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 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
900 /* Constant values are always considered to be part of the set. */
901 if (value_id_constant_p (val))
904 /* If the value membership changed, add the expression. */
905 if (bitmap_set_bit (&set->values, val))
906 bitmap_set_bit (&set->expressions, expr->id);
909 /* Print out EXPR to outfile. */
912 print_pre_expr (FILE *outfile, const pre_expr expr)
917 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
920 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
925 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
926 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
927 for (i = 0; i < nary->length; i++)
929 print_generic_expr (outfile, nary->op[i], 0);
930 if (i != (unsigned) nary->length - 1)
931 fprintf (outfile, ",");
933 fprintf (outfile, "}");
939 vn_reference_op_t vro;
941 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
942 fprintf (outfile, "{");
944 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
947 bool closebrace = false;
948 if (vro->opcode != SSA_NAME
949 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
951 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
954 fprintf (outfile, "<");
960 print_generic_expr (outfile, vro->op0, 0);
963 fprintf (outfile, ",");
964 print_generic_expr (outfile, vro->op1, 0);
968 fprintf (outfile, ",");
969 print_generic_expr (outfile, vro->op2, 0);
973 fprintf (outfile, ">");
974 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
975 fprintf (outfile, ",");
977 fprintf (outfile, "}");
980 fprintf (outfile, "@");
981 print_generic_expr (outfile, ref->vuse, 0);
987 void debug_pre_expr (pre_expr);
989 /* Like print_pre_expr but always prints to stderr. */
991 debug_pre_expr (pre_expr e)
993 print_pre_expr (stderr, e);
994 fprintf (stderr, "\n");
997 /* Print out SET to OUTFILE. */
1000 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1001 const char *setname, int blockindex)
1003 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1010 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1012 const pre_expr expr = expression_for_id (i);
1015 fprintf (outfile, ", ");
1017 print_pre_expr (outfile, expr);
1019 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1022 fprintf (outfile, " }\n");
1025 void debug_bitmap_set (bitmap_set_t);
1028 debug_bitmap_set (bitmap_set_t set)
1030 print_bitmap_set (stderr, set, "debug", 0);
1033 /* Print out the expressions that have VAL to OUTFILE. */
1036 print_value_expressions (FILE *outfile, unsigned int val)
1038 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1042 sprintf (s, "%04d", val);
1043 print_bitmap_set (outfile, set, s, 0);
1049 debug_value_expressions (unsigned int val)
1051 print_value_expressions (stderr, val);
1054 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1058 get_or_alloc_expr_for_constant (tree constant)
1060 unsigned int result_id;
1061 unsigned int value_id;
1062 struct pre_expr_d expr;
1065 expr.kind = CONSTANT;
1066 PRE_EXPR_CONSTANT (&expr) = constant;
1067 result_id = lookup_expression_id (&expr);
1069 return expression_for_id (result_id);
1071 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1072 newexpr->kind = CONSTANT;
1073 PRE_EXPR_CONSTANT (newexpr) = constant;
1074 alloc_expression_id (newexpr);
1075 value_id = get_or_alloc_constant_value_id (constant);
1076 add_to_value (value_id, newexpr);
1080 /* Given a value id V, find the actual tree representing the constant
1081 value if there is one, and return it. Return NULL if we can't find
1085 get_constant_for_value_id (unsigned int v)
1087 if (value_id_constant_p (v))
1091 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1093 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1095 pre_expr expr = expression_for_id (i);
1096 if (expr->kind == CONSTANT)
1097 return PRE_EXPR_CONSTANT (expr);
1103 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1104 Currently only supports constants and SSA_NAMES. */
1106 get_or_alloc_expr_for (tree t)
1108 if (TREE_CODE (t) == SSA_NAME)
1109 return get_or_alloc_expr_for_name (t);
1110 else if (is_gimple_min_invariant (t))
1111 return get_or_alloc_expr_for_constant (t);
1114 /* More complex expressions can result from SCCVN expression
1115 simplification that inserts values for them. As they all
1116 do not have VOPs the get handled by the nary ops struct. */
1117 vn_nary_op_t result;
1118 unsigned int result_id;
1119 vn_nary_op_lookup (t, &result);
1122 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1124 PRE_EXPR_NARY (e) = result;
1125 result_id = lookup_expression_id (e);
1128 pool_free (pre_expr_pool, e);
1129 e = expression_for_id (result_id);
1132 alloc_expression_id (e);
1139 /* Return the folded version of T if T, when folded, is a gimple
1140 min_invariant. Otherwise, return T. */
1143 fully_constant_expression (pre_expr e)
1151 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1152 switch (TREE_CODE_CLASS (nary->opcode))
1154 case tcc_expression:
1155 if (nary->opcode == TRUTH_NOT_EXPR)
1157 if (nary->opcode != TRUTH_AND_EXPR
1158 && nary->opcode != TRUTH_OR_EXPR
1159 && nary->opcode != TRUTH_XOR_EXPR)
1163 case tcc_comparison:
1165 /* We have to go from trees to pre exprs to value ids to
1167 tree naryop0 = nary->op[0];
1168 tree naryop1 = nary->op[1];
1170 if (!is_gimple_min_invariant (naryop0))
1172 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1173 unsigned int vrep0 = get_expr_value_id (rep0);
1174 tree const0 = get_constant_for_value_id (vrep0);
1176 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1178 if (!is_gimple_min_invariant (naryop1))
1180 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1181 unsigned int vrep1 = get_expr_value_id (rep1);
1182 tree const1 = get_constant_for_value_id (vrep1);
1184 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1186 result = fold_binary (nary->opcode, nary->type,
1188 if (result && is_gimple_min_invariant (result))
1189 return get_or_alloc_expr_for_constant (result);
1190 /* We might have simplified the expression to a
1191 SSA_NAME for example from x_1 * 1. But we cannot
1192 insert a PHI for x_1 unconditionally as x_1 might
1193 not be available readily. */
1197 if (nary->opcode != REALPART_EXPR
1198 && nary->opcode != IMAGPART_EXPR
1199 && nary->opcode != VIEW_CONVERT_EXPR)
1205 /* We have to go from trees to pre exprs to value ids to
1207 tree naryop0 = nary->op[0];
1208 tree const0, result;
1209 if (is_gimple_min_invariant (naryop0))
1213 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1214 unsigned int vrep0 = get_expr_value_id (rep0);
1215 const0 = get_constant_for_value_id (vrep0);
1220 tree type1 = TREE_TYPE (nary->op[0]);
1221 const0 = fold_convert (type1, const0);
1222 result = fold_unary (nary->opcode, nary->type, const0);
1224 if (result && is_gimple_min_invariant (result))
1225 return get_or_alloc_expr_for_constant (result);
1234 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1236 if ((folded = fully_constant_vn_reference_p (ref)))
1237 return get_or_alloc_expr_for_constant (folded);
1246 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1247 it has the value it would have in BLOCK. Set *SAME_VALID to true
1248 in case the new vuse doesn't change the value id of the OPERANDS. */
1251 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1252 alias_set_type set, tree type, tree vuse,
1253 basic_block phiblock,
1254 basic_block block, bool *same_valid)
1256 gimple phi = SSA_NAME_DEF_STMT (vuse);
1263 if (gimple_bb (phi) != phiblock)
1266 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1268 /* Use the alias-oracle to find either the PHI node in this block,
1269 the first VUSE used in this block that is equivalent to vuse or
1270 the first VUSE which definition in this block kills the value. */
1271 if (gimple_code (phi) == GIMPLE_PHI)
1272 e = find_edge (block, phiblock);
1273 else if (use_oracle)
1274 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1276 vuse = gimple_vuse (phi);
1277 phi = SSA_NAME_DEF_STMT (vuse);
1278 if (gimple_bb (phi) != phiblock)
1280 if (gimple_code (phi) == GIMPLE_PHI)
1282 e = find_edge (block, phiblock);
1293 bitmap visited = NULL;
1294 /* Try to find a vuse that dominates this phi node by skipping
1295 non-clobbering statements. */
1296 vuse = get_continuation_for_phi (phi, &ref, &visited);
1298 BITMAP_FREE (visited);
1304 /* If we didn't find any, the value ID can't stay the same,
1305 but return the translated vuse. */
1306 *same_valid = false;
1307 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1309 /* ??? We would like to return vuse here as this is the canonical
1310 upmost vdef that this reference is associated with. But during
1311 insertion of the references into the hash tables we only ever
1312 directly insert with their direct gimple_vuse, hence returning
1313 something else would make us not find the other expression. */
1314 return PHI_ARG_DEF (phi, e->dest_idx);
1320 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1321 SET2. This is used to avoid making a set consisting of the union
1322 of PA_IN and ANTIC_IN during insert. */
1324 static inline pre_expr
1325 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1329 result = bitmap_find_leader (set1, val, NULL);
1330 if (!result && set2)
1331 result = bitmap_find_leader (set2, val, NULL);
1335 /* Get the tree type for our PRE expression e. */
1338 get_expr_type (const pre_expr e)
1343 return TREE_TYPE (PRE_EXPR_NAME (e));
1345 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1347 return PRE_EXPR_REFERENCE (e)->type;
1349 return PRE_EXPR_NARY (e)->type;
1354 /* Get a representative SSA_NAME for a given expression.
1355 Since all of our sub-expressions are treated as values, we require
1356 them to be SSA_NAME's for simplicity.
1357 Prior versions of GVNPRE used to use "value handles" here, so that
1358 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1359 either case, the operands are really values (IE we do not expect
1360 them to be usable without finding leaders). */
1363 get_representative_for (const pre_expr e)
1367 unsigned int value_id = get_expr_value_id (e);
1372 return PRE_EXPR_NAME (e);
1374 return PRE_EXPR_CONSTANT (e);
1378 /* Go through all of the expressions representing this value
1379 and pick out an SSA_NAME. */
1382 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1384 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1386 pre_expr rep = expression_for_id (i);
1387 if (rep->kind == NAME)
1388 return PRE_EXPR_NAME (rep);
1393 /* If we reached here we couldn't find an SSA_NAME. This can
1394 happen when we've discovered a value that has never appeared in
1395 the program as set to an SSA_NAME, most likely as the result of
1400 "Could not find SSA_NAME representative for expression:");
1401 print_pre_expr (dump_file, e);
1402 fprintf (dump_file, "\n");
1405 exprtype = get_expr_type (e);
1407 /* Build and insert the assignment of the end result to the temporary
1408 that we will return. */
1409 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1411 pretemp = create_tmp_reg (exprtype, "pretmp");
1412 get_var_ann (pretemp);
1415 name = make_ssa_name (pretemp, gimple_build_nop ());
1416 VN_INFO_GET (name)->value_id = value_id;
1417 if (e->kind == CONSTANT)
1418 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1420 VN_INFO (name)->valnum = name;
1422 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1425 fprintf (dump_file, "Created SSA_NAME representative ");
1426 print_generic_expr (dump_file, name, 0);
1427 fprintf (dump_file, " for expression:");
1428 print_pre_expr (dump_file, e);
1429 fprintf (dump_file, "\n");
1438 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1439 basic_block pred, basic_block phiblock);
1441 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1442 the phis in PRED. Return NULL if we can't find a leader for each part
1443 of the translated expression. */
1446 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1447 basic_block pred, basic_block phiblock)
1454 bool changed = false;
1455 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1456 struct vn_nary_op_s newnary;
1457 /* The NARY structure is only guaranteed to have been
1458 allocated to the nary->length operands. */
1459 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1460 - sizeof (tree) * (4 - nary->length)));
1462 for (i = 0; i < newnary.length; i++)
1464 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1468 pre_expr leader, result;
1469 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1470 leader = find_leader_in_sets (op_val_id, set1, set2);
1471 result = phi_translate (leader, set1, set2, pred, phiblock);
1472 if (result && result != leader)
1474 tree name = get_representative_for (result);
1477 newnary.op[i] = name;
1482 changed |= newnary.op[i] != nary->op[i];
1488 unsigned int new_val_id;
1490 tree result = vn_nary_op_lookup_pieces (newnary.length,
1498 if (result && is_gimple_min_invariant (result))
1499 return get_or_alloc_expr_for_constant (result);
1501 expr = (pre_expr) pool_alloc (pre_expr_pool);
1506 PRE_EXPR_NARY (expr) = nary;
1507 constant = fully_constant_expression (expr);
1508 if (constant != expr)
1511 new_val_id = nary->value_id;
1512 get_or_alloc_expression_id (expr);
1516 new_val_id = get_next_value_id ();
1517 VEC_safe_grow_cleared (bitmap_set_t, heap,
1519 get_max_value_id() + 1);
1520 nary = vn_nary_op_insert_pieces (newnary.length,
1527 result, new_val_id);
1528 PRE_EXPR_NARY (expr) = nary;
1529 constant = fully_constant_expression (expr);
1530 if (constant != expr)
1532 get_or_alloc_expression_id (expr);
1534 add_to_value (new_val_id, expr);
1542 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1543 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1544 tree vuse = ref->vuse;
1545 tree newvuse = vuse;
1546 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1547 bool changed = false, same_valid = true;
1548 unsigned int i, j, n;
1549 vn_reference_op_t operand;
1550 vn_reference_t newref;
1553 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1558 tree type = operand->type;
1559 vn_reference_op_s newop = *operand;
1560 op[0] = operand->op0;
1561 op[1] = operand->op1;
1562 op[2] = operand->op2;
1563 for (n = 0; n < 3; ++n)
1565 unsigned int op_val_id;
1568 if (TREE_CODE (op[n]) != SSA_NAME)
1570 /* We can't possibly insert these. */
1572 && !is_gimple_min_invariant (op[n]))
1576 op_val_id = VN_INFO (op[n])->value_id;
1577 leader = find_leader_in_sets (op_val_id, set1, set2);
1580 /* Make sure we do not recursively translate ourselves
1581 like for translating a[n_1] with the leader for
1582 n_1 being a[n_1]. */
1583 if (get_expression_id (leader) != get_expression_id (expr))
1585 opresult = phi_translate (leader, set1, set2,
1589 if (opresult != leader)
1591 tree name = get_representative_for (opresult);
1594 changed |= name != op[n];
1602 VEC_free (vn_reference_op_s, heap, newoperands);
1606 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1607 /* We may have changed from an SSA_NAME to a constant */
1608 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1609 newop.opcode = TREE_CODE (op[0]);
1614 /* If it transforms a non-constant ARRAY_REF into a constant
1615 one, adjust the constant offset. */
1616 if (newop.opcode == ARRAY_REF
1618 && TREE_CODE (op[0]) == INTEGER_CST
1619 && TREE_CODE (op[1]) == INTEGER_CST
1620 && TREE_CODE (op[2]) == INTEGER_CST)
1622 double_int off = tree_to_double_int (op[0]);
1623 off = double_int_add (off,
1625 (tree_to_double_int (op[1])));
1626 off = double_int_mul (off, tree_to_double_int (op[2]));
1627 if (double_int_fits_in_shwi_p (off))
1628 newop.off = off.low;
1630 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1631 /* If it transforms from an SSA_NAME to an address, fold with
1632 a preceding indirect reference. */
1633 if (j > 0 && op[0] && TREE_CODE (op[0]) == ADDR_EXPR
1634 && VEC_index (vn_reference_op_s,
1635 newoperands, j - 1)->opcode == MEM_REF)
1636 vn_reference_fold_indirect (&newoperands, &j);
1638 if (i != VEC_length (vn_reference_op_s, operands))
1641 VEC_free (vn_reference_op_s, heap, newoperands);
1647 newvuse = translate_vuse_through_block (newoperands,
1648 ref->set, ref->type,
1649 vuse, phiblock, pred,
1651 if (newvuse == NULL_TREE)
1653 VEC_free (vn_reference_op_s, heap, newoperands);
1658 if (changed || newvuse != vuse)
1660 unsigned int new_val_id;
1662 bool converted = false;
1664 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1669 VEC_free (vn_reference_op_s, heap, newoperands);
1672 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1674 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1677 else if (!result && newref
1678 && !useless_type_conversion_p (ref->type, newref->type))
1680 VEC_free (vn_reference_op_s, heap, newoperands);
1684 if (result && is_gimple_min_invariant (result))
1686 gcc_assert (!newoperands);
1687 return get_or_alloc_expr_for_constant (result);
1690 expr = (pre_expr) pool_alloc (pre_expr_pool);
1691 expr->kind = REFERENCE;
1699 gcc_assert (CONVERT_EXPR_P (result)
1700 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1702 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1704 TREE_OPERAND (result, 0),
1705 NULL_TREE, NULL_TREE,
1708 if (nresult && is_gimple_min_invariant (nresult))
1709 return get_or_alloc_expr_for_constant (nresult);
1714 PRE_EXPR_NARY (expr) = nary;
1715 constant = fully_constant_expression (expr);
1716 if (constant != expr)
1719 new_val_id = nary->value_id;
1720 get_or_alloc_expression_id (expr);
1724 new_val_id = get_next_value_id ();
1725 VEC_safe_grow_cleared (bitmap_set_t, heap,
1727 get_max_value_id() + 1);
1728 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1730 TREE_OPERAND (result, 0),
1731 NULL_TREE, NULL_TREE,
1732 NULL_TREE, NULL_TREE,
1734 PRE_EXPR_NARY (expr) = nary;
1735 constant = fully_constant_expression (expr);
1736 if (constant != expr)
1738 get_or_alloc_expression_id (expr);
1743 PRE_EXPR_REFERENCE (expr) = newref;
1744 constant = fully_constant_expression (expr);
1745 if (constant != expr)
1748 new_val_id = newref->value_id;
1749 get_or_alloc_expression_id (expr);
1753 if (changed || !same_valid)
1755 new_val_id = get_next_value_id ();
1756 VEC_safe_grow_cleared (bitmap_set_t, heap,
1758 get_max_value_id() + 1);
1761 new_val_id = ref->value_id;
1762 newref = vn_reference_insert_pieces (newvuse, ref->set,
1765 result, new_val_id);
1767 PRE_EXPR_REFERENCE (expr) = newref;
1768 constant = fully_constant_expression (expr);
1769 if (constant != expr)
1771 get_or_alloc_expression_id (expr);
1773 add_to_value (new_val_id, expr);
1775 VEC_free (vn_reference_op_s, heap, newoperands);
1785 tree name = PRE_EXPR_NAME (expr);
1787 def_stmt = SSA_NAME_DEF_STMT (name);
1788 if (gimple_code (def_stmt) == GIMPLE_PHI
1789 && gimple_bb (def_stmt) == phiblock)
1794 e = find_edge (pred, gimple_bb (phi));
1797 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1800 if (TREE_CODE (def) == SSA_NAME)
1801 def = VN_INFO (def)->valnum;
1803 /* Handle constant. */
1804 if (is_gimple_min_invariant (def))
1805 return get_or_alloc_expr_for_constant (def);
1807 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1810 newexpr = get_or_alloc_expr_for_name (def);
1821 /* Wrapper around phi_translate_1 providing caching functionality. */
1824 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1825 basic_block pred, basic_block phiblock)
1832 /* Constants contain no values that need translation. */
1833 if (expr->kind == CONSTANT)
1836 if (value_id_constant_p (get_expr_value_id (expr)))
1839 if (expr->kind != NAME)
1841 phitrans = phi_trans_lookup (expr, pred);
1847 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1849 /* Don't add empty translations to the cache. Neither add
1850 translations of NAMEs as those are cheap to translate. */
1852 && expr->kind != NAME)
1853 phi_trans_add (expr, phitrans, pred);
1859 /* For each expression in SET, translate the values through phi nodes
1860 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1861 expressions in DEST. */
1864 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1865 basic_block phiblock)
1867 VEC (pre_expr, heap) *exprs;
1871 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1873 bitmap_set_copy (dest, set);
1877 exprs = sorted_array_from_bitmap_set (set);
1878 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1880 pre_expr translated;
1881 translated = phi_translate (expr, set, NULL, pred, phiblock);
1885 /* We might end up with multiple expressions from SET being
1886 translated to the same value. In this case we do not want
1887 to retain the NARY or REFERENCE expression but prefer a NAME
1888 which would be the leader. */
1889 if (translated->kind == NAME)
1890 bitmap_value_replace_in_set (dest, translated);
1892 bitmap_value_insert_into_set (dest, translated);
1894 VEC_free (pre_expr, heap, exprs);
1897 /* Find the leader for a value (i.e., the name representing that
1898 value) in a given set, and return it. If STMT is non-NULL it
1899 makes sure the defining statement for the leader dominates it.
1900 Return NULL if no leader is found. */
1903 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1905 if (value_id_constant_p (val))
1909 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1911 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1913 pre_expr expr = expression_for_id (i);
1914 if (expr->kind == CONSTANT)
1918 if (bitmap_set_contains_value (set, val))
1920 /* Rather than walk the entire bitmap of expressions, and see
1921 whether any of them has the value we are looking for, we look
1922 at the reverse mapping, which tells us the set of expressions
1923 that have a given value (IE value->expressions with that
1924 value) and see if any of those expressions are in our set.
1925 The number of expressions per value is usually significantly
1926 less than the number of expressions in the set. In fact, for
1927 large testcases, doing it this way is roughly 5-10x faster
1928 than walking the bitmap.
1929 If this is somehow a significant lose for some cases, we can
1930 choose which set to walk based on which set is smaller. */
1933 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1935 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1936 &set->expressions, 0, i, bi)
1938 pre_expr val = expression_for_id (i);
1939 /* At the point where stmt is not null, there should always
1940 be an SSA_NAME first in the list of expressions. */
1943 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1944 if (gimple_code (def_stmt) != GIMPLE_PHI
1945 && gimple_bb (def_stmt) == gimple_bb (stmt)
1946 /* PRE insertions are at the end of the basic-block
1948 && (gimple_uid (def_stmt) == 0
1949 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1958 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1959 BLOCK by seeing if it is not killed in the block. Note that we are
1960 only determining whether there is a store that kills it. Because
1961 of the order in which clean iterates over values, we are guaranteed
1962 that altered operands will have caused us to be eliminated from the
1963 ANTIC_IN set already. */
1966 value_dies_in_block_x (pre_expr expr, basic_block block)
1968 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1969 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1971 gimple_stmt_iterator gsi;
1972 unsigned id = get_expression_id (expr);
1979 /* Lookup a previously calculated result. */
1980 if (EXPR_DIES (block)
1981 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1982 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1984 /* A memory expression {e, VUSE} dies in the block if there is a
1985 statement that may clobber e. If, starting statement walk from the
1986 top of the basic block, a statement uses VUSE there can be no kill
1987 inbetween that use and the original statement that loaded {e, VUSE},
1988 so we can stop walking. */
1989 ref.base = NULL_TREE;
1990 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1992 tree def_vuse, def_vdef;
1993 def = gsi_stmt (gsi);
1994 def_vuse = gimple_vuse (def);
1995 def_vdef = gimple_vdef (def);
1997 /* Not a memory statement. */
2001 /* Not a may-def. */
2004 /* A load with the same VUSE, we're done. */
2005 if (def_vuse == vuse)
2011 /* Init ref only if we really need it. */
2012 if (ref.base == NULL_TREE
2013 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
2019 /* If the statement may clobber expr, it dies. */
2020 if (stmt_may_clobber_ref_p_1 (def, &ref))
2027 /* Remember the result. */
2028 if (!EXPR_DIES (block))
2029 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2030 bitmap_set_bit (EXPR_DIES (block), id * 2);
2032 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2038 #define union_contains_value(SET1, SET2, VAL) \
2039 (bitmap_set_contains_value ((SET1), (VAL)) \
2040 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
2042 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2045 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2046 vn_reference_op_t vro)
2048 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2050 struct pre_expr_d temp;
2053 PRE_EXPR_NAME (&temp) = vro->op0;
2054 temp.id = lookup_expression_id (&temp);
2057 if (!union_contains_value (set1, set2,
2058 get_expr_value_id (&temp)))
2061 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2063 struct pre_expr_d temp;
2066 PRE_EXPR_NAME (&temp) = vro->op1;
2067 temp.id = lookup_expression_id (&temp);
2070 if (!union_contains_value (set1, set2,
2071 get_expr_value_id (&temp)))
2075 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2077 struct pre_expr_d temp;
2080 PRE_EXPR_NAME (&temp) = vro->op2;
2081 temp.id = lookup_expression_id (&temp);
2084 if (!union_contains_value (set1, set2,
2085 get_expr_value_id (&temp)))
2092 /* Determine if the expression EXPR is valid in SET1 U SET2.
2093 ONLY SET2 CAN BE NULL.
2094 This means that we have a leader for each part of the expression
2095 (if it consists of values), or the expression is an SSA_NAME.
2096 For loads/calls, we also see if the vuse is killed in this block. */
2099 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2105 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2109 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2110 for (i = 0; i < nary->length; i++)
2112 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2114 struct pre_expr_d temp;
2117 PRE_EXPR_NAME (&temp) = nary->op[i];
2118 temp.id = lookup_expression_id (&temp);
2121 if (!union_contains_value (set1, set2,
2122 get_expr_value_id (&temp)))
2126 /* If the NARY may trap make sure the block does not contain
2127 a possible exit point.
2128 ??? This is overly conservative if we translate AVAIL_OUT
2129 as the available expression might be after the exit point. */
2130 if (BB_MAY_NOTRETURN (block)
2131 && vn_nary_may_trap (nary))
2138 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2139 vn_reference_op_t vro;
2142 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2144 if (!vro_valid_in_sets (set1, set2, vro))
2149 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2150 if (!gimple_nop_p (def_stmt)
2151 && gimple_bb (def_stmt) != block
2152 && !dominated_by_p (CDI_DOMINATORS,
2153 block, gimple_bb (def_stmt)))
2156 return !value_dies_in_block_x (expr, block);
2163 /* Clean the set of expressions that are no longer valid in SET1 or
2164 SET2. This means expressions that are made up of values we have no
2165 leaders for in SET1 or SET2. This version is used for partial
2166 anticipation, which means it is not valid in either ANTIC_IN or
2170 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2172 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2176 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2178 if (!valid_in_sets (set1, set2, expr, block))
2179 bitmap_remove_from_set (set1, expr);
2181 VEC_free (pre_expr, heap, exprs);
2184 /* Clean the set of expressions that are no longer valid in SET. This
2185 means expressions that are made up of values we have no leaders for
2189 clean (bitmap_set_t set, basic_block block)
2191 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2195 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2197 if (!valid_in_sets (set, NULL, expr, block))
2198 bitmap_remove_from_set (set, expr);
2200 VEC_free (pre_expr, heap, exprs);
2203 static sbitmap has_abnormal_preds;
2205 /* List of blocks that may have changed during ANTIC computation and
2206 thus need to be iterated over. */
2208 static sbitmap changed_blocks;
2210 /* Decide whether to defer a block for a later iteration, or PHI
2211 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2212 should defer the block, and true if we processed it. */
2215 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2216 basic_block block, basic_block phiblock)
2218 if (!BB_VISITED (phiblock))
2220 SET_BIT (changed_blocks, block->index);
2221 BB_VISITED (block) = 0;
2222 BB_DEFERRED (block) = 1;
2226 phi_translate_set (dest, source, block, phiblock);
2230 /* Compute the ANTIC set for BLOCK.
2232 If succs(BLOCK) > 1 then
2233 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2234 else if succs(BLOCK) == 1 then
2235 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2237 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2241 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2243 bool changed = false;
2244 bitmap_set_t S, old, ANTIC_OUT;
2250 old = ANTIC_OUT = S = NULL;
2251 BB_VISITED (block) = 1;
2253 /* If any edges from predecessors are abnormal, antic_in is empty,
2255 if (block_has_abnormal_pred_edge)
2256 goto maybe_dump_sets;
2258 old = ANTIC_IN (block);
2259 ANTIC_OUT = bitmap_set_new ();
2261 /* If the block has no successors, ANTIC_OUT is empty. */
2262 if (EDGE_COUNT (block->succs) == 0)
2264 /* If we have one successor, we could have some phi nodes to
2265 translate through. */
2266 else if (single_succ_p (block))
2268 basic_block succ_bb = single_succ (block);
2270 /* We trade iterations of the dataflow equations for having to
2271 phi translate the maximal set, which is incredibly slow
2272 (since the maximal set often has 300+ members, even when you
2273 have a small number of blocks).
2274 Basically, we defer the computation of ANTIC for this block
2275 until we have processed it's successor, which will inevitably
2276 have a *much* smaller set of values to phi translate once
2277 clean has been run on it.
2278 The cost of doing this is that we technically perform more
2279 iterations, however, they are lower cost iterations.
2281 Timings for PRE on tramp3d-v4:
2282 without maximal set fix: 11 seconds
2283 with maximal set fix/without deferring: 26 seconds
2284 with maximal set fix/with deferring: 11 seconds
2287 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2291 goto maybe_dump_sets;
2294 /* If we have multiple successors, we take the intersection of all of
2295 them. Note that in the case of loop exit phi nodes, we may have
2296 phis to translate through. */
2299 VEC(basic_block, heap) * worklist;
2301 basic_block bprime, first = NULL;
2303 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2304 FOR_EACH_EDGE (e, ei, block->succs)
2307 && BB_VISITED (e->dest))
2309 else if (BB_VISITED (e->dest))
2310 VEC_quick_push (basic_block, worklist, e->dest);
2313 /* Of multiple successors we have to have visited one already. */
2316 SET_BIT (changed_blocks, block->index);
2317 BB_VISITED (block) = 0;
2318 BB_DEFERRED (block) = 1;
2320 VEC_free (basic_block, heap, worklist);
2321 goto maybe_dump_sets;
2324 if (!gimple_seq_empty_p (phi_nodes (first)))
2325 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2327 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2329 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2331 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2333 bitmap_set_t tmp = bitmap_set_new ();
2334 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2335 bitmap_set_and (ANTIC_OUT, tmp);
2336 bitmap_set_free (tmp);
2339 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2341 VEC_free (basic_block, heap, worklist);
2344 /* Generate ANTIC_OUT - TMP_GEN. */
2345 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2347 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2348 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2351 /* Then union in the ANTIC_OUT - TMP_GEN values,
2352 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2353 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2354 bitmap_value_insert_into_set (ANTIC_IN (block),
2355 expression_for_id (bii));
2357 clean (ANTIC_IN (block), block);
2359 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2362 SET_BIT (changed_blocks, block->index);
2363 FOR_EACH_EDGE (e, ei, block->preds)
2364 SET_BIT (changed_blocks, e->src->index);
2367 RESET_BIT (changed_blocks, block->index);
2370 if (dump_file && (dump_flags & TDF_DETAILS))
2372 if (!BB_DEFERRED (block) || BB_VISITED (block))
2375 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2377 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2381 print_bitmap_set (dump_file, S, "S", block->index);
2386 "Block %d was deferred for a future iteration.\n",
2391 bitmap_set_free (old);
2393 bitmap_set_free (S);
2395 bitmap_set_free (ANTIC_OUT);
2399 /* Compute PARTIAL_ANTIC for BLOCK.
2401 If succs(BLOCK) > 1 then
2402 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2403 in ANTIC_OUT for all succ(BLOCK)
2404 else if succs(BLOCK) == 1 then
2405 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2407 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2412 compute_partial_antic_aux (basic_block block,
2413 bool block_has_abnormal_pred_edge)
2415 bool changed = false;
2416 bitmap_set_t old_PA_IN;
2417 bitmap_set_t PA_OUT;
2420 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2422 old_PA_IN = PA_OUT = NULL;
2424 /* If any edges from predecessors are abnormal, antic_in is empty,
2426 if (block_has_abnormal_pred_edge)
2427 goto maybe_dump_sets;
2429 /* If there are too many partially anticipatable values in the
2430 block, phi_translate_set can take an exponential time: stop
2431 before the translation starts. */
2433 && single_succ_p (block)
2434 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2435 goto maybe_dump_sets;
2437 old_PA_IN = PA_IN (block);
2438 PA_OUT = bitmap_set_new ();
2440 /* If the block has no successors, ANTIC_OUT is empty. */
2441 if (EDGE_COUNT (block->succs) == 0)
2443 /* If we have one successor, we could have some phi nodes to
2444 translate through. Note that we can't phi translate across DFS
2445 back edges in partial antic, because it uses a union operation on
2446 the successors. For recurrences like IV's, we will end up
2447 generating a new value in the set on each go around (i + 3 (VH.1)
2448 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2449 else if (single_succ_p (block))
2451 basic_block succ = single_succ (block);
2452 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2453 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2455 /* If we have multiple successors, we take the union of all of
2459 VEC(basic_block, heap) * worklist;
2463 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2464 FOR_EACH_EDGE (e, ei, block->succs)
2466 if (e->flags & EDGE_DFS_BACK)
2468 VEC_quick_push (basic_block, worklist, e->dest);
2470 if (VEC_length (basic_block, worklist) > 0)
2472 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2477 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2478 bitmap_value_insert_into_set (PA_OUT,
2479 expression_for_id (i));
2480 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2482 bitmap_set_t pa_in = bitmap_set_new ();
2483 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2484 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2485 bitmap_value_insert_into_set (PA_OUT,
2486 expression_for_id (i));
2487 bitmap_set_free (pa_in);
2490 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2491 bitmap_value_insert_into_set (PA_OUT,
2492 expression_for_id (i));
2495 VEC_free (basic_block, heap, worklist);
2498 /* PA_IN starts with PA_OUT - TMP_GEN.
2499 Then we subtract things from ANTIC_IN. */
2500 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2502 /* For partial antic, we want to put back in the phi results, since
2503 we will properly avoid making them partially antic over backedges. */
2504 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2505 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2507 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2508 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2510 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2512 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2515 SET_BIT (changed_blocks, block->index);
2516 FOR_EACH_EDGE (e, ei, block->preds)
2517 SET_BIT (changed_blocks, e->src->index);
2520 RESET_BIT (changed_blocks, block->index);
2523 if (dump_file && (dump_flags & TDF_DETAILS))
2526 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2528 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2531 bitmap_set_free (old_PA_IN);
2533 bitmap_set_free (PA_OUT);
2537 /* Compute ANTIC and partial ANTIC sets. */
2540 compute_antic (void)
2542 bool changed = true;
2543 int num_iterations = 0;
2547 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2548 We pre-build the map of blocks with incoming abnormal edges here. */
2549 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2550 sbitmap_zero (has_abnormal_preds);
2557 FOR_EACH_EDGE (e, ei, block->preds)
2559 e->flags &= ~EDGE_DFS_BACK;
2560 if (e->flags & EDGE_ABNORMAL)
2562 SET_BIT (has_abnormal_preds, block->index);
2567 BB_VISITED (block) = 0;
2568 BB_DEFERRED (block) = 0;
2570 /* While we are here, give empty ANTIC_IN sets to each block. */
2571 ANTIC_IN (block) = bitmap_set_new ();
2572 PA_IN (block) = bitmap_set_new ();
2575 /* At the exit block we anticipate nothing. */
2576 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2577 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2578 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2580 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2581 sbitmap_ones (changed_blocks);
2584 if (dump_file && (dump_flags & TDF_DETAILS))
2585 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2586 /* ??? We need to clear our PHI translation cache here as the
2587 ANTIC sets shrink and we restrict valid translations to
2588 those having operands with leaders in ANTIC. Same below
2589 for PA ANTIC computation. */
2592 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2594 if (TEST_BIT (changed_blocks, postorder[i]))
2596 basic_block block = BASIC_BLOCK (postorder[i]);
2597 changed |= compute_antic_aux (block,
2598 TEST_BIT (has_abnormal_preds,
2602 /* Theoretically possible, but *highly* unlikely. */
2603 gcc_checking_assert (num_iterations < 500);
2606 statistics_histogram_event (cfun, "compute_antic iterations",
2609 if (do_partial_partial)
2611 sbitmap_ones (changed_blocks);
2612 mark_dfs_back_edges ();
2617 if (dump_file && (dump_flags & TDF_DETAILS))
2618 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2621 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2623 if (TEST_BIT (changed_blocks, postorder[i]))
2625 basic_block block = BASIC_BLOCK (postorder[i]);
2627 |= compute_partial_antic_aux (block,
2628 TEST_BIT (has_abnormal_preds,
2632 /* Theoretically possible, but *highly* unlikely. */
2633 gcc_checking_assert (num_iterations < 500);
2635 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2638 sbitmap_free (has_abnormal_preds);
2639 sbitmap_free (changed_blocks);
2642 /* Return true if we can value number the call in STMT. This is true
2643 if we have a pure or constant call. */
2646 can_value_number_call (gimple stmt)
2648 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2653 /* Return true if OP is a tree which we can perform PRE on.
2654 This may not match the operations we can value number, but in
2655 a perfect world would. */
2658 can_PRE_operation (tree op)
2660 return UNARY_CLASS_P (op)
2661 || BINARY_CLASS_P (op)
2662 || COMPARISON_CLASS_P (op)
2663 || TREE_CODE (op) == MEM_REF
2664 || TREE_CODE (op) == COMPONENT_REF
2665 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2666 || TREE_CODE (op) == CALL_EXPR
2667 || TREE_CODE (op) == ARRAY_REF;
2671 /* Inserted expressions are placed onto this worklist, which is used
2672 for performing quick dead code elimination of insertions we made
2673 that didn't turn out to be necessary. */
2674 static bitmap inserted_exprs;
2676 /* Pool allocated fake store expressions are placed onto this
2677 worklist, which, after performing dead code elimination, is walked
2678 to see which expressions need to be put into GC'able memory */
2679 static VEC(gimple, heap) *need_creation;
2681 /* The actual worker for create_component_ref_by_pieces. */
2684 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2685 unsigned int *operand, gimple_seq *stmts,
2688 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2692 switch (currop->opcode)
2696 tree folded, sc = NULL_TREE;
2697 unsigned int nargs = 0;
2699 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2703 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2704 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2710 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2711 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2715 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2716 ref->operands) - 1);
2717 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2719 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2729 folded = build_call_array (currop->type,
2730 (TREE_CODE (fn) == FUNCTION_DECL
2731 ? build_fold_addr_expr (fn) : fn),
2735 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2741 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2743 tree offset = currop->op0;
2746 if (TREE_CODE (baseop) == ADDR_EXPR
2747 && handled_component_p (TREE_OPERAND (baseop, 0)))
2751 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2754 offset = int_const_binop (PLUS_EXPR, offset,
2755 build_int_cst (TREE_TYPE (offset),
2757 baseop = build_fold_addr_expr (base);
2759 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2762 case TARGET_MEM_REF:
2764 pre_expr op0expr, op1expr;
2765 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2766 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2768 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2774 op0expr = get_or_alloc_expr_for (currop->op0);
2775 genop0 = find_or_generate_expression (block, op0expr,
2782 op1expr = get_or_alloc_expr_for (nextop->op0);
2783 genop1 = find_or_generate_expression (block, op1expr,
2788 return build5 (TARGET_MEM_REF, currop->type,
2789 baseop, currop->op2, genop0, currop->op1, genop1);
2795 gcc_assert (is_gimple_min_invariant (currop->op0));
2801 case VIEW_CONVERT_EXPR:
2804 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2809 folded = fold_build1 (currop->opcode, currop->type,
2817 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2819 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2820 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2826 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2829 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2832 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2837 /* For array ref vn_reference_op's, operand 1 of the array ref
2838 is op0 of the reference op and operand 3 of the array ref is
2840 case ARRAY_RANGE_REF:
2844 tree genop1 = currop->op0;
2846 tree genop2 = currop->op1;
2848 tree genop3 = currop->op2;
2850 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2854 op1expr = get_or_alloc_expr_for (genop1);
2855 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2860 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2861 /* Drop zero minimum index if redundant. */
2862 if (integer_zerop (genop2)
2864 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2868 op2expr = get_or_alloc_expr_for (genop2);
2869 genop2 = find_or_generate_expression (block, op2expr, stmts,
2877 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2878 /* We can't always put a size in units of the element alignment
2879 here as the element alignment may be not visible. See
2880 PR43783. Simply drop the element size for constant
2882 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2886 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2887 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2888 op3expr = get_or_alloc_expr_for (genop3);
2889 genop3 = find_or_generate_expression (block, op3expr, stmts,
2895 return build4 (currop->opcode, currop->type, genop0, genop1,
2902 tree genop2 = currop->op1;
2904 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2908 /* op1 should be a FIELD_DECL, which are represented by
2913 op2expr = get_or_alloc_expr_for (genop2);
2914 genop2 = find_or_generate_expression (block, op2expr, stmts,
2920 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2926 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2927 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2948 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2949 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2950 trying to rename aggregates into ssa form directly, which is a no no.
2952 Thus, this routine doesn't create temporaries, it just builds a
2953 single access expression for the array, calling
2954 find_or_generate_expression to build the innermost pieces.
2956 This function is a subroutine of create_expression_by_pieces, and
2957 should not be called on it's own unless you really know what you
2961 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2962 gimple_seq *stmts, gimple domstmt)
2964 unsigned int op = 0;
2965 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2968 /* Find a leader for an expression, or generate one using
2969 create_expression_by_pieces if it's ANTIC but
2971 BLOCK is the basic_block we are looking for leaders in.
2972 EXPR is the expression to find a leader or generate for.
2973 STMTS is the statement list to put the inserted expressions on.
2974 Returns the SSA_NAME of the LHS of the generated expression or the
2976 DOMSTMT if non-NULL is a statement that should be dominated by
2977 all uses in the generated expression. If DOMSTMT is non-NULL this
2978 routine can fail and return NULL_TREE. Otherwise it will assert
2982 find_or_generate_expression (basic_block block, pre_expr expr,
2983 gimple_seq *stmts, gimple domstmt)
2985 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2986 get_expr_value_id (expr), domstmt);
2990 if (leader->kind == NAME)
2991 genop = PRE_EXPR_NAME (leader);
2992 else if (leader->kind == CONSTANT)
2993 genop = PRE_EXPR_CONSTANT (leader);
2996 /* If it's still NULL, it must be a complex expression, so generate
2997 it recursively. Not so if inserting expressions for values generated
3002 bitmap_set_t exprset;
3003 unsigned int lookfor = get_expr_value_id (expr);
3004 bool handled = false;
3008 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3009 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3011 pre_expr temp = expression_for_id (i);
3012 if (temp->kind != NAME)
3015 genop = create_expression_by_pieces (block, temp, stmts,
3017 get_expr_type (expr));
3021 if (!handled && domstmt)
3024 gcc_assert (handled);
3029 #define NECESSARY GF_PLF_1
3031 /* Create an expression in pieces, so that we can handle very complex
3032 expressions that may be ANTIC, but not necessary GIMPLE.
3033 BLOCK is the basic block the expression will be inserted into,
3034 EXPR is the expression to insert (in value form)
3035 STMTS is a statement list to append the necessary insertions into.
3037 This function will die if we hit some value that shouldn't be
3038 ANTIC but is (IE there is no leader for it, or its components).
3039 This function may also generate expressions that are themselves
3040 partially or fully redundant. Those that are will be either made
3041 fully redundant during the next iteration of insert (for partially
3042 redundant ones), or eliminated by eliminate (for fully redundant
3045 If DOMSTMT is non-NULL then we make sure that all uses in the
3046 expressions dominate that statement. In this case the function
3047 can return NULL_TREE to signal failure. */
3050 create_expression_by_pieces (basic_block block, pre_expr expr,
3051 gimple_seq *stmts, gimple domstmt, tree type)
3055 gimple_seq forced_stmts = NULL;
3056 unsigned int value_id;
3057 gimple_stmt_iterator gsi;
3058 tree exprtype = type ? type : get_expr_type (expr);
3064 /* We may hit the NAME/CONSTANT case if we have to convert types
3065 that value numbering saw through. */
3067 folded = PRE_EXPR_NAME (expr);
3070 folded = PRE_EXPR_CONSTANT (expr);
3074 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3075 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3080 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3081 switch (nary->length)
3085 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3086 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
3087 tree genop1 = find_or_generate_expression (block, op1,
3089 tree genop2 = find_or_generate_expression (block, op2,
3091 if (!genop1 || !genop2)
3093 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
3094 may be a constant with the wrong type. */
3095 if (nary->opcode == POINTER_PLUS_EXPR)
3097 genop1 = fold_convert (nary->type, genop1);
3098 genop2 = fold_convert (sizetype, genop2);
3102 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3103 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
3106 folded = fold_build2 (nary->opcode, nary->type,
3112 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3113 tree genop1 = find_or_generate_expression (block, op1,
3117 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3119 folded = fold_build1 (nary->opcode, nary->type,
3132 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3133 folded = fold_convert (exprtype, folded);
3135 /* Force the generated expression to be a sequence of GIMPLE
3137 We have to call unshare_expr because force_gimple_operand may
3138 modify the tree we pass to it. */
3139 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3142 /* If we have any intermediate expressions to the value sets, add them
3143 to the value sets and chain them in the instruction stream. */
3146 gsi = gsi_start (forced_stmts);
3147 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3149 gimple stmt = gsi_stmt (gsi);
3150 tree forcedname = gimple_get_lhs (stmt);
3153 if (TREE_CODE (forcedname) == SSA_NAME)
3155 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3156 VN_INFO_GET (forcedname)->valnum = forcedname;
3157 VN_INFO (forcedname)->value_id = get_next_value_id ();
3158 nameexpr = get_or_alloc_expr_for_name (forcedname);
3159 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3161 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3162 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3164 mark_symbols_for_renaming (stmt);
3166 gimple_seq_add_seq (stmts, forced_stmts);
3169 /* Build and insert the assignment of the end result to the temporary
3170 that we will return. */
3171 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3173 pretemp = create_tmp_reg (exprtype, "pretmp");
3174 get_var_ann (pretemp);
3178 add_referenced_var (temp);
3180 newstmt = gimple_build_assign (temp, folded);
3181 name = make_ssa_name (temp, newstmt);
3182 gimple_assign_set_lhs (newstmt, name);
3183 gimple_set_plf (newstmt, NECESSARY, false);
3185 gimple_seq_add_stmt (stmts, newstmt);
3186 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3188 /* All the symbols in NEWEXPR should be put into SSA form. */
3189 mark_symbols_for_renaming (newstmt);
3191 /* Add a value number to the temporary.
3192 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3193 we are creating the expression by pieces, and this particular piece of
3194 the expression may have been represented. There is no harm in replacing
3196 VN_INFO_GET (name)->valnum = name;
3197 value_id = get_expr_value_id (expr);
3198 VN_INFO (name)->value_id = value_id;
3199 nameexpr = get_or_alloc_expr_for_name (name);
3200 add_to_value (value_id, nameexpr);
3201 if (NEW_SETS (block))
3202 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3203 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3205 pre_stats.insertions++;
3206 if (dump_file && (dump_flags & TDF_DETAILS))
3208 fprintf (dump_file, "Inserted ");
3209 print_gimple_stmt (dump_file, newstmt, 0, 0);
3210 fprintf (dump_file, " in predecessor %d\n", block->index);
3217 /* Returns true if we want to inhibit the insertions of PHI nodes
3218 for the given EXPR for basic block BB (a member of a loop).
3219 We want to do this, when we fear that the induction variable we
3220 create might inhibit vectorization. */
3223 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3225 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3226 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3227 vn_reference_op_t op;
3230 /* If we aren't going to vectorize we don't inhibit anything. */
3231 if (!flag_tree_vectorize)
3234 /* Otherwise we inhibit the insertion when the address of the
3235 memory reference is a simple induction variable. In other
3236 cases the vectorizer won't do anything anyway (either it's
3237 loop invariant or a complicated expression). */
3238 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3243 case ARRAY_RANGE_REF:
3244 if (TREE_CODE (op->op0) != SSA_NAME)
3249 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3251 /* Default defs are loop invariant. */
3254 /* Defined outside this loop, also loop invariant. */
3255 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3257 /* If it's a simple induction variable inhibit insertion,
3258 the vectorizer might be interested in this one. */
3259 if (simple_iv (bb->loop_father, bb->loop_father,
3260 op->op0, &iv, true))
3262 /* No simple IV, vectorizer can't do anything, hence no
3263 reason to inhibit the transformation for this operand. */
3273 /* Insert the to-be-made-available values of expression EXPRNUM for each
3274 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3275 merge the result with a phi node, given the same value number as
3276 NODE. Return true if we have inserted new stuff. */
3279 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3282 pre_expr expr = expression_for_id (exprnum);
3284 unsigned int val = get_expr_value_id (expr);
3286 bool insertions = false;
3291 tree type = get_expr_type (expr);
3295 if (dump_file && (dump_flags & TDF_DETAILS))
3297 fprintf (dump_file, "Found partial redundancy for expression ");
3298 print_pre_expr (dump_file, expr);
3299 fprintf (dump_file, " (%04d)\n", val);
3302 /* Make sure we aren't creating an induction variable. */
3303 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3305 bool firstinsideloop = false;
3306 bool secondinsideloop = false;
3307 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3308 EDGE_PRED (block, 0)->src);
3309 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3310 EDGE_PRED (block, 1)->src);
3311 /* Induction variables only have one edge inside the loop. */
3312 if ((firstinsideloop ^ secondinsideloop)
3313 && (expr->kind != REFERENCE
3314 || inhibit_phi_insertion (block, expr)))
3316 if (dump_file && (dump_flags & TDF_DETAILS))
3317 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3322 /* Make the necessary insertions. */
3323 FOR_EACH_EDGE (pred, ei, block->preds)
3325 gimple_seq stmts = NULL;
3328 eprime = avail[bprime->index];
3330 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3332 builtexpr = create_expression_by_pieces (bprime,
3336 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3337 gsi_insert_seq_on_edge (pred, stmts);
3338 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3341 else if (eprime->kind == CONSTANT)
3343 /* Constants may not have the right type, fold_convert
3344 should give us back a constant with the right type.
3346 tree constant = PRE_EXPR_CONSTANT (eprime);
3347 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3349 tree builtexpr = fold_convert (type, constant);
3350 if (!is_gimple_min_invariant (builtexpr))
3352 tree forcedexpr = force_gimple_operand (builtexpr,
3355 if (!is_gimple_min_invariant (forcedexpr))
3357 if (forcedexpr != builtexpr)
3359 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3360 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3364 gimple_stmt_iterator gsi;
3365 gsi = gsi_start (stmts);
3366 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3368 gimple stmt = gsi_stmt (gsi);
3369 tree lhs = gimple_get_lhs (stmt);
3370 if (TREE_CODE (lhs) == SSA_NAME)
3371 bitmap_set_bit (inserted_exprs,
3372 SSA_NAME_VERSION (lhs));
3373 gimple_set_plf (stmt, NECESSARY, false);
3375 gsi_insert_seq_on_edge (pred, stmts);
3377 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3381 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3384 else if (eprime->kind == NAME)
3386 /* We may have to do a conversion because our value
3387 numbering can look through types in certain cases, but
3388 our IL requires all operands of a phi node have the same
3390 tree name = PRE_EXPR_NAME (eprime);
3391 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3395 builtexpr = fold_convert (type, name);
3396 forcedexpr = force_gimple_operand (builtexpr,
3400 if (forcedexpr != name)
3402 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3403 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3408 gimple_stmt_iterator gsi;
3409 gsi = gsi_start (stmts);
3410 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3412 gimple stmt = gsi_stmt (gsi);
3413 tree lhs = gimple_get_lhs (stmt);
3414 if (TREE_CODE (lhs) == SSA_NAME)
3415 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3416 gimple_set_plf (stmt, NECESSARY, false);
3418 gsi_insert_seq_on_edge (pred, stmts);
3420 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3424 /* If we didn't want a phi node, and we made insertions, we still have
3425 inserted new stuff, and thus return true. If we didn't want a phi node,
3426 and didn't make insertions, we haven't added anything new, so return
3428 if (nophi && insertions)
3430 else if (nophi && !insertions)
3433 /* Now build a phi for the new variable. */
3434 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3436 prephitemp = create_tmp_var (type, "prephitmp");
3437 get_var_ann (prephitemp);
3441 add_referenced_var (temp);
3443 if (TREE_CODE (type) == COMPLEX_TYPE
3444 || TREE_CODE (type) == VECTOR_TYPE)
3445 DECL_GIMPLE_REG_P (temp) = 1;
3446 phi = create_phi_node (temp, block);
3448 gimple_set_plf (phi, NECESSARY, false);
3449 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3450 VN_INFO (gimple_phi_result (phi))->value_id = val;
3451 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3452 FOR_EACH_EDGE (pred, ei, block->preds)
3454 pre_expr ae = avail[pred->src->index];
3455 gcc_assert (get_expr_type (ae) == type
3456 || useless_type_conversion_p (type, get_expr_type (ae)));
3457 if (ae->kind == CONSTANT)
3458 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3460 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3464 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3465 add_to_value (val, newphi);
3467 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3468 this insertion, since we test for the existence of this value in PHI_GEN
3469 before proceeding with the partial redundancy checks in insert_aux.
3471 The value may exist in AVAIL_OUT, in particular, it could be represented
3472 by the expression we are trying to eliminate, in which case we want the
3473 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3476 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3477 this block, because if it did, it would have existed in our dominator's
3478 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3481 bitmap_insert_into_set (PHI_GEN (block), newphi);
3482 bitmap_value_replace_in_set (AVAIL_OUT (block),
3484 bitmap_insert_into_set (NEW_SETS (block),
3487 if (dump_file && (dump_flags & TDF_DETAILS))
3489 fprintf (dump_file, "Created phi ");
3490 print_gimple_stmt (dump_file, phi, 0, 0);
3491 fprintf (dump_file, " in block %d\n", block->index);
3499 /* Perform insertion of partially redundant values.
3500 For BLOCK, do the following:
3501 1. Propagate the NEW_SETS of the dominator into the current block.
3502 If the block has multiple predecessors,
3503 2a. Iterate over the ANTIC expressions for the block to see if
3504 any of them are partially redundant.
3505 2b. If so, insert them into the necessary predecessors to make
3506 the expression fully redundant.
3507 2c. Insert a new PHI merging the values of the predecessors.
3508 2d. Insert the new PHI, and the new expressions, into the
3510 3. Recursively call ourselves on the dominator children of BLOCK.
3512 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3513 do_regular_insertion and do_partial_insertion.
3518 do_regular_insertion (basic_block block, basic_block dom)
3520 bool new_stuff = false;
3521 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3525 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3527 if (expr->kind != NAME)
3531 bool by_some = false;
3532 bool cant_insert = false;
3533 bool all_same = true;
3534 pre_expr first_s = NULL;
3537 pre_expr eprime = NULL;
3539 pre_expr edoubleprime = NULL;
3540 bool do_insertion = false;
3542 val = get_expr_value_id (expr);
3543 if (bitmap_set_contains_value (PHI_GEN (block), val))
3545 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3547 if (dump_file && (dump_flags & TDF_DETAILS))
3548 fprintf (dump_file, "Found fully redundant value\n");
3552 avail = XCNEWVEC (pre_expr, last_basic_block);
3553 FOR_EACH_EDGE (pred, ei, block->preds)
3555 unsigned int vprime;
3557 /* We should never run insertion for the exit block
3558 and so not come across fake pred edges. */
3559 gcc_assert (!(pred->flags & EDGE_FAKE));
3561 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3564 /* eprime will generally only be NULL if the
3565 value of the expression, translated
3566 through the PHI for this predecessor, is
3567 undefined. If that is the case, we can't
3568 make the expression fully redundant,
3569 because its value is undefined along a
3570 predecessor path. We can thus break out
3571 early because it doesn't matter what the
3572 rest of the results are. */
3579 eprime = fully_constant_expression (eprime);
3580 vprime = get_expr_value_id (eprime);
3581 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3583 if (edoubleprime == NULL)
3585 avail[bprime->index] = eprime;
3590 avail[bprime->index] = edoubleprime;
3592 /* We want to perform insertions to remove a redundancy on
3593 a path in the CFG we want to optimize for speed. */
3594 if (optimize_edge_for_speed_p (pred))
3595 do_insertion = true;
3596 if (first_s == NULL)
3597 first_s = edoubleprime;
3598 else if (!pre_expr_eq (first_s, edoubleprime))
3602 /* If we can insert it, it's not the same value
3603 already existing along every predecessor, and
3604 it's defined by some predecessor, it is
3605 partially redundant. */
3606 if (!cant_insert && !all_same && by_some)
3610 if (dump_file && (dump_flags & TDF_DETAILS))
3612 fprintf (dump_file, "Skipping partial redundancy for "
3614 print_pre_expr (dump_file, expr);
3615 fprintf (dump_file, " (%04d), no redundancy on to be "
3616 "optimized for speed edge\n", val);
3619 else if (dbg_cnt (treepre_insert)
3620 && insert_into_preds_of_block (block,
3621 get_expression_id (expr),
3625 /* If all edges produce the same value and that value is
3626 an invariant, then the PHI has the same value on all
3627 edges. Note this. */
3628 else if (!cant_insert && all_same && eprime
3629 && (edoubleprime->kind == CONSTANT
3630 || edoubleprime->kind == NAME)
3631 && !value_id_constant_p (val))
3635 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3636 value_expressions, val);
3638 unsigned int new_val = get_expr_value_id (edoubleprime);
3639 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3641 pre_expr expr = expression_for_id (j);
3643 if (expr->kind == NAME)
3645 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3646 /* Just reset the value id and valnum so it is
3647 the same as the constant we have discovered. */
3648 if (edoubleprime->kind == CONSTANT)
3650 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3651 pre_stats.constified++;
3654 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3655 info->value_id = new_val;
3663 VEC_free (pre_expr, heap, exprs);
3668 /* Perform insertion for partially anticipatable expressions. There
3669 is only one case we will perform insertion for these. This case is
3670 if the expression is partially anticipatable, and fully available.
3671 In this case, we know that putting it earlier will enable us to
3672 remove the later computation. */
3676 do_partial_partial_insertion (basic_block block, basic_block dom)
3678 bool new_stuff = false;
3679 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3683 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3685 if (expr->kind != NAME)
3690 bool cant_insert = false;
3693 pre_expr eprime = NULL;
3696 val = get_expr_value_id (expr);
3697 if (bitmap_set_contains_value (PHI_GEN (block), val))
3699 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3702 avail = XCNEWVEC (pre_expr, last_basic_block);
3703 FOR_EACH_EDGE (pred, ei, block->preds)
3705 unsigned int vprime;
3706 pre_expr edoubleprime;
3708 /* We should never run insertion for the exit block
3709 and so not come across fake pred edges. */
3710 gcc_assert (!(pred->flags & EDGE_FAKE));
3712 eprime = phi_translate (expr, ANTIC_IN (block),
3716 /* eprime will generally only be NULL if the
3717 value of the expression, translated
3718 through the PHI for this predecessor, is
3719 undefined. If that is the case, we can't
3720 make the expression fully redundant,
3721 because its value is undefined along a
3722 predecessor path. We can thus break out
3723 early because it doesn't matter what the
3724 rest of the results are. */
3731 eprime = fully_constant_expression (eprime);
3732 vprime = get_expr_value_id (eprime);
3733 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3735 if (edoubleprime == NULL)
3741 avail[bprime->index] = edoubleprime;
3745 /* If we can insert it, it's not the same value
3746 already existing along every predecessor, and
3747 it's defined by some predecessor, it is
3748 partially redundant. */
3749 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3751 pre_stats.pa_insert++;
3752 if (insert_into_preds_of_block (block, get_expression_id (expr),
3760 VEC_free (pre_expr, heap, exprs);
3765 insert_aux (basic_block block)
3768 bool new_stuff = false;
3773 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3778 bitmap_set_t newset = NEW_SETS (dom);
3781 /* Note that we need to value_replace both NEW_SETS, and
3782 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3783 represented by some non-simple expression here that we want
3784 to replace it with. */
3785 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3787 pre_expr expr = expression_for_id (i);
3788 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3789 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3792 if (!single_pred_p (block))
3794 new_stuff |= do_regular_insertion (block, dom);
3795 if (do_partial_partial)
3796 new_stuff |= do_partial_partial_insertion (block, dom);
3800 for (son = first_dom_son (CDI_DOMINATORS, block);
3802 son = next_dom_son (CDI_DOMINATORS, son))
3804 new_stuff |= insert_aux (son);
3810 /* Perform insertion of partially redundant values. */
3815 bool new_stuff = true;
3817 int num_iterations = 0;
3820 NEW_SETS (bb) = bitmap_set_new ();
3825 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3827 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3831 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3834 add_to_exp_gen (basic_block block, tree op)
3839 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3841 result = get_or_alloc_expr_for_name (op);
3842 bitmap_value_insert_into_set (EXP_GEN (block), result);
3846 /* Create value ids for PHI in BLOCK. */
3849 make_values_for_phi (gimple phi, basic_block block)
3851 tree result = gimple_phi_result (phi);
3853 /* We have no need for virtual phis, as they don't represent
3854 actual computations. */
3855 if (is_gimple_reg (result))
3857 pre_expr e = get_or_alloc_expr_for_name (result);
3858 add_to_value (get_expr_value_id (e), e);
3859 bitmap_insert_into_set (PHI_GEN (block), e);
3860 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3864 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3866 tree arg = gimple_phi_arg_def (phi, i);
3867 if (TREE_CODE (arg) == SSA_NAME)
3869 e = get_or_alloc_expr_for_name (arg);
3870 add_to_value (get_expr_value_id (e), e);
3877 /* Compute the AVAIL set for all basic blocks.
3879 This function performs value numbering of the statements in each basic
3880 block. The AVAIL sets are built from information we glean while doing
3881 this value numbering, since the AVAIL sets contain only one entry per
3884 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3885 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3888 compute_avail (void)
3891 basic_block block, son;
3892 basic_block *worklist;
3896 /* We pretend that default definitions are defined in the entry block.
3897 This includes function arguments and the static chain decl. */
3898 for (i = 1; i < num_ssa_names; ++i)
3900 tree name = ssa_name (i);
3903 || !SSA_NAME_IS_DEFAULT_DEF (name)
3904 || has_zero_uses (name)
3905 || !is_gimple_reg (name))
3908 e = get_or_alloc_expr_for_name (name);
3909 add_to_value (get_expr_value_id (e), e);
3911 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3912 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3915 /* Allocate the worklist. */
3916 worklist = XNEWVEC (basic_block, n_basic_blocks);
3918 /* Seed the algorithm by putting the dominator children of the entry
3919 block on the worklist. */
3920 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3922 son = next_dom_son (CDI_DOMINATORS, son))
3923 worklist[sp++] = son;
3925 /* Loop until the worklist is empty. */
3928 gimple_stmt_iterator gsi;
3931 unsigned int stmt_uid = 1;
3933 /* Pick a block from the worklist. */
3934 block = worklist[--sp];
3936 /* Initially, the set of available values in BLOCK is that of
3937 its immediate dominator. */
3938 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3940 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3942 /* Generate values for PHI nodes. */
3943 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3944 make_values_for_phi (gsi_stmt (gsi), block);
3946 BB_MAY_NOTRETURN (block) = 0;
3948 /* Now compute value numbers and populate value sets with all
3949 the expressions computed in BLOCK. */
3950 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3955 stmt = gsi_stmt (gsi);
3956 gimple_set_uid (stmt, stmt_uid++);
3958 /* Cache whether the basic-block has any non-visible side-effect
3960 If this isn't a call or it is the last stmt in the
3961 basic-block then the CFG represents things correctly. */
3962 if (is_gimple_call (stmt)
3963 && !stmt_ends_bb_p (stmt))
3965 /* Non-looping const functions always return normally.
3966 Otherwise the call might not return or have side-effects
3967 that forbids hoisting possibly trapping expressions
3969 int flags = gimple_call_flags (stmt);
3970 if (!(flags & ECF_CONST)
3971 || (flags & ECF_LOOPING_CONST_OR_PURE))
3972 BB_MAY_NOTRETURN (block) = 1;
3975 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3977 pre_expr e = get_or_alloc_expr_for_name (op);
3979 add_to_value (get_expr_value_id (e), e);
3981 bitmap_insert_into_set (TMP_GEN (block), e);
3982 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3985 if (gimple_has_volatile_ops (stmt)
3986 || stmt_could_throw_p (stmt))
3989 switch (gimple_code (stmt))
3992 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3993 add_to_exp_gen (block, op);
4000 vn_reference_op_t vro;
4001 pre_expr result = NULL;
4002 VEC(vn_reference_op_s, heap) *ops = NULL;
4004 if (!can_value_number_call (stmt))
4007 copy_reference_ops_from_call (stmt, &ops);
4008 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4009 gimple_expr_type (stmt),
4010 ops, &ref, VN_NOWALK);
4011 VEC_free (vn_reference_op_s, heap, ops);
4015 for (i = 0; VEC_iterate (vn_reference_op_s,
4019 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4020 add_to_exp_gen (block, vro->op0);
4021 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4022 add_to_exp_gen (block, vro->op1);
4023 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4024 add_to_exp_gen (block, vro->op2);
4026 result = (pre_expr) pool_alloc (pre_expr_pool);
4027 result->kind = REFERENCE;
4029 PRE_EXPR_REFERENCE (result) = ref;
4031 get_or_alloc_expression_id (result);
4032 add_to_value (get_expr_value_id (result), result);
4034 bitmap_value_insert_into_set (EXP_GEN (block), result);
4040 pre_expr result = NULL;
4041 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4045 case tcc_comparison:
4050 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4051 gimple_assign_rhs_code (stmt),
4052 gimple_expr_type (stmt),
4053 gimple_assign_rhs1 (stmt),
4054 gimple_assign_rhs2 (stmt),
4055 NULL_TREE, NULL_TREE, &nary);
4060 for (i = 0; i < nary->length; i++)
4061 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4062 add_to_exp_gen (block, nary->op[i]);
4064 result = (pre_expr) pool_alloc (pre_expr_pool);
4065 result->kind = NARY;
4067 PRE_EXPR_NARY (result) = nary;
4071 case tcc_declaration:
4076 vn_reference_op_t vro;
4078 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4084 for (i = 0; VEC_iterate (vn_reference_op_s,
4088 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4089 add_to_exp_gen (block, vro->op0);
4090 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4091 add_to_exp_gen (block, vro->op1);
4092 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4093 add_to_exp_gen (block, vro->op2);
4095 result = (pre_expr) pool_alloc (pre_expr_pool);
4096 result->kind = REFERENCE;
4098 PRE_EXPR_REFERENCE (result) = ref;
4103 /* For any other statement that we don't
4104 recognize, simply add all referenced
4105 SSA_NAMEs to EXP_GEN. */
4106 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4107 add_to_exp_gen (block, op);
4111 get_or_alloc_expression_id (result);
4112 add_to_value (get_expr_value_id (result), result);
4114 bitmap_value_insert_into_set (EXP_GEN (block), result);
4123 /* Put the dominator children of BLOCK on the worklist of blocks
4124 to compute available sets for. */
4125 for (son = first_dom_son (CDI_DOMINATORS, block);
4127 son = next_dom_son (CDI_DOMINATORS, son))
4128 worklist[sp++] = son;
4134 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4135 than the available expressions for it. The insertion point is
4136 right before the first use in STMT. Returns the SSA_NAME that should
4137 be used for replacement. */
4140 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4142 basic_block bb = gimple_bb (stmt);
4143 gimple_stmt_iterator gsi;
4144 gimple_seq stmts = NULL;
4148 /* First create a value expression from the expression we want
4149 to insert and associate it with the value handle for SSA_VN. */
4150 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4154 /* Then use create_expression_by_pieces to generate a valid
4155 expression to insert at this point of the IL stream. */
4156 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4157 if (expr == NULL_TREE)
4159 gsi = gsi_for_stmt (stmt);
4160 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4165 /* Eliminate fully redundant computations. */
4170 VEC (gimple, heap) *to_remove = NULL;
4172 unsigned int todo = 0;
4173 gimple_stmt_iterator gsi;
4179 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4181 stmt = gsi_stmt (gsi);
4183 /* Lookup the RHS of the expression, see if we have an
4184 available computation for it. If so, replace the RHS with
4185 the available computation. */
4186 if (gimple_has_lhs (stmt)
4187 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4188 && !gimple_assign_ssa_name_copy_p (stmt)
4189 && (!gimple_assign_single_p (stmt)
4190 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4191 && !gimple_has_volatile_ops (stmt)
4192 && !has_zero_uses (gimple_get_lhs (stmt)))
4194 tree lhs = gimple_get_lhs (stmt);
4195 tree rhs = NULL_TREE;
4197 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4198 pre_expr sprimeexpr;
4200 if (gimple_assign_single_p (stmt))
4201 rhs = gimple_assign_rhs1 (stmt);
4203 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4204 get_expr_value_id (lhsexpr),
4209 if (sprimeexpr->kind == CONSTANT)
4210 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4211 else if (sprimeexpr->kind == NAME)
4212 sprime = PRE_EXPR_NAME (sprimeexpr);
4217 /* If there is no existing leader but SCCVN knows this
4218 value is constant, use that constant. */
4219 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4221 sprime = VN_INFO (lhs)->valnum;
4222 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4223 TREE_TYPE (sprime)))
4224 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4226 if (dump_file && (dump_flags & TDF_DETAILS))
4228 fprintf (dump_file, "Replaced ");
4229 print_gimple_expr (dump_file, stmt, 0, 0);
4230 fprintf (dump_file, " with ");
4231 print_generic_expr (dump_file, sprime, 0);
4232 fprintf (dump_file, " in ");
4233 print_gimple_stmt (dump_file, stmt, 0, 0);
4235 pre_stats.eliminations++;
4236 propagate_tree_value_into_stmt (&gsi, sprime);
4237 stmt = gsi_stmt (gsi);
4242 /* If there is no existing usable leader but SCCVN thinks
4243 it has an expression it wants to use as replacement,
4245 if (!sprime || sprime == lhs)
4247 tree val = VN_INFO (lhs)->valnum;
4249 && TREE_CODE (val) == SSA_NAME
4250 && VN_INFO (val)->needs_insertion
4251 && can_PRE_operation (vn_get_expr_for (val)))
4252 sprime = do_SCCVN_insertion (stmt, val);
4256 && (rhs == NULL_TREE
4257 || TREE_CODE (rhs) != SSA_NAME
4258 || may_propagate_copy (rhs, sprime)))
4260 bool can_make_abnormal_goto
4261 = is_gimple_call (stmt)
4262 && stmt_can_make_abnormal_goto (stmt);
4264 gcc_assert (sprime != rhs);
4266 if (dump_file && (dump_flags & TDF_DETAILS))
4268 fprintf (dump_file, "Replaced ");
4269 print_gimple_expr (dump_file, stmt, 0, 0);
4270 fprintf (dump_file, " with ");
4271 print_generic_expr (dump_file, sprime, 0);
4272 fprintf (dump_file, " in ");
4273 print_gimple_stmt (dump_file, stmt, 0, 0);
4276 if (TREE_CODE (sprime) == SSA_NAME)
4277 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4279 /* We need to make sure the new and old types actually match,
4280 which may require adding a simple cast, which fold_convert
4282 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4283 && !useless_type_conversion_p (gimple_expr_type (stmt),
4284 TREE_TYPE (sprime)))
4285 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4287 pre_stats.eliminations++;
4288 propagate_tree_value_into_stmt (&gsi, sprime);
4289 stmt = gsi_stmt (gsi);
4292 /* If we removed EH side-effects from the statement, clean
4293 its EH information. */
4294 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4296 bitmap_set_bit (need_eh_cleanup,
4297 gimple_bb (stmt)->index);
4298 if (dump_file && (dump_flags & TDF_DETAILS))
4299 fprintf (dump_file, " Removed EH side-effects.\n");
4302 /* Likewise for AB side-effects. */
4303 if (can_make_abnormal_goto
4304 && !stmt_can_make_abnormal_goto (stmt))
4306 bitmap_set_bit (need_ab_cleanup,
4307 gimple_bb (stmt)->index);
4308 if (dump_file && (dump_flags & TDF_DETAILS))
4309 fprintf (dump_file, " Removed AB side-effects.\n");
4313 /* If the statement is a scalar store, see if the expression
4314 has the same value number as its rhs. If so, the store is
4316 else if (gimple_assign_single_p (stmt)
4317 && !is_gimple_reg (gimple_assign_lhs (stmt))
4318 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4319 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4321 tree rhs = gimple_assign_rhs1 (stmt);
4323 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4324 gimple_vuse (stmt), VN_WALK, NULL);
4325 if (TREE_CODE (rhs) == SSA_NAME)
4326 rhs = VN_INFO (rhs)->valnum;
4328 && operand_equal_p (val, rhs, 0))
4330 if (dump_file && (dump_flags & TDF_DETAILS))
4332 fprintf (dump_file, "Deleted redundant store ");
4333 print_gimple_stmt (dump_file, stmt, 0, 0);
4336 /* Queue stmt for removal. */
4337 VEC_safe_push (gimple, heap, to_remove, stmt);
4340 /* Visit COND_EXPRs and fold the comparison with the
4341 available value-numbers. */
4342 else if (gimple_code (stmt) == GIMPLE_COND)
4344 tree op0 = gimple_cond_lhs (stmt);
4345 tree op1 = gimple_cond_rhs (stmt);
4348 if (TREE_CODE (op0) == SSA_NAME)
4349 op0 = VN_INFO (op0)->valnum;
4350 if (TREE_CODE (op1) == SSA_NAME)
4351 op1 = VN_INFO (op1)->valnum;
4352 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4354 if (result && TREE_CODE (result) == INTEGER_CST)
4356 if (integer_zerop (result))
4357 gimple_cond_make_false (stmt);
4359 gimple_cond_make_true (stmt);
4361 todo = TODO_cleanup_cfg;
4364 /* Visit indirect calls and turn them into direct calls if
4366 if (is_gimple_call (stmt)
4367 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4369 tree fn = VN_INFO (gimple_call_fn (stmt))->valnum;
4370 if (TREE_CODE (fn) == ADDR_EXPR
4371 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
4373 bool can_make_abnormal_goto
4374 = stmt_can_make_abnormal_goto (stmt);
4375 bool was_noreturn = gimple_call_noreturn_p (stmt);
4377 if (dump_file && (dump_flags & TDF_DETAILS))
4379 fprintf (dump_file, "Replacing call target with ");
4380 print_generic_expr (dump_file, fn, 0);
4381 fprintf (dump_file, " in ");
4382 print_gimple_stmt (dump_file, stmt, 0, 0);
4385 gimple_call_set_fn (stmt, fn);
4388 /* When changing a call into a noreturn call, cfg cleanup
4389 is needed to fix up the noreturn call. */
4390 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4391 todo |= TODO_cleanup_cfg;
4393 /* If we removed EH side-effects from the statement, clean
4394 its EH information. */
4395 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4397 bitmap_set_bit (need_eh_cleanup,
4398 gimple_bb (stmt)->index);
4399 if (dump_file && (dump_flags & TDF_DETAILS))
4400 fprintf (dump_file, " Removed EH side-effects.\n");
4403 /* Likewise for AB side-effects. */
4404 if (can_make_abnormal_goto
4405 && !stmt_can_make_abnormal_goto (stmt))
4407 bitmap_set_bit (need_ab_cleanup,
4408 gimple_bb (stmt)->index);
4409 if (dump_file && (dump_flags & TDF_DETAILS))
4410 fprintf (dump_file, " Removed AB side-effects.\n");
4413 /* Changing an indirect call to a direct call may
4414 have exposed different semantics. This may
4415 require an SSA update. */
4416 todo |= TODO_update_ssa_only_virtuals;
4421 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4423 gimple stmt, phi = gsi_stmt (gsi);
4424 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4425 pre_expr sprimeexpr, resexpr;
4426 gimple_stmt_iterator gsi2;
4428 /* We want to perform redundant PHI elimination. Do so by
4429 replacing the PHI with a single copy if possible.
4430 Do not touch inserted, single-argument or virtual PHIs. */
4431 if (gimple_phi_num_args (phi) == 1
4432 || !is_gimple_reg (res))
4438 resexpr = get_or_alloc_expr_for_name (res);
4439 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4440 get_expr_value_id (resexpr), NULL);
4443 if (sprimeexpr->kind == CONSTANT)
4444 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4445 else if (sprimeexpr->kind == NAME)
4446 sprime = PRE_EXPR_NAME (sprimeexpr);
4450 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4452 sprime = VN_INFO (res)->valnum;
4453 if (!useless_type_conversion_p (TREE_TYPE (res),
4454 TREE_TYPE (sprime)))
4455 sprime = fold_convert (TREE_TYPE (res), sprime);
4464 if (dump_file && (dump_flags & TDF_DETAILS))
4466 fprintf (dump_file, "Replaced redundant PHI node defining ");
4467 print_generic_expr (dump_file, res, 0);
4468 fprintf (dump_file, " with ");
4469 print_generic_expr (dump_file, sprime, 0);
4470 fprintf (dump_file, "\n");
4473 remove_phi_node (&gsi, false);
4475 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4476 && TREE_CODE (sprime) == SSA_NAME)
4477 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4479 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4480 sprime = fold_convert (TREE_TYPE (res), sprime);
4481 stmt = gimple_build_assign (res, sprime);
4482 SSA_NAME_DEF_STMT (res) = stmt;
4483 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4485 gsi2 = gsi_after_labels (b);
4486 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4487 /* Queue the copy for eventual removal. */
4488 VEC_safe_push (gimple, heap, to_remove, stmt);
4489 /* If we inserted this PHI node ourself, it's not an elimination. */
4490 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4493 pre_stats.eliminations++;
4497 /* We cannot remove stmts during BB walk, especially not release SSA
4498 names there as this confuses the VN machinery. The stmts ending
4499 up in to_remove are either stores or simple copies. */
4500 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4502 tree lhs = gimple_assign_lhs (stmt);
4503 tree rhs = gimple_assign_rhs1 (stmt);
4504 use_operand_p use_p;
4507 /* If there is a single use only, propagate the equivalency
4508 instead of keeping the copy. */
4509 if (TREE_CODE (lhs) == SSA_NAME
4510 && TREE_CODE (rhs) == SSA_NAME
4511 && single_imm_use (lhs, &use_p, &use_stmt)
4512 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4514 SET_USE (use_p, rhs);
4515 update_stmt (use_stmt);
4516 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4517 && TREE_CODE (rhs) == SSA_NAME)
4518 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4521 /* If this is a store or a now unused copy, remove it. */
4522 if (TREE_CODE (lhs) != SSA_NAME
4523 || has_zero_uses (lhs))
4525 basic_block bb = gimple_bb (stmt);
4526 gsi = gsi_for_stmt (stmt);
4527 unlink_stmt_vdef (stmt);
4528 gsi_remove (&gsi, true);
4529 if (gimple_purge_dead_eh_edges (bb))
4530 todo |= TODO_cleanup_cfg;
4531 if (TREE_CODE (lhs) == SSA_NAME)
4532 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4533 release_defs (stmt);
4536 VEC_free (gimple, heap, to_remove);
4541 /* Borrow a bit of tree-ssa-dce.c for the moment.
4542 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4543 this may be a bit faster, and we may want critical edges kept split. */
4545 /* If OP's defining statement has not already been determined to be necessary,
4546 mark that statement necessary. Return the stmt, if it is newly
4549 static inline gimple
4550 mark_operand_necessary (tree op)
4556 if (TREE_CODE (op) != SSA_NAME)
4559 stmt = SSA_NAME_DEF_STMT (op);
4562 if (gimple_plf (stmt, NECESSARY)
4563 || gimple_nop_p (stmt))
4566 gimple_set_plf (stmt, NECESSARY, true);
4570 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4571 to insert PHI nodes sometimes, and because value numbering of casts isn't
4572 perfect, we sometimes end up inserting dead code. This simple DCE-like
4573 pass removes any insertions we made that weren't actually used. */
4576 remove_dead_inserted_code (void)
4583 worklist = BITMAP_ALLOC (NULL);
4584 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4586 t = SSA_NAME_DEF_STMT (ssa_name (i));
4587 if (gimple_plf (t, NECESSARY))
4588 bitmap_set_bit (worklist, i);
4590 while (!bitmap_empty_p (worklist))
4592 i = bitmap_first_set_bit (worklist);
4593 bitmap_clear_bit (worklist, i);
4594 t = SSA_NAME_DEF_STMT (ssa_name (i));
4596 /* PHI nodes are somewhat special in that each PHI alternative has
4597 data and control dependencies. All the statements feeding the
4598 PHI node's arguments are always necessary. */
4599 if (gimple_code (t) == GIMPLE_PHI)
4603 for (k = 0; k < gimple_phi_num_args (t); k++)
4605 tree arg = PHI_ARG_DEF (t, k);
4606 if (TREE_CODE (arg) == SSA_NAME)
4608 gimple n = mark_operand_necessary (arg);
4610 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4616 /* Propagate through the operands. Examine all the USE, VUSE and
4617 VDEF operands in this statement. Mark all the statements
4618 which feed this statement's uses as necessary. */
4622 /* The operands of VDEF expressions are also needed as they
4623 represent potential definitions that may reach this
4624 statement (VDEF operands allow us to follow def-def
4627 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4629 gimple n = mark_operand_necessary (use);
4631 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4636 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4638 t = SSA_NAME_DEF_STMT (ssa_name (i));
4639 if (!gimple_plf (t, NECESSARY))
4641 gimple_stmt_iterator gsi;
4643 if (dump_file && (dump_flags & TDF_DETAILS))
4645 fprintf (dump_file, "Removing unnecessary insertion:");
4646 print_gimple_stmt (dump_file, t, 0, 0);
4649 gsi = gsi_for_stmt (t);
4650 if (gimple_code (t) == GIMPLE_PHI)
4651 remove_phi_node (&gsi, true);
4654 gsi_remove (&gsi, true);
4659 BITMAP_FREE (worklist);
4662 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4663 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4664 the number of visited blocks. */
4667 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4669 edge_iterator *stack;
4671 int post_order_num = 0;
4674 if (include_entry_exit)
4675 post_order[post_order_num++] = EXIT_BLOCK;
4677 /* Allocate stack for back-tracking up CFG. */
4678 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4681 /* Allocate bitmap to track nodes that have been visited. */
4682 visited = sbitmap_alloc (last_basic_block);
4684 /* None of the nodes in the CFG have been visited yet. */
4685 sbitmap_zero (visited);
4687 /* Push the last edge on to the stack. */
4688 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4696 /* Look at the edge on the top of the stack. */
4698 src = ei_edge (ei)->src;
4699 dest = ei_edge (ei)->dest;
4701 /* Check if the edge destination has been visited yet. */
4702 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4704 /* Mark that we have visited the destination. */
4705 SET_BIT (visited, src->index);
4707 if (EDGE_COUNT (src->preds) > 0)
4708 /* Since the DEST node has been visited for the first
4709 time, check its successors. */
4710 stack[sp++] = ei_start (src->preds);
4712 post_order[post_order_num++] = src->index;
4716 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4717 post_order[post_order_num++] = dest->index;
4719 if (!ei_one_before_end_p (ei))
4720 ei_next (&stack[sp - 1]);
4726 if (include_entry_exit)
4727 post_order[post_order_num++] = ENTRY_BLOCK;
4730 sbitmap_free (visited);
4731 return post_order_num;
4735 /* Initialize data structures used by PRE. */
4738 init_pre (bool do_fre)
4742 next_expression_id = 1;
4744 VEC_safe_push (pre_expr, heap, expressions, NULL);
4745 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4746 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4747 get_max_value_id() + 1);
4752 inserted_exprs = BITMAP_ALLOC (NULL);
4753 need_creation = NULL;
4754 pretemp = NULL_TREE;
4755 storetemp = NULL_TREE;
4756 prephitemp = NULL_TREE;
4758 connect_infinite_loops_to_exit ();
4759 memset (&pre_stats, 0, sizeof (pre_stats));
4762 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4763 my_rev_post_order_compute (postorder, false);
4765 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4767 calculate_dominance_info (CDI_POST_DOMINATORS);
4768 calculate_dominance_info (CDI_DOMINATORS);
4770 bitmap_obstack_initialize (&grand_bitmap_obstack);
4771 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4772 expr_pred_trans_eq, free);
4773 expression_to_id = htab_create (num_ssa_names * 3,
4776 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4777 sizeof (struct bitmap_set), 30);
4778 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4779 sizeof (struct pre_expr_d), 30);
4782 EXP_GEN (bb) = bitmap_set_new ();
4783 PHI_GEN (bb) = bitmap_set_new ();
4784 TMP_GEN (bb) = bitmap_set_new ();
4785 AVAIL_OUT (bb) = bitmap_set_new ();
4788 need_eh_cleanup = BITMAP_ALLOC (NULL);
4789 need_ab_cleanup = BITMAP_ALLOC (NULL);
4793 /* Deallocate data structures used by PRE. */
4796 fini_pre (bool do_fre)
4799 VEC_free (bitmap_set_t, heap, value_expressions);
4800 BITMAP_FREE (inserted_exprs);
4801 VEC_free (gimple, heap, need_creation);
4802 bitmap_obstack_release (&grand_bitmap_obstack);
4803 free_alloc_pool (bitmap_set_pool);
4804 free_alloc_pool (pre_expr_pool);
4805 htab_delete (phi_translate_table);
4806 htab_delete (expression_to_id);
4807 VEC_free (unsigned, heap, name_to_id);
4809 free_aux_for_blocks ();
4811 free_dominance_info (CDI_POST_DOMINATORS);
4813 if (!bitmap_empty_p (need_eh_cleanup))
4815 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4816 cleanup_tree_cfg ();
4819 BITMAP_FREE (need_eh_cleanup);
4821 if (!bitmap_empty_p (need_ab_cleanup))
4823 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4824 cleanup_tree_cfg ();
4827 BITMAP_FREE (need_ab_cleanup);
4830 loop_optimizer_finalize ();
4833 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4834 only wants to do full redundancy elimination. */
4837 execute_pre (bool do_fre)
4839 unsigned int todo = 0;
4841 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4843 /* This has to happen before SCCVN runs because
4844 loop_optimizer_init may create new phis, etc. */
4846 loop_optimizer_init (LOOPS_NORMAL);
4848 if (!run_scc_vn (do_fre ? VN_WALKREWRITE : VN_WALK))
4851 loop_optimizer_finalize ();
4859 /* Collect and value number expressions computed in each basic block. */
4862 if (dump_file && (dump_flags & TDF_DETAILS))
4868 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4869 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4870 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4871 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4875 /* Insert can get quite slow on an incredibly large number of basic
4876 blocks due to some quadratic behavior. Until this behavior is
4877 fixed, don't run it when he have an incredibly large number of
4878 bb's. If we aren't going to run insert, there is no point in
4879 computing ANTIC, either, even though it's plenty fast. */
4880 if (!do_fre && n_basic_blocks < 4000)
4886 /* Make sure to remove fake edges before committing our inserts.
4887 This makes sure we don't end up with extra critical edges that
4888 we would need to split. */
4889 remove_fake_exit_edges ();
4890 gsi_commit_edge_inserts ();
4892 /* Remove all the redundant expressions. */
4893 todo |= eliminate ();
4895 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4896 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4897 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4898 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4899 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4901 clear_expression_ids ();
4905 remove_dead_inserted_code ();
4906 todo |= TODO_verify_flow;
4915 /* Gate and execute functions for PRE. */
4920 return execute_pre (false);
4926 return flag_tree_pre != 0;
4929 struct gimple_opt_pass pass_pre =
4934 gate_pre, /* gate */
4935 do_pre, /* execute */
4938 0, /* static_pass_number */
4939 TV_TREE_PRE, /* tv_id */
4940 PROP_no_crit_edges | PROP_cfg
4941 | PROP_ssa, /* properties_required */
4942 0, /* properties_provided */
4943 0, /* properties_destroyed */
4944 TODO_rebuild_alias, /* todo_flags_start */
4945 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4946 | TODO_verify_ssa /* todo_flags_finish */
4951 /* Gate and execute functions for FRE. */
4956 return execute_pre (true);
4962 return flag_tree_fre != 0;
4965 struct gimple_opt_pass pass_fre =
4970 gate_fre, /* gate */
4971 execute_fre, /* execute */
4974 0, /* static_pass_number */
4975 TV_TREE_FRE, /* tv_id */
4976 PROP_cfg | PROP_ssa, /* properties_required */
4977 0, /* properties_provided */
4978 0, /* properties_destroyed */
4979 0, /* todo_flags_start */
4980 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */