,
1721 if (nresult && is_gimple_min_invariant (nresult))
1722 return get_or_alloc_expr_for_constant (nresult);
1727 PRE_EXPR_NARY (expr) = nary;
1728 constant = fully_constant_expression (expr);
1729 if (constant != expr)
1732 new_val_id = nary->value_id;
1733 get_or_alloc_expression_id (expr);
1737 new_val_id = get_next_value_id ();
1738 VEC_safe_grow_cleared (bitmap_set_t, heap,
1740 get_max_value_id() + 1);
1741 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1743 TREE_OPERAND (result, 0),
1744 NULL_TREE, NULL_TREE,
1745 NULL_TREE, NULL_TREE,
1747 PRE_EXPR_NARY (expr) = nary;
1748 constant = fully_constant_expression (expr);
1749 if (constant != expr)
1751 get_or_alloc_expression_id (expr);
1756 PRE_EXPR_REFERENCE (expr) = newref;
1757 constant = fully_constant_expression (expr);
1758 if (constant != expr)
1761 new_val_id = newref->value_id;
1762 get_or_alloc_expression_id (expr);
1766 if (changed || !same_valid)
1768 new_val_id = get_next_value_id ();
1769 VEC_safe_grow_cleared (bitmap_set_t, heap,
1771 get_max_value_id() + 1);
1774 new_val_id = ref->value_id;
1775 newref = vn_reference_insert_pieces (newvuse, ref->set,
1778 result, new_val_id);
1780 PRE_EXPR_REFERENCE (expr) = newref;
1781 constant = fully_constant_expression (expr);
1782 if (constant != expr)
1784 get_or_alloc_expression_id (expr);
1786 add_to_value (new_val_id, expr);
1788 VEC_free (vn_reference_op_s, heap, newoperands);
1798 tree name = PRE_EXPR_NAME (expr);
1800 def_stmt = SSA_NAME_DEF_STMT (name);
1801 if (gimple_code (def_stmt) == GIMPLE_PHI
1802 && gimple_bb (def_stmt) == phiblock)
1807 e = find_edge (pred, gimple_bb (phi));
1810 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1813 if (TREE_CODE (def) == SSA_NAME)
1814 def = VN_INFO (def)->valnum;
1816 /* Handle constant. */
1817 if (is_gimple_min_invariant (def))
1818 return get_or_alloc_expr_for_constant (def);
1820 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1823 newexpr = get_or_alloc_expr_for_name (def);
1834 /* Wrapper around phi_translate_1 providing caching functionality. */
1837 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1838 basic_block pred, basic_block phiblock)
1845 /* Constants contain no values that need translation. */
1846 if (expr->kind == CONSTANT)
1849 if (value_id_constant_p (get_expr_value_id (expr)))
1852 if (expr->kind != NAME)
1854 phitrans = phi_trans_lookup (expr, pred);
1860 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1862 /* Don't add empty translations to the cache. Neither add
1863 translations of NAMEs as those are cheap to translate. */
1865 && expr->kind != NAME)
1866 phi_trans_add (expr, phitrans, pred);
1872 /* For each expression in SET, translate the values through phi nodes
1873 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1874 expressions in DEST. */
1877 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1878 basic_block phiblock)
1880 VEC (pre_expr, heap) *exprs;
1884 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1886 bitmap_set_copy (dest, set);
1890 exprs = sorted_array_from_bitmap_set (set);
1891 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1893 pre_expr translated;
1894 translated = phi_translate (expr, set, NULL, pred, phiblock);
1898 /* We might end up with multiple expressions from SET being
1899 translated to the same value. In this case we do not want
1900 to retain the NARY or REFERENCE expression but prefer a NAME
1901 which would be the leader. */
1902 if (translated->kind == NAME)
1903 bitmap_value_replace_in_set (dest, translated);
1905 bitmap_value_insert_into_set (dest, translated);
1907 VEC_free (pre_expr, heap, exprs);
1910 /* Find the leader for a value (i.e., the name representing that
1911 value) in a given set, and return it. If STMT is non-NULL it
1912 makes sure the defining statement for the leader dominates it.
1913 Return NULL if no leader is found. */
1916 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1918 if (value_id_constant_p (val))
1922 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1924 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1926 pre_expr expr = expression_for_id (i);
1927 if (expr->kind == CONSTANT)
1931 if (bitmap_set_contains_value (set, val))
1933 /* Rather than walk the entire bitmap of expressions, and see
1934 whether any of them has the value we are looking for, we look
1935 at the reverse mapping, which tells us the set of expressions
1936 that have a given value (IE value->expressions with that
1937 value) and see if any of those expressions are in our set.
1938 The number of expressions per value is usually significantly
1939 less than the number of expressions in the set. In fact, for
1940 large testcases, doing it this way is roughly 5-10x faster
1941 than walking the bitmap.
1942 If this is somehow a significant lose for some cases, we can
1943 choose which set to walk based on which set is smaller. */
1946 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1948 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1949 &set->expressions, 0, i, bi)
1951 pre_expr val = expression_for_id (i);
1952 /* At the point where stmt is not null, there should always
1953 be an SSA_NAME first in the list of expressions. */
1956 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1957 if (gimple_code (def_stmt) != GIMPLE_PHI
1958 && gimple_bb (def_stmt) == gimple_bb (stmt)
1959 /* PRE insertions are at the end of the basic-block
1961 && (gimple_uid (def_stmt) == 0
1962 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1971 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1972 BLOCK by seeing if it is not killed in the block. Note that we are
1973 only determining whether there is a store that kills it. Because
1974 of the order in which clean iterates over values, we are guaranteed
1975 that altered operands will have caused us to be eliminated from the
1976 ANTIC_IN set already. */
1979 value_dies_in_block_x (pre_expr expr, basic_block block)
1981 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1982 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1984 gimple_stmt_iterator gsi;
1985 unsigned id = get_expression_id (expr);
1992 /* Lookup a previously calculated result. */
1993 if (EXPR_DIES (block)
1994 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1995 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1997 /* A memory expression {e, VUSE} dies in the block if there is a
1998 statement that may clobber e. If, starting statement walk from the
1999 top of the basic block, a statement uses VUSE there can be no kill
2000 inbetween that use and the original statement that loaded {e, VUSE},
2001 so we can stop walking. */
2002 ref.base = NULL_TREE;
2003 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
2005 tree def_vuse, def_vdef;
2006 def = gsi_stmt (gsi);
2007 def_vuse = gimple_vuse (def);
2008 def_vdef = gimple_vdef (def);
2010 /* Not a memory statement. */
2014 /* Not a may-def. */
2017 /* A load with the same VUSE, we're done. */
2018 if (def_vuse == vuse)
2024 /* Init ref only if we really need it. */
2025 if (ref.base == NULL_TREE
2026 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
2032 /* If the statement may clobber expr, it dies. */
2033 if (stmt_may_clobber_ref_p_1 (def, &ref))
2040 /* Remember the result. */
2041 if (!EXPR_DIES (block))
2042 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2043 bitmap_set_bit (EXPR_DIES (block), id * 2);
2045 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2051 #define union_contains_value(SET1, SET2, VAL) \
2052 (bitmap_set_contains_value ((SET1), (VAL)) \
2053 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
2055 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2058 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2059 vn_reference_op_t vro)
2061 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2063 struct pre_expr_d temp;
2066 PRE_EXPR_NAME (&temp) = vro->op0;
2067 temp.id = lookup_expression_id (&temp);
2070 if (!union_contains_value (set1, set2,
2071 get_expr_value_id (&temp)))
2074 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2076 struct pre_expr_d temp;
2079 PRE_EXPR_NAME (&temp) = vro->op1;
2080 temp.id = lookup_expression_id (&temp);
2083 if (!union_contains_value (set1, set2,
2084 get_expr_value_id (&temp)))
2088 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2090 struct pre_expr_d temp;
2093 PRE_EXPR_NAME (&temp) = vro->op2;
2094 temp.id = lookup_expression_id (&temp);
2097 if (!union_contains_value (set1, set2,
2098 get_expr_value_id (&temp)))
2105 /* Determine if the expression EXPR is valid in SET1 U SET2.
2106 ONLY SET2 CAN BE NULL.
2107 This means that we have a leader for each part of the expression
2108 (if it consists of values), or the expression is an SSA_NAME.
2109 For loads/calls, we also see if the vuse is killed in this block. */
2112 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2118 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2122 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2123 for (i = 0; i < nary->length; i++)
2125 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2127 struct pre_expr_d temp;
2130 PRE_EXPR_NAME (&temp) = nary->op[i];
2131 temp.id = lookup_expression_id (&temp);
2134 if (!union_contains_value (set1, set2,
2135 get_expr_value_id (&temp)))
2139 /* If the NARY may trap make sure the block does not contain
2140 a possible exit point.
2141 ??? This is overly conservative if we translate AVAIL_OUT
2142 as the available expression might be after the exit point. */
2143 if (BB_MAY_NOTRETURN (block)
2144 && vn_nary_may_trap (nary))
2151 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2152 vn_reference_op_t vro;
2155 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2157 if (!vro_valid_in_sets (set1, set2, vro))
2162 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2163 if (!gimple_nop_p (def_stmt)
2164 && gimple_bb (def_stmt) != block
2165 && !dominated_by_p (CDI_DOMINATORS,
2166 block, gimple_bb (def_stmt)))
2169 return !value_dies_in_block_x (expr, block);
2176 /* Clean the set of expressions that are no longer valid in SET1 or
2177 SET2. This means expressions that are made up of values we have no
2178 leaders for in SET1 or SET2. This version is used for partial
2179 anticipation, which means it is not valid in either ANTIC_IN or
2183 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2185 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2189 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2191 if (!valid_in_sets (set1, set2, expr, block))
2192 bitmap_remove_from_set (set1, expr);
2194 VEC_free (pre_expr, heap, exprs);
2197 /* Clean the set of expressions that are no longer valid in SET. This
2198 means expressions that are made up of values we have no leaders for
2202 clean (bitmap_set_t set, basic_block block)
2204 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2208 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2210 if (!valid_in_sets (set, NULL, expr, block))
2211 bitmap_remove_from_set (set, expr);
2213 VEC_free (pre_expr, heap, exprs);
2216 static sbitmap has_abnormal_preds;
2218 /* List of blocks that may have changed during ANTIC computation and
2219 thus need to be iterated over. */
2221 static sbitmap changed_blocks;
2223 /* Decide whether to defer a block for a later iteration, or PHI
2224 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2225 should defer the block, and true if we processed it. */
2228 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2229 basic_block block, basic_block phiblock)
2231 if (!BB_VISITED (phiblock))
2233 SET_BIT (changed_blocks, block->index);
2234 BB_VISITED (block) = 0;
2235 BB_DEFERRED (block) = 1;
2239 phi_translate_set (dest, source, block, phiblock);
2243 /* Compute the ANTIC set for BLOCK.
2245 If succs(BLOCK) > 1 then
2246 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2247 else if succs(BLOCK) == 1 then
2248 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2250 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2254 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2256 bool changed = false;
2257 bitmap_set_t S, old, ANTIC_OUT;
2263 old = ANTIC_OUT = S = NULL;
2264 BB_VISITED (block) = 1;
2266 /* If any edges from predecessors are abnormal, antic_in is empty,
2268 if (block_has_abnormal_pred_edge)
2269 goto maybe_dump_sets;
2271 old = ANTIC_IN (block);
2272 ANTIC_OUT = bitmap_set_new ();
2274 /* If the block has no successors, ANTIC_OUT is empty. */
2275 if (EDGE_COUNT (block->succs) == 0)
2277 /* If we have one successor, we could have some phi nodes to
2278 translate through. */
2279 else if (single_succ_p (block))
2281 basic_block succ_bb = single_succ (block);
2283 /* We trade iterations of the dataflow equations for having to
2284 phi translate the maximal set, which is incredibly slow
2285 (since the maximal set often has 300+ members, even when you
2286 have a small number of blocks).
2287 Basically, we defer the computation of ANTIC for this block
2288 until we have processed it's successor, which will inevitably
2289 have a *much* smaller set of values to phi translate once
2290 clean has been run on it.
2291 The cost of doing this is that we technically perform more
2292 iterations, however, they are lower cost iterations.
2294 Timings for PRE on tramp3d-v4:
2295 without maximal set fix: 11 seconds
2296 with maximal set fix/without deferring: 26 seconds
2297 with maximal set fix/with deferring: 11 seconds
2300 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2304 goto maybe_dump_sets;
2307 /* If we have multiple successors, we take the intersection of all of
2308 them. Note that in the case of loop exit phi nodes, we may have
2309 phis to translate through. */
2312 VEC(basic_block, heap) * worklist;
2314 basic_block bprime, first = NULL;
2316 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2317 FOR_EACH_EDGE (e, ei, block->succs)
2320 && BB_VISITED (e->dest))
2322 else if (BB_VISITED (e->dest))
2323 VEC_quick_push (basic_block, worklist, e->dest);
2326 /* Of multiple successors we have to have visited one already. */
2329 SET_BIT (changed_blocks, block->index);
2330 BB_VISITED (block) = 0;
2331 BB_DEFERRED (block) = 1;
2333 VEC_free (basic_block, heap, worklist);
2334 goto maybe_dump_sets;
2337 if (!gimple_seq_empty_p (phi_nodes (first)))
2338 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2340 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2342 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2344 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2346 bitmap_set_t tmp = bitmap_set_new ();
2347 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2348 bitmap_set_and (ANTIC_OUT, tmp);
2349 bitmap_set_free (tmp);
2352 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2354 VEC_free (basic_block, heap, worklist);
2357 /* Generate ANTIC_OUT - TMP_GEN. */
2358 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2360 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2361 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2364 /* Then union in the ANTIC_OUT - TMP_GEN values,
2365 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2366 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2367 bitmap_value_insert_into_set (ANTIC_IN (block),
2368 expression_for_id (bii));
2370 clean (ANTIC_IN (block), block);
2372 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2375 SET_BIT (changed_blocks, block->index);
2376 FOR_EACH_EDGE (e, ei, block->preds)
2377 SET_BIT (changed_blocks, e->src->index);
2380 RESET_BIT (changed_blocks, block->index);
2383 if (dump_file && (dump_flags & TDF_DETAILS))
2385 if (!BB_DEFERRED (block) || BB_VISITED (block))
2388 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2390 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2394 print_bitmap_set (dump_file, S, "S", block->index);
2399 "Block %d was deferred for a future iteration.\n",
2404 bitmap_set_free (old);
2406 bitmap_set_free (S);
2408 bitmap_set_free (ANTIC_OUT);
2412 /* Compute PARTIAL_ANTIC for BLOCK.
2414 If succs(BLOCK) > 1 then
2415 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2416 in ANTIC_OUT for all succ(BLOCK)
2417 else if succs(BLOCK) == 1 then
2418 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2420 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2425 compute_partial_antic_aux (basic_block block,
2426 bool block_has_abnormal_pred_edge)
2428 bool changed = false;
2429 bitmap_set_t old_PA_IN;
2430 bitmap_set_t PA_OUT;
2433 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2435 old_PA_IN = PA_OUT = NULL;
2437 /* If any edges from predecessors are abnormal, antic_in is empty,
2439 if (block_has_abnormal_pred_edge)
2440 goto maybe_dump_sets;
2442 /* If there are too many partially anticipatable values in the
2443 block, phi_translate_set can take an exponential time: stop
2444 before the translation starts. */
2446 && single_succ_p (block)
2447 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2448 goto maybe_dump_sets;
2450 old_PA_IN = PA_IN (block);
2451 PA_OUT = bitmap_set_new ();
2453 /* If the block has no successors, ANTIC_OUT is empty. */
2454 if (EDGE_COUNT (block->succs) == 0)
2456 /* If we have one successor, we could have some phi nodes to
2457 translate through. Note that we can't phi translate across DFS
2458 back edges in partial antic, because it uses a union operation on
2459 the successors. For recurrences like IV's, we will end up
2460 generating a new value in the set on each go around (i + 3 (VH.1)
2461 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2462 else if (single_succ_p (block))
2464 basic_block succ = single_succ (block);
2465 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2466 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2468 /* If we have multiple successors, we take the union of all of
2472 VEC(basic_block, heap) * worklist;
2476 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2477 FOR_EACH_EDGE (e, ei, block->succs)
2479 if (e->flags & EDGE_DFS_BACK)
2481 VEC_quick_push (basic_block, worklist, e->dest);
2483 if (VEC_length (basic_block, worklist) > 0)
2485 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2490 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2491 bitmap_value_insert_into_set (PA_OUT,
2492 expression_for_id (i));
2493 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2495 bitmap_set_t pa_in = bitmap_set_new ();
2496 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2497 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2498 bitmap_value_insert_into_set (PA_OUT,
2499 expression_for_id (i));
2500 bitmap_set_free (pa_in);
2503 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2504 bitmap_value_insert_into_set (PA_OUT,
2505 expression_for_id (i));
2508 VEC_free (basic_block, heap, worklist);
2511 /* PA_IN starts with PA_OUT - TMP_GEN.
2512 Then we subtract things from ANTIC_IN. */
2513 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2515 /* For partial antic, we want to put back in the phi results, since
2516 we will properly avoid making them partially antic over backedges. */
2517 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2518 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2520 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2521 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2523 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2525 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2528 SET_BIT (changed_blocks, block->index);
2529 FOR_EACH_EDGE (e, ei, block->preds)
2530 SET_BIT (changed_blocks, e->src->index);
2533 RESET_BIT (changed_blocks, block->index);
2536 if (dump_file && (dump_flags & TDF_DETAILS))
2539 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2541 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2544 bitmap_set_free (old_PA_IN);
2546 bitmap_set_free (PA_OUT);
2550 /* Compute ANTIC and partial ANTIC sets. */
2553 compute_antic (void)
2555 bool changed = true;
2556 int num_iterations = 0;
2560 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2561 We pre-build the map of blocks with incoming abnormal edges here. */
2562 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2563 sbitmap_zero (has_abnormal_preds);
2570 FOR_EACH_EDGE (e, ei, block->preds)
2572 e->flags &= ~EDGE_DFS_BACK;
2573 if (e->flags & EDGE_ABNORMAL)
2575 SET_BIT (has_abnormal_preds, block->index);
2580 BB_VISITED (block) = 0;
2581 BB_DEFERRED (block) = 0;
2583 /* While we are here, give empty ANTIC_IN sets to each block. */
2584 ANTIC_IN (block) = bitmap_set_new ();
2585 PA_IN (block) = bitmap_set_new ();
2588 /* At the exit block we anticipate nothing. */
2589 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2590 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2591 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2593 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2594 sbitmap_ones (changed_blocks);
2597 if (dump_file && (dump_flags & TDF_DETAILS))
2598 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2601 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2603 if (TEST_BIT (changed_blocks, postorder[i]))
2605 basic_block block = BASIC_BLOCK (postorder[i]);
2606 changed |= compute_antic_aux (block,
2607 TEST_BIT (has_abnormal_preds,
2611 #ifdef ENABLE_CHECKING
2612 /* Theoretically possible, but *highly* unlikely. */
2613 gcc_assert (num_iterations < 500);
2617 statistics_histogram_event (cfun, "compute_antic iterations",
2620 if (do_partial_partial)
2622 sbitmap_ones (changed_blocks);
2623 mark_dfs_back_edges ();
2628 if (dump_file && (dump_flags & TDF_DETAILS))
2629 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2632 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2634 if (TEST_BIT (changed_blocks, postorder[i]))
2636 basic_block block = BASIC_BLOCK (postorder[i]);
2638 |= compute_partial_antic_aux (block,
2639 TEST_BIT (has_abnormal_preds,
2643 #ifdef ENABLE_CHECKING
2644 /* Theoretically possible, but *highly* unlikely. */
2645 gcc_assert (num_iterations < 500);
2648 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2651 sbitmap_free (has_abnormal_preds);
2652 sbitmap_free (changed_blocks);
2655 /* Return true if we can value number the call in STMT. This is true
2656 if we have a pure or constant call. */
2659 can_value_number_call (gimple stmt)
2661 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2666 /* Return true if OP is a tree which we can perform PRE on.
2667 This may not match the operations we can value number, but in
2668 a perfect world would. */
2671 can_PRE_operation (tree op)
2673 return UNARY_CLASS_P (op)
2674 || BINARY_CLASS_P (op)
2675 || COMPARISON_CLASS_P (op)
2676 || TREE_CODE (op) == MEM_REF
2677 || TREE_CODE (op) == COMPONENT_REF
2678 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2679 || TREE_CODE (op) == CALL_EXPR
2680 || TREE_CODE (op) == ARRAY_REF;
2684 /* Inserted expressions are placed onto this worklist, which is used
2685 for performing quick dead code elimination of insertions we made
2686 that didn't turn out to be necessary. */
2687 static bitmap inserted_exprs;
2689 /* Pool allocated fake store expressions are placed onto this
2690 worklist, which, after performing dead code elimination, is walked
2691 to see which expressions need to be put into GC'able memory */
2692 static VEC(gimple, heap) *need_creation;
2694 /* The actual worker for create_component_ref_by_pieces. */
2697 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2698 unsigned int *operand, gimple_seq *stmts,
2701 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2705 switch (currop->opcode)
2709 tree folded, sc = NULL_TREE;
2710 unsigned int nargs = 0;
2712 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2716 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2717 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2723 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2724 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2728 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2729 ref->operands) - 1);
2730 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2732 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2742 folded = build_call_array (currop->type,
2743 (TREE_CODE (fn) == FUNCTION_DECL
2744 ? build_fold_addr_expr (fn) : fn),
2748 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2754 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2756 tree offset = currop->op0;
2759 if (TREE_CODE (baseop) == ADDR_EXPR
2760 && handled_component_p (TREE_OPERAND (baseop, 0)))
2764 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2767 offset = int_const_binop (PLUS_EXPR, offset,
2768 build_int_cst (TREE_TYPE (offset),
2770 baseop = build_fold_addr_expr (base);
2772 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2775 case TARGET_MEM_REF:
2777 pre_expr op0expr, op1expr;
2778 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2779 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2781 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2787 op0expr = get_or_alloc_expr_for (currop->op0);
2788 genop0 = find_or_generate_expression (block, op0expr,
2795 op1expr = get_or_alloc_expr_for (nextop->op0);
2796 genop1 = find_or_generate_expression (block, op1expr,
2801 return build5 (TARGET_MEM_REF, currop->type,
2802 baseop, currop->op2, genop0, currop->op1, genop1);
2808 gcc_assert (is_gimple_min_invariant (currop->op0));
2814 case VIEW_CONVERT_EXPR:
2817 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2822 folded = fold_build1 (currop->opcode, currop->type,
2830 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2832 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2833 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2839 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2842 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2845 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2850 /* For array ref vn_reference_op's, operand 1 of the array ref
2851 is op0 of the reference op and operand 3 of the array ref is
2853 case ARRAY_RANGE_REF:
2857 tree genop1 = currop->op0;
2859 tree genop2 = currop->op1;
2861 tree genop3 = currop->op2;
2863 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2867 op1expr = get_or_alloc_expr_for (genop1);
2868 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2873 /* Drop zero minimum index. */
2874 if (tree_int_cst_equal (genop2, integer_zero_node))
2878 op2expr = get_or_alloc_expr_for (genop2);
2879 genop2 = find_or_generate_expression (block, op2expr, stmts,
2887 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2888 /* We can't always put a size in units of the element alignment
2889 here as the element alignment may be not visible. See
2890 PR43783. Simply drop the element size for constant
2892 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2896 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2897 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2898 op3expr = get_or_alloc_expr_for (genop3);
2899 genop3 = find_or_generate_expression (block, op3expr, stmts,
2905 return build4 (currop->opcode, currop->type, genop0, genop1,
2912 tree genop2 = currop->op1;
2914 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2918 /* op1 should be a FIELD_DECL, which are represented by
2923 op2expr = get_or_alloc_expr_for (genop2);
2924 genop2 = find_or_generate_expression (block, op2expr, stmts,
2930 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2936 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2937 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2958 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2959 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2960 trying to rename aggregates into ssa form directly, which is a no no.
2962 Thus, this routine doesn't create temporaries, it just builds a
2963 single access expression for the array, calling
2964 find_or_generate_expression to build the innermost pieces.
2966 This function is a subroutine of create_expression_by_pieces, and
2967 should not be called on it's own unless you really know what you
2971 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2972 gimple_seq *stmts, gimple domstmt)
2974 unsigned int op = 0;
2975 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2978 /* Find a leader for an expression, or generate one using
2979 create_expression_by_pieces if it's ANTIC but
2981 BLOCK is the basic_block we are looking for leaders in.
2982 EXPR is the expression to find a leader or generate for.
2983 STMTS is the statement list to put the inserted expressions on.
2984 Returns the SSA_NAME of the LHS of the generated expression or the
2986 DOMSTMT if non-NULL is a statement that should be dominated by
2987 all uses in the generated expression. If DOMSTMT is non-NULL this
2988 routine can fail and return NULL_TREE. Otherwise it will assert
2992 find_or_generate_expression (basic_block block, pre_expr expr,
2993 gimple_seq *stmts, gimple domstmt)
2995 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2996 get_expr_value_id (expr), domstmt);
3000 if (leader->kind == NAME)
3001 genop = PRE_EXPR_NAME (leader);
3002 else if (leader->kind == CONSTANT)
3003 genop = PRE_EXPR_CONSTANT (leader);
3006 /* If it's still NULL, it must be a complex expression, so generate
3007 it recursively. Not so if inserting expressions for values generated
3012 bitmap_set_t exprset;
3013 unsigned int lookfor = get_expr_value_id (expr);
3014 bool handled = false;
3018 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3019 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3021 pre_expr temp = expression_for_id (i);
3022 if (temp->kind != NAME)
3025 genop = create_expression_by_pieces (block, temp, stmts,
3027 get_expr_type (expr));
3031 if (!handled && domstmt)
3034 gcc_assert (handled);
3039 #define NECESSARY GF_PLF_1
3041 /* Create an expression in pieces, so that we can handle very complex
3042 expressions that may be ANTIC, but not necessary GIMPLE.
3043 BLOCK is the basic block the expression will be inserted into,
3044 EXPR is the expression to insert (in value form)
3045 STMTS is a statement list to append the necessary insertions into.
3047 This function will die if we hit some value that shouldn't be
3048 ANTIC but is (IE there is no leader for it, or its components).
3049 This function may also generate expressions that are themselves
3050 partially or fully redundant. Those that are will be either made
3051 fully redundant during the next iteration of insert (for partially
3052 redundant ones), or eliminated by eliminate (for fully redundant
3055 If DOMSTMT is non-NULL then we make sure that all uses in the
3056 expressions dominate that statement. In this case the function
3057 can return NULL_TREE to signal failure. */
3060 create_expression_by_pieces (basic_block block, pre_expr expr,
3061 gimple_seq *stmts, gimple domstmt, tree type)
3065 gimple_seq forced_stmts = NULL;
3066 unsigned int value_id;
3067 gimple_stmt_iterator gsi;
3068 tree exprtype = type ? type : get_expr_type (expr);
3074 /* We may hit the NAME/CONSTANT case if we have to convert types
3075 that value numbering saw through. */
3077 folded = PRE_EXPR_NAME (expr);
3080 folded = PRE_EXPR_CONSTANT (expr);
3084 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3085 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3090 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3091 switch (nary->length)
3095 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3096 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
3097 tree genop1 = find_or_generate_expression (block, op1,
3099 tree genop2 = find_or_generate_expression (block, op2,
3101 if (!genop1 || !genop2)
3103 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
3104 may be a constant with the wrong type. */
3105 if (nary->opcode == POINTER_PLUS_EXPR)
3107 genop1 = fold_convert (nary->type, genop1);
3108 genop2 = fold_convert (sizetype, genop2);
3112 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3113 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
3116 folded = fold_build2 (nary->opcode, nary->type,
3122 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3123 tree genop1 = find_or_generate_expression (block, op1,
3127 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3129 folded = fold_build1 (nary->opcode, nary->type,
3142 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3143 folded = fold_convert (exprtype, folded);
3145 /* Force the generated expression to be a sequence of GIMPLE
3147 We have to call unshare_expr because force_gimple_operand may
3148 modify the tree we pass to it. */
3149 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3152 /* If we have any intermediate expressions to the value sets, add them
3153 to the value sets and chain them in the instruction stream. */
3156 gsi = gsi_start (forced_stmts);
3157 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3159 gimple stmt = gsi_stmt (gsi);
3160 tree forcedname = gimple_get_lhs (stmt);
3163 if (TREE_CODE (forcedname) == SSA_NAME)
3165 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3166 VN_INFO_GET (forcedname)->valnum = forcedname;
3167 VN_INFO (forcedname)->value_id = get_next_value_id ();
3168 nameexpr = get_or_alloc_expr_for_name (forcedname);
3169 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3171 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3172 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3174 mark_symbols_for_renaming (stmt);
3176 gimple_seq_add_seq (stmts, forced_stmts);
3179 /* Build and insert the assignment of the end result to the temporary
3180 that we will return. */
3181 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3183 pretemp = create_tmp_reg (exprtype, "pretmp");
3184 get_var_ann (pretemp);
3188 add_referenced_var (temp);
3190 newstmt = gimple_build_assign (temp, folded);
3191 name = make_ssa_name (temp, newstmt);
3192 gimple_assign_set_lhs (newstmt, name);
3193 gimple_set_plf (newstmt, NECESSARY, false);
3195 gimple_seq_add_stmt (stmts, newstmt);
3196 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3198 /* All the symbols in NEWEXPR should be put into SSA form. */
3199 mark_symbols_for_renaming (newstmt);
3201 /* Add a value number to the temporary.
3202 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3203 we are creating the expression by pieces, and this particular piece of
3204 the expression may have been represented. There is no harm in replacing
3206 VN_INFO_GET (name)->valnum = name;
3207 value_id = get_expr_value_id (expr);
3208 VN_INFO (name)->value_id = value_id;
3209 nameexpr = get_or_alloc_expr_for_name (name);
3210 add_to_value (value_id, nameexpr);
3211 if (NEW_SETS (block))
3212 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3213 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3215 pre_stats.insertions++;
3216 if (dump_file && (dump_flags & TDF_DETAILS))
3218 fprintf (dump_file, "Inserted ");
3219 print_gimple_stmt (dump_file, newstmt, 0, 0);
3220 fprintf (dump_file, " in predecessor %d\n", block->index);
3227 /* Returns true if we want to inhibit the insertions of PHI nodes
3228 for the given EXPR for basic block BB (a member of a loop).
3229 We want to do this, when we fear that the induction variable we
3230 create might inhibit vectorization. */
3233 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3235 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3236 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3237 vn_reference_op_t op;
3240 /* If we aren't going to vectorize we don't inhibit anything. */
3241 if (!flag_tree_vectorize)
3244 /* Otherwise we inhibit the insertion when the address of the
3245 memory reference is a simple induction variable. In other
3246 cases the vectorizer won't do anything anyway (either it's
3247 loop invariant or a complicated expression). */
3248 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3253 case ARRAY_RANGE_REF:
3254 if (TREE_CODE (op->op0) != SSA_NAME)
3259 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3261 /* Default defs are loop invariant. */
3264 /* Defined outside this loop, also loop invariant. */
3265 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3267 /* If it's a simple induction variable inhibit insertion,
3268 the vectorizer might be interested in this one. */
3269 if (simple_iv (bb->loop_father, bb->loop_father,
3270 op->op0, &iv, true))
3272 /* No simple IV, vectorizer can't do anything, hence no
3273 reason to inhibit the transformation for this operand. */
3283 /* Insert the to-be-made-available values of expression EXPRNUM for each
3284 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3285 merge the result with a phi node, given the same value number as
3286 NODE. Return true if we have inserted new stuff. */
3289 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3292 pre_expr expr = expression_for_id (exprnum);
3294 unsigned int val = get_expr_value_id (expr);
3296 bool insertions = false;
3301 tree type = get_expr_type (expr);
3305 if (dump_file && (dump_flags & TDF_DETAILS))
3307 fprintf (dump_file, "Found partial redundancy for expression ");
3308 print_pre_expr (dump_file, expr);
3309 fprintf (dump_file, " (%04d)\n", val);
3312 /* Make sure we aren't creating an induction variable. */
3313 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3315 bool firstinsideloop = false;
3316 bool secondinsideloop = false;
3317 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3318 EDGE_PRED (block, 0)->src);
3319 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3320 EDGE_PRED (block, 1)->src);
3321 /* Induction variables only have one edge inside the loop. */
3322 if ((firstinsideloop ^ secondinsideloop)
3323 && (expr->kind != REFERENCE
3324 || inhibit_phi_insertion (block, expr)))
3326 if (dump_file && (dump_flags & TDF_DETAILS))
3327 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3332 /* Make the necessary insertions. */
3333 FOR_EACH_EDGE (pred, ei, block->preds)
3335 gimple_seq stmts = NULL;
3338 eprime = avail[bprime->index];
3340 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3342 builtexpr = create_expression_by_pieces (bprime,
3346 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3347 gsi_insert_seq_on_edge (pred, stmts);
3348 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3351 else if (eprime->kind == CONSTANT)
3353 /* Constants may not have the right type, fold_convert
3354 should give us back a constant with the right type.
3356 tree constant = PRE_EXPR_CONSTANT (eprime);
3357 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3359 tree builtexpr = fold_convert (type, constant);
3360 if (!is_gimple_min_invariant (builtexpr))
3362 tree forcedexpr = force_gimple_operand (builtexpr,
3365 if (!is_gimple_min_invariant (forcedexpr))
3367 if (forcedexpr != builtexpr)
3369 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3370 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3374 gimple_stmt_iterator gsi;
3375 gsi = gsi_start (stmts);
3376 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3378 gimple stmt = gsi_stmt (gsi);
3379 tree lhs = gimple_get_lhs (stmt);
3380 if (TREE_CODE (lhs) == SSA_NAME)
3381 bitmap_set_bit (inserted_exprs,
3382 SSA_NAME_VERSION (lhs));
3383 gimple_set_plf (stmt, NECESSARY, false);
3385 gsi_insert_seq_on_edge (pred, stmts);
3387 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3391 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3394 else if (eprime->kind == NAME)
3396 /* We may have to do a conversion because our value
3397 numbering can look through types in certain cases, but
3398 our IL requires all operands of a phi node have the same
3400 tree name = PRE_EXPR_NAME (eprime);
3401 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3405 builtexpr = fold_convert (type, name);
3406 forcedexpr = force_gimple_operand (builtexpr,
3410 if (forcedexpr != name)
3412 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3413 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3418 gimple_stmt_iterator gsi;
3419 gsi = gsi_start (stmts);
3420 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3422 gimple stmt = gsi_stmt (gsi);
3423 tree lhs = gimple_get_lhs (stmt);
3424 if (TREE_CODE (lhs) == SSA_NAME)
3425 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3426 gimple_set_plf (stmt, NECESSARY, false);
3428 gsi_insert_seq_on_edge (pred, stmts);
3430 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3434 /* If we didn't want a phi node, and we made insertions, we still have
3435 inserted new stuff, and thus return true. If we didn't want a phi node,
3436 and didn't make insertions, we haven't added anything new, so return
3438 if (nophi && insertions)
3440 else if (nophi && !insertions)
3443 /* Now build a phi for the new variable. */
3444 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3446 prephitemp = create_tmp_var (type, "prephitmp");
3447 get_var_ann (prephitemp);
3451 add_referenced_var (temp);
3453 if (TREE_CODE (type) == COMPLEX_TYPE
3454 || TREE_CODE (type) == VECTOR_TYPE)
3455 DECL_GIMPLE_REG_P (temp) = 1;
3456 phi = create_phi_node (temp, block);
3458 gimple_set_plf (phi, NECESSARY, false);
3459 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3460 VN_INFO (gimple_phi_result (phi))->value_id = val;
3461 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3462 FOR_EACH_EDGE (pred, ei, block->preds)
3464 pre_expr ae = avail[pred->src->index];
3465 gcc_assert (get_expr_type (ae) == type
3466 || useless_type_conversion_p (type, get_expr_type (ae)));
3467 if (ae->kind == CONSTANT)
3468 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3470 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3474 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3475 add_to_value (val, newphi);
3477 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3478 this insertion, since we test for the existence of this value in PHI_GEN
3479 before proceeding with the partial redundancy checks in insert_aux.
3481 The value may exist in AVAIL_OUT, in particular, it could be represented
3482 by the expression we are trying to eliminate, in which case we want the
3483 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3486 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3487 this block, because if it did, it would have existed in our dominator's
3488 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3491 bitmap_insert_into_set (PHI_GEN (block), newphi);
3492 bitmap_value_replace_in_set (AVAIL_OUT (block),
3494 bitmap_insert_into_set (NEW_SETS (block),
3497 if (dump_file && (dump_flags & TDF_DETAILS))
3499 fprintf (dump_file, "Created phi ");
3500 print_gimple_stmt (dump_file, phi, 0, 0);
3501 fprintf (dump_file, " in block %d\n", block->index);
3509 /* Perform insertion of partially redundant values.
3510 For BLOCK, do the following:
3511 1. Propagate the NEW_SETS of the dominator into the current block.
3512 If the block has multiple predecessors,
3513 2a. Iterate over the ANTIC expressions for the block to see if
3514 any of them are partially redundant.
3515 2b. If so, insert them into the necessary predecessors to make
3516 the expression fully redundant.
3517 2c. Insert a new PHI merging the values of the predecessors.
3518 2d. Insert the new PHI, and the new expressions, into the
3520 3. Recursively call ourselves on the dominator children of BLOCK.
3522 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3523 do_regular_insertion and do_partial_insertion.
3528 do_regular_insertion (basic_block block, basic_block dom)
3530 bool new_stuff = false;
3531 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3535 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3537 if (expr->kind != NAME)
3541 bool by_some = false;
3542 bool cant_insert = false;
3543 bool all_same = true;
3544 pre_expr first_s = NULL;
3547 pre_expr eprime = NULL;
3549 pre_expr edoubleprime = NULL;
3550 bool do_insertion = false;
3552 val = get_expr_value_id (expr);
3553 if (bitmap_set_contains_value (PHI_GEN (block), val))
3555 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3557 if (dump_file && (dump_flags & TDF_DETAILS))
3558 fprintf (dump_file, "Found fully redundant value\n");
3562 avail = XCNEWVEC (pre_expr, last_basic_block);
3563 FOR_EACH_EDGE (pred, ei, block->preds)
3565 unsigned int vprime;
3567 /* We should never run insertion for the exit block
3568 and so not come across fake pred edges. */
3569 gcc_assert (!(pred->flags & EDGE_FAKE));
3571 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3574 /* eprime will generally only be NULL if the
3575 value of the expression, translated
3576 through the PHI for this predecessor, is
3577 undefined. If that is the case, we can't
3578 make the expression fully redundant,
3579 because its value is undefined along a
3580 predecessor path. We can thus break out
3581 early because it doesn't matter what the
3582 rest of the results are. */
3589 eprime = fully_constant_expression (eprime);
3590 vprime = get_expr_value_id (eprime);
3591 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3593 if (edoubleprime == NULL)
3595 avail[bprime->index] = eprime;
3600 avail[bprime->index] = edoubleprime;
3602 /* We want to perform insertions to remove a redundancy on
3603 a path in the CFG we want to optimize for speed. */
3604 if (optimize_edge_for_speed_p (pred))
3605 do_insertion = true;
3606 if (first_s == NULL)
3607 first_s = edoubleprime;
3608 else if (!pre_expr_eq (first_s, edoubleprime))
3612 /* If we can insert it, it's not the same value
3613 already existing along every predecessor, and
3614 it's defined by some predecessor, it is
3615 partially redundant. */
3616 if (!cant_insert && !all_same && by_some && do_insertion
3617 && dbg_cnt (treepre_insert))
3619 if (insert_into_preds_of_block (block, get_expression_id (expr),
3623 /* If all edges produce the same value and that value is
3624 an invariant, then the PHI has the same value on all
3625 edges. Note this. */
3626 else if (!cant_insert && all_same && eprime
3627 && (edoubleprime->kind == CONSTANT
3628 || edoubleprime->kind == NAME)
3629 && !value_id_constant_p (val))
3633 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3634 value_expressions, val);
3636 unsigned int new_val = get_expr_value_id (edoubleprime);
3637 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3639 pre_expr expr = expression_for_id (j);
3641 if (expr->kind == NAME)
3643 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3644 /* Just reset the value id and valnum so it is
3645 the same as the constant we have discovered. */
3646 if (edoubleprime->kind == CONSTANT)
3648 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3649 pre_stats.constified++;
3652 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3653 info->value_id = new_val;
3661 VEC_free (pre_expr, heap, exprs);
3666 /* Perform insertion for partially anticipatable expressions. There
3667 is only one case we will perform insertion for these. This case is
3668 if the expression is partially anticipatable, and fully available.
3669 In this case, we know that putting it earlier will enable us to
3670 remove the later computation. */
3674 do_partial_partial_insertion (basic_block block, basic_block dom)
3676 bool new_stuff = false;
3677 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3681 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3683 if (expr->kind != NAME)
3688 bool cant_insert = false;
3691 pre_expr eprime = NULL;
3694 val = get_expr_value_id (expr);
3695 if (bitmap_set_contains_value (PHI_GEN (block), val))
3697 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3700 avail = XCNEWVEC (pre_expr, last_basic_block);
3701 FOR_EACH_EDGE (pred, ei, block->preds)
3703 unsigned int vprime;
3704 pre_expr edoubleprime;
3706 /* We should never run insertion for the exit block
3707 and so not come across fake pred edges. */
3708 gcc_assert (!(pred->flags & EDGE_FAKE));
3710 eprime = phi_translate (expr, ANTIC_IN (block),
3714 /* eprime will generally only be NULL if the
3715 value of the expression, translated
3716 through the PHI for this predecessor, is
3717 undefined. If that is the case, we can't
3718 make the expression fully redundant,
3719 because its value is undefined along a
3720 predecessor path. We can thus break out
3721 early because it doesn't matter what the
3722 rest of the results are. */
3729 eprime = fully_constant_expression (eprime);
3730 vprime = get_expr_value_id (eprime);
3731 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3733 if (edoubleprime == NULL)
3739 avail[bprime->index] = edoubleprime;
3743 /* If we can insert it, it's not the same value
3744 already existing along every predecessor, and
3745 it's defined by some predecessor, it is
3746 partially redundant. */
3747 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3749 pre_stats.pa_insert++;
3750 if (insert_into_preds_of_block (block, get_expression_id (expr),
3758 VEC_free (pre_expr, heap, exprs);
3763 insert_aux (basic_block block)
3766 bool new_stuff = false;
3771 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3776 bitmap_set_t newset = NEW_SETS (dom);
3779 /* Note that we need to value_replace both NEW_SETS, and
3780 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3781 represented by some non-simple expression here that we want
3782 to replace it with. */
3783 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3785 pre_expr expr = expression_for_id (i);
3786 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3787 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3790 if (!single_pred_p (block))
3792 new_stuff |= do_regular_insertion (block, dom);
3793 if (do_partial_partial)
3794 new_stuff |= do_partial_partial_insertion (block, dom);
3798 for (son = first_dom_son (CDI_DOMINATORS, block);
3800 son = next_dom_son (CDI_DOMINATORS, son))
3802 new_stuff |= insert_aux (son);
3808 /* Perform insertion of partially redundant values. */
3813 bool new_stuff = true;
3815 int num_iterations = 0;
3818 NEW_SETS (bb) = bitmap_set_new ();
3823 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3825 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3829 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3832 add_to_exp_gen (basic_block block, tree op)
3837 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3839 result = get_or_alloc_expr_for_name (op);
3840 bitmap_value_insert_into_set (EXP_GEN (block), result);
3844 /* Create value ids for PHI in BLOCK. */
3847 make_values_for_phi (gimple phi, basic_block block)
3849 tree result = gimple_phi_result (phi);
3851 /* We have no need for virtual phis, as they don't represent
3852 actual computations. */
3853 if (is_gimple_reg (result))
3855 pre_expr e = get_or_alloc_expr_for_name (result);
3856 add_to_value (get_expr_value_id (e), e);
3857 bitmap_insert_into_set (PHI_GEN (block), e);
3858 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3862 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3864 tree arg = gimple_phi_arg_def (phi, i);
3865 if (TREE_CODE (arg) == SSA_NAME)
3867 e = get_or_alloc_expr_for_name (arg);
3868 add_to_value (get_expr_value_id (e), e);
3875 /* Compute the AVAIL set for all basic blocks.
3877 This function performs value numbering of the statements in each basic
3878 block. The AVAIL sets are built from information we glean while doing
3879 this value numbering, since the AVAIL sets contain only one entry per
3882 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3883 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3886 compute_avail (void)
3889 basic_block block, son;
3890 basic_block *worklist;
3894 /* We pretend that default definitions are defined in the entry block.
3895 This includes function arguments and the static chain decl. */
3896 for (i = 1; i < num_ssa_names; ++i)
3898 tree name = ssa_name (i);
3901 || !SSA_NAME_IS_DEFAULT_DEF (name)
3902 || has_zero_uses (name)
3903 || !is_gimple_reg (name))
3906 e = get_or_alloc_expr_for_name (name);
3907 add_to_value (get_expr_value_id (e), e);
3909 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3910 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3913 /* Allocate the worklist. */
3914 worklist = XNEWVEC (basic_block, n_basic_blocks);
3916 /* Seed the algorithm by putting the dominator children of the entry
3917 block on the worklist. */
3918 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3920 son = next_dom_son (CDI_DOMINATORS, son))
3921 worklist[sp++] = son;
3923 /* Loop until the worklist is empty. */
3926 gimple_stmt_iterator gsi;
3929 unsigned int stmt_uid = 1;
3931 /* Pick a block from the worklist. */
3932 block = worklist[--sp];
3934 /* Initially, the set of available values in BLOCK is that of
3935 its immediate dominator. */
3936 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3938 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3940 /* Generate values for PHI nodes. */
3941 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3942 make_values_for_phi (gsi_stmt (gsi), block);
3944 BB_MAY_NOTRETURN (block) = 0;
3946 /* Now compute value numbers and populate value sets with all
3947 the expressions computed in BLOCK. */
3948 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3953 stmt = gsi_stmt (gsi);
3954 gimple_set_uid (stmt, stmt_uid++);
3956 /* Cache whether the basic-block has any non-visible side-effect
3958 If this isn't a call or it is the last stmt in the
3959 basic-block then the CFG represents things correctly. */
3960 if (is_gimple_call (stmt)
3961 && !stmt_ends_bb_p (stmt))
3963 /* Non-looping const functions always return normally.
3964 Otherwise the call might not return or have side-effects
3965 that forbids hoisting possibly trapping expressions
3967 int flags = gimple_call_flags (stmt);
3968 if (!(flags & ECF_CONST)
3969 || (flags & ECF_LOOPING_CONST_OR_PURE))
3970 BB_MAY_NOTRETURN (block) = 1;
3973 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3975 pre_expr e = get_or_alloc_expr_for_name (op);
3977 add_to_value (get_expr_value_id (e), e);
3979 bitmap_insert_into_set (TMP_GEN (block), e);
3980 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3983 if (gimple_has_volatile_ops (stmt)
3984 || stmt_could_throw_p (stmt))
3987 switch (gimple_code (stmt))
3990 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3991 add_to_exp_gen (block, op);
3998 vn_reference_op_t vro;
3999 pre_expr result = NULL;
4000 VEC(vn_reference_op_s, heap) *ops = NULL;
4002 if (!can_value_number_call (stmt))
4005 copy_reference_ops_from_call (stmt, &ops);
4006 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4007 gimple_expr_type (stmt),
4009 VEC_free (vn_reference_op_s, heap, ops);
4013 for (i = 0; VEC_iterate (vn_reference_op_s,
4017 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4018 add_to_exp_gen (block, vro->op0);
4019 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4020 add_to_exp_gen (block, vro->op1);
4021 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4022 add_to_exp_gen (block, vro->op2);
4024 result = (pre_expr) pool_alloc (pre_expr_pool);
4025 result->kind = REFERENCE;
4027 PRE_EXPR_REFERENCE (result) = ref;
4029 get_or_alloc_expression_id (result);
4030 add_to_value (get_expr_value_id (result), result);
4032 bitmap_value_insert_into_set (EXP_GEN (block), result);
4038 pre_expr result = NULL;
4039 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4043 case tcc_comparison:
4048 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4049 gimple_assign_rhs_code (stmt),
4050 gimple_expr_type (stmt),
4051 gimple_assign_rhs1 (stmt),
4052 gimple_assign_rhs2 (stmt),
4053 NULL_TREE, NULL_TREE, &nary);
4058 for (i = 0; i < nary->length; i++)
4059 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4060 add_to_exp_gen (block, nary->op[i]);
4062 result = (pre_expr) pool_alloc (pre_expr_pool);
4063 result->kind = NARY;
4065 PRE_EXPR_NARY (result) = nary;
4069 case tcc_declaration:
4074 vn_reference_op_t vro;
4076 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4082 for (i = 0; VEC_iterate (vn_reference_op_s,
4086 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4087 add_to_exp_gen (block, vro->op0);
4088 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4089 add_to_exp_gen (block, vro->op1);
4090 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4091 add_to_exp_gen (block, vro->op2);
4093 result = (pre_expr) pool_alloc (pre_expr_pool);
4094 result->kind = REFERENCE;
4096 PRE_EXPR_REFERENCE (result) = ref;
4101 /* For any other statement that we don't
4102 recognize, simply add all referenced
4103 SSA_NAMEs to EXP_GEN. */
4104 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4105 add_to_exp_gen (block, op);
4109 get_or_alloc_expression_id (result);
4110 add_to_value (get_expr_value_id (result), result);
4112 bitmap_value_insert_into_set (EXP_GEN (block), result);
4121 /* Put the dominator children of BLOCK on the worklist of blocks
4122 to compute available sets for. */
4123 for (son = first_dom_son (CDI_DOMINATORS, block);
4125 son = next_dom_son (CDI_DOMINATORS, son))
4126 worklist[sp++] = son;
4132 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4133 than the available expressions for it. The insertion point is
4134 right before the first use in STMT. Returns the SSA_NAME that should
4135 be used for replacement. */
4138 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4140 basic_block bb = gimple_bb (stmt);
4141 gimple_stmt_iterator gsi;
4142 gimple_seq stmts = NULL;
4146 /* First create a value expression from the expression we want
4147 to insert and associate it with the value handle for SSA_VN. */
4148 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4152 /* Then use create_expression_by_pieces to generate a valid
4153 expression to insert at this point of the IL stream. */
4154 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4155 if (expr == NULL_TREE)
4157 gsi = gsi_for_stmt (stmt);
4158 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4163 /* Eliminate fully redundant computations. */
4168 VEC (gimple, heap) *to_remove = NULL;
4170 unsigned int todo = 0;
4171 gimple_stmt_iterator gsi;
4177 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4179 stmt = gsi_stmt (gsi);
4181 /* Lookup the RHS of the expression, see if we have an
4182 available computation for it. If so, replace the RHS with
4183 the available computation. */
4184 if (gimple_has_lhs (stmt)
4185 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4186 && !gimple_assign_ssa_name_copy_p (stmt)
4187 && (!gimple_assign_single_p (stmt)
4188 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4189 && !gimple_has_volatile_ops (stmt)
4190 && !has_zero_uses (gimple_get_lhs (stmt)))
4192 tree lhs = gimple_get_lhs (stmt);
4193 tree rhs = NULL_TREE;
4195 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4196 pre_expr sprimeexpr;
4198 if (gimple_assign_single_p (stmt))
4199 rhs = gimple_assign_rhs1 (stmt);
4201 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4202 get_expr_value_id (lhsexpr),
4207 if (sprimeexpr->kind == CONSTANT)
4208 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4209 else if (sprimeexpr->kind == NAME)
4210 sprime = PRE_EXPR_NAME (sprimeexpr);
4215 /* If there is no existing leader but SCCVN knows this
4216 value is constant, use that constant. */
4217 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4219 sprime = VN_INFO (lhs)->valnum;
4220 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4221 TREE_TYPE (sprime)))
4222 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4224 if (dump_file && (dump_flags & TDF_DETAILS))
4226 fprintf (dump_file, "Replaced ");
4227 print_gimple_expr (dump_file, stmt, 0, 0);
4228 fprintf (dump_file, " with ");
4229 print_generic_expr (dump_file, sprime, 0);
4230 fprintf (dump_file, " in ");
4231 print_gimple_stmt (dump_file, stmt, 0, 0);
4233 pre_stats.eliminations++;
4234 propagate_tree_value_into_stmt (&gsi, sprime);
4235 stmt = gsi_stmt (gsi);
4240 /* If there is no existing usable leader but SCCVN thinks
4241 it has an expression it wants to use as replacement,
4243 if (!sprime || sprime == lhs)
4245 tree val = VN_INFO (lhs)->valnum;
4247 && TREE_CODE (val) == SSA_NAME
4248 && VN_INFO (val)->needs_insertion
4249 && can_PRE_operation (vn_get_expr_for (val)))
4250 sprime = do_SCCVN_insertion (stmt, val);
4254 && (rhs == NULL_TREE
4255 || TREE_CODE (rhs) != SSA_NAME
4256 || may_propagate_copy (rhs, sprime)))
4258 bool can_make_abnormal_goto
4259 = is_gimple_call (stmt)
4260 && stmt_can_make_abnormal_goto (stmt);
4262 gcc_assert (sprime != rhs);
4264 if (dump_file && (dump_flags & TDF_DETAILS))
4266 fprintf (dump_file, "Replaced ");
4267 print_gimple_expr (dump_file, stmt, 0, 0);
4268 fprintf (dump_file, " with ");
4269 print_generic_expr (dump_file, sprime, 0);
4270 fprintf (dump_file, " in ");
4271 print_gimple_stmt (dump_file, stmt, 0, 0);
4274 if (TREE_CODE (sprime) == SSA_NAME)
4275 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4277 /* We need to make sure the new and old types actually match,
4278 which may require adding a simple cast, which fold_convert
4280 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4281 && !useless_type_conversion_p (gimple_expr_type (stmt),
4282 TREE_TYPE (sprime)))
4283 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4285 pre_stats.eliminations++;
4286 propagate_tree_value_into_stmt (&gsi, sprime);
4287 stmt = gsi_stmt (gsi);
4290 /* If we removed EH side-effects from the statement, clean
4291 its EH information. */
4292 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4294 bitmap_set_bit (need_eh_cleanup,
4295 gimple_bb (stmt)->index);
4296 if (dump_file && (dump_flags & TDF_DETAILS))
4297 fprintf (dump_file, " Removed EH side-effects.\n");
4300 /* Likewise for AB side-effects. */
4301 if (can_make_abnormal_goto
4302 && !stmt_can_make_abnormal_goto (stmt))
4304 bitmap_set_bit (need_ab_cleanup,
4305 gimple_bb (stmt)->index);
4306 if (dump_file && (dump_flags & TDF_DETAILS))
4307 fprintf (dump_file, " Removed AB 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 (is_gimple_call (stmt)
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 bool can_make_abnormal_goto
4372 = stmt_can_make_abnormal_goto (stmt);
4374 if (dump_file && (dump_flags & TDF_DETAILS))
4376 fprintf (dump_file, "Replacing call target with ");
4377 print_generic_expr (dump_file, fn, 0);
4378 fprintf (dump_file, " in ");
4379 print_gimple_stmt (dump_file, stmt, 0, 0);
4382 gimple_call_set_fn (stmt, fn);
4385 /* If we removed EH side-effects from the statement, clean
4386 its EH information. */
4387 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4389 bitmap_set_bit (need_eh_cleanup,
4390 gimple_bb (stmt)->index);
4391 if (dump_file && (dump_flags & TDF_DETAILS))
4392 fprintf (dump_file, " Removed EH side-effects.\n");
4395 /* Likewise for AB side-effects. */
4396 if (can_make_abnormal_goto
4397 && !stmt_can_make_abnormal_goto (stmt))
4399 bitmap_set_bit (need_ab_cleanup,
4400 gimple_bb (stmt)->index);
4401 if (dump_file && (dump_flags & TDF_DETAILS))
4402 fprintf (dump_file, " Removed AB side-effects.\n");
4405 /* Changing an indirect call to a direct call may
4406 have exposed different semantics. This may
4407 require an SSA update. */
4408 todo |= TODO_update_ssa_only_virtuals;
4413 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4415 gimple stmt, phi = gsi_stmt (gsi);
4416 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4417 pre_expr sprimeexpr, resexpr;
4418 gimple_stmt_iterator gsi2;
4420 /* We want to perform redundant PHI elimination. Do so by
4421 replacing the PHI with a single copy if possible.
4422 Do not touch inserted, single-argument or virtual PHIs. */
4423 if (gimple_phi_num_args (phi) == 1
4424 || !is_gimple_reg (res))
4430 resexpr = get_or_alloc_expr_for_name (res);
4431 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4432 get_expr_value_id (resexpr), NULL);
4435 if (sprimeexpr->kind == CONSTANT)
4436 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4437 else if (sprimeexpr->kind == NAME)
4438 sprime = PRE_EXPR_NAME (sprimeexpr);
4442 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4444 sprime = VN_INFO (res)->valnum;
4445 if (!useless_type_conversion_p (TREE_TYPE (res),
4446 TREE_TYPE (sprime)))
4447 sprime = fold_convert (TREE_TYPE (res), sprime);
4456 if (dump_file && (dump_flags & TDF_DETAILS))
4458 fprintf (dump_file, "Replaced redundant PHI node defining ");
4459 print_generic_expr (dump_file, res, 0);
4460 fprintf (dump_file, " with ");
4461 print_generic_expr (dump_file, sprime, 0);
4462 fprintf (dump_file, "\n");
4465 remove_phi_node (&gsi, false);
4467 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4468 && TREE_CODE (sprime) == SSA_NAME)
4469 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4471 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4472 sprime = fold_convert (TREE_TYPE (res), sprime);
4473 stmt = gimple_build_assign (res, sprime);
4474 SSA_NAME_DEF_STMT (res) = stmt;
4475 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4477 gsi2 = gsi_after_labels (b);
4478 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4479 /* Queue the copy for eventual removal. */
4480 VEC_safe_push (gimple, heap, to_remove, stmt);
4481 /* If we inserted this PHI node ourself, it's not an elimination. */
4482 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4485 pre_stats.eliminations++;
4489 /* We cannot remove stmts during BB walk, especially not release SSA
4490 names there as this confuses the VN machinery. The stmts ending
4491 up in to_remove are either stores or simple copies. */
4492 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4494 tree lhs = gimple_assign_lhs (stmt);
4495 tree rhs = gimple_assign_rhs1 (stmt);
4496 use_operand_p use_p;
4499 /* If there is a single use only, propagate the equivalency
4500 instead of keeping the copy. */
4501 if (TREE_CODE (lhs) == SSA_NAME
4502 && TREE_CODE (rhs) == SSA_NAME
4503 && single_imm_use (lhs, &use_p, &use_stmt)
4504 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4506 SET_USE (use_p, rhs);
4507 update_stmt (use_stmt);
4508 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4509 && TREE_CODE (rhs) == SSA_NAME)
4510 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4513 /* If this is a store or a now unused copy, remove it. */
4514 if (TREE_CODE (lhs) != SSA_NAME
4515 || has_zero_uses (lhs))
4517 basic_block bb = gimple_bb (stmt);
4518 gsi = gsi_for_stmt (stmt);
4519 unlink_stmt_vdef (stmt);
4520 gsi_remove (&gsi, true);
4521 if (gimple_purge_dead_eh_edges (bb))
4522 todo |= TODO_cleanup_cfg;
4523 if (TREE_CODE (lhs) == SSA_NAME)
4524 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4525 release_defs (stmt);
4528 VEC_free (gimple, heap, to_remove);
4533 /* Borrow a bit of tree-ssa-dce.c for the moment.
4534 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4535 this may be a bit faster, and we may want critical edges kept split. */
4537 /* If OP's defining statement has not already been determined to be necessary,
4538 mark that statement necessary. Return the stmt, if it is newly
4541 static inline gimple
4542 mark_operand_necessary (tree op)
4548 if (TREE_CODE (op) != SSA_NAME)
4551 stmt = SSA_NAME_DEF_STMT (op);
4554 if (gimple_plf (stmt, NECESSARY)
4555 || gimple_nop_p (stmt))
4558 gimple_set_plf (stmt, NECESSARY, true);
4562 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4563 to insert PHI nodes sometimes, and because value numbering of casts isn't
4564 perfect, we sometimes end up inserting dead code. This simple DCE-like
4565 pass removes any insertions we made that weren't actually used. */
4568 remove_dead_inserted_code (void)
4575 worklist = BITMAP_ALLOC (NULL);
4576 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4578 t = SSA_NAME_DEF_STMT (ssa_name (i));
4579 if (gimple_plf (t, NECESSARY))
4580 bitmap_set_bit (worklist, i);
4582 while (!bitmap_empty_p (worklist))
4584 i = bitmap_first_set_bit (worklist);
4585 bitmap_clear_bit (worklist, i);
4586 t = SSA_NAME_DEF_STMT (ssa_name (i));
4588 /* PHI nodes are somewhat special in that each PHI alternative has
4589 data and control dependencies. All the statements feeding the
4590 PHI node's arguments are always necessary. */
4591 if (gimple_code (t) == GIMPLE_PHI)
4595 for (k = 0; k < gimple_phi_num_args (t); k++)
4597 tree arg = PHI_ARG_DEF (t, k);
4598 if (TREE_CODE (arg) == SSA_NAME)
4600 gimple n = mark_operand_necessary (arg);
4602 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4608 /* Propagate through the operands. Examine all the USE, VUSE and
4609 VDEF operands in this statement. Mark all the statements
4610 which feed this statement's uses as necessary. */
4614 /* The operands of VDEF expressions are also needed as they
4615 represent potential definitions that may reach this
4616 statement (VDEF operands allow us to follow def-def
4619 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4621 gimple n = mark_operand_necessary (use);
4623 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4628 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4630 t = SSA_NAME_DEF_STMT (ssa_name (i));
4631 if (!gimple_plf (t, NECESSARY))
4633 gimple_stmt_iterator gsi;
4635 if (dump_file && (dump_flags & TDF_DETAILS))
4637 fprintf (dump_file, "Removing unnecessary insertion:");
4638 print_gimple_stmt (dump_file, t, 0, 0);
4641 gsi = gsi_for_stmt (t);
4642 if (gimple_code (t) == GIMPLE_PHI)
4643 remove_phi_node (&gsi, true);
4646 gsi_remove (&gsi, true);
4651 BITMAP_FREE (worklist);
4654 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4655 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4656 the number of visited blocks. */
4659 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4661 edge_iterator *stack;
4663 int post_order_num = 0;
4666 if (include_entry_exit)
4667 post_order[post_order_num++] = EXIT_BLOCK;
4669 /* Allocate stack for back-tracking up CFG. */
4670 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4673 /* Allocate bitmap to track nodes that have been visited. */
4674 visited = sbitmap_alloc (last_basic_block);
4676 /* None of the nodes in the CFG have been visited yet. */
4677 sbitmap_zero (visited);
4679 /* Push the last edge on to the stack. */
4680 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4688 /* Look at the edge on the top of the stack. */
4690 src = ei_edge (ei)->src;
4691 dest = ei_edge (ei)->dest;
4693 /* Check if the edge destination has been visited yet. */
4694 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4696 /* Mark that we have visited the destination. */
4697 SET_BIT (visited, src->index);
4699 if (EDGE_COUNT (src->preds) > 0)
4700 /* Since the DEST node has been visited for the first
4701 time, check its successors. */
4702 stack[sp++] = ei_start (src->preds);
4704 post_order[post_order_num++] = src->index;
4708 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4709 post_order[post_order_num++] = dest->index;
4711 if (!ei_one_before_end_p (ei))
4712 ei_next (&stack[sp - 1]);
4718 if (include_entry_exit)
4719 post_order[post_order_num++] = ENTRY_BLOCK;
4722 sbitmap_free (visited);
4723 return post_order_num;
4727 /* Initialize data structures used by PRE. */
4730 init_pre (bool do_fre)
4734 next_expression_id = 1;
4736 VEC_safe_push (pre_expr, heap, expressions, NULL);
4737 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4738 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4739 get_max_value_id() + 1);
4744 inserted_exprs = BITMAP_ALLOC (NULL);
4745 need_creation = NULL;
4746 pretemp = NULL_TREE;
4747 storetemp = NULL_TREE;
4748 prephitemp = NULL_TREE;
4750 connect_infinite_loops_to_exit ();
4751 memset (&pre_stats, 0, sizeof (pre_stats));
4754 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4755 my_rev_post_order_compute (postorder, false);
4757 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4759 calculate_dominance_info (CDI_POST_DOMINATORS);
4760 calculate_dominance_info (CDI_DOMINATORS);
4762 bitmap_obstack_initialize (&grand_bitmap_obstack);
4763 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4764 expr_pred_trans_eq, free);
4765 expression_to_id = htab_create (num_ssa_names * 3,
4768 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4769 sizeof (struct bitmap_set), 30);
4770 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4771 sizeof (struct pre_expr_d), 30);
4774 EXP_GEN (bb) = bitmap_set_new ();
4775 PHI_GEN (bb) = bitmap_set_new ();
4776 TMP_GEN (bb) = bitmap_set_new ();
4777 AVAIL_OUT (bb) = bitmap_set_new ();
4780 need_eh_cleanup = BITMAP_ALLOC (NULL);
4781 need_ab_cleanup = BITMAP_ALLOC (NULL);
4785 /* Deallocate data structures used by PRE. */
4788 fini_pre (bool do_fre)
4791 VEC_free (bitmap_set_t, heap, value_expressions);
4792 BITMAP_FREE (inserted_exprs);
4793 VEC_free (gimple, heap, need_creation);
4794 bitmap_obstack_release (&grand_bitmap_obstack);
4795 free_alloc_pool (bitmap_set_pool);
4796 free_alloc_pool (pre_expr_pool);
4797 htab_delete (phi_translate_table);
4798 htab_delete (expression_to_id);
4799 VEC_free (unsigned, heap, name_to_id);
4801 free_aux_for_blocks ();
4803 free_dominance_info (CDI_POST_DOMINATORS);
4805 if (!bitmap_empty_p (need_eh_cleanup))
4807 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4808 cleanup_tree_cfg ();
4811 BITMAP_FREE (need_eh_cleanup);
4813 if (!bitmap_empty_p (need_ab_cleanup))
4815 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4816 cleanup_tree_cfg ();
4819 BITMAP_FREE (need_ab_cleanup);
4822 loop_optimizer_finalize ();
4825 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4826 only wants to do full redundancy elimination. */
4829 execute_pre (bool do_fre)
4831 unsigned int todo = 0;
4833 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4835 /* This has to happen before SCCVN runs because
4836 loop_optimizer_init may create new phis, etc. */
4838 loop_optimizer_init (LOOPS_NORMAL);
4843 loop_optimizer_finalize ();
4851 /* Collect and value number expressions computed in each basic block. */
4854 if (dump_file && (dump_flags & TDF_DETAILS))
4860 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4861 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4862 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4863 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4867 /* Insert can get quite slow on an incredibly large number of basic
4868 blocks due to some quadratic behavior. Until this behavior is
4869 fixed, don't run it when he have an incredibly large number of
4870 bb's. If we aren't going to run insert, there is no point in
4871 computing ANTIC, either, even though it's plenty fast. */
4872 if (!do_fre && n_basic_blocks < 4000)
4878 /* Make sure to remove fake edges before committing our inserts.
4879 This makes sure we don't end up with extra critical edges that
4880 we would need to split. */
4881 remove_fake_exit_edges ();
4882 gsi_commit_edge_inserts ();
4884 /* Remove all the redundant expressions. */
4885 todo |= eliminate ();
4887 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4888 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4889 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4890 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4891 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4893 clear_expression_ids ();
4896 remove_dead_inserted_code ();
4904 /* Gate and execute functions for PRE. */
4909 return execute_pre (false);
4915 return flag_tree_pre != 0;
4918 struct gimple_opt_pass pass_pre =
4923 gate_pre, /* gate */
4924 do_pre, /* execute */
4927 0, /* static_pass_number */
4928 TV_TREE_PRE, /* tv_id */
4929 PROP_no_crit_edges | PROP_cfg
4930 | PROP_ssa, /* properties_required */
4931 0, /* properties_provided */
4932 0, /* properties_destroyed */
4933 TODO_rebuild_alias, /* todo_flags_start */
4934 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4935 | TODO_verify_ssa /* todo_flags_finish */
4940 /* Gate and execute functions for FRE. */
4945 return execute_pre (true);
4951 return flag_tree_fre != 0;
4954 struct gimple_opt_pass pass_fre =
4959 gate_fre, /* gate */
4960 execute_fre, /* execute */
4963 0, /* static_pass_number */
4964 TV_TREE_FRE, /* tv_id */
4965 PROP_cfg | PROP_ssa, /* properties_required */
4966 0, /* properties_provided */
4967 0, /* properties_destroyed */
4968 0, /* todo_flags_start */
4969 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */