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;
1549 vn_reference_op_t operand;
1550 vn_reference_t newref;
1553 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1557 tree oldop0 = operand->op0;
1558 tree oldop1 = operand->op1;
1559 tree oldop2 = operand->op2;
1563 tree type = operand->type;
1564 vn_reference_op_s newop = *operand;
1566 if (op0 && TREE_CODE (op0) == SSA_NAME)
1568 unsigned int op_val_id = VN_INFO (op0)->value_id;
1569 leader = find_leader_in_sets (op_val_id, set1, set2);
1570 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1571 if (opresult && opresult != leader)
1573 tree name = get_representative_for (opresult);
1581 changed |= op0 != oldop0;
1583 if (op1 && TREE_CODE (op1) == SSA_NAME)
1585 unsigned int op_val_id = VN_INFO (op1)->value_id;
1586 leader = find_leader_in_sets (op_val_id, set1, set2);
1587 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1588 if (opresult && opresult != leader)
1590 tree name = get_representative_for (opresult);
1598 /* We can't possibly insert these. */
1599 else if (op1 && !is_gimple_min_invariant (op1))
1601 changed |= op1 != oldop1;
1602 if (op2 && TREE_CODE (op2) == SSA_NAME)
1604 unsigned int op_val_id = VN_INFO (op2)->value_id;
1605 leader = find_leader_in_sets (op_val_id, set1, set2);
1606 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1607 if (opresult && opresult != leader)
1609 tree name = get_representative_for (opresult);
1617 /* We can't possibly insert these. */
1618 else if (op2 && !is_gimple_min_invariant (op2))
1620 changed |= op2 != oldop2;
1623 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1624 /* We may have changed from an SSA_NAME to a constant */
1625 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1626 newop.opcode = TREE_CODE (op0);
1631 /* If it transforms a non-constant ARRAY_REF into a constant
1632 one, adjust the constant offset. */
1633 if (newop.opcode == ARRAY_REF
1635 && TREE_CODE (op0) == INTEGER_CST
1636 && TREE_CODE (op1) == INTEGER_CST
1637 && TREE_CODE (op2) == INTEGER_CST)
1639 double_int off = tree_to_double_int (op0);
1640 off = double_int_add (off,
1642 (tree_to_double_int (op1)));
1643 off = double_int_mul (off, tree_to_double_int (op2));
1644 if (double_int_fits_in_shwi_p (off))
1645 newop.off = off.low;
1647 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1648 /* If it transforms from an SSA_NAME to an address, fold with
1649 a preceding indirect reference. */
1650 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1651 && VEC_index (vn_reference_op_s,
1652 newoperands, j - 1)->opcode == MEM_REF)
1653 vn_reference_fold_indirect (&newoperands, &j);
1655 if (i != VEC_length (vn_reference_op_s, operands))
1658 VEC_free (vn_reference_op_s, heap, newoperands);
1664 newvuse = translate_vuse_through_block (newoperands,
1665 ref->set, ref->type,
1666 vuse, phiblock, pred,
1668 if (newvuse == NULL_TREE)
1670 VEC_free (vn_reference_op_s, heap, newoperands);
1675 if (changed || newvuse != vuse)
1677 unsigned int new_val_id;
1679 bool converted = false;
1681 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1686 VEC_free (vn_reference_op_s, heap, newoperands);
1689 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1691 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1695 if (result && is_gimple_min_invariant (result))
1697 gcc_assert (!newoperands);
1698 return get_or_alloc_expr_for_constant (result);
1701 expr = (pre_expr) pool_alloc (pre_expr_pool);
1702 expr->kind = REFERENCE;
1710 gcc_assert (CONVERT_EXPR_P (result)
1711 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1713 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1715 TREE_OPERAND (result, 0),
1716 NULL_TREE, NULL_TREE,
1719 if (nresult && is_gimple_min_invariant (nresult))
1720 return get_or_alloc_expr_for_constant (nresult);
1725 PRE_EXPR_NARY (expr) = nary;
1726 constant = fully_constant_expression (expr);
1727 if (constant != expr)
1730 new_val_id = nary->value_id;
1731 get_or_alloc_expression_id (expr);
1735 new_val_id = get_next_value_id ();
1736 VEC_safe_grow_cleared (bitmap_set_t, heap,
1738 get_max_value_id() + 1);
1739 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1741 TREE_OPERAND (result, 0),
1742 NULL_TREE, NULL_TREE,
1743 NULL_TREE, NULL_TREE,
1745 PRE_EXPR_NARY (expr) = nary;
1746 constant = fully_constant_expression (expr);
1747 if (constant != expr)
1749 get_or_alloc_expression_id (expr);
1754 PRE_EXPR_REFERENCE (expr) = newref;
1755 constant = fully_constant_expression (expr);
1756 if (constant != expr)
1759 new_val_id = newref->value_id;
1760 get_or_alloc_expression_id (expr);
1764 if (changed || !same_valid)
1766 new_val_id = get_next_value_id ();
1767 VEC_safe_grow_cleared (bitmap_set_t, heap,
1769 get_max_value_id() + 1);
1772 new_val_id = ref->value_id;
1773 newref = vn_reference_insert_pieces (newvuse, ref->set,
1776 result, new_val_id);
1778 PRE_EXPR_REFERENCE (expr) = newref;
1779 constant = fully_constant_expression (expr);
1780 if (constant != expr)
1782 get_or_alloc_expression_id (expr);
1784 add_to_value (new_val_id, expr);
1786 VEC_free (vn_reference_op_s, heap, newoperands);
1796 tree name = PRE_EXPR_NAME (expr);
1798 def_stmt = SSA_NAME_DEF_STMT (name);
1799 if (gimple_code (def_stmt) == GIMPLE_PHI
1800 && gimple_bb (def_stmt) == phiblock)
1805 e = find_edge (pred, gimple_bb (phi));
1808 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1811 if (TREE_CODE (def) == SSA_NAME)
1812 def = VN_INFO (def)->valnum;
1814 /* Handle constant. */
1815 if (is_gimple_min_invariant (def))
1816 return get_or_alloc_expr_for_constant (def);
1818 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1821 newexpr = get_or_alloc_expr_for_name (def);
1832 /* Wrapper around phi_translate_1 providing caching functionality. */
1835 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1836 basic_block pred, basic_block phiblock)
1843 /* Constants contain no values that need translation. */
1844 if (expr->kind == CONSTANT)
1847 if (value_id_constant_p (get_expr_value_id (expr)))
1850 if (expr->kind != NAME)
1852 phitrans = phi_trans_lookup (expr, pred);
1858 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1860 /* Don't add empty translations to the cache. Neither add
1861 translations of NAMEs as those are cheap to translate. */
1863 && expr->kind != NAME)
1864 phi_trans_add (expr, phitrans, pred);
1870 /* For each expression in SET, translate the values through phi nodes
1871 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1872 expressions in DEST. */
1875 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1876 basic_block phiblock)
1878 VEC (pre_expr, heap) *exprs;
1882 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1884 bitmap_set_copy (dest, set);
1888 exprs = sorted_array_from_bitmap_set (set);
1889 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1891 pre_expr translated;
1892 translated = phi_translate (expr, set, NULL, pred, phiblock);
1896 /* We might end up with multiple expressions from SET being
1897 translated to the same value. In this case we do not want
1898 to retain the NARY or REFERENCE expression but prefer a NAME
1899 which would be the leader. */
1900 if (translated->kind == NAME)
1901 bitmap_value_replace_in_set (dest, translated);
1903 bitmap_value_insert_into_set (dest, translated);
1905 VEC_free (pre_expr, heap, exprs);
1908 /* Find the leader for a value (i.e., the name representing that
1909 value) in a given set, and return it. If STMT is non-NULL it
1910 makes sure the defining statement for the leader dominates it.
1911 Return NULL if no leader is found. */
1914 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1916 if (value_id_constant_p (val))
1920 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1922 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1924 pre_expr expr = expression_for_id (i);
1925 if (expr->kind == CONSTANT)
1929 if (bitmap_set_contains_value (set, val))
1931 /* Rather than walk the entire bitmap of expressions, and see
1932 whether any of them has the value we are looking for, we look
1933 at the reverse mapping, which tells us the set of expressions
1934 that have a given value (IE value->expressions with that
1935 value) and see if any of those expressions are in our set.
1936 The number of expressions per value is usually significantly
1937 less than the number of expressions in the set. In fact, for
1938 large testcases, doing it this way is roughly 5-10x faster
1939 than walking the bitmap.
1940 If this is somehow a significant lose for some cases, we can
1941 choose which set to walk based on which set is smaller. */
1944 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1946 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1947 &set->expressions, 0, i, bi)
1949 pre_expr val = expression_for_id (i);
1950 /* At the point where stmt is not null, there should always
1951 be an SSA_NAME first in the list of expressions. */
1954 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1955 if (gimple_code (def_stmt) != GIMPLE_PHI
1956 && gimple_bb (def_stmt) == gimple_bb (stmt)
1957 /* PRE insertions are at the end of the basic-block
1959 && (gimple_uid (def_stmt) == 0
1960 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1969 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1970 BLOCK by seeing if it is not killed in the block. Note that we are
1971 only determining whether there is a store that kills it. Because
1972 of the order in which clean iterates over values, we are guaranteed
1973 that altered operands will have caused us to be eliminated from the
1974 ANTIC_IN set already. */
1977 value_dies_in_block_x (pre_expr expr, basic_block block)
1979 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1980 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1982 gimple_stmt_iterator gsi;
1983 unsigned id = get_expression_id (expr);
1990 /* Lookup a previously calculated result. */
1991 if (EXPR_DIES (block)
1992 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1993 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1995 /* A memory expression {e, VUSE} dies in the block if there is a
1996 statement that may clobber e. If, starting statement walk from the
1997 top of the basic block, a statement uses VUSE there can be no kill
1998 inbetween that use and the original statement that loaded {e, VUSE},
1999 so we can stop walking. */
2000 ref.base = NULL_TREE;
2001 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
2003 tree def_vuse, def_vdef;
2004 def = gsi_stmt (gsi);
2005 def_vuse = gimple_vuse (def);
2006 def_vdef = gimple_vdef (def);
2008 /* Not a memory statement. */
2012 /* Not a may-def. */
2015 /* A load with the same VUSE, we're done. */
2016 if (def_vuse == vuse)
2022 /* Init ref only if we really need it. */
2023 if (ref.base == NULL_TREE
2024 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
2030 /* If the statement may clobber expr, it dies. */
2031 if (stmt_may_clobber_ref_p_1 (def, &ref))
2038 /* Remember the result. */
2039 if (!EXPR_DIES (block))
2040 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2041 bitmap_set_bit (EXPR_DIES (block), id * 2);
2043 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2049 #define union_contains_value(SET1, SET2, VAL) \
2050 (bitmap_set_contains_value ((SET1), (VAL)) \
2051 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
2053 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2056 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2057 vn_reference_op_t vro)
2059 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2061 struct pre_expr_d temp;
2064 PRE_EXPR_NAME (&temp) = vro->op0;
2065 temp.id = lookup_expression_id (&temp);
2068 if (!union_contains_value (set1, set2,
2069 get_expr_value_id (&temp)))
2072 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2074 struct pre_expr_d temp;
2077 PRE_EXPR_NAME (&temp) = vro->op1;
2078 temp.id = lookup_expression_id (&temp);
2081 if (!union_contains_value (set1, set2,
2082 get_expr_value_id (&temp)))
2086 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2088 struct pre_expr_d temp;
2091 PRE_EXPR_NAME (&temp) = vro->op2;
2092 temp.id = lookup_expression_id (&temp);
2095 if (!union_contains_value (set1, set2,
2096 get_expr_value_id (&temp)))
2103 /* Determine if the expression EXPR is valid in SET1 U SET2.
2104 ONLY SET2 CAN BE NULL.
2105 This means that we have a leader for each part of the expression
2106 (if it consists of values), or the expression is an SSA_NAME.
2107 For loads/calls, we also see if the vuse is killed in this block. */
2110 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2116 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2120 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2121 for (i = 0; i < nary->length; i++)
2123 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2125 struct pre_expr_d temp;
2128 PRE_EXPR_NAME (&temp) = nary->op[i];
2129 temp.id = lookup_expression_id (&temp);
2132 if (!union_contains_value (set1, set2,
2133 get_expr_value_id (&temp)))
2137 /* If the NARY may trap make sure the block does not contain
2138 a possible exit point.
2139 ??? This is overly conservative if we translate AVAIL_OUT
2140 as the available expression might be after the exit point. */
2141 if (BB_MAY_NOTRETURN (block)
2142 && vn_nary_may_trap (nary))
2149 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2150 vn_reference_op_t vro;
2153 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2155 if (!vro_valid_in_sets (set1, set2, vro))
2160 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2161 if (!gimple_nop_p (def_stmt)
2162 && gimple_bb (def_stmt) != block
2163 && !dominated_by_p (CDI_DOMINATORS,
2164 block, gimple_bb (def_stmt)))
2167 return !value_dies_in_block_x (expr, block);
2174 /* Clean the set of expressions that are no longer valid in SET1 or
2175 SET2. This means expressions that are made up of values we have no
2176 leaders for in SET1 or SET2. This version is used for partial
2177 anticipation, which means it is not valid in either ANTIC_IN or
2181 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2183 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2187 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2189 if (!valid_in_sets (set1, set2, expr, block))
2190 bitmap_remove_from_set (set1, expr);
2192 VEC_free (pre_expr, heap, exprs);
2195 /* Clean the set of expressions that are no longer valid in SET. This
2196 means expressions that are made up of values we have no leaders for
2200 clean (bitmap_set_t set, basic_block block)
2202 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2206 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2208 if (!valid_in_sets (set, NULL, expr, block))
2209 bitmap_remove_from_set (set, expr);
2211 VEC_free (pre_expr, heap, exprs);
2214 static sbitmap has_abnormal_preds;
2216 /* List of blocks that may have changed during ANTIC computation and
2217 thus need to be iterated over. */
2219 static sbitmap changed_blocks;
2221 /* Decide whether to defer a block for a later iteration, or PHI
2222 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2223 should defer the block, and true if we processed it. */
2226 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2227 basic_block block, basic_block phiblock)
2229 if (!BB_VISITED (phiblock))
2231 SET_BIT (changed_blocks, block->index);
2232 BB_VISITED (block) = 0;
2233 BB_DEFERRED (block) = 1;
2237 phi_translate_set (dest, source, block, phiblock);
2241 /* Compute the ANTIC set for BLOCK.
2243 If succs(BLOCK) > 1 then
2244 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2245 else if succs(BLOCK) == 1 then
2246 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2248 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2252 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2254 bool changed = false;
2255 bitmap_set_t S, old, ANTIC_OUT;
2261 old = ANTIC_OUT = S = NULL;
2262 BB_VISITED (block) = 1;
2264 /* If any edges from predecessors are abnormal, antic_in is empty,
2266 if (block_has_abnormal_pred_edge)
2267 goto maybe_dump_sets;
2269 old = ANTIC_IN (block);
2270 ANTIC_OUT = bitmap_set_new ();
2272 /* If the block has no successors, ANTIC_OUT is empty. */
2273 if (EDGE_COUNT (block->succs) == 0)
2275 /* If we have one successor, we could have some phi nodes to
2276 translate through. */
2277 else if (single_succ_p (block))
2279 basic_block succ_bb = single_succ (block);
2281 /* We trade iterations of the dataflow equations for having to
2282 phi translate the maximal set, which is incredibly slow
2283 (since the maximal set often has 300+ members, even when you
2284 have a small number of blocks).
2285 Basically, we defer the computation of ANTIC for this block
2286 until we have processed it's successor, which will inevitably
2287 have a *much* smaller set of values to phi translate once
2288 clean has been run on it.
2289 The cost of doing this is that we technically perform more
2290 iterations, however, they are lower cost iterations.
2292 Timings for PRE on tramp3d-v4:
2293 without maximal set fix: 11 seconds
2294 with maximal set fix/without deferring: 26 seconds
2295 with maximal set fix/with deferring: 11 seconds
2298 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2302 goto maybe_dump_sets;
2305 /* If we have multiple successors, we take the intersection of all of
2306 them. Note that in the case of loop exit phi nodes, we may have
2307 phis to translate through. */
2310 VEC(basic_block, heap) * worklist;
2312 basic_block bprime, first = NULL;
2314 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2315 FOR_EACH_EDGE (e, ei, block->succs)
2318 && BB_VISITED (e->dest))
2320 else if (BB_VISITED (e->dest))
2321 VEC_quick_push (basic_block, worklist, e->dest);
2324 /* Of multiple successors we have to have visited one already. */
2327 SET_BIT (changed_blocks, block->index);
2328 BB_VISITED (block) = 0;
2329 BB_DEFERRED (block) = 1;
2331 VEC_free (basic_block, heap, worklist);
2332 goto maybe_dump_sets;
2335 if (!gimple_seq_empty_p (phi_nodes (first)))
2336 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2338 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2340 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2342 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2344 bitmap_set_t tmp = bitmap_set_new ();
2345 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2346 bitmap_set_and (ANTIC_OUT, tmp);
2347 bitmap_set_free (tmp);
2350 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2352 VEC_free (basic_block, heap, worklist);
2355 /* Generate ANTIC_OUT - TMP_GEN. */
2356 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2358 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2359 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2362 /* Then union in the ANTIC_OUT - TMP_GEN values,
2363 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2364 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2365 bitmap_value_insert_into_set (ANTIC_IN (block),
2366 expression_for_id (bii));
2368 clean (ANTIC_IN (block), block);
2370 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2373 SET_BIT (changed_blocks, block->index);
2374 FOR_EACH_EDGE (e, ei, block->preds)
2375 SET_BIT (changed_blocks, e->src->index);
2378 RESET_BIT (changed_blocks, block->index);
2381 if (dump_file && (dump_flags & TDF_DETAILS))
2383 if (!BB_DEFERRED (block) || BB_VISITED (block))
2386 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2388 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2392 print_bitmap_set (dump_file, S, "S", block->index);
2397 "Block %d was deferred for a future iteration.\n",
2402 bitmap_set_free (old);
2404 bitmap_set_free (S);
2406 bitmap_set_free (ANTIC_OUT);
2410 /* Compute PARTIAL_ANTIC for BLOCK.
2412 If succs(BLOCK) > 1 then
2413 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2414 in ANTIC_OUT for all succ(BLOCK)
2415 else if succs(BLOCK) == 1 then
2416 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2418 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2423 compute_partial_antic_aux (basic_block block,
2424 bool block_has_abnormal_pred_edge)
2426 bool changed = false;
2427 bitmap_set_t old_PA_IN;
2428 bitmap_set_t PA_OUT;
2431 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2433 old_PA_IN = PA_OUT = NULL;
2435 /* If any edges from predecessors are abnormal, antic_in is empty,
2437 if (block_has_abnormal_pred_edge)
2438 goto maybe_dump_sets;
2440 /* If there are too many partially anticipatable values in the
2441 block, phi_translate_set can take an exponential time: stop
2442 before the translation starts. */
2444 && single_succ_p (block)
2445 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2446 goto maybe_dump_sets;
2448 old_PA_IN = PA_IN (block);
2449 PA_OUT = bitmap_set_new ();
2451 /* If the block has no successors, ANTIC_OUT is empty. */
2452 if (EDGE_COUNT (block->succs) == 0)
2454 /* If we have one successor, we could have some phi nodes to
2455 translate through. Note that we can't phi translate across DFS
2456 back edges in partial antic, because it uses a union operation on
2457 the successors. For recurrences like IV's, we will end up
2458 generating a new value in the set on each go around (i + 3 (VH.1)
2459 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2460 else if (single_succ_p (block))
2462 basic_block succ = single_succ (block);
2463 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2464 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2466 /* If we have multiple successors, we take the union of all of
2470 VEC(basic_block, heap) * worklist;
2474 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2475 FOR_EACH_EDGE (e, ei, block->succs)
2477 if (e->flags & EDGE_DFS_BACK)
2479 VEC_quick_push (basic_block, worklist, e->dest);
2481 if (VEC_length (basic_block, worklist) > 0)
2483 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2488 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2489 bitmap_value_insert_into_set (PA_OUT,
2490 expression_for_id (i));
2491 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2493 bitmap_set_t pa_in = bitmap_set_new ();
2494 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2495 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2496 bitmap_value_insert_into_set (PA_OUT,
2497 expression_for_id (i));
2498 bitmap_set_free (pa_in);
2501 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2502 bitmap_value_insert_into_set (PA_OUT,
2503 expression_for_id (i));
2506 VEC_free (basic_block, heap, worklist);
2509 /* PA_IN starts with PA_OUT - TMP_GEN.
2510 Then we subtract things from ANTIC_IN. */
2511 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2513 /* For partial antic, we want to put back in the phi results, since
2514 we will properly avoid making them partially antic over backedges. */
2515 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2516 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2518 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2519 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2521 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2523 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2526 SET_BIT (changed_blocks, block->index);
2527 FOR_EACH_EDGE (e, ei, block->preds)
2528 SET_BIT (changed_blocks, e->src->index);
2531 RESET_BIT (changed_blocks, block->index);
2534 if (dump_file && (dump_flags & TDF_DETAILS))
2537 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2539 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2542 bitmap_set_free (old_PA_IN);
2544 bitmap_set_free (PA_OUT);
2548 /* Compute ANTIC and partial ANTIC sets. */
2551 compute_antic (void)
2553 bool changed = true;
2554 int num_iterations = 0;
2558 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2559 We pre-build the map of blocks with incoming abnormal edges here. */
2560 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2561 sbitmap_zero (has_abnormal_preds);
2568 FOR_EACH_EDGE (e, ei, block->preds)
2570 e->flags &= ~EDGE_DFS_BACK;
2571 if (e->flags & EDGE_ABNORMAL)
2573 SET_BIT (has_abnormal_preds, block->index);
2578 BB_VISITED (block) = 0;
2579 BB_DEFERRED (block) = 0;
2581 /* While we are here, give empty ANTIC_IN sets to each block. */
2582 ANTIC_IN (block) = bitmap_set_new ();
2583 PA_IN (block) = bitmap_set_new ();
2586 /* At the exit block we anticipate nothing. */
2587 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2588 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2589 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2591 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2592 sbitmap_ones (changed_blocks);
2595 if (dump_file && (dump_flags & TDF_DETAILS))
2596 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2599 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2601 if (TEST_BIT (changed_blocks, postorder[i]))
2603 basic_block block = BASIC_BLOCK (postorder[i]);
2604 changed |= compute_antic_aux (block,
2605 TEST_BIT (has_abnormal_preds,
2609 /* Theoretically possible, but *highly* unlikely. */
2610 gcc_checking_assert (num_iterations < 500);
2613 statistics_histogram_event (cfun, "compute_antic iterations",
2616 if (do_partial_partial)
2618 sbitmap_ones (changed_blocks);
2619 mark_dfs_back_edges ();
2624 if (dump_file && (dump_flags & TDF_DETAILS))
2625 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2628 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2630 if (TEST_BIT (changed_blocks, postorder[i]))
2632 basic_block block = BASIC_BLOCK (postorder[i]);
2634 |= compute_partial_antic_aux (block,
2635 TEST_BIT (has_abnormal_preds,
2639 /* Theoretically possible, but *highly* unlikely. */
2640 gcc_checking_assert (num_iterations < 500);
2642 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2645 sbitmap_free (has_abnormal_preds);
2646 sbitmap_free (changed_blocks);
2649 /* Return true if we can value number the call in STMT. This is true
2650 if we have a pure or constant call. */
2653 can_value_number_call (gimple stmt)
2655 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2660 /* Return true if OP is a tree which we can perform PRE on.
2661 This may not match the operations we can value number, but in
2662 a perfect world would. */
2665 can_PRE_operation (tree op)
2667 return UNARY_CLASS_P (op)
2668 || BINARY_CLASS_P (op)
2669 || COMPARISON_CLASS_P (op)
2670 || TREE_CODE (op) == MEM_REF
2671 || TREE_CODE (op) == COMPONENT_REF
2672 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2673 || TREE_CODE (op) == CALL_EXPR
2674 || TREE_CODE (op) == ARRAY_REF;
2678 /* Inserted expressions are placed onto this worklist, which is used
2679 for performing quick dead code elimination of insertions we made
2680 that didn't turn out to be necessary. */
2681 static bitmap inserted_exprs;
2683 /* Pool allocated fake store expressions are placed onto this
2684 worklist, which, after performing dead code elimination, is walked
2685 to see which expressions need to be put into GC'able memory */
2686 static VEC(gimple, heap) *need_creation;
2688 /* The actual worker for create_component_ref_by_pieces. */
2691 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2692 unsigned int *operand, gimple_seq *stmts,
2695 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2699 switch (currop->opcode)
2703 tree folded, sc = NULL_TREE;
2704 unsigned int nargs = 0;
2706 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2710 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2711 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2717 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2718 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2722 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2723 ref->operands) - 1);
2724 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2726 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2736 folded = build_call_array (currop->type,
2737 (TREE_CODE (fn) == FUNCTION_DECL
2738 ? build_fold_addr_expr (fn) : fn),
2742 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2748 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2750 tree offset = currop->op0;
2753 if (TREE_CODE (baseop) == ADDR_EXPR
2754 && handled_component_p (TREE_OPERAND (baseop, 0)))
2758 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2761 offset = int_const_binop (PLUS_EXPR, offset,
2762 build_int_cst (TREE_TYPE (offset),
2764 baseop = build_fold_addr_expr (base);
2766 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2769 case TARGET_MEM_REF:
2771 pre_expr op0expr, op1expr;
2772 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2773 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2775 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2781 op0expr = get_or_alloc_expr_for (currop->op0);
2782 genop0 = find_or_generate_expression (block, op0expr,
2789 op1expr = get_or_alloc_expr_for (nextop->op0);
2790 genop1 = find_or_generate_expression (block, op1expr,
2795 return build5 (TARGET_MEM_REF, currop->type,
2796 baseop, currop->op2, genop0, currop->op1, genop1);
2802 gcc_assert (is_gimple_min_invariant (currop->op0));
2808 case VIEW_CONVERT_EXPR:
2811 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2816 folded = fold_build1 (currop->opcode, currop->type,
2824 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2826 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2827 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2833 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2836 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2839 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2844 /* For array ref vn_reference_op's, operand 1 of the array ref
2845 is op0 of the reference op and operand 3 of the array ref is
2847 case ARRAY_RANGE_REF:
2851 tree genop1 = currop->op0;
2853 tree genop2 = currop->op1;
2855 tree genop3 = currop->op2;
2857 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2861 op1expr = get_or_alloc_expr_for (genop1);
2862 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2867 /* Drop zero minimum index. */
2868 if (tree_int_cst_equal (genop2, integer_zero_node))
2872 op2expr = get_or_alloc_expr_for (genop2);
2873 genop2 = find_or_generate_expression (block, op2expr, stmts,
2881 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2882 /* We can't always put a size in units of the element alignment
2883 here as the element alignment may be not visible. See
2884 PR43783. Simply drop the element size for constant
2886 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2890 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2891 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2892 op3expr = get_or_alloc_expr_for (genop3);
2893 genop3 = find_or_generate_expression (block, op3expr, stmts,
2899 return build4 (currop->opcode, currop->type, genop0, genop1,
2906 tree genop2 = currop->op1;
2908 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2912 /* op1 should be a FIELD_DECL, which are represented by
2917 op2expr = get_or_alloc_expr_for (genop2);
2918 genop2 = find_or_generate_expression (block, op2expr, stmts,
2924 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2930 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2931 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2952 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2953 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2954 trying to rename aggregates into ssa form directly, which is a no no.
2956 Thus, this routine doesn't create temporaries, it just builds a
2957 single access expression for the array, calling
2958 find_or_generate_expression to build the innermost pieces.
2960 This function is a subroutine of create_expression_by_pieces, and
2961 should not be called on it's own unless you really know what you
2965 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2966 gimple_seq *stmts, gimple domstmt)
2968 unsigned int op = 0;
2969 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2972 /* Find a leader for an expression, or generate one using
2973 create_expression_by_pieces if it's ANTIC but
2975 BLOCK is the basic_block we are looking for leaders in.
2976 EXPR is the expression to find a leader or generate for.
2977 STMTS is the statement list to put the inserted expressions on.
2978 Returns the SSA_NAME of the LHS of the generated expression or the
2980 DOMSTMT if non-NULL is a statement that should be dominated by
2981 all uses in the generated expression. If DOMSTMT is non-NULL this
2982 routine can fail and return NULL_TREE. Otherwise it will assert
2986 find_or_generate_expression (basic_block block, pre_expr expr,
2987 gimple_seq *stmts, gimple domstmt)
2989 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2990 get_expr_value_id (expr), domstmt);
2994 if (leader->kind == NAME)
2995 genop = PRE_EXPR_NAME (leader);
2996 else if (leader->kind == CONSTANT)
2997 genop = PRE_EXPR_CONSTANT (leader);
3000 /* If it's still NULL, it must be a complex expression, so generate
3001 it recursively. Not so if inserting expressions for values generated
3006 bitmap_set_t exprset;
3007 unsigned int lookfor = get_expr_value_id (expr);
3008 bool handled = false;
3012 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3013 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3015 pre_expr temp = expression_for_id (i);
3016 if (temp->kind != NAME)
3019 genop = create_expression_by_pieces (block, temp, stmts,
3021 get_expr_type (expr));
3025 if (!handled && domstmt)
3028 gcc_assert (handled);
3033 #define NECESSARY GF_PLF_1
3035 /* Create an expression in pieces, so that we can handle very complex
3036 expressions that may be ANTIC, but not necessary GIMPLE.
3037 BLOCK is the basic block the expression will be inserted into,
3038 EXPR is the expression to insert (in value form)
3039 STMTS is a statement list to append the necessary insertions into.
3041 This function will die if we hit some value that shouldn't be
3042 ANTIC but is (IE there is no leader for it, or its components).
3043 This function may also generate expressions that are themselves
3044 partially or fully redundant. Those that are will be either made
3045 fully redundant during the next iteration of insert (for partially
3046 redundant ones), or eliminated by eliminate (for fully redundant
3049 If DOMSTMT is non-NULL then we make sure that all uses in the
3050 expressions dominate that statement. In this case the function
3051 can return NULL_TREE to signal failure. */
3054 create_expression_by_pieces (basic_block block, pre_expr expr,
3055 gimple_seq *stmts, gimple domstmt, tree type)
3059 gimple_seq forced_stmts = NULL;
3060 unsigned int value_id;
3061 gimple_stmt_iterator gsi;
3062 tree exprtype = type ? type : get_expr_type (expr);
3068 /* We may hit the NAME/CONSTANT case if we have to convert types
3069 that value numbering saw through. */
3071 folded = PRE_EXPR_NAME (expr);
3074 folded = PRE_EXPR_CONSTANT (expr);
3078 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3079 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3084 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3085 switch (nary->length)
3089 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3090 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
3091 tree genop1 = find_or_generate_expression (block, op1,
3093 tree genop2 = find_or_generate_expression (block, op2,
3095 if (!genop1 || !genop2)
3097 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
3098 may be a constant with the wrong type. */
3099 if (nary->opcode == POINTER_PLUS_EXPR)
3101 genop1 = fold_convert (nary->type, genop1);
3102 genop2 = fold_convert (sizetype, genop2);
3106 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3107 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
3110 folded = fold_build2 (nary->opcode, nary->type,
3116 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3117 tree genop1 = find_or_generate_expression (block, op1,
3121 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3123 folded = fold_build1 (nary->opcode, nary->type,
3136 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3137 folded = fold_convert (exprtype, folded);
3139 /* Force the generated expression to be a sequence of GIMPLE
3141 We have to call unshare_expr because force_gimple_operand may
3142 modify the tree we pass to it. */
3143 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3146 /* If we have any intermediate expressions to the value sets, add them
3147 to the value sets and chain them in the instruction stream. */
3150 gsi = gsi_start (forced_stmts);
3151 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3153 gimple stmt = gsi_stmt (gsi);
3154 tree forcedname = gimple_get_lhs (stmt);
3157 if (TREE_CODE (forcedname) == SSA_NAME)
3159 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3160 VN_INFO_GET (forcedname)->valnum = forcedname;
3161 VN_INFO (forcedname)->value_id = get_next_value_id ();
3162 nameexpr = get_or_alloc_expr_for_name (forcedname);
3163 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3165 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3166 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3168 mark_symbols_for_renaming (stmt);
3170 gimple_seq_add_seq (stmts, forced_stmts);
3173 /* Build and insert the assignment of the end result to the temporary
3174 that we will return. */
3175 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3177 pretemp = create_tmp_reg (exprtype, "pretmp");
3178 get_var_ann (pretemp);
3182 add_referenced_var (temp);
3184 newstmt = gimple_build_assign (temp, folded);
3185 name = make_ssa_name (temp, newstmt);
3186 gimple_assign_set_lhs (newstmt, name);
3187 gimple_set_plf (newstmt, NECESSARY, false);
3189 gimple_seq_add_stmt (stmts, newstmt);
3190 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3192 /* All the symbols in NEWEXPR should be put into SSA form. */
3193 mark_symbols_for_renaming (newstmt);
3195 /* Add a value number to the temporary.
3196 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3197 we are creating the expression by pieces, and this particular piece of
3198 the expression may have been represented. There is no harm in replacing
3200 VN_INFO_GET (name)->valnum = name;
3201 value_id = get_expr_value_id (expr);
3202 VN_INFO (name)->value_id = value_id;
3203 nameexpr = get_or_alloc_expr_for_name (name);
3204 add_to_value (value_id, nameexpr);
3205 if (NEW_SETS (block))
3206 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3207 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3209 pre_stats.insertions++;
3210 if (dump_file && (dump_flags & TDF_DETAILS))
3212 fprintf (dump_file, "Inserted ");
3213 print_gimple_stmt (dump_file, newstmt, 0, 0);
3214 fprintf (dump_file, " in predecessor %d\n", block->index);
3221 /* Returns true if we want to inhibit the insertions of PHI nodes
3222 for the given EXPR for basic block BB (a member of a loop).
3223 We want to do this, when we fear that the induction variable we
3224 create might inhibit vectorization. */
3227 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3229 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3230 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3231 vn_reference_op_t op;
3234 /* If we aren't going to vectorize we don't inhibit anything. */
3235 if (!flag_tree_vectorize)
3238 /* Otherwise we inhibit the insertion when the address of the
3239 memory reference is a simple induction variable. In other
3240 cases the vectorizer won't do anything anyway (either it's
3241 loop invariant or a complicated expression). */
3242 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3247 case ARRAY_RANGE_REF:
3248 if (TREE_CODE (op->op0) != SSA_NAME)
3253 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3255 /* Default defs are loop invariant. */
3258 /* Defined outside this loop, also loop invariant. */
3259 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3261 /* If it's a simple induction variable inhibit insertion,
3262 the vectorizer might be interested in this one. */
3263 if (simple_iv (bb->loop_father, bb->loop_father,
3264 op->op0, &iv, true))
3266 /* No simple IV, vectorizer can't do anything, hence no
3267 reason to inhibit the transformation for this operand. */
3277 /* Insert the to-be-made-available values of expression EXPRNUM for each
3278 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3279 merge the result with a phi node, given the same value number as
3280 NODE. Return true if we have inserted new stuff. */
3283 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3286 pre_expr expr = expression_for_id (exprnum);
3288 unsigned int val = get_expr_value_id (expr);
3290 bool insertions = false;
3295 tree type = get_expr_type (expr);
3299 if (dump_file && (dump_flags & TDF_DETAILS))
3301 fprintf (dump_file, "Found partial redundancy for expression ");
3302 print_pre_expr (dump_file, expr);
3303 fprintf (dump_file, " (%04d)\n", val);
3306 /* Make sure we aren't creating an induction variable. */
3307 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3309 bool firstinsideloop = false;
3310 bool secondinsideloop = false;
3311 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3312 EDGE_PRED (block, 0)->src);
3313 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3314 EDGE_PRED (block, 1)->src);
3315 /* Induction variables only have one edge inside the loop. */
3316 if ((firstinsideloop ^ secondinsideloop)
3317 && (expr->kind != REFERENCE
3318 || inhibit_phi_insertion (block, expr)))
3320 if (dump_file && (dump_flags & TDF_DETAILS))
3321 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3326 /* Make the necessary insertions. */
3327 FOR_EACH_EDGE (pred, ei, block->preds)
3329 gimple_seq stmts = NULL;
3332 eprime = avail[bprime->index];
3334 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3336 builtexpr = create_expression_by_pieces (bprime,
3340 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3341 gsi_insert_seq_on_edge (pred, stmts);
3342 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3345 else if (eprime->kind == CONSTANT)
3347 /* Constants may not have the right type, fold_convert
3348 should give us back a constant with the right type.
3350 tree constant = PRE_EXPR_CONSTANT (eprime);
3351 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3353 tree builtexpr = fold_convert (type, constant);
3354 if (!is_gimple_min_invariant (builtexpr))
3356 tree forcedexpr = force_gimple_operand (builtexpr,
3359 if (!is_gimple_min_invariant (forcedexpr))
3361 if (forcedexpr != builtexpr)
3363 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3364 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3368 gimple_stmt_iterator gsi;
3369 gsi = gsi_start (stmts);
3370 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3372 gimple stmt = gsi_stmt (gsi);
3373 tree lhs = gimple_get_lhs (stmt);
3374 if (TREE_CODE (lhs) == SSA_NAME)
3375 bitmap_set_bit (inserted_exprs,
3376 SSA_NAME_VERSION (lhs));
3377 gimple_set_plf (stmt, NECESSARY, false);
3379 gsi_insert_seq_on_edge (pred, stmts);
3381 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3385 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3388 else if (eprime->kind == NAME)
3390 /* We may have to do a conversion because our value
3391 numbering can look through types in certain cases, but
3392 our IL requires all operands of a phi node have the same
3394 tree name = PRE_EXPR_NAME (eprime);
3395 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3399 builtexpr = fold_convert (type, name);
3400 forcedexpr = force_gimple_operand (builtexpr,
3404 if (forcedexpr != name)
3406 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3407 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3412 gimple_stmt_iterator gsi;
3413 gsi = gsi_start (stmts);
3414 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3416 gimple stmt = gsi_stmt (gsi);
3417 tree lhs = gimple_get_lhs (stmt);
3418 if (TREE_CODE (lhs) == SSA_NAME)
3419 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3420 gimple_set_plf (stmt, NECESSARY, false);
3422 gsi_insert_seq_on_edge (pred, stmts);
3424 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3428 /* If we didn't want a phi node, and we made insertions, we still have
3429 inserted new stuff, and thus return true. If we didn't want a phi node,
3430 and didn't make insertions, we haven't added anything new, so return
3432 if (nophi && insertions)
3434 else if (nophi && !insertions)
3437 /* Now build a phi for the new variable. */
3438 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3440 prephitemp = create_tmp_var (type, "prephitmp");
3441 get_var_ann (prephitemp);
3445 add_referenced_var (temp);
3447 if (TREE_CODE (type) == COMPLEX_TYPE
3448 || TREE_CODE (type) == VECTOR_TYPE)
3449 DECL_GIMPLE_REG_P (temp) = 1;
3450 phi = create_phi_node (temp, block);
3452 gimple_set_plf (phi, NECESSARY, false);
3453 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3454 VN_INFO (gimple_phi_result (phi))->value_id = val;
3455 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3456 FOR_EACH_EDGE (pred, ei, block->preds)
3458 pre_expr ae = avail[pred->src->index];
3459 gcc_assert (get_expr_type (ae) == type
3460 || useless_type_conversion_p (type, get_expr_type (ae)));
3461 if (ae->kind == CONSTANT)
3462 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3464 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3468 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3469 add_to_value (val, newphi);
3471 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3472 this insertion, since we test for the existence of this value in PHI_GEN
3473 before proceeding with the partial redundancy checks in insert_aux.
3475 The value may exist in AVAIL_OUT, in particular, it could be represented
3476 by the expression we are trying to eliminate, in which case we want the
3477 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3480 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3481 this block, because if it did, it would have existed in our dominator's
3482 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3485 bitmap_insert_into_set (PHI_GEN (block), newphi);
3486 bitmap_value_replace_in_set (AVAIL_OUT (block),
3488 bitmap_insert_into_set (NEW_SETS (block),
3491 if (dump_file && (dump_flags & TDF_DETAILS))
3493 fprintf (dump_file, "Created phi ");
3494 print_gimple_stmt (dump_file, phi, 0, 0);
3495 fprintf (dump_file, " in block %d\n", block->index);
3503 /* Perform insertion of partially redundant values.
3504 For BLOCK, do the following:
3505 1. Propagate the NEW_SETS of the dominator into the current block.
3506 If the block has multiple predecessors,
3507 2a. Iterate over the ANTIC expressions for the block to see if
3508 any of them are partially redundant.
3509 2b. If so, insert them into the necessary predecessors to make
3510 the expression fully redundant.
3511 2c. Insert a new PHI merging the values of the predecessors.
3512 2d. Insert the new PHI, and the new expressions, into the
3514 3. Recursively call ourselves on the dominator children of BLOCK.
3516 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3517 do_regular_insertion and do_partial_insertion.
3522 do_regular_insertion (basic_block block, basic_block dom)
3524 bool new_stuff = false;
3525 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3529 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3531 if (expr->kind != NAME)
3535 bool by_some = false;
3536 bool cant_insert = false;
3537 bool all_same = true;
3538 pre_expr first_s = NULL;
3541 pre_expr eprime = NULL;
3543 pre_expr edoubleprime = NULL;
3544 bool do_insertion = false;
3546 val = get_expr_value_id (expr);
3547 if (bitmap_set_contains_value (PHI_GEN (block), val))
3549 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3551 if (dump_file && (dump_flags & TDF_DETAILS))
3552 fprintf (dump_file, "Found fully redundant value\n");
3556 avail = XCNEWVEC (pre_expr, last_basic_block);
3557 FOR_EACH_EDGE (pred, ei, block->preds)
3559 unsigned int vprime;
3561 /* We should never run insertion for the exit block
3562 and so not come across fake pred edges. */
3563 gcc_assert (!(pred->flags & EDGE_FAKE));
3565 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3568 /* eprime will generally only be NULL if the
3569 value of the expression, translated
3570 through the PHI for this predecessor, is
3571 undefined. If that is the case, we can't
3572 make the expression fully redundant,
3573 because its value is undefined along a
3574 predecessor path. We can thus break out
3575 early because it doesn't matter what the
3576 rest of the results are. */
3583 eprime = fully_constant_expression (eprime);
3584 vprime = get_expr_value_id (eprime);
3585 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3587 if (edoubleprime == NULL)
3589 avail[bprime->index] = eprime;
3594 avail[bprime->index] = edoubleprime;
3596 /* We want to perform insertions to remove a redundancy on
3597 a path in the CFG we want to optimize for speed. */
3598 if (optimize_edge_for_speed_p (pred))
3599 do_insertion = true;
3600 if (first_s == NULL)
3601 first_s = edoubleprime;
3602 else if (!pre_expr_eq (first_s, edoubleprime))
3606 /* If we can insert it, it's not the same value
3607 already existing along every predecessor, and
3608 it's defined by some predecessor, it is
3609 partially redundant. */
3610 if (!cant_insert && !all_same && by_some && do_insertion
3611 && dbg_cnt (treepre_insert))
3613 if (insert_into_preds_of_block (block, get_expression_id (expr),
3617 /* If all edges produce the same value and that value is
3618 an invariant, then the PHI has the same value on all
3619 edges. Note this. */
3620 else if (!cant_insert && all_same && eprime
3621 && (edoubleprime->kind == CONSTANT
3622 || edoubleprime->kind == NAME)
3623 && !value_id_constant_p (val))
3627 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3628 value_expressions, val);
3630 unsigned int new_val = get_expr_value_id (edoubleprime);
3631 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3633 pre_expr expr = expression_for_id (j);
3635 if (expr->kind == NAME)
3637 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3638 /* Just reset the value id and valnum so it is
3639 the same as the constant we have discovered. */
3640 if (edoubleprime->kind == CONSTANT)
3642 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3643 pre_stats.constified++;
3646 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3647 info->value_id = new_val;
3655 VEC_free (pre_expr, heap, exprs);
3660 /* Perform insertion for partially anticipatable expressions. There
3661 is only one case we will perform insertion for these. This case is
3662 if the expression is partially anticipatable, and fully available.
3663 In this case, we know that putting it earlier will enable us to
3664 remove the later computation. */
3668 do_partial_partial_insertion (basic_block block, basic_block dom)
3670 bool new_stuff = false;
3671 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3675 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3677 if (expr->kind != NAME)
3682 bool cant_insert = false;
3685 pre_expr eprime = NULL;
3688 val = get_expr_value_id (expr);
3689 if (bitmap_set_contains_value (PHI_GEN (block), val))
3691 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3694 avail = XCNEWVEC (pre_expr, last_basic_block);
3695 FOR_EACH_EDGE (pred, ei, block->preds)
3697 unsigned int vprime;
3698 pre_expr edoubleprime;
3700 /* We should never run insertion for the exit block
3701 and so not come across fake pred edges. */
3702 gcc_assert (!(pred->flags & EDGE_FAKE));
3704 eprime = phi_translate (expr, ANTIC_IN (block),
3708 /* eprime will generally only be NULL if the
3709 value of the expression, translated
3710 through the PHI for this predecessor, is
3711 undefined. If that is the case, we can't
3712 make the expression fully redundant,
3713 because its value is undefined along a
3714 predecessor path. We can thus break out
3715 early because it doesn't matter what the
3716 rest of the results are. */
3723 eprime = fully_constant_expression (eprime);
3724 vprime = get_expr_value_id (eprime);
3725 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3727 if (edoubleprime == NULL)
3733 avail[bprime->index] = edoubleprime;
3737 /* If we can insert it, it's not the same value
3738 already existing along every predecessor, and
3739 it's defined by some predecessor, it is
3740 partially redundant. */
3741 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3743 pre_stats.pa_insert++;
3744 if (insert_into_preds_of_block (block, get_expression_id (expr),
3752 VEC_free (pre_expr, heap, exprs);
3757 insert_aux (basic_block block)
3760 bool new_stuff = false;
3765 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3770 bitmap_set_t newset = NEW_SETS (dom);
3773 /* Note that we need to value_replace both NEW_SETS, and
3774 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3775 represented by some non-simple expression here that we want
3776 to replace it with. */
3777 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3779 pre_expr expr = expression_for_id (i);
3780 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3781 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3784 if (!single_pred_p (block))
3786 new_stuff |= do_regular_insertion (block, dom);
3787 if (do_partial_partial)
3788 new_stuff |= do_partial_partial_insertion (block, dom);
3792 for (son = first_dom_son (CDI_DOMINATORS, block);
3794 son = next_dom_son (CDI_DOMINATORS, son))
3796 new_stuff |= insert_aux (son);
3802 /* Perform insertion of partially redundant values. */
3807 bool new_stuff = true;
3809 int num_iterations = 0;
3812 NEW_SETS (bb) = bitmap_set_new ();
3817 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3819 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3823 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3826 add_to_exp_gen (basic_block block, tree op)
3831 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3833 result = get_or_alloc_expr_for_name (op);
3834 bitmap_value_insert_into_set (EXP_GEN (block), result);
3838 /* Create value ids for PHI in BLOCK. */
3841 make_values_for_phi (gimple phi, basic_block block)
3843 tree result = gimple_phi_result (phi);
3845 /* We have no need for virtual phis, as they don't represent
3846 actual computations. */
3847 if (is_gimple_reg (result))
3849 pre_expr e = get_or_alloc_expr_for_name (result);
3850 add_to_value (get_expr_value_id (e), e);
3851 bitmap_insert_into_set (PHI_GEN (block), e);
3852 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3856 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3858 tree arg = gimple_phi_arg_def (phi, i);
3859 if (TREE_CODE (arg) == SSA_NAME)
3861 e = get_or_alloc_expr_for_name (arg);
3862 add_to_value (get_expr_value_id (e), e);
3869 /* Compute the AVAIL set for all basic blocks.
3871 This function performs value numbering of the statements in each basic
3872 block. The AVAIL sets are built from information we glean while doing
3873 this value numbering, since the AVAIL sets contain only one entry per
3876 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3877 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3880 compute_avail (void)
3883 basic_block block, son;
3884 basic_block *worklist;
3888 /* We pretend that default definitions are defined in the entry block.
3889 This includes function arguments and the static chain decl. */
3890 for (i = 1; i < num_ssa_names; ++i)
3892 tree name = ssa_name (i);
3895 || !SSA_NAME_IS_DEFAULT_DEF (name)
3896 || has_zero_uses (name)
3897 || !is_gimple_reg (name))
3900 e = get_or_alloc_expr_for_name (name);
3901 add_to_value (get_expr_value_id (e), e);
3903 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3904 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3907 /* Allocate the worklist. */
3908 worklist = XNEWVEC (basic_block, n_basic_blocks);
3910 /* Seed the algorithm by putting the dominator children of the entry
3911 block on the worklist. */
3912 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3914 son = next_dom_son (CDI_DOMINATORS, son))
3915 worklist[sp++] = son;
3917 /* Loop until the worklist is empty. */
3920 gimple_stmt_iterator gsi;
3923 unsigned int stmt_uid = 1;
3925 /* Pick a block from the worklist. */
3926 block = worklist[--sp];
3928 /* Initially, the set of available values in BLOCK is that of
3929 its immediate dominator. */
3930 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3932 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3934 /* Generate values for PHI nodes. */
3935 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3936 make_values_for_phi (gsi_stmt (gsi), block);
3938 BB_MAY_NOTRETURN (block) = 0;
3940 /* Now compute value numbers and populate value sets with all
3941 the expressions computed in BLOCK. */
3942 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3947 stmt = gsi_stmt (gsi);
3948 gimple_set_uid (stmt, stmt_uid++);
3950 /* Cache whether the basic-block has any non-visible side-effect
3952 If this isn't a call or it is the last stmt in the
3953 basic-block then the CFG represents things correctly. */
3954 if (is_gimple_call (stmt)
3955 && !stmt_ends_bb_p (stmt))
3957 /* Non-looping const functions always return normally.
3958 Otherwise the call might not return or have side-effects
3959 that forbids hoisting possibly trapping expressions
3961 int flags = gimple_call_flags (stmt);
3962 if (!(flags & ECF_CONST)
3963 || (flags & ECF_LOOPING_CONST_OR_PURE))
3964 BB_MAY_NOTRETURN (block) = 1;
3967 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3969 pre_expr e = get_or_alloc_expr_for_name (op);
3971 add_to_value (get_expr_value_id (e), e);
3973 bitmap_insert_into_set (TMP_GEN (block), e);
3974 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3977 if (gimple_has_volatile_ops (stmt)
3978 || stmt_could_throw_p (stmt))
3981 switch (gimple_code (stmt))
3984 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3985 add_to_exp_gen (block, op);
3992 vn_reference_op_t vro;
3993 pre_expr result = NULL;
3994 VEC(vn_reference_op_s, heap) *ops = NULL;
3996 if (!can_value_number_call (stmt))
3999 copy_reference_ops_from_call (stmt, &ops);
4000 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4001 gimple_expr_type (stmt),
4003 VEC_free (vn_reference_op_s, heap, ops);
4007 for (i = 0; VEC_iterate (vn_reference_op_s,
4011 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4012 add_to_exp_gen (block, vro->op0);
4013 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4014 add_to_exp_gen (block, vro->op1);
4015 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4016 add_to_exp_gen (block, vro->op2);
4018 result = (pre_expr) pool_alloc (pre_expr_pool);
4019 result->kind = REFERENCE;
4021 PRE_EXPR_REFERENCE (result) = ref;
4023 get_or_alloc_expression_id (result);
4024 add_to_value (get_expr_value_id (result), result);
4026 bitmap_value_insert_into_set (EXP_GEN (block), result);
4032 pre_expr result = NULL;
4033 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4037 case tcc_comparison:
4042 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4043 gimple_assign_rhs_code (stmt),
4044 gimple_expr_type (stmt),
4045 gimple_assign_rhs1 (stmt),
4046 gimple_assign_rhs2 (stmt),
4047 NULL_TREE, NULL_TREE, &nary);
4052 for (i = 0; i < nary->length; i++)
4053 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4054 add_to_exp_gen (block, nary->op[i]);
4056 result = (pre_expr) pool_alloc (pre_expr_pool);
4057 result->kind = NARY;
4059 PRE_EXPR_NARY (result) = nary;
4063 case tcc_declaration:
4068 vn_reference_op_t vro;
4070 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4076 for (i = 0; VEC_iterate (vn_reference_op_s,
4080 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4081 add_to_exp_gen (block, vro->op0);
4082 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4083 add_to_exp_gen (block, vro->op1);
4084 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4085 add_to_exp_gen (block, vro->op2);
4087 result = (pre_expr) pool_alloc (pre_expr_pool);
4088 result->kind = REFERENCE;
4090 PRE_EXPR_REFERENCE (result) = ref;
4095 /* For any other statement that we don't
4096 recognize, simply add all referenced
4097 SSA_NAMEs to EXP_GEN. */
4098 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4099 add_to_exp_gen (block, op);
4103 get_or_alloc_expression_id (result);
4104 add_to_value (get_expr_value_id (result), result);
4106 bitmap_value_insert_into_set (EXP_GEN (block), result);
4115 /* Put the dominator children of BLOCK on the worklist of blocks
4116 to compute available sets for. */
4117 for (son = first_dom_son (CDI_DOMINATORS, block);
4119 son = next_dom_son (CDI_DOMINATORS, son))
4120 worklist[sp++] = son;
4126 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4127 than the available expressions for it. The insertion point is
4128 right before the first use in STMT. Returns the SSA_NAME that should
4129 be used for replacement. */
4132 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4134 basic_block bb = gimple_bb (stmt);
4135 gimple_stmt_iterator gsi;
4136 gimple_seq stmts = NULL;
4140 /* First create a value expression from the expression we want
4141 to insert and associate it with the value handle for SSA_VN. */
4142 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4146 /* Then use create_expression_by_pieces to generate a valid
4147 expression to insert at this point of the IL stream. */
4148 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4149 if (expr == NULL_TREE)
4151 gsi = gsi_for_stmt (stmt);
4152 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4157 /* Eliminate fully redundant computations. */
4162 VEC (gimple, heap) *to_remove = NULL;
4164 unsigned int todo = 0;
4165 gimple_stmt_iterator gsi;
4171 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4173 stmt = gsi_stmt (gsi);
4175 /* Lookup the RHS of the expression, see if we have an
4176 available computation for it. If so, replace the RHS with
4177 the available computation. */
4178 if (gimple_has_lhs (stmt)
4179 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4180 && !gimple_assign_ssa_name_copy_p (stmt)
4181 && (!gimple_assign_single_p (stmt)
4182 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4183 && !gimple_has_volatile_ops (stmt)
4184 && !has_zero_uses (gimple_get_lhs (stmt)))
4186 tree lhs = gimple_get_lhs (stmt);
4187 tree rhs = NULL_TREE;
4189 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4190 pre_expr sprimeexpr;
4192 if (gimple_assign_single_p (stmt))
4193 rhs = gimple_assign_rhs1 (stmt);
4195 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4196 get_expr_value_id (lhsexpr),
4201 if (sprimeexpr->kind == CONSTANT)
4202 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4203 else if (sprimeexpr->kind == NAME)
4204 sprime = PRE_EXPR_NAME (sprimeexpr);
4209 /* If there is no existing leader but SCCVN knows this
4210 value is constant, use that constant. */
4211 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4213 sprime = VN_INFO (lhs)->valnum;
4214 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4215 TREE_TYPE (sprime)))
4216 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4218 if (dump_file && (dump_flags & TDF_DETAILS))
4220 fprintf (dump_file, "Replaced ");
4221 print_gimple_expr (dump_file, stmt, 0, 0);
4222 fprintf (dump_file, " with ");
4223 print_generic_expr (dump_file, sprime, 0);
4224 fprintf (dump_file, " in ");
4225 print_gimple_stmt (dump_file, stmt, 0, 0);
4227 pre_stats.eliminations++;
4228 propagate_tree_value_into_stmt (&gsi, sprime);
4229 stmt = gsi_stmt (gsi);
4234 /* If there is no existing usable leader but SCCVN thinks
4235 it has an expression it wants to use as replacement,
4237 if (!sprime || sprime == lhs)
4239 tree val = VN_INFO (lhs)->valnum;
4241 && TREE_CODE (val) == SSA_NAME
4242 && VN_INFO (val)->needs_insertion
4243 && can_PRE_operation (vn_get_expr_for (val)))
4244 sprime = do_SCCVN_insertion (stmt, val);
4248 && (rhs == NULL_TREE
4249 || TREE_CODE (rhs) != SSA_NAME
4250 || may_propagate_copy (rhs, sprime)))
4252 bool can_make_abnormal_goto
4253 = is_gimple_call (stmt)
4254 && stmt_can_make_abnormal_goto (stmt);
4256 gcc_assert (sprime != rhs);
4258 if (dump_file && (dump_flags & TDF_DETAILS))
4260 fprintf (dump_file, "Replaced ");
4261 print_gimple_expr (dump_file, stmt, 0, 0);
4262 fprintf (dump_file, " with ");
4263 print_generic_expr (dump_file, sprime, 0);
4264 fprintf (dump_file, " in ");
4265 print_gimple_stmt (dump_file, stmt, 0, 0);
4268 if (TREE_CODE (sprime) == SSA_NAME)
4269 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4271 /* We need to make sure the new and old types actually match,
4272 which may require adding a simple cast, which fold_convert
4274 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4275 && !useless_type_conversion_p (gimple_expr_type (stmt),
4276 TREE_TYPE (sprime)))
4277 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4279 pre_stats.eliminations++;
4280 propagate_tree_value_into_stmt (&gsi, sprime);
4281 stmt = gsi_stmt (gsi);
4284 /* If we removed EH side-effects from the statement, clean
4285 its EH information. */
4286 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4288 bitmap_set_bit (need_eh_cleanup,
4289 gimple_bb (stmt)->index);
4290 if (dump_file && (dump_flags & TDF_DETAILS))
4291 fprintf (dump_file, " Removed EH side-effects.\n");
4294 /* Likewise for AB side-effects. */
4295 if (can_make_abnormal_goto
4296 && !stmt_can_make_abnormal_goto (stmt))
4298 bitmap_set_bit (need_ab_cleanup,
4299 gimple_bb (stmt)->index);
4300 if (dump_file && (dump_flags & TDF_DETAILS))
4301 fprintf (dump_file, " Removed AB side-effects.\n");
4305 /* If the statement is a scalar store, see if the expression
4306 has the same value number as its rhs. If so, the store is
4308 else if (gimple_assign_single_p (stmt)
4309 && !is_gimple_reg (gimple_assign_lhs (stmt))
4310 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4311 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4313 tree rhs = gimple_assign_rhs1 (stmt);
4315 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4316 gimple_vuse (stmt), true, NULL);
4317 if (TREE_CODE (rhs) == SSA_NAME)
4318 rhs = VN_INFO (rhs)->valnum;
4320 && operand_equal_p (val, rhs, 0))
4322 if (dump_file && (dump_flags & TDF_DETAILS))
4324 fprintf (dump_file, "Deleted redundant store ");
4325 print_gimple_stmt (dump_file, stmt, 0, 0);
4328 /* Queue stmt for removal. */
4329 VEC_safe_push (gimple, heap, to_remove, stmt);
4332 /* Visit COND_EXPRs and fold the comparison with the
4333 available value-numbers. */
4334 else if (gimple_code (stmt) == GIMPLE_COND)
4336 tree op0 = gimple_cond_lhs (stmt);
4337 tree op1 = gimple_cond_rhs (stmt);
4340 if (TREE_CODE (op0) == SSA_NAME)
4341 op0 = VN_INFO (op0)->valnum;
4342 if (TREE_CODE (op1) == SSA_NAME)
4343 op1 = VN_INFO (op1)->valnum;
4344 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4346 if (result && TREE_CODE (result) == INTEGER_CST)
4348 if (integer_zerop (result))
4349 gimple_cond_make_false (stmt);
4351 gimple_cond_make_true (stmt);
4353 todo = TODO_cleanup_cfg;
4356 /* Visit indirect calls and turn them into direct calls if
4358 if (is_gimple_call (stmt)
4359 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4361 tree fn = VN_INFO (gimple_call_fn (stmt))->valnum;
4362 if (TREE_CODE (fn) == ADDR_EXPR
4363 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
4365 bool can_make_abnormal_goto
4366 = stmt_can_make_abnormal_goto (stmt);
4367 bool was_noreturn = gimple_call_noreturn_p (stmt);
4369 if (dump_file && (dump_flags & TDF_DETAILS))
4371 fprintf (dump_file, "Replacing call target with ");
4372 print_generic_expr (dump_file, fn, 0);
4373 fprintf (dump_file, " in ");
4374 print_gimple_stmt (dump_file, stmt, 0, 0);
4377 gimple_call_set_fn (stmt, fn);
4380 /* When changing a call into a noreturn call, cfg cleanup
4381 is needed to fix up the noreturn call. */
4382 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4383 todo |= TODO_cleanup_cfg;
4385 /* If we removed EH side-effects from the statement, clean
4386 its EH information. */
4387 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4389 bitmap_set_bit (need_eh_cleanup,
4390 gimple_bb (stmt)->index);
4391 if (dump_file && (dump_flags & TDF_DETAILS))
4392 fprintf (dump_file, " Removed EH side-effects.\n");
4395 /* Likewise for AB side-effects. */
4396 if (can_make_abnormal_goto
4397 && !stmt_can_make_abnormal_goto (stmt))
4399 bitmap_set_bit (need_ab_cleanup,
4400 gimple_bb (stmt)->index);
4401 if (dump_file && (dump_flags & TDF_DETAILS))
4402 fprintf (dump_file, " Removed AB side-effects.\n");
4405 /* Changing an indirect call to a direct call may
4406 have exposed different semantics. This may
4407 require an SSA update. */
4408 todo |= TODO_update_ssa_only_virtuals;
4413 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4415 gimple stmt, phi = gsi_stmt (gsi);
4416 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4417 pre_expr sprimeexpr, resexpr;
4418 gimple_stmt_iterator gsi2;
4420 /* We want to perform redundant PHI elimination. Do so by
4421 replacing the PHI with a single copy if possible.
4422 Do not touch inserted, single-argument or virtual PHIs. */
4423 if (gimple_phi_num_args (phi) == 1
4424 || !is_gimple_reg (res))
4430 resexpr = get_or_alloc_expr_for_name (res);
4431 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4432 get_expr_value_id (resexpr), NULL);
4435 if (sprimeexpr->kind == CONSTANT)
4436 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4437 else if (sprimeexpr->kind == NAME)
4438 sprime = PRE_EXPR_NAME (sprimeexpr);
4442 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4444 sprime = VN_INFO (res)->valnum;
4445 if (!useless_type_conversion_p (TREE_TYPE (res),
4446 TREE_TYPE (sprime)))
4447 sprime = fold_convert (TREE_TYPE (res), sprime);
4456 if (dump_file && (dump_flags & TDF_DETAILS))
4458 fprintf (dump_file, "Replaced redundant PHI node defining ");
4459 print_generic_expr (dump_file, res, 0);
4460 fprintf (dump_file, " with ");
4461 print_generic_expr (dump_file, sprime, 0);
4462 fprintf (dump_file, "\n");
4465 remove_phi_node (&gsi, false);
4467 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4468 && TREE_CODE (sprime) == SSA_NAME)
4469 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4471 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4472 sprime = fold_convert (TREE_TYPE (res), sprime);
4473 stmt = gimple_build_assign (res, sprime);
4474 SSA_NAME_DEF_STMT (res) = stmt;
4475 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4477 gsi2 = gsi_after_labels (b);
4478 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4479 /* Queue the copy for eventual removal. */
4480 VEC_safe_push (gimple, heap, to_remove, stmt);
4481 /* If we inserted this PHI node ourself, it's not an elimination. */
4482 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4485 pre_stats.eliminations++;
4489 /* We cannot remove stmts during BB walk, especially not release SSA
4490 names there as this confuses the VN machinery. The stmts ending
4491 up in to_remove are either stores or simple copies. */
4492 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4494 tree lhs = gimple_assign_lhs (stmt);
4495 tree rhs = gimple_assign_rhs1 (stmt);
4496 use_operand_p use_p;
4499 /* If there is a single use only, propagate the equivalency
4500 instead of keeping the copy. */
4501 if (TREE_CODE (lhs) == SSA_NAME
4502 && TREE_CODE (rhs) == SSA_NAME
4503 && single_imm_use (lhs, &use_p, &use_stmt)
4504 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4506 SET_USE (use_p, rhs);
4507 update_stmt (use_stmt);
4508 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4509 && TREE_CODE (rhs) == SSA_NAME)
4510 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4513 /* If this is a store or a now unused copy, remove it. */
4514 if (TREE_CODE (lhs) != SSA_NAME
4515 || has_zero_uses (lhs))
4517 basic_block bb = gimple_bb (stmt);
4518 gsi = gsi_for_stmt (stmt);
4519 unlink_stmt_vdef (stmt);
4520 gsi_remove (&gsi, true);
4521 if (gimple_purge_dead_eh_edges (bb))
4522 todo |= TODO_cleanup_cfg;
4523 if (TREE_CODE (lhs) == SSA_NAME)
4524 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4525 release_defs (stmt);
4528 VEC_free (gimple, heap, to_remove);
4533 /* Borrow a bit of tree-ssa-dce.c for the moment.
4534 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4535 this may be a bit faster, and we may want critical edges kept split. */
4537 /* If OP's defining statement has not already been determined to be necessary,
4538 mark that statement necessary. Return the stmt, if it is newly
4541 static inline gimple
4542 mark_operand_necessary (tree op)
4548 if (TREE_CODE (op) != SSA_NAME)
4551 stmt = SSA_NAME_DEF_STMT (op);
4554 if (gimple_plf (stmt, NECESSARY)
4555 || gimple_nop_p (stmt))
4558 gimple_set_plf (stmt, NECESSARY, true);
4562 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4563 to insert PHI nodes sometimes, and because value numbering of casts isn't
4564 perfect, we sometimes end up inserting dead code. This simple DCE-like
4565 pass removes any insertions we made that weren't actually used. */
4568 remove_dead_inserted_code (void)
4575 worklist = BITMAP_ALLOC (NULL);
4576 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4578 t = SSA_NAME_DEF_STMT (ssa_name (i));
4579 if (gimple_plf (t, NECESSARY))
4580 bitmap_set_bit (worklist, i);
4582 while (!bitmap_empty_p (worklist))
4584 i = bitmap_first_set_bit (worklist);
4585 bitmap_clear_bit (worklist, i);
4586 t = SSA_NAME_DEF_STMT (ssa_name (i));
4588 /* PHI nodes are somewhat special in that each PHI alternative has
4589 data and control dependencies. All the statements feeding the
4590 PHI node's arguments are always necessary. */
4591 if (gimple_code (t) == GIMPLE_PHI)
4595 for (k = 0; k < gimple_phi_num_args (t); k++)
4597 tree arg = PHI_ARG_DEF (t, k);
4598 if (TREE_CODE (arg) == SSA_NAME)
4600 gimple n = mark_operand_necessary (arg);
4602 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4608 /* Propagate through the operands. Examine all the USE, VUSE and
4609 VDEF operands in this statement. Mark all the statements
4610 which feed this statement's uses as necessary. */
4614 /* The operands of VDEF expressions are also needed as they
4615 represent potential definitions that may reach this
4616 statement (VDEF operands allow us to follow def-def
4619 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4621 gimple n = mark_operand_necessary (use);
4623 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4628 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4630 t = SSA_NAME_DEF_STMT (ssa_name (i));
4631 if (!gimple_plf (t, NECESSARY))
4633 gimple_stmt_iterator gsi;
4635 if (dump_file && (dump_flags & TDF_DETAILS))
4637 fprintf (dump_file, "Removing unnecessary insertion:");
4638 print_gimple_stmt (dump_file, t, 0, 0);
4641 gsi = gsi_for_stmt (t);
4642 if (gimple_code (t) == GIMPLE_PHI)
4643 remove_phi_node (&gsi, true);
4646 gsi_remove (&gsi, true);
4651 BITMAP_FREE (worklist);
4654 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4655 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4656 the number of visited blocks. */
4659 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4661 edge_iterator *stack;
4663 int post_order_num = 0;
4666 if (include_entry_exit)
4667 post_order[post_order_num++] = EXIT_BLOCK;
4669 /* Allocate stack for back-tracking up CFG. */
4670 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4673 /* Allocate bitmap to track nodes that have been visited. */
4674 visited = sbitmap_alloc (last_basic_block);
4676 /* None of the nodes in the CFG have been visited yet. */
4677 sbitmap_zero (visited);
4679 /* Push the last edge on to the stack. */
4680 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4688 /* Look at the edge on the top of the stack. */
4690 src = ei_edge (ei)->src;
4691 dest = ei_edge (ei)->dest;
4693 /* Check if the edge destination has been visited yet. */
4694 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4696 /* Mark that we have visited the destination. */
4697 SET_BIT (visited, src->index);
4699 if (EDGE_COUNT (src->preds) > 0)
4700 /* Since the DEST node has been visited for the first
4701 time, check its successors. */
4702 stack[sp++] = ei_start (src->preds);
4704 post_order[post_order_num++] = src->index;
4708 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4709 post_order[post_order_num++] = dest->index;
4711 if (!ei_one_before_end_p (ei))
4712 ei_next (&stack[sp - 1]);
4718 if (include_entry_exit)
4719 post_order[post_order_num++] = ENTRY_BLOCK;
4722 sbitmap_free (visited);
4723 return post_order_num;
4727 /* Initialize data structures used by PRE. */
4730 init_pre (bool do_fre)
4734 next_expression_id = 1;
4736 VEC_safe_push (pre_expr, heap, expressions, NULL);
4737 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4738 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4739 get_max_value_id() + 1);
4744 inserted_exprs = BITMAP_ALLOC (NULL);
4745 need_creation = NULL;
4746 pretemp = NULL_TREE;
4747 storetemp = NULL_TREE;
4748 prephitemp = NULL_TREE;
4750 connect_infinite_loops_to_exit ();
4751 memset (&pre_stats, 0, sizeof (pre_stats));
4754 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4755 my_rev_post_order_compute (postorder, false);
4757 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4759 calculate_dominance_info (CDI_POST_DOMINATORS);
4760 calculate_dominance_info (CDI_DOMINATORS);
4762 bitmap_obstack_initialize (&grand_bitmap_obstack);
4763 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4764 expr_pred_trans_eq, free);
4765 expression_to_id = htab_create (num_ssa_names * 3,
4768 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4769 sizeof (struct bitmap_set), 30);
4770 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4771 sizeof (struct pre_expr_d), 30);
4774 EXP_GEN (bb) = bitmap_set_new ();
4775 PHI_GEN (bb) = bitmap_set_new ();
4776 TMP_GEN (bb) = bitmap_set_new ();
4777 AVAIL_OUT (bb) = bitmap_set_new ();
4780 need_eh_cleanup = BITMAP_ALLOC (NULL);
4781 need_ab_cleanup = BITMAP_ALLOC (NULL);
4785 /* Deallocate data structures used by PRE. */
4788 fini_pre (bool do_fre)
4791 VEC_free (bitmap_set_t, heap, value_expressions);
4792 BITMAP_FREE (inserted_exprs);
4793 VEC_free (gimple, heap, need_creation);
4794 bitmap_obstack_release (&grand_bitmap_obstack);
4795 free_alloc_pool (bitmap_set_pool);
4796 free_alloc_pool (pre_expr_pool);
4797 htab_delete (phi_translate_table);
4798 htab_delete (expression_to_id);
4799 VEC_free (unsigned, heap, name_to_id);
4801 free_aux_for_blocks ();
4803 free_dominance_info (CDI_POST_DOMINATORS);
4805 if (!bitmap_empty_p (need_eh_cleanup))
4807 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4808 cleanup_tree_cfg ();
4811 BITMAP_FREE (need_eh_cleanup);
4813 if (!bitmap_empty_p (need_ab_cleanup))
4815 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4816 cleanup_tree_cfg ();
4819 BITMAP_FREE (need_ab_cleanup);
4822 loop_optimizer_finalize ();
4825 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4826 only wants to do full redundancy elimination. */
4829 execute_pre (bool do_fre)
4831 unsigned int todo = 0;
4833 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4835 /* This has to happen before SCCVN runs because
4836 loop_optimizer_init may create new phis, etc. */
4838 loop_optimizer_init (LOOPS_NORMAL);
4843 loop_optimizer_finalize ();
4851 /* Collect and value number expressions computed in each basic block. */
4854 if (dump_file && (dump_flags & TDF_DETAILS))
4860 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4861 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4862 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4863 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4867 /* Insert can get quite slow on an incredibly large number of basic
4868 blocks due to some quadratic behavior. Until this behavior is
4869 fixed, don't run it when he have an incredibly large number of
4870 bb's. If we aren't going to run insert, there is no point in
4871 computing ANTIC, either, even though it's plenty fast. */
4872 if (!do_fre && n_basic_blocks < 4000)
4878 /* Make sure to remove fake edges before committing our inserts.
4879 This makes sure we don't end up with extra critical edges that
4880 we would need to split. */
4881 remove_fake_exit_edges ();
4882 gsi_commit_edge_inserts ();
4884 /* Remove all the redundant expressions. */
4885 todo |= eliminate ();
4887 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4888 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4889 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4890 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4891 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4893 clear_expression_ids ();
4896 remove_dead_inserted_code ();
4904 /* Gate and execute functions for PRE. */
4909 return execute_pre (false);
4915 return flag_tree_pre != 0;
4918 struct gimple_opt_pass pass_pre =
4923 gate_pre, /* gate */
4924 do_pre, /* execute */
4927 0, /* static_pass_number */
4928 TV_TREE_PRE, /* tv_id */
4929 PROP_no_crit_edges | PROP_cfg
4930 | PROP_ssa, /* properties_required */
4931 0, /* properties_provided */
4932 0, /* properties_destroyed */
4933 TODO_rebuild_alias, /* todo_flags_start */
4934 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4935 | TODO_verify_ssa /* todo_flags_finish */
4940 /* Gate and execute functions for FRE. */
4945 return execute_pre (true);
4951 return flag_tree_fre != 0;
4954 struct gimple_opt_pass pass_fre =
4959 gate_fre, /* gate */
4960 execute_fre, /* execute */
4963 0, /* static_pass_number */
4964 TV_TREE_FRE, /* tv_id */
4965 PROP_cfg | PROP_ssa, /* properties_required */
4966 0, /* properties_provided */
4967 0, /* properties_destroyed */
4968 0, /* todo_flags_start */
4969 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */