gcc/ChangeLog
[official-gcc.git] / gcc / tree-ssa-pre.c
blob345ebcceaad4e40f4ff6fa86c76a2412d62fb512
1 /* SSA-PRE for trees.
2 Copyright (C) 2001-2013 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
4 <stevenb@suse.de>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
11 any later version.
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "tree.h"
27 #include "basic-block.h"
28 #include "gimple-pretty-print.h"
29 #include "tree-inline.h"
30 #include "tree-flow.h"
31 #include "gimple.h"
32 #include "hash-table.h"
33 #include "tree-iterator.h"
34 #include "alloc-pool.h"
35 #include "obstack.h"
36 #include "tree-pass.h"
37 #include "flags.h"
38 #include "bitmap.h"
39 #include "langhooks.h"
40 #include "cfgloop.h"
41 #include "tree-ssa-sccvn.h"
42 #include "tree-scalar-evolution.h"
43 #include "params.h"
44 #include "dbgcnt.h"
45 #include "domwalk.h"
46 #include "ipa-prop.h"
48 /* TODO:
50 1. Avail sets can be shared by making an avail_find_leader that
51 walks up the dominator tree and looks in those avail sets.
52 This might affect code optimality, it's unclear right now.
53 2. Strength reduction can be performed by anticipating expressions
54 we can repair later on.
55 3. We can do back-substitution or smarter value numbering to catch
56 commutative expressions split up over multiple statements.
59 /* For ease of terminology, "expression node" in the below refers to
60 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
61 represent the actual statement containing the expressions we care about,
62 and we cache the value number by putting it in the expression. */
64 /* Basic algorithm
66 First we walk the statements to generate the AVAIL sets, the
67 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
68 generation of values/expressions by a given block. We use them
69 when computing the ANTIC sets. The AVAIL sets consist of
70 SSA_NAME's that represent values, so we know what values are
71 available in what blocks. AVAIL is a forward dataflow problem. In
72 SSA, values are never killed, so we don't need a kill set, or a
73 fixpoint iteration, in order to calculate the AVAIL sets. In
74 traditional parlance, AVAIL sets tell us the downsafety of the
75 expressions/values.
77 Next, we generate the ANTIC sets. These sets represent the
78 anticipatable expressions. ANTIC is a backwards dataflow
79 problem. An expression is anticipatable in a given block if it could
80 be generated in that block. This means that if we had to perform
81 an insertion in that block, of the value of that expression, we
82 could. Calculating the ANTIC sets requires phi translation of
83 expressions, because the flow goes backwards through phis. We must
84 iterate to a fixpoint of the ANTIC sets, because we have a kill
85 set. Even in SSA form, values are not live over the entire
86 function, only from their definition point onwards. So we have to
87 remove values from the ANTIC set once we go past the definition
88 point of the leaders that make them up.
89 compute_antic/compute_antic_aux performs this computation.
91 Third, we perform insertions to make partially redundant
92 expressions fully redundant.
94 An expression is partially redundant (excluding partial
95 anticipation) if:
97 1. It is AVAIL in some, but not all, of the predecessors of a
98 given block.
99 2. It is ANTIC in all the predecessors.
101 In order to make it fully redundant, we insert the expression into
102 the predecessors where it is not available, but is ANTIC.
104 For the partial anticipation case, we only perform insertion if it
105 is partially anticipated in some block, and fully available in all
106 of the predecessors.
108 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
109 performs these steps.
111 Fourth, we eliminate fully redundant expressions.
112 This is a simple statement walk that replaces redundant
113 calculations with the now available values. */
115 /* Representations of value numbers:
117 Value numbers are represented by a representative SSA_NAME. We
118 will create fake SSA_NAME's in situations where we need a
119 representative but do not have one (because it is a complex
120 expression). In order to facilitate storing the value numbers in
121 bitmaps, and keep the number of wasted SSA_NAME's down, we also
122 associate a value_id with each value number, and create full blown
123 ssa_name's only where we actually need them (IE in operands of
124 existing expressions).
126 Theoretically you could replace all the value_id's with
127 SSA_NAME_VERSION, but this would allocate a large number of
128 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
129 It would also require an additional indirection at each point we
130 use the value id. */
132 /* Representation of expressions on value numbers:
134 Expressions consisting of value numbers are represented the same
135 way as our VN internally represents them, with an additional
136 "pre_expr" wrapping around them in order to facilitate storing all
137 of the expressions in the same sets. */
139 /* Representation of sets:
141 The dataflow sets do not need to be sorted in any particular order
142 for the majority of their lifetime, are simply represented as two
143 bitmaps, one that keeps track of values present in the set, and one
144 that keeps track of expressions present in the set.
146 When we need them in topological order, we produce it on demand by
147 transforming the bitmap into an array and sorting it into topo
148 order. */
150 /* Type of expression, used to know which member of the PRE_EXPR union
151 is valid. */
153 enum pre_expr_kind
155 NAME,
156 NARY,
157 REFERENCE,
158 CONSTANT
161 typedef union pre_expr_union_d
163 tree name;
164 tree constant;
165 vn_nary_op_t nary;
166 vn_reference_t reference;
167 } pre_expr_union;
169 typedef struct pre_expr_d : typed_noop_remove <pre_expr_d>
171 enum pre_expr_kind kind;
172 unsigned int id;
173 pre_expr_union u;
175 /* hash_table support. */
176 typedef pre_expr_d value_type;
177 typedef pre_expr_d compare_type;
178 static inline hashval_t hash (const pre_expr_d *);
179 static inline int equal (const pre_expr_d *, const pre_expr_d *);
180 } *pre_expr;
182 #define PRE_EXPR_NAME(e) (e)->u.name
183 #define PRE_EXPR_NARY(e) (e)->u.nary
184 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
185 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
187 /* Compare E1 and E1 for equality. */
189 inline int
190 pre_expr_d::equal (const value_type *e1, const compare_type *e2)
192 if (e1->kind != e2->kind)
193 return false;
195 switch (e1->kind)
197 case CONSTANT:
198 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
199 PRE_EXPR_CONSTANT (e2));
200 case NAME:
201 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
202 case NARY:
203 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
204 case REFERENCE:
205 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
206 PRE_EXPR_REFERENCE (e2));
207 default:
208 gcc_unreachable ();
212 /* Hash E. */
214 inline hashval_t
215 pre_expr_d::hash (const value_type *e)
217 switch (e->kind)
219 case CONSTANT:
220 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
221 case NAME:
222 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
223 case NARY:
224 return PRE_EXPR_NARY (e)->hashcode;
225 case REFERENCE:
226 return PRE_EXPR_REFERENCE (e)->hashcode;
227 default:
228 gcc_unreachable ();
232 /* Next global expression id number. */
233 static unsigned int next_expression_id;
235 /* Mapping from expression to id number we can use in bitmap sets. */
236 static vec<pre_expr> expressions;
237 static hash_table <pre_expr_d> expression_to_id;
238 static vec<unsigned> name_to_id;
240 /* Allocate an expression id for EXPR. */
242 static inline unsigned int
243 alloc_expression_id (pre_expr expr)
245 struct pre_expr_d **slot;
246 /* Make sure we won't overflow. */
247 gcc_assert (next_expression_id + 1 > next_expression_id);
248 expr->id = next_expression_id++;
249 expressions.safe_push (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 unsigned old_len = name_to_id.length ();
257 name_to_id.reserve (num_ssa_names - old_len);
258 name_to_id.safe_grow_cleared (num_ssa_names);
259 gcc_assert (name_to_id[version] == 0);
260 name_to_id[version] = expr->id;
262 else
264 slot = expression_to_id.find_slot (expr, INSERT);
265 gcc_assert (!*slot);
266 *slot = expr;
268 return next_expression_id - 1;
271 /* Return the expression id for tree EXPR. */
273 static inline unsigned int
274 get_expression_id (const pre_expr expr)
276 return expr->id;
279 static inline unsigned int
280 lookup_expression_id (const pre_expr expr)
282 struct pre_expr_d **slot;
284 if (expr->kind == NAME)
286 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
287 if (name_to_id.length () <= version)
288 return 0;
289 return name_to_id[version];
291 else
293 slot = expression_to_id.find_slot (expr, NO_INSERT);
294 if (!slot)
295 return 0;
296 return ((pre_expr)*slot)->id;
300 /* Return the existing expression id for EXPR, or create one if one
301 does not exist yet. */
303 static inline unsigned int
304 get_or_alloc_expression_id (pre_expr expr)
306 unsigned int id = lookup_expression_id (expr);
307 if (id == 0)
308 return alloc_expression_id (expr);
309 return expr->id = id;
312 /* Return the expression that has expression id ID */
314 static inline pre_expr
315 expression_for_id (unsigned int id)
317 return expressions[id];
320 /* Free the expression id field in all of our expressions,
321 and then destroy the expressions array. */
323 static void
324 clear_expression_ids (void)
326 expressions.release ();
329 static alloc_pool pre_expr_pool;
331 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
333 static pre_expr
334 get_or_alloc_expr_for_name (tree name)
336 struct pre_expr_d expr;
337 pre_expr result;
338 unsigned int result_id;
340 expr.kind = NAME;
341 expr.id = 0;
342 PRE_EXPR_NAME (&expr) = name;
343 result_id = lookup_expression_id (&expr);
344 if (result_id != 0)
345 return expression_for_id (result_id);
347 result = (pre_expr) pool_alloc (pre_expr_pool);
348 result->kind = NAME;
349 PRE_EXPR_NAME (result) = name;
350 alloc_expression_id (result);
351 return result;
354 /* An unordered bitmap set. One bitmap tracks values, the other,
355 expressions. */
356 typedef struct bitmap_set
358 bitmap_head expressions;
359 bitmap_head values;
360 } *bitmap_set_t;
362 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
363 EXECUTE_IF_SET_IN_BITMAP(&(set)->expressions, 0, (id), (bi))
365 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
366 EXECUTE_IF_SET_IN_BITMAP(&(set)->values, 0, (id), (bi))
368 /* Mapping from value id to expressions with that value_id. */
369 static vec<bitmap> value_expressions;
371 /* Sets that we need to keep track of. */
372 typedef struct bb_bitmap_sets
374 /* The EXP_GEN set, which represents expressions/values generated in
375 a basic block. */
376 bitmap_set_t exp_gen;
378 /* The PHI_GEN set, which represents PHI results generated in a
379 basic block. */
380 bitmap_set_t phi_gen;
382 /* The TMP_GEN set, which represents results/temporaries generated
383 in a basic block. IE the LHS of an expression. */
384 bitmap_set_t tmp_gen;
386 /* The AVAIL_OUT set, which represents which values are available in
387 a given basic block. */
388 bitmap_set_t avail_out;
390 /* The ANTIC_IN set, which represents which values are anticipatable
391 in a given basic block. */
392 bitmap_set_t antic_in;
394 /* The PA_IN set, which represents which values are
395 partially anticipatable in a given basic block. */
396 bitmap_set_t pa_in;
398 /* The NEW_SETS set, which is used during insertion to augment the
399 AVAIL_OUT set of blocks with the new insertions performed during
400 the current iteration. */
401 bitmap_set_t new_sets;
403 /* A cache for value_dies_in_block_x. */
404 bitmap expr_dies;
406 /* True if we have visited this block during ANTIC calculation. */
407 unsigned int visited : 1;
409 /* True we have deferred processing this block during ANTIC
410 calculation until its successor is processed. */
411 unsigned int deferred : 1;
413 /* True when the block contains a call that might not return. */
414 unsigned int contains_may_not_return_call : 1;
415 } *bb_value_sets_t;
417 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
418 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
419 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
420 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
421 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
422 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
423 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
424 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
425 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
426 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
427 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
430 /* Basic block list in postorder. */
431 static int *postorder;
432 static int postorder_num;
434 /* This structure is used to keep track of statistics on what
435 optimization PRE was able to perform. */
436 static struct
438 /* The number of RHS computations eliminated by PRE. */
439 int eliminations;
441 /* The number of new expressions/temporaries generated by PRE. */
442 int insertions;
444 /* The number of inserts found due to partial anticipation */
445 int pa_insert;
447 /* The number of new PHI nodes added by PRE. */
448 int phis;
449 } pre_stats;
451 static bool do_partial_partial;
452 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int);
453 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
454 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
455 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
456 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
457 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
458 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
459 unsigned int, bool);
460 static bitmap_set_t bitmap_set_new (void);
461 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
462 tree);
463 static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
464 static unsigned int get_expr_value_id (pre_expr);
466 /* We can add and remove elements and entries to and from sets
467 and hash tables, so we use alloc pools for them. */
469 static alloc_pool bitmap_set_pool;
470 static bitmap_obstack grand_bitmap_obstack;
472 /* Set of blocks with statements that have had their EH properties changed. */
473 static bitmap need_eh_cleanup;
475 /* Set of blocks with statements that have had their AB properties changed. */
476 static bitmap need_ab_cleanup;
478 /* A three tuple {e, pred, v} used to cache phi translations in the
479 phi_translate_table. */
481 typedef struct expr_pred_trans_d : typed_free_remove<expr_pred_trans_d>
483 /* The expression. */
484 pre_expr e;
486 /* The predecessor block along which we translated the expression. */
487 basic_block pred;
489 /* The value that resulted from the translation. */
490 pre_expr v;
492 /* The hashcode for the expression, pred pair. This is cached for
493 speed reasons. */
494 hashval_t hashcode;
496 /* hash_table support. */
497 typedef expr_pred_trans_d value_type;
498 typedef expr_pred_trans_d compare_type;
499 static inline hashval_t hash (const value_type *);
500 static inline int equal (const value_type *, const compare_type *);
501 } *expr_pred_trans_t;
502 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
504 inline hashval_t
505 expr_pred_trans_d::hash (const expr_pred_trans_d *e)
507 return e->hashcode;
510 inline int
511 expr_pred_trans_d::equal (const value_type *ve1,
512 const compare_type *ve2)
514 basic_block b1 = ve1->pred;
515 basic_block b2 = ve2->pred;
517 /* If they are not translations for the same basic block, they can't
518 be equal. */
519 if (b1 != b2)
520 return false;
521 return pre_expr_d::equal (ve1->e, ve2->e);
524 /* The phi_translate_table caches phi translations for a given
525 expression and predecessor. */
526 static hash_table <expr_pred_trans_d> phi_translate_table;
528 /* Search in the phi translation table for the translation of
529 expression E in basic block PRED.
530 Return the translated value, if found, NULL otherwise. */
532 static inline pre_expr
533 phi_trans_lookup (pre_expr e, basic_block pred)
535 expr_pred_trans_t *slot;
536 struct expr_pred_trans_d ept;
538 ept.e = e;
539 ept.pred = pred;
540 ept.hashcode = iterative_hash_hashval_t (pre_expr_d::hash (e), pred->index);
541 slot = phi_translate_table.find_slot_with_hash (&ept, ept.hashcode,
542 NO_INSERT);
543 if (!slot)
544 return NULL;
545 else
546 return (*slot)->v;
550 /* Add the tuple mapping from {expression E, basic block PRED} to
551 value V, to the phi translation table. */
553 static inline void
554 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
556 expr_pred_trans_t *slot;
557 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
558 new_pair->e = e;
559 new_pair->pred = pred;
560 new_pair->v = v;
561 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_d::hash (e),
562 pred->index);
564 slot = phi_translate_table.find_slot_with_hash (new_pair,
565 new_pair->hashcode, INSERT);
566 free (*slot);
567 *slot = new_pair;
571 /* Add expression E to the expression set of value id V. */
573 static void
574 add_to_value (unsigned int v, pre_expr e)
576 bitmap set;
578 gcc_checking_assert (get_expr_value_id (e) == v);
580 if (v >= value_expressions.length ())
582 value_expressions.safe_grow_cleared (v + 1);
585 set = value_expressions[v];
586 if (!set)
588 set = BITMAP_ALLOC (&grand_bitmap_obstack);
589 value_expressions[v] = set;
592 bitmap_set_bit (set, get_or_alloc_expression_id (e));
595 /* Create a new bitmap set and return it. */
597 static bitmap_set_t
598 bitmap_set_new (void)
600 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
601 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
602 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
603 return ret;
606 /* Return the value id for a PRE expression EXPR. */
608 static unsigned int
609 get_expr_value_id (pre_expr expr)
611 unsigned int id;
612 switch (expr->kind)
614 case CONSTANT:
615 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
616 break;
617 case NAME:
618 id = VN_INFO (PRE_EXPR_NAME (expr))->value_id;
619 break;
620 case NARY:
621 id = PRE_EXPR_NARY (expr)->value_id;
622 break;
623 case REFERENCE:
624 id = PRE_EXPR_REFERENCE (expr)->value_id;
625 break;
626 default:
627 gcc_unreachable ();
629 /* ??? We cannot assert that expr has a value-id (it can be 0), because
630 we assign value-ids only to expressions that have a result
631 in set_hashtable_value_ids. */
632 return id;
635 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
637 static tree
638 sccvn_valnum_from_value_id (unsigned int val)
640 bitmap_iterator bi;
641 unsigned int i;
642 bitmap exprset = value_expressions[val];
643 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
645 pre_expr vexpr = expression_for_id (i);
646 if (vexpr->kind == NAME)
647 return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
648 else if (vexpr->kind == CONSTANT)
649 return PRE_EXPR_CONSTANT (vexpr);
651 return NULL_TREE;
654 /* Remove an expression EXPR from a bitmapped set. */
656 static void
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));
667 static void
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. */
682 static void
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. */
690 static void
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. */
699 static void
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>
710 sorted_array_from_bitmap_set (bitmap_set_t set)
712 unsigned int i, j;
713 bitmap_iterator bi, bj;
714 vec<pre_expr> result;
716 /* Pre-allocate roughly enough space for the array. */
717 result.create (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
727 topological order.
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 exprset = value_expressions[i];
732 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj)
734 if (bitmap_bit_p (&set->expressions, j))
735 result.safe_push (expression_for_id (j));
739 return result;
742 /* Perform bitmapped set operation DEST &= ORIG. */
744 static void
745 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
747 bitmap_iterator bi;
748 unsigned int i;
750 if (dest != orig)
752 bitmap_head temp;
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. */
770 static bitmap_set_t
771 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
773 bitmap_set_t result = bitmap_set_new ();
774 bitmap_iterator bi;
775 unsigned int i;
777 bitmap_and_compl (&result->expressions, &dest->expressions,
778 &orig->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);
787 return result;
790 /* Subtract all the values in bitmap set B from bitmap set A. */
792 static void
793 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
795 unsigned int i;
796 bitmap_iterator bi;
797 bitmap_head temp;
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. */
814 static bool
815 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
817 if (value_id_constant_p (value_id))
818 return true;
820 if (!set || bitmap_empty_p (&set->expressions))
821 return false;
823 return bitmap_bit_p (&set->values, value_id);
826 static inline bool
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. */
834 static void
835 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
836 const pre_expr expr)
838 bitmap exprset;
839 unsigned int i;
840 bitmap_iterator bi;
842 if (value_id_constant_p (lookfor))
843 return;
845 if (!bitmap_set_contains_value (set, lookfor))
846 return;
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 = value_expressions[lookfor];
858 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
860 if (bitmap_clear_bit (&set->expressions, i))
862 bitmap_set_bit (&set->expressions, get_expression_id (expr));
863 return;
867 gcc_unreachable ();
870 /* Return true if two bitmap sets are equal. */
872 static bool
873 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
875 return bitmap_equal_p (&a->values, &b->values);
878 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
879 and add it otherwise. */
881 static void
882 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
884 unsigned int val = get_expr_value_id (expr);
886 if (bitmap_set_contains_value (set, val))
887 bitmap_set_replace_value (set, val, expr);
888 else
889 bitmap_insert_into_set (set, expr);
892 /* Insert EXPR into SET if EXPR's value is not already present in
893 SET. */
895 static void
896 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
898 unsigned int val = get_expr_value_id (expr);
900 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
902 /* Constant values are always considered to be part of the set. */
903 if (value_id_constant_p (val))
904 return;
906 /* If the value membership changed, add the expression. */
907 if (bitmap_set_bit (&set->values, val))
908 bitmap_set_bit (&set->expressions, expr->id);
911 /* Print out EXPR to outfile. */
913 static void
914 print_pre_expr (FILE *outfile, const pre_expr expr)
916 switch (expr->kind)
918 case CONSTANT:
919 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
920 break;
921 case NAME:
922 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
923 break;
924 case NARY:
926 unsigned int i;
927 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
928 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
929 for (i = 0; i < nary->length; i++)
931 print_generic_expr (outfile, nary->op[i], 0);
932 if (i != (unsigned) nary->length - 1)
933 fprintf (outfile, ",");
935 fprintf (outfile, "}");
937 break;
939 case REFERENCE:
941 vn_reference_op_t vro;
942 unsigned int i;
943 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
944 fprintf (outfile, "{");
945 for (i = 0;
946 ref->operands.iterate (i, &vro);
947 i++)
949 bool closebrace = false;
950 if (vro->opcode != SSA_NAME
951 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
953 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
954 if (vro->op0)
956 fprintf (outfile, "<");
957 closebrace = true;
960 if (vro->op0)
962 print_generic_expr (outfile, vro->op0, 0);
963 if (vro->op1)
965 fprintf (outfile, ",");
966 print_generic_expr (outfile, vro->op1, 0);
968 if (vro->op2)
970 fprintf (outfile, ",");
971 print_generic_expr (outfile, vro->op2, 0);
974 if (closebrace)
975 fprintf (outfile, ">");
976 if (i != ref->operands.length () - 1)
977 fprintf (outfile, ",");
979 fprintf (outfile, "}");
980 if (ref->vuse)
982 fprintf (outfile, "@");
983 print_generic_expr (outfile, ref->vuse, 0);
986 break;
989 void debug_pre_expr (pre_expr);
991 /* Like print_pre_expr but always prints to stderr. */
992 DEBUG_FUNCTION void
993 debug_pre_expr (pre_expr e)
995 print_pre_expr (stderr, e);
996 fprintf (stderr, "\n");
999 /* Print out SET to OUTFILE. */
1001 static void
1002 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1003 const char *setname, int blockindex)
1005 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1006 if (set)
1008 bool first = true;
1009 unsigned i;
1010 bitmap_iterator bi;
1012 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1014 const pre_expr expr = expression_for_id (i);
1016 if (!first)
1017 fprintf (outfile, ", ");
1018 first = false;
1019 print_pre_expr (outfile, expr);
1021 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1024 fprintf (outfile, " }\n");
1027 void debug_bitmap_set (bitmap_set_t);
1029 DEBUG_FUNCTION void
1030 debug_bitmap_set (bitmap_set_t set)
1032 print_bitmap_set (stderr, set, "debug", 0);
1035 void debug_bitmap_sets_for (basic_block);
1037 DEBUG_FUNCTION void
1038 debug_bitmap_sets_for (basic_block bb)
1040 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1041 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1042 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1043 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1044 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1045 if (do_partial_partial)
1046 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1047 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1050 /* Print out the expressions that have VAL to OUTFILE. */
1052 static void
1053 print_value_expressions (FILE *outfile, unsigned int val)
1055 bitmap set = value_expressions[val];
1056 if (set)
1058 bitmap_set x;
1059 char s[10];
1060 sprintf (s, "%04d", val);
1061 x.expressions = *set;
1062 print_bitmap_set (outfile, &x, s, 0);
1067 DEBUG_FUNCTION void
1068 debug_value_expressions (unsigned int val)
1070 print_value_expressions (stderr, val);
1073 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1074 represent it. */
1076 static pre_expr
1077 get_or_alloc_expr_for_constant (tree constant)
1079 unsigned int result_id;
1080 unsigned int value_id;
1081 struct pre_expr_d expr;
1082 pre_expr newexpr;
1084 expr.kind = CONSTANT;
1085 PRE_EXPR_CONSTANT (&expr) = constant;
1086 result_id = lookup_expression_id (&expr);
1087 if (result_id != 0)
1088 return expression_for_id (result_id);
1090 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1091 newexpr->kind = CONSTANT;
1092 PRE_EXPR_CONSTANT (newexpr) = constant;
1093 alloc_expression_id (newexpr);
1094 value_id = get_or_alloc_constant_value_id (constant);
1095 add_to_value (value_id, newexpr);
1096 return newexpr;
1099 /* Given a value id V, find the actual tree representing the constant
1100 value if there is one, and return it. Return NULL if we can't find
1101 a constant. */
1103 static tree
1104 get_constant_for_value_id (unsigned int v)
1106 if (value_id_constant_p (v))
1108 unsigned int i;
1109 bitmap_iterator bi;
1110 bitmap exprset = value_expressions[v];
1112 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1114 pre_expr expr = expression_for_id (i);
1115 if (expr->kind == CONSTANT)
1116 return PRE_EXPR_CONSTANT (expr);
1119 return NULL;
1122 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1123 Currently only supports constants and SSA_NAMES. */
1124 static pre_expr
1125 get_or_alloc_expr_for (tree t)
1127 if (TREE_CODE (t) == SSA_NAME)
1128 return get_or_alloc_expr_for_name (t);
1129 else if (is_gimple_min_invariant (t))
1130 return get_or_alloc_expr_for_constant (t);
1131 else
1133 /* More complex expressions can result from SCCVN expression
1134 simplification that inserts values for them. As they all
1135 do not have VOPs the get handled by the nary ops struct. */
1136 vn_nary_op_t result;
1137 unsigned int result_id;
1138 vn_nary_op_lookup (t, &result);
1139 if (result != NULL)
1141 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1142 e->kind = NARY;
1143 PRE_EXPR_NARY (e) = result;
1144 result_id = lookup_expression_id (e);
1145 if (result_id != 0)
1147 pool_free (pre_expr_pool, e);
1148 e = expression_for_id (result_id);
1149 return e;
1151 alloc_expression_id (e);
1152 return e;
1155 return NULL;
1158 /* Return the folded version of T if T, when folded, is a gimple
1159 min_invariant. Otherwise, return T. */
1161 static pre_expr
1162 fully_constant_expression (pre_expr e)
1164 switch (e->kind)
1166 case CONSTANT:
1167 return e;
1168 case NARY:
1170 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1171 switch (TREE_CODE_CLASS (nary->opcode))
1173 case tcc_binary:
1174 case tcc_comparison:
1176 /* We have to go from trees to pre exprs to value ids to
1177 constants. */
1178 tree naryop0 = nary->op[0];
1179 tree naryop1 = nary->op[1];
1180 tree result;
1181 if (!is_gimple_min_invariant (naryop0))
1183 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1184 unsigned int vrep0 = get_expr_value_id (rep0);
1185 tree const0 = get_constant_for_value_id (vrep0);
1186 if (const0)
1187 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1189 if (!is_gimple_min_invariant (naryop1))
1191 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1192 unsigned int vrep1 = get_expr_value_id (rep1);
1193 tree const1 = get_constant_for_value_id (vrep1);
1194 if (const1)
1195 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1197 result = fold_binary (nary->opcode, nary->type,
1198 naryop0, naryop1);
1199 if (result && is_gimple_min_invariant (result))
1200 return get_or_alloc_expr_for_constant (result);
1201 /* We might have simplified the expression to a
1202 SSA_NAME for example from x_1 * 1. But we cannot
1203 insert a PHI for x_1 unconditionally as x_1 might
1204 not be available readily. */
1205 return e;
1207 case tcc_reference:
1208 if (nary->opcode != REALPART_EXPR
1209 && nary->opcode != IMAGPART_EXPR
1210 && nary->opcode != VIEW_CONVERT_EXPR)
1211 return e;
1212 /* Fallthrough. */
1213 case tcc_unary:
1215 /* We have to go from trees to pre exprs to value ids to
1216 constants. */
1217 tree naryop0 = nary->op[0];
1218 tree const0, result;
1219 if (is_gimple_min_invariant (naryop0))
1220 const0 = naryop0;
1221 else
1223 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1224 unsigned int vrep0 = get_expr_value_id (rep0);
1225 const0 = get_constant_for_value_id (vrep0);
1227 result = NULL;
1228 if (const0)
1230 tree type1 = TREE_TYPE (nary->op[0]);
1231 const0 = fold_convert (type1, const0);
1232 result = fold_unary (nary->opcode, nary->type, const0);
1234 if (result && is_gimple_min_invariant (result))
1235 return get_or_alloc_expr_for_constant (result);
1236 return e;
1238 default:
1239 return e;
1242 case REFERENCE:
1244 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1245 tree folded;
1246 if ((folded = fully_constant_vn_reference_p (ref)))
1247 return get_or_alloc_expr_for_constant (folded);
1248 return e;
1250 default:
1251 return e;
1253 return e;
1256 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1257 it has the value it would have in BLOCK. Set *SAME_VALID to true
1258 in case the new vuse doesn't change the value id of the OPERANDS. */
1260 static tree
1261 translate_vuse_through_block (vec<vn_reference_op_s> operands,
1262 alias_set_type set, tree type, tree vuse,
1263 basic_block phiblock,
1264 basic_block block, bool *same_valid)
1266 gimple phi = SSA_NAME_DEF_STMT (vuse);
1267 ao_ref ref;
1268 edge e = NULL;
1269 bool use_oracle;
1271 *same_valid = true;
1273 if (gimple_bb (phi) != phiblock)
1274 return vuse;
1276 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1278 /* Use the alias-oracle to find either the PHI node in this block,
1279 the first VUSE used in this block that is equivalent to vuse or
1280 the first VUSE which definition in this block kills the value. */
1281 if (gimple_code (phi) == GIMPLE_PHI)
1282 e = find_edge (block, phiblock);
1283 else if (use_oracle)
1284 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1286 vuse = gimple_vuse (phi);
1287 phi = SSA_NAME_DEF_STMT (vuse);
1288 if (gimple_bb (phi) != phiblock)
1289 return vuse;
1290 if (gimple_code (phi) == GIMPLE_PHI)
1292 e = find_edge (block, phiblock);
1293 break;
1296 else
1297 return NULL_TREE;
1299 if (e)
1301 if (use_oracle)
1303 bitmap visited = NULL;
1304 unsigned int cnt;
1305 /* Try to find a vuse that dominates this phi node by skipping
1306 non-clobbering statements. */
1307 vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false);
1308 if (visited)
1309 BITMAP_FREE (visited);
1311 else
1312 vuse = NULL_TREE;
1313 if (!vuse)
1315 /* If we didn't find any, the value ID can't stay the same,
1316 but return the translated vuse. */
1317 *same_valid = false;
1318 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1320 /* ??? We would like to return vuse here as this is the canonical
1321 upmost vdef that this reference is associated with. But during
1322 insertion of the references into the hash tables we only ever
1323 directly insert with their direct gimple_vuse, hence returning
1324 something else would make us not find the other expression. */
1325 return PHI_ARG_DEF (phi, e->dest_idx);
1328 return NULL_TREE;
1331 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1332 SET2. This is used to avoid making a set consisting of the union
1333 of PA_IN and ANTIC_IN during insert. */
1335 static inline pre_expr
1336 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1338 pre_expr result;
1340 result = bitmap_find_leader (set1, val);
1341 if (!result && set2)
1342 result = bitmap_find_leader (set2, val);
1343 return result;
1346 /* Get the tree type for our PRE expression e. */
1348 static tree
1349 get_expr_type (const pre_expr e)
1351 switch (e->kind)
1353 case NAME:
1354 return TREE_TYPE (PRE_EXPR_NAME (e));
1355 case CONSTANT:
1356 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1357 case REFERENCE:
1358 return PRE_EXPR_REFERENCE (e)->type;
1359 case NARY:
1360 return PRE_EXPR_NARY (e)->type;
1362 gcc_unreachable();
1365 /* Get a representative SSA_NAME for a given expression.
1366 Since all of our sub-expressions are treated as values, we require
1367 them to be SSA_NAME's for simplicity.
1368 Prior versions of GVNPRE used to use "value handles" here, so that
1369 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1370 either case, the operands are really values (IE we do not expect
1371 them to be usable without finding leaders). */
1373 static tree
1374 get_representative_for (const pre_expr e)
1376 tree name;
1377 unsigned int value_id = get_expr_value_id (e);
1379 switch (e->kind)
1381 case NAME:
1382 return PRE_EXPR_NAME (e);
1383 case CONSTANT:
1384 return PRE_EXPR_CONSTANT (e);
1385 case NARY:
1386 case REFERENCE:
1388 /* Go through all of the expressions representing this value
1389 and pick out an SSA_NAME. */
1390 unsigned int i;
1391 bitmap_iterator bi;
1392 bitmap exprs = value_expressions[value_id];
1393 EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1395 pre_expr rep = expression_for_id (i);
1396 if (rep->kind == NAME)
1397 return PRE_EXPR_NAME (rep);
1398 else if (rep->kind == CONSTANT)
1399 return PRE_EXPR_CONSTANT (rep);
1402 break;
1405 /* If we reached here we couldn't find an SSA_NAME. This can
1406 happen when we've discovered a value that has never appeared in
1407 the program as set to an SSA_NAME, as the result of phi translation.
1408 Create one here.
1409 ??? We should be able to re-use this when we insert the statement
1410 to compute it. */
1411 name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1412 VN_INFO_GET (name)->value_id = value_id;
1413 VN_INFO (name)->valnum = name;
1414 /* ??? For now mark this SSA name for release by SCCVN. */
1415 VN_INFO (name)->needs_insertion = true;
1416 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1417 if (dump_file && (dump_flags & TDF_DETAILS))
1419 fprintf (dump_file, "Created SSA_NAME representative ");
1420 print_generic_expr (dump_file, name, 0);
1421 fprintf (dump_file, " for expression:");
1422 print_pre_expr (dump_file, e);
1423 fprintf (dump_file, "\n");
1426 return name;
1431 static pre_expr
1432 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1433 basic_block pred, basic_block phiblock);
1435 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1436 the phis in PRED. Return NULL if we can't find a leader for each part
1437 of the translated expression. */
1439 static pre_expr
1440 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1441 basic_block pred, basic_block phiblock)
1443 switch (expr->kind)
1445 case NARY:
1447 unsigned int i;
1448 bool changed = false;
1449 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1450 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1451 sizeof_vn_nary_op (nary->length));
1452 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1454 for (i = 0; i < newnary->length; i++)
1456 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1457 continue;
1458 else
1460 pre_expr leader, result;
1461 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1462 leader = find_leader_in_sets (op_val_id, set1, set2);
1463 result = phi_translate (leader, set1, set2, pred, phiblock);
1464 if (result && result != leader)
1466 tree name = get_representative_for (result);
1467 if (!name)
1468 return NULL;
1469 newnary->op[i] = name;
1471 else if (!result)
1472 return NULL;
1474 changed |= newnary->op[i] != nary->op[i];
1477 if (changed)
1479 pre_expr constant;
1480 unsigned int new_val_id;
1482 tree result = vn_nary_op_lookup_pieces (newnary->length,
1483 newnary->opcode,
1484 newnary->type,
1485 &newnary->op[0],
1486 &nary);
1487 if (result && is_gimple_min_invariant (result))
1488 return get_or_alloc_expr_for_constant (result);
1490 expr = (pre_expr) pool_alloc (pre_expr_pool);
1491 expr->kind = NARY;
1492 expr->id = 0;
1493 if (nary)
1495 PRE_EXPR_NARY (expr) = nary;
1496 constant = fully_constant_expression (expr);
1497 if (constant != expr)
1498 return constant;
1500 new_val_id = nary->value_id;
1501 get_or_alloc_expression_id (expr);
1503 else
1505 new_val_id = get_next_value_id ();
1506 value_expressions.safe_grow_cleared (get_max_value_id() + 1);
1507 nary = vn_nary_op_insert_pieces (newnary->length,
1508 newnary->opcode,
1509 newnary->type,
1510 &newnary->op[0],
1511 result, new_val_id);
1512 PRE_EXPR_NARY (expr) = nary;
1513 constant = fully_constant_expression (expr);
1514 if (constant != expr)
1515 return constant;
1516 get_or_alloc_expression_id (expr);
1518 add_to_value (new_val_id, expr);
1520 return expr;
1522 break;
1524 case REFERENCE:
1526 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1527 vec<vn_reference_op_s> operands = ref->operands;
1528 tree vuse = ref->vuse;
1529 tree newvuse = vuse;
1530 vec<vn_reference_op_s> newoperands = vNULL;
1531 bool changed = false, same_valid = true;
1532 unsigned int i, j, n;
1533 vn_reference_op_t operand;
1534 vn_reference_t newref;
1536 for (i = 0, j = 0;
1537 operands.iterate (i, &operand); i++, j++)
1539 pre_expr opresult;
1540 pre_expr leader;
1541 tree op[3];
1542 tree type = operand->type;
1543 vn_reference_op_s newop = *operand;
1544 op[0] = operand->op0;
1545 op[1] = operand->op1;
1546 op[2] = operand->op2;
1547 for (n = 0; n < 3; ++n)
1549 unsigned int op_val_id;
1550 if (!op[n])
1551 continue;
1552 if (TREE_CODE (op[n]) != SSA_NAME)
1554 /* We can't possibly insert these. */
1555 if (n != 0
1556 && !is_gimple_min_invariant (op[n]))
1557 break;
1558 continue;
1560 op_val_id = VN_INFO (op[n])->value_id;
1561 leader = find_leader_in_sets (op_val_id, set1, set2);
1562 if (!leader)
1563 break;
1564 /* Make sure we do not recursively translate ourselves
1565 like for translating a[n_1] with the leader for
1566 n_1 being a[n_1]. */
1567 if (get_expression_id (leader) != get_expression_id (expr))
1569 opresult = phi_translate (leader, set1, set2,
1570 pred, phiblock);
1571 if (!opresult)
1572 break;
1573 if (opresult != leader)
1575 tree name = get_representative_for (opresult);
1576 if (!name)
1577 break;
1578 changed |= name != op[n];
1579 op[n] = name;
1583 if (n != 3)
1585 newoperands.release ();
1586 return NULL;
1588 if (!newoperands.exists ())
1589 newoperands = operands.copy ();
1590 /* We may have changed from an SSA_NAME to a constant */
1591 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1592 newop.opcode = TREE_CODE (op[0]);
1593 newop.type = type;
1594 newop.op0 = op[0];
1595 newop.op1 = op[1];
1596 newop.op2 = op[2];
1597 /* If it transforms a non-constant ARRAY_REF into a constant
1598 one, adjust the constant offset. */
1599 if (newop.opcode == ARRAY_REF
1600 && newop.off == -1
1601 && TREE_CODE (op[0]) == INTEGER_CST
1602 && TREE_CODE (op[1]) == INTEGER_CST
1603 && TREE_CODE (op[2]) == INTEGER_CST)
1605 double_int off = tree_to_double_int (op[0]);
1606 off += -tree_to_double_int (op[1]);
1607 off *= tree_to_double_int (op[2]);
1608 if (off.fits_shwi ())
1609 newop.off = off.low;
1611 newoperands[j] = newop;
1612 /* If it transforms from an SSA_NAME to an address, fold with
1613 a preceding indirect reference. */
1614 if (j > 0 && op[0] && TREE_CODE (op[0]) == ADDR_EXPR
1615 && newoperands[j - 1].opcode == MEM_REF)
1616 vn_reference_fold_indirect (&newoperands, &j);
1618 if (i != operands.length ())
1620 newoperands.release ();
1621 return NULL;
1624 if (vuse)
1626 newvuse = translate_vuse_through_block (newoperands,
1627 ref->set, ref->type,
1628 vuse, phiblock, pred,
1629 &same_valid);
1630 if (newvuse == NULL_TREE)
1632 newoperands.release ();
1633 return NULL;
1637 if (changed || newvuse != vuse)
1639 unsigned int new_val_id;
1640 pre_expr constant;
1642 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1643 ref->type,
1644 newoperands,
1645 &newref, VN_WALK);
1646 if (result)
1647 newoperands.release ();
1649 /* We can always insert constants, so if we have a partial
1650 redundant constant load of another type try to translate it
1651 to a constant of appropriate type. */
1652 if (result && is_gimple_min_invariant (result))
1654 tree tem = result;
1655 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1657 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1658 if (tem && !is_gimple_min_invariant (tem))
1659 tem = NULL_TREE;
1661 if (tem)
1662 return get_or_alloc_expr_for_constant (tem);
1665 /* If we'd have to convert things we would need to validate
1666 if we can insert the translated expression. So fail
1667 here for now - we cannot insert an alias with a different
1668 type in the VN tables either, as that would assert. */
1669 if (result
1670 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1671 return NULL;
1672 else if (!result && newref
1673 && !useless_type_conversion_p (ref->type, newref->type))
1675 newoperands.release ();
1676 return NULL;
1679 expr = (pre_expr) pool_alloc (pre_expr_pool);
1680 expr->kind = REFERENCE;
1681 expr->id = 0;
1683 if (newref)
1685 PRE_EXPR_REFERENCE (expr) = newref;
1686 constant = fully_constant_expression (expr);
1687 if (constant != expr)
1688 return constant;
1690 new_val_id = newref->value_id;
1691 get_or_alloc_expression_id (expr);
1693 else
1695 if (changed || !same_valid)
1697 new_val_id = get_next_value_id ();
1698 value_expressions.safe_grow_cleared(get_max_value_id() + 1);
1700 else
1701 new_val_id = ref->value_id;
1702 newref = vn_reference_insert_pieces (newvuse, ref->set,
1703 ref->type,
1704 newoperands,
1705 result, new_val_id);
1706 newoperands.create (0);
1707 PRE_EXPR_REFERENCE (expr) = newref;
1708 constant = fully_constant_expression (expr);
1709 if (constant != expr)
1710 return constant;
1711 get_or_alloc_expression_id (expr);
1713 add_to_value (new_val_id, expr);
1715 newoperands.release ();
1716 return expr;
1718 break;
1720 case NAME:
1722 tree name = PRE_EXPR_NAME (expr);
1723 gimple def_stmt = SSA_NAME_DEF_STMT (name);
1724 /* If the SSA name is defined by a PHI node in this block,
1725 translate it. */
1726 if (gimple_code (def_stmt) == GIMPLE_PHI
1727 && gimple_bb (def_stmt) == phiblock)
1729 edge e = find_edge (pred, gimple_bb (def_stmt));
1730 tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1732 /* Handle constant. */
1733 if (is_gimple_min_invariant (def))
1734 return get_or_alloc_expr_for_constant (def);
1736 return get_or_alloc_expr_for_name (def);
1738 /* Otherwise return it unchanged - it will get cleaned if its
1739 value is not available in PREDs AVAIL_OUT set of expressions. */
1740 return expr;
1743 default:
1744 gcc_unreachable ();
1748 /* Wrapper around phi_translate_1 providing caching functionality. */
1750 static pre_expr
1751 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1752 basic_block pred, basic_block phiblock)
1754 pre_expr phitrans;
1756 if (!expr)
1757 return NULL;
1759 /* Constants contain no values that need translation. */
1760 if (expr->kind == CONSTANT)
1761 return expr;
1763 if (value_id_constant_p (get_expr_value_id (expr)))
1764 return expr;
1766 if (expr->kind != NAME)
1768 phitrans = phi_trans_lookup (expr, pred);
1769 if (phitrans)
1770 return phitrans;
1773 /* Translate. */
1774 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1776 /* Don't add empty translations to the cache. Neither add
1777 translations of NAMEs as those are cheap to translate. */
1778 if (phitrans
1779 && expr->kind != NAME)
1780 phi_trans_add (expr, phitrans, pred);
1782 return phitrans;
1786 /* For each expression in SET, translate the values through phi nodes
1787 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1788 expressions in DEST. */
1790 static void
1791 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1792 basic_block phiblock)
1794 vec<pre_expr> exprs;
1795 pre_expr expr;
1796 int i;
1798 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1800 bitmap_set_copy (dest, set);
1801 return;
1804 exprs = sorted_array_from_bitmap_set (set);
1805 FOR_EACH_VEC_ELT (exprs, i, expr)
1807 pre_expr translated;
1808 translated = phi_translate (expr, set, NULL, pred, phiblock);
1809 if (!translated)
1810 continue;
1812 /* We might end up with multiple expressions from SET being
1813 translated to the same value. In this case we do not want
1814 to retain the NARY or REFERENCE expression but prefer a NAME
1815 which would be the leader. */
1816 if (translated->kind == NAME)
1817 bitmap_value_replace_in_set (dest, translated);
1818 else
1819 bitmap_value_insert_into_set (dest, translated);
1821 exprs.release ();
1824 /* Find the leader for a value (i.e., the name representing that
1825 value) in a given set, and return it. If STMT is non-NULL it
1826 makes sure the defining statement for the leader dominates it.
1827 Return NULL if no leader is found. */
1829 static pre_expr
1830 bitmap_find_leader (bitmap_set_t set, unsigned int val)
1832 if (value_id_constant_p (val))
1834 unsigned int i;
1835 bitmap_iterator bi;
1836 bitmap exprset = value_expressions[val];
1838 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1840 pre_expr expr = expression_for_id (i);
1841 if (expr->kind == CONSTANT)
1842 return expr;
1845 if (bitmap_set_contains_value (set, val))
1847 /* Rather than walk the entire bitmap of expressions, and see
1848 whether any of them has the value we are looking for, we look
1849 at the reverse mapping, which tells us the set of expressions
1850 that have a given value (IE value->expressions with that
1851 value) and see if any of those expressions are in our set.
1852 The number of expressions per value is usually significantly
1853 less than the number of expressions in the set. In fact, for
1854 large testcases, doing it this way is roughly 5-10x faster
1855 than walking the bitmap.
1856 If this is somehow a significant lose for some cases, we can
1857 choose which set to walk based on which set is smaller. */
1858 unsigned int i;
1859 bitmap_iterator bi;
1860 bitmap exprset = value_expressions[val];
1862 EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1863 return expression_for_id (i);
1865 return NULL;
1868 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1869 BLOCK by seeing if it is not killed in the block. Note that we are
1870 only determining whether there is a store that kills it. Because
1871 of the order in which clean iterates over values, we are guaranteed
1872 that altered operands will have caused us to be eliminated from the
1873 ANTIC_IN set already. */
1875 static bool
1876 value_dies_in_block_x (pre_expr expr, basic_block block)
1878 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1879 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1880 gimple def;
1881 gimple_stmt_iterator gsi;
1882 unsigned id = get_expression_id (expr);
1883 bool res = false;
1884 ao_ref ref;
1886 if (!vuse)
1887 return false;
1889 /* Lookup a previously calculated result. */
1890 if (EXPR_DIES (block)
1891 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1892 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1894 /* A memory expression {e, VUSE} dies in the block if there is a
1895 statement that may clobber e. If, starting statement walk from the
1896 top of the basic block, a statement uses VUSE there can be no kill
1897 inbetween that use and the original statement that loaded {e, VUSE},
1898 so we can stop walking. */
1899 ref.base = NULL_TREE;
1900 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1902 tree def_vuse, def_vdef;
1903 def = gsi_stmt (gsi);
1904 def_vuse = gimple_vuse (def);
1905 def_vdef = gimple_vdef (def);
1907 /* Not a memory statement. */
1908 if (!def_vuse)
1909 continue;
1911 /* Not a may-def. */
1912 if (!def_vdef)
1914 /* A load with the same VUSE, we're done. */
1915 if (def_vuse == vuse)
1916 break;
1918 continue;
1921 /* Init ref only if we really need it. */
1922 if (ref.base == NULL_TREE
1923 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1924 refx->operands))
1926 res = true;
1927 break;
1929 /* If the statement may clobber expr, it dies. */
1930 if (stmt_may_clobber_ref_p_1 (def, &ref))
1932 res = true;
1933 break;
1937 /* Remember the result. */
1938 if (!EXPR_DIES (block))
1939 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1940 bitmap_set_bit (EXPR_DIES (block), id * 2);
1941 if (res)
1942 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1944 return res;
1948 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1949 contains its value-id. */
1951 static bool
1952 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1954 if (op && TREE_CODE (op) == SSA_NAME)
1956 unsigned int value_id = VN_INFO (op)->value_id;
1957 if (!(bitmap_set_contains_value (set1, value_id)
1958 || (set2 && bitmap_set_contains_value (set2, value_id))))
1959 return false;
1961 return true;
1964 /* Determine if the expression EXPR is valid in SET1 U SET2.
1965 ONLY SET2 CAN BE NULL.
1966 This means that we have a leader for each part of the expression
1967 (if it consists of values), or the expression is an SSA_NAME.
1968 For loads/calls, we also see if the vuse is killed in this block. */
1970 static bool
1971 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
1972 basic_block block)
1974 switch (expr->kind)
1976 case NAME:
1977 return bitmap_find_leader (AVAIL_OUT (block),
1978 get_expr_value_id (expr)) != NULL;
1979 case NARY:
1981 unsigned int i;
1982 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1983 for (i = 0; i < nary->length; i++)
1984 if (!op_valid_in_sets (set1, set2, nary->op[i]))
1985 return false;
1986 return true;
1988 break;
1989 case REFERENCE:
1991 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1992 vn_reference_op_t vro;
1993 unsigned int i;
1995 FOR_EACH_VEC_ELT (ref->operands, i, vro)
1997 if (!op_valid_in_sets (set1, set2, vro->op0)
1998 || !op_valid_in_sets (set1, set2, vro->op1)
1999 || !op_valid_in_sets (set1, set2, vro->op2))
2000 return false;
2002 return true;
2004 default:
2005 gcc_unreachable ();
2009 /* Clean the set of expressions that are no longer valid in SET1 or
2010 SET2. This means expressions that are made up of values we have no
2011 leaders for in SET1 or SET2. This version is used for partial
2012 anticipation, which means it is not valid in either ANTIC_IN or
2013 PA_IN. */
2015 static void
2016 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2018 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
2019 pre_expr expr;
2020 int i;
2022 FOR_EACH_VEC_ELT (exprs, i, expr)
2024 if (!valid_in_sets (set1, set2, expr, block))
2025 bitmap_remove_from_set (set1, expr);
2027 exprs.release ();
2030 /* Clean the set of expressions that are no longer valid in SET. This
2031 means expressions that are made up of values we have no leaders for
2032 in SET. */
2034 static void
2035 clean (bitmap_set_t set, basic_block block)
2037 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set);
2038 pre_expr expr;
2039 int i;
2041 FOR_EACH_VEC_ELT (exprs, i, expr)
2043 if (!valid_in_sets (set, NULL, expr, block))
2044 bitmap_remove_from_set (set, expr);
2046 exprs.release ();
2049 /* Clean the set of expressions that are no longer valid in SET because
2050 they are clobbered in BLOCK or because they trap and may not be executed. */
2052 static void
2053 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2055 bitmap_iterator bi;
2056 unsigned i;
2058 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2060 pre_expr expr = expression_for_id (i);
2061 if (expr->kind == REFERENCE)
2063 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2064 if (ref->vuse)
2066 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2067 if (!gimple_nop_p (def_stmt)
2068 && ((gimple_bb (def_stmt) != block
2069 && !dominated_by_p (CDI_DOMINATORS,
2070 block, gimple_bb (def_stmt)))
2071 || (gimple_bb (def_stmt) == block
2072 && value_dies_in_block_x (expr, block))))
2073 bitmap_remove_from_set (set, expr);
2076 else if (expr->kind == NARY)
2078 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2079 /* If the NARY may trap make sure the block does not contain
2080 a possible exit point.
2081 ??? This is overly conservative if we translate AVAIL_OUT
2082 as the available expression might be after the exit point. */
2083 if (BB_MAY_NOTRETURN (block)
2084 && vn_nary_may_trap (nary))
2085 bitmap_remove_from_set (set, expr);
2090 static sbitmap has_abnormal_preds;
2092 /* List of blocks that may have changed during ANTIC computation and
2093 thus need to be iterated over. */
2095 static sbitmap changed_blocks;
2097 /* Decide whether to defer a block for a later iteration, or PHI
2098 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2099 should defer the block, and true if we processed it. */
2101 static bool
2102 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2103 basic_block block, basic_block phiblock)
2105 if (!BB_VISITED (phiblock))
2107 bitmap_set_bit (changed_blocks, block->index);
2108 BB_VISITED (block) = 0;
2109 BB_DEFERRED (block) = 1;
2110 return false;
2112 else
2113 phi_translate_set (dest, source, block, phiblock);
2114 return true;
2117 /* Compute the ANTIC set for BLOCK.
2119 If succs(BLOCK) > 1 then
2120 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2121 else if succs(BLOCK) == 1 then
2122 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2124 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2127 static bool
2128 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2130 bool changed = false;
2131 bitmap_set_t S, old, ANTIC_OUT;
2132 bitmap_iterator bi;
2133 unsigned int bii;
2134 edge e;
2135 edge_iterator ei;
2137 old = ANTIC_OUT = S = NULL;
2138 BB_VISITED (block) = 1;
2140 /* If any edges from predecessors are abnormal, antic_in is empty,
2141 so do nothing. */
2142 if (block_has_abnormal_pred_edge)
2143 goto maybe_dump_sets;
2145 old = ANTIC_IN (block);
2146 ANTIC_OUT = bitmap_set_new ();
2148 /* If the block has no successors, ANTIC_OUT is empty. */
2149 if (EDGE_COUNT (block->succs) == 0)
2151 /* If we have one successor, we could have some phi nodes to
2152 translate through. */
2153 else if (single_succ_p (block))
2155 basic_block succ_bb = single_succ (block);
2157 /* We trade iterations of the dataflow equations for having to
2158 phi translate the maximal set, which is incredibly slow
2159 (since the maximal set often has 300+ members, even when you
2160 have a small number of blocks).
2161 Basically, we defer the computation of ANTIC for this block
2162 until we have processed it's successor, which will inevitably
2163 have a *much* smaller set of values to phi translate once
2164 clean has been run on it.
2165 The cost of doing this is that we technically perform more
2166 iterations, however, they are lower cost iterations.
2168 Timings for PRE on tramp3d-v4:
2169 without maximal set fix: 11 seconds
2170 with maximal set fix/without deferring: 26 seconds
2171 with maximal set fix/with deferring: 11 seconds
2174 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2175 block, succ_bb))
2177 changed = true;
2178 goto maybe_dump_sets;
2181 /* If we have multiple successors, we take the intersection of all of
2182 them. Note that in the case of loop exit phi nodes, we may have
2183 phis to translate through. */
2184 else
2186 vec<basic_block> worklist;
2187 size_t i;
2188 basic_block bprime, first = NULL;
2190 worklist.create (EDGE_COUNT (block->succs));
2191 FOR_EACH_EDGE (e, ei, block->succs)
2193 if (!first
2194 && BB_VISITED (e->dest))
2195 first = e->dest;
2196 else if (BB_VISITED (e->dest))
2197 worklist.quick_push (e->dest);
2200 /* Of multiple successors we have to have visited one already. */
2201 if (!first)
2203 bitmap_set_bit (changed_blocks, block->index);
2204 BB_VISITED (block) = 0;
2205 BB_DEFERRED (block) = 1;
2206 changed = true;
2207 worklist.release ();
2208 goto maybe_dump_sets;
2211 if (!gimple_seq_empty_p (phi_nodes (first)))
2212 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2213 else
2214 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2216 FOR_EACH_VEC_ELT (worklist, i, bprime)
2218 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2220 bitmap_set_t tmp = bitmap_set_new ();
2221 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2222 bitmap_set_and (ANTIC_OUT, tmp);
2223 bitmap_set_free (tmp);
2225 else
2226 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2228 worklist.release ();
2231 /* Prune expressions that are clobbered in block and thus become
2232 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2233 prune_clobbered_mems (ANTIC_OUT, block);
2235 /* Generate ANTIC_OUT - TMP_GEN. */
2236 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2238 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2239 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2240 TMP_GEN (block));
2242 /* Then union in the ANTIC_OUT - TMP_GEN values,
2243 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2244 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2245 bitmap_value_insert_into_set (ANTIC_IN (block),
2246 expression_for_id (bii));
2248 clean (ANTIC_IN (block), block);
2250 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2252 changed = true;
2253 bitmap_set_bit (changed_blocks, block->index);
2254 FOR_EACH_EDGE (e, ei, block->preds)
2255 bitmap_set_bit (changed_blocks, e->src->index);
2257 else
2258 bitmap_clear_bit (changed_blocks, block->index);
2260 maybe_dump_sets:
2261 if (dump_file && (dump_flags & TDF_DETAILS))
2263 if (!BB_DEFERRED (block) || BB_VISITED (block))
2265 if (ANTIC_OUT)
2266 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2268 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2269 block->index);
2271 if (S)
2272 print_bitmap_set (dump_file, S, "S", block->index);
2274 else
2276 fprintf (dump_file,
2277 "Block %d was deferred for a future iteration.\n",
2278 block->index);
2281 if (old)
2282 bitmap_set_free (old);
2283 if (S)
2284 bitmap_set_free (S);
2285 if (ANTIC_OUT)
2286 bitmap_set_free (ANTIC_OUT);
2287 return changed;
2290 /* Compute PARTIAL_ANTIC for BLOCK.
2292 If succs(BLOCK) > 1 then
2293 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2294 in ANTIC_OUT for all succ(BLOCK)
2295 else if succs(BLOCK) == 1 then
2296 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2298 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2299 - ANTIC_IN[BLOCK])
2302 static bool
2303 compute_partial_antic_aux (basic_block block,
2304 bool block_has_abnormal_pred_edge)
2306 bool changed = false;
2307 bitmap_set_t old_PA_IN;
2308 bitmap_set_t PA_OUT;
2309 edge e;
2310 edge_iterator ei;
2311 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2313 old_PA_IN = PA_OUT = NULL;
2315 /* If any edges from predecessors are abnormal, antic_in is empty,
2316 so do nothing. */
2317 if (block_has_abnormal_pred_edge)
2318 goto maybe_dump_sets;
2320 /* If there are too many partially anticipatable values in the
2321 block, phi_translate_set can take an exponential time: stop
2322 before the translation starts. */
2323 if (max_pa
2324 && single_succ_p (block)
2325 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2326 goto maybe_dump_sets;
2328 old_PA_IN = PA_IN (block);
2329 PA_OUT = bitmap_set_new ();
2331 /* If the block has no successors, ANTIC_OUT is empty. */
2332 if (EDGE_COUNT (block->succs) == 0)
2334 /* If we have one successor, we could have some phi nodes to
2335 translate through. Note that we can't phi translate across DFS
2336 back edges in partial antic, because it uses a union operation on
2337 the successors. For recurrences like IV's, we will end up
2338 generating a new value in the set on each go around (i + 3 (VH.1)
2339 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2340 else if (single_succ_p (block))
2342 basic_block succ = single_succ (block);
2343 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2344 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2346 /* If we have multiple successors, we take the union of all of
2347 them. */
2348 else
2350 vec<basic_block> worklist;
2351 size_t i;
2352 basic_block bprime;
2354 worklist.create (EDGE_COUNT (block->succs));
2355 FOR_EACH_EDGE (e, ei, block->succs)
2357 if (e->flags & EDGE_DFS_BACK)
2358 continue;
2359 worklist.quick_push (e->dest);
2361 if (worklist.length () > 0)
2363 FOR_EACH_VEC_ELT (worklist, i, bprime)
2365 unsigned int i;
2366 bitmap_iterator bi;
2368 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2369 bitmap_value_insert_into_set (PA_OUT,
2370 expression_for_id (i));
2371 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2373 bitmap_set_t pa_in = bitmap_set_new ();
2374 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2375 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2376 bitmap_value_insert_into_set (PA_OUT,
2377 expression_for_id (i));
2378 bitmap_set_free (pa_in);
2380 else
2381 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2382 bitmap_value_insert_into_set (PA_OUT,
2383 expression_for_id (i));
2386 worklist.release ();
2389 /* Prune expressions that are clobbered in block and thus become
2390 invalid if translated from PA_OUT to PA_IN. */
2391 prune_clobbered_mems (PA_OUT, block);
2393 /* PA_IN starts with PA_OUT - TMP_GEN.
2394 Then we subtract things from ANTIC_IN. */
2395 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2397 /* For partial antic, we want to put back in the phi results, since
2398 we will properly avoid making them partially antic over backedges. */
2399 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2400 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2402 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2403 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2405 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2407 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2409 changed = true;
2410 bitmap_set_bit (changed_blocks, block->index);
2411 FOR_EACH_EDGE (e, ei, block->preds)
2412 bitmap_set_bit (changed_blocks, e->src->index);
2414 else
2415 bitmap_clear_bit (changed_blocks, block->index);
2417 maybe_dump_sets:
2418 if (dump_file && (dump_flags & TDF_DETAILS))
2420 if (PA_OUT)
2421 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2423 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2425 if (old_PA_IN)
2426 bitmap_set_free (old_PA_IN);
2427 if (PA_OUT)
2428 bitmap_set_free (PA_OUT);
2429 return changed;
2432 /* Compute ANTIC and partial ANTIC sets. */
2434 static void
2435 compute_antic (void)
2437 bool changed = true;
2438 int num_iterations = 0;
2439 basic_block block;
2440 int i;
2442 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2443 We pre-build the map of blocks with incoming abnormal edges here. */
2444 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2445 bitmap_clear (has_abnormal_preds);
2447 FOR_ALL_BB (block)
2449 edge_iterator ei;
2450 edge e;
2452 FOR_EACH_EDGE (e, ei, block->preds)
2454 e->flags &= ~EDGE_DFS_BACK;
2455 if (e->flags & EDGE_ABNORMAL)
2457 bitmap_set_bit (has_abnormal_preds, block->index);
2458 break;
2462 BB_VISITED (block) = 0;
2463 BB_DEFERRED (block) = 0;
2465 /* While we are here, give empty ANTIC_IN sets to each block. */
2466 ANTIC_IN (block) = bitmap_set_new ();
2467 PA_IN (block) = bitmap_set_new ();
2470 /* At the exit block we anticipate nothing. */
2471 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2473 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2474 bitmap_ones (changed_blocks);
2475 while (changed)
2477 if (dump_file && (dump_flags & TDF_DETAILS))
2478 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2479 /* ??? We need to clear our PHI translation cache here as the
2480 ANTIC sets shrink and we restrict valid translations to
2481 those having operands with leaders in ANTIC. Same below
2482 for PA ANTIC computation. */
2483 num_iterations++;
2484 changed = false;
2485 for (i = postorder_num - 1; i >= 0; i--)
2487 if (bitmap_bit_p (changed_blocks, postorder[i]))
2489 basic_block block = BASIC_BLOCK (postorder[i]);
2490 changed |= compute_antic_aux (block,
2491 bitmap_bit_p (has_abnormal_preds,
2492 block->index));
2495 /* Theoretically possible, but *highly* unlikely. */
2496 gcc_checking_assert (num_iterations < 500);
2499 statistics_histogram_event (cfun, "compute_antic iterations",
2500 num_iterations);
2502 if (do_partial_partial)
2504 bitmap_ones (changed_blocks);
2505 mark_dfs_back_edges ();
2506 num_iterations = 0;
2507 changed = true;
2508 while (changed)
2510 if (dump_file && (dump_flags & TDF_DETAILS))
2511 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2512 num_iterations++;
2513 changed = false;
2514 for (i = postorder_num - 1 ; i >= 0; i--)
2516 if (bitmap_bit_p (changed_blocks, postorder[i]))
2518 basic_block block = BASIC_BLOCK (postorder[i]);
2519 changed
2520 |= compute_partial_antic_aux (block,
2521 bitmap_bit_p (has_abnormal_preds,
2522 block->index));
2525 /* Theoretically possible, but *highly* unlikely. */
2526 gcc_checking_assert (num_iterations < 500);
2528 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2529 num_iterations);
2531 sbitmap_free (has_abnormal_preds);
2532 sbitmap_free (changed_blocks);
2536 /* Inserted expressions are placed onto this worklist, which is used
2537 for performing quick dead code elimination of insertions we made
2538 that didn't turn out to be necessary. */
2539 static bitmap inserted_exprs;
2541 /* The actual worker for create_component_ref_by_pieces. */
2543 static tree
2544 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2545 unsigned int *operand, gimple_seq *stmts)
2547 vn_reference_op_t currop = &ref->operands[*operand];
2548 tree genop;
2549 ++*operand;
2550 switch (currop->opcode)
2552 case CALL_EXPR:
2554 tree folded, sc = NULL_TREE;
2555 unsigned int nargs = 0;
2556 tree fn, *args;
2557 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2558 fn = currop->op0;
2559 else
2560 fn = find_or_generate_expression (block, currop->op0, stmts);
2561 if (!fn)
2562 return NULL_TREE;
2563 if (currop->op1)
2565 sc = find_or_generate_expression (block, currop->op1, stmts);
2566 if (!sc)
2567 return NULL_TREE;
2569 args = XNEWVEC (tree, ref->operands.length () - 1);
2570 while (*operand < ref->operands.length ())
2572 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2573 operand, stmts);
2574 if (!args[nargs])
2575 return NULL_TREE;
2576 nargs++;
2578 folded = build_call_array (currop->type,
2579 (TREE_CODE (fn) == FUNCTION_DECL
2580 ? build_fold_addr_expr (fn) : fn),
2581 nargs, args);
2582 free (args);
2583 if (sc)
2584 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2585 return folded;
2588 case MEM_REF:
2590 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2591 stmts);
2592 if (!baseop)
2593 return NULL_TREE;
2594 tree offset = currop->op0;
2595 if (TREE_CODE (baseop) == ADDR_EXPR
2596 && handled_component_p (TREE_OPERAND (baseop, 0)))
2598 HOST_WIDE_INT off;
2599 tree base;
2600 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2601 &off);
2602 gcc_assert (base);
2603 offset = int_const_binop (PLUS_EXPR, offset,
2604 build_int_cst (TREE_TYPE (offset),
2605 off));
2606 baseop = build_fold_addr_expr (base);
2608 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2611 case TARGET_MEM_REF:
2613 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2614 vn_reference_op_t nextop = &ref->operands[++*operand];
2615 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2616 stmts);
2617 if (!baseop)
2618 return NULL_TREE;
2619 if (currop->op0)
2621 genop0 = find_or_generate_expression (block, currop->op0, stmts);
2622 if (!genop0)
2623 return NULL_TREE;
2625 if (nextop->op0)
2627 genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2628 if (!genop1)
2629 return NULL_TREE;
2631 return build5 (TARGET_MEM_REF, currop->type,
2632 baseop, currop->op2, genop0, currop->op1, genop1);
2635 case ADDR_EXPR:
2636 if (currop->op0)
2638 gcc_assert (is_gimple_min_invariant (currop->op0));
2639 return currop->op0;
2641 /* Fallthrough. */
2642 case REALPART_EXPR:
2643 case IMAGPART_EXPR:
2644 case VIEW_CONVERT_EXPR:
2646 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2647 stmts);
2648 if (!genop0)
2649 return NULL_TREE;
2650 return fold_build1 (currop->opcode, currop->type, genop0);
2653 case WITH_SIZE_EXPR:
2655 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2656 stmts);
2657 if (!genop0)
2658 return NULL_TREE;
2659 tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2660 if (!genop1)
2661 return NULL_TREE;
2662 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2665 case BIT_FIELD_REF:
2667 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2668 stmts);
2669 if (!genop0)
2670 return NULL_TREE;
2671 tree op1 = currop->op0;
2672 tree op2 = currop->op1;
2673 return fold_build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2676 /* For array ref vn_reference_op's, operand 1 of the array ref
2677 is op0 of the reference op and operand 3 of the array ref is
2678 op1. */
2679 case ARRAY_RANGE_REF:
2680 case ARRAY_REF:
2682 tree genop0;
2683 tree genop1 = currop->op0;
2684 tree genop2 = currop->op1;
2685 tree genop3 = currop->op2;
2686 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2687 stmts);
2688 if (!genop0)
2689 return NULL_TREE;
2690 genop1 = find_or_generate_expression (block, genop1, stmts);
2691 if (!genop1)
2692 return NULL_TREE;
2693 if (genop2)
2695 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2696 /* Drop zero minimum index if redundant. */
2697 if (integer_zerop (genop2)
2698 && (!domain_type
2699 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2700 genop2 = NULL_TREE;
2701 else
2703 genop2 = find_or_generate_expression (block, genop2, stmts);
2704 if (!genop2)
2705 return NULL_TREE;
2708 if (genop3)
2710 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2711 /* We can't always put a size in units of the element alignment
2712 here as the element alignment may be not visible. See
2713 PR43783. Simply drop the element size for constant
2714 sizes. */
2715 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2716 genop3 = NULL_TREE;
2717 else
2719 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2720 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2721 genop3 = find_or_generate_expression (block, genop3, stmts);
2722 if (!genop3)
2723 return NULL_TREE;
2726 return build4 (currop->opcode, currop->type, genop0, genop1,
2727 genop2, genop3);
2729 case COMPONENT_REF:
2731 tree op0;
2732 tree op1;
2733 tree genop2 = currop->op1;
2734 op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2735 if (!op0)
2736 return NULL_TREE;
2737 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2738 op1 = currop->op0;
2739 if (genop2)
2741 genop2 = find_or_generate_expression (block, genop2, stmts);
2742 if (!genop2)
2743 return NULL_TREE;
2745 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2748 case SSA_NAME:
2750 genop = find_or_generate_expression (block, currop->op0, stmts);
2751 return genop;
2753 case STRING_CST:
2754 case INTEGER_CST:
2755 case COMPLEX_CST:
2756 case VECTOR_CST:
2757 case REAL_CST:
2758 case CONSTRUCTOR:
2759 case VAR_DECL:
2760 case PARM_DECL:
2761 case CONST_DECL:
2762 case RESULT_DECL:
2763 case FUNCTION_DECL:
2764 return currop->op0;
2766 default:
2767 gcc_unreachable ();
2771 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2772 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2773 trying to rename aggregates into ssa form directly, which is a no no.
2775 Thus, this routine doesn't create temporaries, it just builds a
2776 single access expression for the array, calling
2777 find_or_generate_expression to build the innermost pieces.
2779 This function is a subroutine of create_expression_by_pieces, and
2780 should not be called on it's own unless you really know what you
2781 are doing. */
2783 static tree
2784 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2785 gimple_seq *stmts)
2787 unsigned int op = 0;
2788 return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2791 /* Find a simple leader for an expression, or generate one using
2792 create_expression_by_pieces from a NARY expression for the value.
2793 BLOCK is the basic_block we are looking for leaders in.
2794 OP is the tree expression to find a leader for or generate.
2795 Returns the leader or NULL_TREE on failure. */
2797 static tree
2798 find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2800 pre_expr expr = get_or_alloc_expr_for (op);
2801 unsigned int lookfor = get_expr_value_id (expr);
2802 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2803 if (leader)
2805 if (leader->kind == NAME)
2806 return PRE_EXPR_NAME (leader);
2807 else if (leader->kind == CONSTANT)
2808 return PRE_EXPR_CONSTANT (leader);
2810 /* Defer. */
2811 return NULL_TREE;
2814 /* It must be a complex expression, so generate it recursively. Note
2815 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2816 where the insert algorithm fails to insert a required expression. */
2817 bitmap exprset = value_expressions[lookfor];
2818 bitmap_iterator bi;
2819 unsigned int i;
2820 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2822 pre_expr temp = expression_for_id (i);
2823 /* We cannot insert random REFERENCE expressions at arbitrary
2824 places. We can insert NARYs which eventually re-materializes
2825 its operand values. */
2826 if (temp->kind == NARY)
2827 return create_expression_by_pieces (block, temp, stmts,
2828 get_expr_type (expr));
2831 /* Defer. */
2832 return NULL_TREE;
2835 #define NECESSARY GF_PLF_1
2837 /* Create an expression in pieces, so that we can handle very complex
2838 expressions that may be ANTIC, but not necessary GIMPLE.
2839 BLOCK is the basic block the expression will be inserted into,
2840 EXPR is the expression to insert (in value form)
2841 STMTS is a statement list to append the necessary insertions into.
2843 This function will die if we hit some value that shouldn't be
2844 ANTIC but is (IE there is no leader for it, or its components).
2845 The function returns NULL_TREE in case a different antic expression
2846 has to be inserted first.
2847 This function may also generate expressions that are themselves
2848 partially or fully redundant. Those that are will be either made
2849 fully redundant during the next iteration of insert (for partially
2850 redundant ones), or eliminated by eliminate (for fully redundant
2851 ones). */
2853 static tree
2854 create_expression_by_pieces (basic_block block, pre_expr expr,
2855 gimple_seq *stmts, tree type)
2857 tree name;
2858 tree folded;
2859 gimple_seq forced_stmts = NULL;
2860 unsigned int value_id;
2861 gimple_stmt_iterator gsi;
2862 tree exprtype = type ? type : get_expr_type (expr);
2863 pre_expr nameexpr;
2864 gimple newstmt;
2866 switch (expr->kind)
2868 /* We may hit the NAME/CONSTANT case if we have to convert types
2869 that value numbering saw through. */
2870 case NAME:
2871 folded = PRE_EXPR_NAME (expr);
2872 break;
2873 case CONSTANT:
2874 folded = PRE_EXPR_CONSTANT (expr);
2875 break;
2876 case REFERENCE:
2878 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2879 folded = create_component_ref_by_pieces (block, ref, stmts);
2880 if (!folded)
2881 return NULL_TREE;
2883 break;
2884 case NARY:
2886 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2887 tree *genop = XALLOCAVEC (tree, nary->length);
2888 unsigned i;
2889 for (i = 0; i < nary->length; ++i)
2891 genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2892 if (!genop[i])
2893 return NULL_TREE;
2894 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2895 may have conversions stripped. */
2896 if (nary->opcode == POINTER_PLUS_EXPR)
2898 if (i == 0)
2899 genop[i] = fold_convert (nary->type, genop[i]);
2900 else if (i == 1)
2901 genop[i] = convert_to_ptrofftype (genop[i]);
2903 else
2904 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]);
2906 if (nary->opcode == CONSTRUCTOR)
2908 vec<constructor_elt, va_gc> *elts = NULL;
2909 for (i = 0; i < nary->length; ++i)
2910 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2911 folded = build_constructor (nary->type, elts);
2913 else
2915 switch (nary->length)
2917 case 1:
2918 folded = fold_build1 (nary->opcode, nary->type,
2919 genop[0]);
2920 break;
2921 case 2:
2922 folded = fold_build2 (nary->opcode, nary->type,
2923 genop[0], genop[1]);
2924 break;
2925 case 3:
2926 folded = fold_build3 (nary->opcode, nary->type,
2927 genop[0], genop[1], genop[2]);
2928 break;
2929 default:
2930 gcc_unreachable ();
2934 break;
2935 default:
2936 gcc_unreachable ();
2939 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2940 folded = fold_convert (exprtype, folded);
2942 /* Force the generated expression to be a sequence of GIMPLE
2943 statements.
2944 We have to call unshare_expr because force_gimple_operand may
2945 modify the tree we pass to it. */
2946 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
2947 false, NULL);
2949 /* If we have any intermediate expressions to the value sets, add them
2950 to the value sets and chain them in the instruction stream. */
2951 if (forced_stmts)
2953 gsi = gsi_start (forced_stmts);
2954 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2956 gimple stmt = gsi_stmt (gsi);
2957 tree forcedname = gimple_get_lhs (stmt);
2958 pre_expr nameexpr;
2960 if (TREE_CODE (forcedname) == SSA_NAME)
2962 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
2963 VN_INFO_GET (forcedname)->valnum = forcedname;
2964 VN_INFO (forcedname)->value_id = get_next_value_id ();
2965 nameexpr = get_or_alloc_expr_for_name (forcedname);
2966 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2967 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2968 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2971 gimple_seq_add_seq (stmts, forced_stmts);
2974 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2975 newstmt = gimple_build_assign (name, folded);
2976 gimple_set_plf (newstmt, NECESSARY, false);
2978 gimple_seq_add_stmt (stmts, newstmt);
2979 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
2981 /* Fold the last statement. */
2982 gsi = gsi_last (*stmts);
2983 if (fold_stmt_inplace (&gsi))
2984 update_stmt (gsi_stmt (gsi));
2986 /* Add a value number to the temporary.
2987 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2988 we are creating the expression by pieces, and this particular piece of
2989 the expression may have been represented. There is no harm in replacing
2990 here. */
2991 value_id = get_expr_value_id (expr);
2992 VN_INFO_GET (name)->value_id = value_id;
2993 VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id);
2994 if (VN_INFO (name)->valnum == NULL_TREE)
2995 VN_INFO (name)->valnum = name;
2996 gcc_assert (VN_INFO (name)->valnum != NULL_TREE);
2997 nameexpr = get_or_alloc_expr_for_name (name);
2998 add_to_value (value_id, nameexpr);
2999 if (NEW_SETS (block))
3000 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3001 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3003 pre_stats.insertions++;
3004 if (dump_file && (dump_flags & TDF_DETAILS))
3006 fprintf (dump_file, "Inserted ");
3007 print_gimple_stmt (dump_file, newstmt, 0, 0);
3008 fprintf (dump_file, " in predecessor %d\n", block->index);
3011 return name;
3015 /* Returns true if we want to inhibit the insertions of PHI nodes
3016 for the given EXPR for basic block BB (a member of a loop).
3017 We want to do this, when we fear that the induction variable we
3018 create might inhibit vectorization. */
3020 static bool
3021 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3023 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3024 vec<vn_reference_op_s> ops = vr->operands;
3025 vn_reference_op_t op;
3026 unsigned i;
3028 /* If we aren't going to vectorize we don't inhibit anything. */
3029 if (!flag_tree_vectorize)
3030 return false;
3032 /* Otherwise we inhibit the insertion when the address of the
3033 memory reference is a simple induction variable. In other
3034 cases the vectorizer won't do anything anyway (either it's
3035 loop invariant or a complicated expression). */
3036 FOR_EACH_VEC_ELT (ops, i, op)
3038 switch (op->opcode)
3040 case CALL_EXPR:
3041 /* Calls are not a problem. */
3042 return false;
3044 case ARRAY_REF:
3045 case ARRAY_RANGE_REF:
3046 if (TREE_CODE (op->op0) != SSA_NAME)
3047 break;
3048 /* Fallthru. */
3049 case SSA_NAME:
3051 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3052 affine_iv iv;
3053 /* Default defs are loop invariant. */
3054 if (!defbb)
3055 break;
3056 /* Defined outside this loop, also loop invariant. */
3057 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3058 break;
3059 /* If it's a simple induction variable inhibit insertion,
3060 the vectorizer might be interested in this one. */
3061 if (simple_iv (bb->loop_father, bb->loop_father,
3062 op->op0, &iv, true))
3063 return true;
3064 /* No simple IV, vectorizer can't do anything, hence no
3065 reason to inhibit the transformation for this operand. */
3066 break;
3068 default:
3069 break;
3072 return false;
3075 /* Insert the to-be-made-available values of expression EXPRNUM for each
3076 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3077 merge the result with a phi node, given the same value number as
3078 NODE. Return true if we have inserted new stuff. */
3080 static bool
3081 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3082 vec<pre_expr> avail)
3084 pre_expr expr = expression_for_id (exprnum);
3085 pre_expr newphi;
3086 unsigned int val = get_expr_value_id (expr);
3087 edge pred;
3088 bool insertions = false;
3089 bool nophi = false;
3090 basic_block bprime;
3091 pre_expr eprime;
3092 edge_iterator ei;
3093 tree type = get_expr_type (expr);
3094 tree temp;
3095 gimple phi;
3097 /* Make sure we aren't creating an induction variable. */
3098 if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
3100 bool firstinsideloop = false;
3101 bool secondinsideloop = false;
3102 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3103 EDGE_PRED (block, 0)->src);
3104 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3105 EDGE_PRED (block, 1)->src);
3106 /* Induction variables only have one edge inside the loop. */
3107 if ((firstinsideloop ^ secondinsideloop)
3108 && (expr->kind != REFERENCE
3109 || inhibit_phi_insertion (block, expr)))
3111 if (dump_file && (dump_flags & TDF_DETAILS))
3112 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3113 nophi = true;
3117 /* Make the necessary insertions. */
3118 FOR_EACH_EDGE (pred, ei, block->preds)
3120 gimple_seq stmts = NULL;
3121 tree builtexpr;
3122 bprime = pred->src;
3123 eprime = avail[pred->dest_idx];
3125 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3127 builtexpr = create_expression_by_pieces (bprime, eprime,
3128 &stmts, type);
3129 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3130 gsi_insert_seq_on_edge (pred, stmts);
3131 if (!builtexpr)
3133 /* We cannot insert a PHI node if we failed to insert
3134 on one edge. */
3135 nophi = true;
3136 continue;
3138 avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3139 insertions = true;
3141 else if (eprime->kind == CONSTANT)
3143 /* Constants may not have the right type, fold_convert
3144 should give us back a constant with the right type. */
3145 tree constant = PRE_EXPR_CONSTANT (eprime);
3146 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3148 tree builtexpr = fold_convert (type, constant);
3149 if (!is_gimple_min_invariant (builtexpr))
3151 tree forcedexpr = force_gimple_operand (builtexpr,
3152 &stmts, true,
3153 NULL);
3154 if (!is_gimple_min_invariant (forcedexpr))
3156 if (forcedexpr != builtexpr)
3158 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3159 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3161 if (stmts)
3163 gimple_stmt_iterator gsi;
3164 gsi = gsi_start (stmts);
3165 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3167 gimple stmt = gsi_stmt (gsi);
3168 tree lhs = gimple_get_lhs (stmt);
3169 if (TREE_CODE (lhs) == SSA_NAME)
3170 bitmap_set_bit (inserted_exprs,
3171 SSA_NAME_VERSION (lhs));
3172 gimple_set_plf (stmt, NECESSARY, false);
3174 gsi_insert_seq_on_edge (pred, stmts);
3176 avail[pred->dest_idx]
3177 = get_or_alloc_expr_for_name (forcedexpr);
3180 else
3181 avail[pred->dest_idx]
3182 = get_or_alloc_expr_for_constant (builtexpr);
3185 else if (eprime->kind == NAME)
3187 /* We may have to do a conversion because our value
3188 numbering can look through types in certain cases, but
3189 our IL requires all operands of a phi node have the same
3190 type. */
3191 tree name = PRE_EXPR_NAME (eprime);
3192 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3194 tree builtexpr;
3195 tree forcedexpr;
3196 builtexpr = fold_convert (type, name);
3197 forcedexpr = force_gimple_operand (builtexpr,
3198 &stmts, true,
3199 NULL);
3201 if (forcedexpr != name)
3203 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3204 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3207 if (stmts)
3209 gimple_stmt_iterator gsi;
3210 gsi = gsi_start (stmts);
3211 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3213 gimple stmt = gsi_stmt (gsi);
3214 tree lhs = gimple_get_lhs (stmt);
3215 if (TREE_CODE (lhs) == SSA_NAME)
3216 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3217 gimple_set_plf (stmt, NECESSARY, false);
3219 gsi_insert_seq_on_edge (pred, stmts);
3221 avail[pred->dest_idx] = get_or_alloc_expr_for_name (forcedexpr);
3225 /* If we didn't want a phi node, and we made insertions, we still have
3226 inserted new stuff, and thus return true. If we didn't want a phi node,
3227 and didn't make insertions, we haven't added anything new, so return
3228 false. */
3229 if (nophi && insertions)
3230 return true;
3231 else if (nophi && !insertions)
3232 return false;
3234 /* Now build a phi for the new variable. */
3235 temp = make_temp_ssa_name (type, NULL, "prephitmp");
3236 phi = create_phi_node (temp, block);
3238 gimple_set_plf (phi, NECESSARY, false);
3239 VN_INFO_GET (temp)->value_id = val;
3240 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3241 if (VN_INFO (temp)->valnum == NULL_TREE)
3242 VN_INFO (temp)->valnum = temp;
3243 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3244 FOR_EACH_EDGE (pred, ei, block->preds)
3246 pre_expr ae = avail[pred->dest_idx];
3247 gcc_assert (get_expr_type (ae) == type
3248 || useless_type_conversion_p (type, get_expr_type (ae)));
3249 if (ae->kind == CONSTANT)
3250 add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3251 pred, UNKNOWN_LOCATION);
3252 else
3253 add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3256 newphi = get_or_alloc_expr_for_name (temp);
3257 add_to_value (val, newphi);
3259 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3260 this insertion, since we test for the existence of this value in PHI_GEN
3261 before proceeding with the partial redundancy checks in insert_aux.
3263 The value may exist in AVAIL_OUT, in particular, it could be represented
3264 by the expression we are trying to eliminate, in which case we want the
3265 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3266 inserted there.
3268 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3269 this block, because if it did, it would have existed in our dominator's
3270 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3273 bitmap_insert_into_set (PHI_GEN (block), newphi);
3274 bitmap_value_replace_in_set (AVAIL_OUT (block),
3275 newphi);
3276 bitmap_insert_into_set (NEW_SETS (block),
3277 newphi);
3279 if (dump_file && (dump_flags & TDF_DETAILS))
3281 fprintf (dump_file, "Created phi ");
3282 print_gimple_stmt (dump_file, phi, 0, 0);
3283 fprintf (dump_file, " in block %d\n", block->index);
3285 pre_stats.phis++;
3286 return true;
3291 /* Perform insertion of partially redundant values.
3292 For BLOCK, do the following:
3293 1. Propagate the NEW_SETS of the dominator into the current block.
3294 If the block has multiple predecessors,
3295 2a. Iterate over the ANTIC expressions for the block to see if
3296 any of them are partially redundant.
3297 2b. If so, insert them into the necessary predecessors to make
3298 the expression fully redundant.
3299 2c. Insert a new PHI merging the values of the predecessors.
3300 2d. Insert the new PHI, and the new expressions, into the
3301 NEW_SETS set.
3302 3. Recursively call ourselves on the dominator children of BLOCK.
3304 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3305 do_regular_insertion and do_partial_insertion.
3309 static bool
3310 do_regular_insertion (basic_block block, basic_block dom)
3312 bool new_stuff = false;
3313 vec<pre_expr> exprs;
3314 pre_expr expr;
3315 vec<pre_expr> avail = vNULL;
3316 int i;
3318 exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3319 avail.safe_grow (EDGE_COUNT (block->preds));
3321 FOR_EACH_VEC_ELT (exprs, i, expr)
3323 if (expr->kind == NARY
3324 || expr->kind == REFERENCE)
3326 unsigned int val;
3327 bool by_some = false;
3328 bool cant_insert = false;
3329 bool all_same = true;
3330 pre_expr first_s = NULL;
3331 edge pred;
3332 basic_block bprime;
3333 pre_expr eprime = NULL;
3334 edge_iterator ei;
3335 pre_expr edoubleprime = NULL;
3336 bool do_insertion = false;
3338 val = get_expr_value_id (expr);
3339 if (bitmap_set_contains_value (PHI_GEN (block), val))
3340 continue;
3341 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3343 if (dump_file && (dump_flags & TDF_DETAILS))
3344 fprintf (dump_file, "Found fully redundant value\n");
3345 continue;
3348 FOR_EACH_EDGE (pred, ei, block->preds)
3350 unsigned int vprime;
3352 /* We should never run insertion for the exit block
3353 and so not come across fake pred edges. */
3354 gcc_assert (!(pred->flags & EDGE_FAKE));
3355 bprime = pred->src;
3356 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3357 bprime, block);
3359 /* eprime will generally only be NULL if the
3360 value of the expression, translated
3361 through the PHI for this predecessor, is
3362 undefined. If that is the case, we can't
3363 make the expression fully redundant,
3364 because its value is undefined along a
3365 predecessor path. We can thus break out
3366 early because it doesn't matter what the
3367 rest of the results are. */
3368 if (eprime == NULL)
3370 avail[pred->dest_idx] = NULL;
3371 cant_insert = true;
3372 break;
3375 eprime = fully_constant_expression (eprime);
3376 vprime = get_expr_value_id (eprime);
3377 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3378 vprime);
3379 if (edoubleprime == NULL)
3381 avail[pred->dest_idx] = eprime;
3382 all_same = false;
3384 else
3386 avail[pred->dest_idx] = edoubleprime;
3387 by_some = true;
3388 /* We want to perform insertions to remove a redundancy on
3389 a path in the CFG we want to optimize for speed. */
3390 if (optimize_edge_for_speed_p (pred))
3391 do_insertion = true;
3392 if (first_s == NULL)
3393 first_s = edoubleprime;
3394 else if (!pre_expr_d::equal (first_s, edoubleprime))
3395 all_same = false;
3398 /* If we can insert it, it's not the same value
3399 already existing along every predecessor, and
3400 it's defined by some predecessor, it is
3401 partially redundant. */
3402 if (!cant_insert && !all_same && by_some)
3404 if (!do_insertion)
3406 if (dump_file && (dump_flags & TDF_DETAILS))
3408 fprintf (dump_file, "Skipping partial redundancy for "
3409 "expression ");
3410 print_pre_expr (dump_file, expr);
3411 fprintf (dump_file, " (%04d), no redundancy on to be "
3412 "optimized for speed edge\n", val);
3415 else if (dbg_cnt (treepre_insert))
3417 if (dump_file && (dump_flags & TDF_DETAILS))
3419 fprintf (dump_file, "Found partial redundancy for "
3420 "expression ");
3421 print_pre_expr (dump_file, expr);
3422 fprintf (dump_file, " (%04d)\n",
3423 get_expr_value_id (expr));
3425 if (insert_into_preds_of_block (block,
3426 get_expression_id (expr),
3427 avail))
3428 new_stuff = true;
3431 /* If all edges produce the same value and that value is
3432 an invariant, then the PHI has the same value on all
3433 edges. Note this. */
3434 else if (!cant_insert && all_same)
3436 gcc_assert (edoubleprime->kind == CONSTANT
3437 || edoubleprime->kind == NAME);
3439 tree temp = make_temp_ssa_name (get_expr_type (expr),
3440 NULL, "pretmp");
3441 gimple assign = gimple_build_assign (temp,
3442 edoubleprime->kind == CONSTANT ? PRE_EXPR_CONSTANT (edoubleprime) : PRE_EXPR_NAME (edoubleprime));
3443 gimple_stmt_iterator gsi = gsi_after_labels (block);
3444 gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3446 gimple_set_plf (assign, NECESSARY, false);
3447 VN_INFO_GET (temp)->value_id = val;
3448 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3449 if (VN_INFO (temp)->valnum == NULL_TREE)
3450 VN_INFO (temp)->valnum = temp;
3451 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3452 pre_expr newe = get_or_alloc_expr_for_name (temp);
3453 add_to_value (val, newe);
3454 bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3455 bitmap_insert_into_set (NEW_SETS (block), newe);
3460 exprs.release ();
3461 avail.release ();
3462 return new_stuff;
3466 /* Perform insertion for partially anticipatable expressions. There
3467 is only one case we will perform insertion for these. This case is
3468 if the expression is partially anticipatable, and fully available.
3469 In this case, we know that putting it earlier will enable us to
3470 remove the later computation. */
3473 static bool
3474 do_partial_partial_insertion (basic_block block, basic_block dom)
3476 bool new_stuff = false;
3477 vec<pre_expr> exprs;
3478 pre_expr expr;
3479 vec<pre_expr> avail = vNULL;
3480 int i;
3482 exprs = sorted_array_from_bitmap_set (PA_IN (block));
3483 avail.safe_grow (EDGE_COUNT (block->preds));
3485 FOR_EACH_VEC_ELT (exprs, i, expr)
3487 if (expr->kind == NARY
3488 || expr->kind == REFERENCE)
3490 unsigned int val;
3491 bool by_all = true;
3492 bool cant_insert = false;
3493 edge pred;
3494 basic_block bprime;
3495 pre_expr eprime = NULL;
3496 edge_iterator ei;
3498 val = get_expr_value_id (expr);
3499 if (bitmap_set_contains_value (PHI_GEN (block), val))
3500 continue;
3501 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3502 continue;
3504 FOR_EACH_EDGE (pred, ei, block->preds)
3506 unsigned int vprime;
3507 pre_expr edoubleprime;
3509 /* We should never run insertion for the exit block
3510 and so not come across fake pred edges. */
3511 gcc_assert (!(pred->flags & EDGE_FAKE));
3512 bprime = pred->src;
3513 eprime = phi_translate (expr, ANTIC_IN (block),
3514 PA_IN (block),
3515 bprime, block);
3517 /* eprime will generally only be NULL if the
3518 value of the expression, translated
3519 through the PHI for this predecessor, is
3520 undefined. If that is the case, we can't
3521 make the expression fully redundant,
3522 because its value is undefined along a
3523 predecessor path. We can thus break out
3524 early because it doesn't matter what the
3525 rest of the results are. */
3526 if (eprime == NULL)
3528 avail[pred->dest_idx] = NULL;
3529 cant_insert = true;
3530 break;
3533 eprime = fully_constant_expression (eprime);
3534 vprime = get_expr_value_id (eprime);
3535 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3536 avail[pred->dest_idx] = edoubleprime;
3537 if (edoubleprime == NULL)
3539 by_all = false;
3540 break;
3544 /* If we can insert it, it's not the same value
3545 already existing along every predecessor, and
3546 it's defined by some predecessor, it is
3547 partially redundant. */
3548 if (!cant_insert && by_all)
3550 edge succ;
3551 bool do_insertion = false;
3553 /* Insert only if we can remove a later expression on a path
3554 that we want to optimize for speed.
3555 The phi node that we will be inserting in BLOCK is not free,
3556 and inserting it for the sake of !optimize_for_speed successor
3557 may cause regressions on the speed path. */
3558 FOR_EACH_EDGE (succ, ei, block->succs)
3560 if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3561 || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3563 if (optimize_edge_for_speed_p (succ))
3564 do_insertion = true;
3568 if (!do_insertion)
3570 if (dump_file && (dump_flags & TDF_DETAILS))
3572 fprintf (dump_file, "Skipping partial partial redundancy "
3573 "for expression ");
3574 print_pre_expr (dump_file, expr);
3575 fprintf (dump_file, " (%04d), not (partially) anticipated "
3576 "on any to be optimized for speed edges\n", val);
3579 else if (dbg_cnt (treepre_insert))
3581 pre_stats.pa_insert++;
3582 if (dump_file && (dump_flags & TDF_DETAILS))
3584 fprintf (dump_file, "Found partial partial redundancy "
3585 "for expression ");
3586 print_pre_expr (dump_file, expr);
3587 fprintf (dump_file, " (%04d)\n",
3588 get_expr_value_id (expr));
3590 if (insert_into_preds_of_block (block,
3591 get_expression_id (expr),
3592 avail))
3593 new_stuff = true;
3599 exprs.release ();
3600 avail.release ();
3601 return new_stuff;
3604 static bool
3605 insert_aux (basic_block block)
3607 basic_block son;
3608 bool new_stuff = false;
3610 if (block)
3612 basic_block dom;
3613 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3614 if (dom)
3616 unsigned i;
3617 bitmap_iterator bi;
3618 bitmap_set_t newset = NEW_SETS (dom);
3619 if (newset)
3621 /* Note that we need to value_replace both NEW_SETS, and
3622 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3623 represented by some non-simple expression here that we want
3624 to replace it with. */
3625 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3627 pre_expr expr = expression_for_id (i);
3628 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3629 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3632 if (!single_pred_p (block))
3634 new_stuff |= do_regular_insertion (block, dom);
3635 if (do_partial_partial)
3636 new_stuff |= do_partial_partial_insertion (block, dom);
3640 for (son = first_dom_son (CDI_DOMINATORS, block);
3641 son;
3642 son = next_dom_son (CDI_DOMINATORS, son))
3644 new_stuff |= insert_aux (son);
3647 return new_stuff;
3650 /* Perform insertion of partially redundant values. */
3652 static void
3653 insert (void)
3655 bool new_stuff = true;
3656 basic_block bb;
3657 int num_iterations = 0;
3659 FOR_ALL_BB (bb)
3660 NEW_SETS (bb) = bitmap_set_new ();
3662 while (new_stuff)
3664 num_iterations++;
3665 if (dump_file && dump_flags & TDF_DETAILS)
3666 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3667 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3669 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3673 /* Compute the AVAIL set for all basic blocks.
3675 This function performs value numbering of the statements in each basic
3676 block. The AVAIL sets are built from information we glean while doing
3677 this value numbering, since the AVAIL sets contain only one entry per
3678 value.
3680 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3681 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3683 static void
3684 compute_avail (void)
3687 basic_block block, son;
3688 basic_block *worklist;
3689 size_t sp = 0;
3690 unsigned i;
3692 /* We pretend that default definitions are defined in the entry block.
3693 This includes function arguments and the static chain decl. */
3694 for (i = 1; i < num_ssa_names; ++i)
3696 tree name = ssa_name (i);
3697 pre_expr e;
3698 if (!name
3699 || !SSA_NAME_IS_DEFAULT_DEF (name)
3700 || has_zero_uses (name)
3701 || virtual_operand_p (name))
3702 continue;
3704 e = get_or_alloc_expr_for_name (name);
3705 add_to_value (get_expr_value_id (e), e);
3706 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3707 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3710 if (dump_file && (dump_flags & TDF_DETAILS))
3712 print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR),
3713 "tmp_gen", ENTRY_BLOCK);
3714 print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR),
3715 "avail_out", ENTRY_BLOCK);
3718 /* Allocate the worklist. */
3719 worklist = XNEWVEC (basic_block, n_basic_blocks);
3721 /* Seed the algorithm by putting the dominator children of the entry
3722 block on the worklist. */
3723 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3724 son;
3725 son = next_dom_son (CDI_DOMINATORS, son))
3726 worklist[sp++] = son;
3728 /* Loop until the worklist is empty. */
3729 while (sp)
3731 gimple_stmt_iterator gsi;
3732 gimple stmt;
3733 basic_block dom;
3735 /* Pick a block from the worklist. */
3736 block = worklist[--sp];
3738 /* Initially, the set of available values in BLOCK is that of
3739 its immediate dominator. */
3740 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3741 if (dom)
3742 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3744 /* Generate values for PHI nodes. */
3745 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3747 tree result = gimple_phi_result (gsi_stmt (gsi));
3749 /* We have no need for virtual phis, as they don't represent
3750 actual computations. */
3751 if (virtual_operand_p (result))
3752 continue;
3754 pre_expr e = get_or_alloc_expr_for_name (result);
3755 add_to_value (get_expr_value_id (e), e);
3756 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3757 bitmap_insert_into_set (PHI_GEN (block), e);
3760 BB_MAY_NOTRETURN (block) = 0;
3762 /* Now compute value numbers and populate value sets with all
3763 the expressions computed in BLOCK. */
3764 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3766 ssa_op_iter iter;
3767 tree op;
3769 stmt = gsi_stmt (gsi);
3771 /* Cache whether the basic-block has any non-visible side-effect
3772 or control flow.
3773 If this isn't a call or it is the last stmt in the
3774 basic-block then the CFG represents things correctly. */
3775 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3777 /* Non-looping const functions always return normally.
3778 Otherwise the call might not return or have side-effects
3779 that forbids hoisting possibly trapping expressions
3780 before it. */
3781 int flags = gimple_call_flags (stmt);
3782 if (!(flags & ECF_CONST)
3783 || (flags & ECF_LOOPING_CONST_OR_PURE))
3784 BB_MAY_NOTRETURN (block) = 1;
3787 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3789 pre_expr e = get_or_alloc_expr_for_name (op);
3791 add_to_value (get_expr_value_id (e), e);
3792 bitmap_insert_into_set (TMP_GEN (block), e);
3793 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3796 if (gimple_has_side_effects (stmt)
3797 || stmt_could_throw_p (stmt)
3798 || is_gimple_debug (stmt))
3799 continue;
3801 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3803 if (ssa_undefined_value_p (op))
3804 continue;
3805 pre_expr e = get_or_alloc_expr_for_name (op);
3806 bitmap_value_insert_into_set (EXP_GEN (block), e);
3809 switch (gimple_code (stmt))
3811 case GIMPLE_RETURN:
3812 continue;
3814 case GIMPLE_CALL:
3816 vn_reference_t ref;
3817 pre_expr result = NULL;
3818 vec<vn_reference_op_s> ops = vNULL;
3820 /* We can value number only calls to real functions. */
3821 if (gimple_call_internal_p (stmt))
3822 continue;
3824 copy_reference_ops_from_call (stmt, &ops);
3825 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3826 gimple_expr_type (stmt),
3827 ops, &ref, VN_NOWALK);
3828 ops.release ();
3829 if (!ref)
3830 continue;
3832 /* If the value of the call is not invalidated in
3833 this block until it is computed, add the expression
3834 to EXP_GEN. */
3835 if (!gimple_vuse (stmt)
3836 || gimple_code
3837 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
3838 || gimple_bb (SSA_NAME_DEF_STMT
3839 (gimple_vuse (stmt))) != block)
3841 result = (pre_expr) pool_alloc (pre_expr_pool);
3842 result->kind = REFERENCE;
3843 result->id = 0;
3844 PRE_EXPR_REFERENCE (result) = ref;
3846 get_or_alloc_expression_id (result);
3847 add_to_value (get_expr_value_id (result), result);
3848 bitmap_value_insert_into_set (EXP_GEN (block), result);
3850 continue;
3853 case GIMPLE_ASSIGN:
3855 pre_expr result = NULL;
3856 switch (vn_get_stmt_kind (stmt))
3858 case VN_NARY:
3860 enum tree_code code = gimple_assign_rhs_code (stmt);
3861 vn_nary_op_t nary;
3863 /* COND_EXPR and VEC_COND_EXPR are awkward in
3864 that they contain an embedded complex expression.
3865 Don't even try to shove those through PRE. */
3866 if (code == COND_EXPR
3867 || code == VEC_COND_EXPR)
3868 continue;
3870 vn_nary_op_lookup_stmt (stmt, &nary);
3871 if (!nary)
3872 continue;
3874 /* If the NARY traps and there was a preceding
3875 point in the block that might not return avoid
3876 adding the nary to EXP_GEN. */
3877 if (BB_MAY_NOTRETURN (block)
3878 && vn_nary_may_trap (nary))
3879 continue;
3881 result = (pre_expr) pool_alloc (pre_expr_pool);
3882 result->kind = NARY;
3883 result->id = 0;
3884 PRE_EXPR_NARY (result) = nary;
3885 break;
3888 case VN_REFERENCE:
3890 vn_reference_t ref;
3891 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3892 gimple_vuse (stmt),
3893 VN_WALK, &ref);
3894 if (!ref)
3895 continue;
3897 /* If the value of the reference is not invalidated in
3898 this block until it is computed, add the expression
3899 to EXP_GEN. */
3900 if (gimple_vuse (stmt))
3902 gimple def_stmt;
3903 bool ok = true;
3904 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
3905 while (!gimple_nop_p (def_stmt)
3906 && gimple_code (def_stmt) != GIMPLE_PHI
3907 && gimple_bb (def_stmt) == block)
3909 if (stmt_may_clobber_ref_p
3910 (def_stmt, gimple_assign_rhs1 (stmt)))
3912 ok = false;
3913 break;
3915 def_stmt
3916 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
3918 if (!ok)
3919 continue;
3922 result = (pre_expr) pool_alloc (pre_expr_pool);
3923 result->kind = REFERENCE;
3924 result->id = 0;
3925 PRE_EXPR_REFERENCE (result) = ref;
3926 break;
3929 default:
3930 continue;
3933 get_or_alloc_expression_id (result);
3934 add_to_value (get_expr_value_id (result), result);
3935 bitmap_value_insert_into_set (EXP_GEN (block), result);
3936 continue;
3938 default:
3939 break;
3943 if (dump_file && (dump_flags & TDF_DETAILS))
3945 print_bitmap_set (dump_file, EXP_GEN (block),
3946 "exp_gen", block->index);
3947 print_bitmap_set (dump_file, PHI_GEN (block),
3948 "phi_gen", block->index);
3949 print_bitmap_set (dump_file, TMP_GEN (block),
3950 "tmp_gen", block->index);
3951 print_bitmap_set (dump_file, AVAIL_OUT (block),
3952 "avail_out", block->index);
3955 /* Put the dominator children of BLOCK on the worklist of blocks
3956 to compute available sets for. */
3957 for (son = first_dom_son (CDI_DOMINATORS, block);
3958 son;
3959 son = next_dom_son (CDI_DOMINATORS, son))
3960 worklist[sp++] = son;
3963 free (worklist);
3967 /* Local state for the eliminate domwalk. */
3968 static vec<gimple> el_to_remove;
3969 static vec<gimple> el_to_update;
3970 static unsigned int el_todo;
3971 static vec<tree> el_avail;
3972 static vec<tree> el_avail_stack;
3974 /* Return a leader for OP that is available at the current point of the
3975 eliminate domwalk. */
3977 static tree
3978 eliminate_avail (tree op)
3980 tree valnum = VN_INFO (op)->valnum;
3981 if (TREE_CODE (valnum) == SSA_NAME)
3983 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
3984 return valnum;
3985 if (el_avail.length () > SSA_NAME_VERSION (valnum))
3986 return el_avail[SSA_NAME_VERSION (valnum)];
3988 else if (is_gimple_min_invariant (valnum))
3989 return valnum;
3990 return NULL_TREE;
3993 /* At the current point of the eliminate domwalk make OP available. */
3995 static void
3996 eliminate_push_avail (tree op)
3998 tree valnum = VN_INFO (op)->valnum;
3999 if (TREE_CODE (valnum) == SSA_NAME)
4001 if (el_avail.length () <= SSA_NAME_VERSION (valnum))
4002 el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1);
4003 el_avail[SSA_NAME_VERSION (valnum)] = op;
4004 el_avail_stack.safe_push (op);
4008 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
4009 the leader for the expression if insertion was successful. */
4011 static tree
4012 eliminate_insert (gimple_stmt_iterator *gsi, tree val)
4014 tree expr = vn_get_expr_for (val);
4015 if (!CONVERT_EXPR_P (expr)
4016 && TREE_CODE (expr) != VIEW_CONVERT_EXPR)
4017 return NULL_TREE;
4019 tree op = TREE_OPERAND (expr, 0);
4020 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op;
4021 if (!leader)
4022 return NULL_TREE;
4024 tree res = make_temp_ssa_name (TREE_TYPE (val), NULL, "pretmp");
4025 gimple tem = gimple_build_assign (res,
4026 fold_build1 (TREE_CODE (expr),
4027 TREE_TYPE (expr), leader));
4028 gsi_insert_before (gsi, tem, GSI_SAME_STMT);
4029 VN_INFO_GET (res)->valnum = val;
4031 if (TREE_CODE (leader) == SSA_NAME)
4032 gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true);
4034 pre_stats.insertions++;
4035 if (dump_file && (dump_flags & TDF_DETAILS))
4037 fprintf (dump_file, "Inserted ");
4038 print_gimple_stmt (dump_file, tem, 0, 0);
4041 return res;
4044 /* Perform elimination for the basic-block B during the domwalk. */
4046 static void
4047 eliminate_bb (dom_walk_data *, basic_block b)
4049 gimple_stmt_iterator gsi;
4050 gimple stmt;
4052 /* Mark new bb. */
4053 el_avail_stack.safe_push (NULL_TREE);
4055 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4057 gimple stmt, phi = gsi_stmt (gsi);
4058 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4059 gimple_stmt_iterator gsi2;
4061 /* We want to perform redundant PHI elimination. Do so by
4062 replacing the PHI with a single copy if possible.
4063 Do not touch inserted, single-argument or virtual PHIs. */
4064 if (gimple_phi_num_args (phi) == 1
4065 || virtual_operand_p (res))
4067 gsi_next (&gsi);
4068 continue;
4071 sprime = eliminate_avail (res);
4072 if (!sprime
4073 || sprime == res)
4075 eliminate_push_avail (res);
4076 gsi_next (&gsi);
4077 continue;
4079 else if (is_gimple_min_invariant (sprime))
4081 if (!useless_type_conversion_p (TREE_TYPE (res),
4082 TREE_TYPE (sprime)))
4083 sprime = fold_convert (TREE_TYPE (res), sprime);
4086 if (dump_file && (dump_flags & TDF_DETAILS))
4088 fprintf (dump_file, "Replaced redundant PHI node defining ");
4089 print_generic_expr (dump_file, res, 0);
4090 fprintf (dump_file, " with ");
4091 print_generic_expr (dump_file, sprime, 0);
4092 fprintf (dump_file, "\n");
4095 remove_phi_node (&gsi, false);
4097 if (inserted_exprs
4098 && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4099 && TREE_CODE (sprime) == SSA_NAME)
4100 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4102 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4103 sprime = fold_convert (TREE_TYPE (res), sprime);
4104 stmt = gimple_build_assign (res, sprime);
4105 SSA_NAME_DEF_STMT (res) = stmt;
4106 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4108 gsi2 = gsi_after_labels (b);
4109 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4110 /* Queue the copy for eventual removal. */
4111 el_to_remove.safe_push (stmt);
4112 /* If we inserted this PHI node ourself, it's not an elimination. */
4113 if (inserted_exprs
4114 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4115 pre_stats.phis--;
4116 else
4117 pre_stats.eliminations++;
4120 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4122 tree lhs = NULL_TREE;
4123 tree rhs = NULL_TREE;
4125 stmt = gsi_stmt (gsi);
4127 if (gimple_has_lhs (stmt))
4128 lhs = gimple_get_lhs (stmt);
4130 if (gimple_assign_single_p (stmt))
4131 rhs = gimple_assign_rhs1 (stmt);
4133 /* Lookup the RHS of the expression, see if we have an
4134 available computation for it. If so, replace the RHS with
4135 the available computation. */
4136 if (gimple_has_lhs (stmt)
4137 && TREE_CODE (lhs) == SSA_NAME
4138 && !gimple_has_volatile_ops (stmt))
4140 tree sprime;
4141 gimple orig_stmt = stmt;
4143 sprime = eliminate_avail (lhs);
4144 /* If there is no usable leader mark lhs as leader for its value. */
4145 if (!sprime)
4146 eliminate_push_avail (lhs);
4148 /* See PR43491. Do not replace a global register variable when
4149 it is a the RHS of an assignment. Do replace local register
4150 variables since gcc does not guarantee a local variable will
4151 be allocated in register.
4152 Do not perform copy propagation or undo constant propagation. */
4153 if (gimple_assign_single_p (stmt)
4154 && (TREE_CODE (rhs) == SSA_NAME
4155 || is_gimple_min_invariant (rhs)
4156 || (TREE_CODE (rhs) == VAR_DECL
4157 && is_global_var (rhs)
4158 && DECL_HARD_REGISTER (rhs))))
4159 continue;
4161 if (!sprime)
4163 /* If there is no existing usable leader but SCCVN thinks
4164 it has an expression it wants to use as replacement,
4165 insert that. */
4166 tree val = VN_INFO (lhs)->valnum;
4167 if (val != VN_TOP
4168 && TREE_CODE (val) == SSA_NAME
4169 && VN_INFO (val)->needs_insertion
4170 && VN_INFO (val)->expr != NULL_TREE
4171 && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE)
4172 eliminate_push_avail (sprime);
4174 else if (is_gimple_min_invariant (sprime))
4176 /* If there is no existing leader but SCCVN knows this
4177 value is constant, use that constant. */
4178 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4179 TREE_TYPE (sprime)))
4180 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4182 if (dump_file && (dump_flags & TDF_DETAILS))
4184 fprintf (dump_file, "Replaced ");
4185 print_gimple_expr (dump_file, stmt, 0, 0);
4186 fprintf (dump_file, " with ");
4187 print_generic_expr (dump_file, sprime, 0);
4188 fprintf (dump_file, " in ");
4189 print_gimple_stmt (dump_file, stmt, 0, 0);
4191 pre_stats.eliminations++;
4192 propagate_tree_value_into_stmt (&gsi, sprime);
4193 stmt = gsi_stmt (gsi);
4194 update_stmt (stmt);
4196 /* If we removed EH side-effects from the statement, clean
4197 its EH information. */
4198 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4200 bitmap_set_bit (need_eh_cleanup,
4201 gimple_bb (stmt)->index);
4202 if (dump_file && (dump_flags & TDF_DETAILS))
4203 fprintf (dump_file, " Removed EH side-effects.\n");
4205 continue;
4208 if (sprime
4209 && sprime != lhs
4210 && (rhs == NULL_TREE
4211 || TREE_CODE (rhs) != SSA_NAME
4212 || may_propagate_copy (rhs, sprime)))
4214 bool can_make_abnormal_goto
4215 = is_gimple_call (stmt)
4216 && stmt_can_make_abnormal_goto (stmt);
4218 gcc_assert (sprime != rhs);
4220 if (dump_file && (dump_flags & TDF_DETAILS))
4222 fprintf (dump_file, "Replaced ");
4223 print_gimple_expr (dump_file, stmt, 0, 0);
4224 fprintf (dump_file, " with ");
4225 print_generic_expr (dump_file, sprime, 0);
4226 fprintf (dump_file, " in ");
4227 print_gimple_stmt (dump_file, stmt, 0, 0);
4230 if (TREE_CODE (sprime) == SSA_NAME)
4231 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4232 NECESSARY, true);
4233 /* We need to make sure the new and old types actually match,
4234 which may require adding a simple cast, which fold_convert
4235 will do for us. */
4236 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4237 && !useless_type_conversion_p (gimple_expr_type (stmt),
4238 TREE_TYPE (sprime)))
4239 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4241 pre_stats.eliminations++;
4242 propagate_tree_value_into_stmt (&gsi, sprime);
4243 stmt = gsi_stmt (gsi);
4244 update_stmt (stmt);
4246 /* If we removed EH side-effects from the statement, clean
4247 its EH information. */
4248 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4250 bitmap_set_bit (need_eh_cleanup,
4251 gimple_bb (stmt)->index);
4252 if (dump_file && (dump_flags & TDF_DETAILS))
4253 fprintf (dump_file, " Removed EH side-effects.\n");
4256 /* Likewise for AB side-effects. */
4257 if (can_make_abnormal_goto
4258 && !stmt_can_make_abnormal_goto (stmt))
4260 bitmap_set_bit (need_ab_cleanup,
4261 gimple_bb (stmt)->index);
4262 if (dump_file && (dump_flags & TDF_DETAILS))
4263 fprintf (dump_file, " Removed AB side-effects.\n");
4267 /* If the statement is a scalar store, see if the expression
4268 has the same value number as its rhs. If so, the store is
4269 dead. */
4270 else if (gimple_assign_single_p (stmt)
4271 && !gimple_has_volatile_ops (stmt)
4272 && !is_gimple_reg (gimple_assign_lhs (stmt))
4273 && (TREE_CODE (rhs) == SSA_NAME
4274 || is_gimple_min_invariant (rhs)))
4276 tree val;
4277 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4278 gimple_vuse (stmt), VN_WALK, NULL);
4279 if (TREE_CODE (rhs) == SSA_NAME)
4280 rhs = VN_INFO (rhs)->valnum;
4281 if (val
4282 && operand_equal_p (val, rhs, 0))
4284 if (dump_file && (dump_flags & TDF_DETAILS))
4286 fprintf (dump_file, "Deleted redundant store ");
4287 print_gimple_stmt (dump_file, stmt, 0, 0);
4290 /* Queue stmt for removal. */
4291 el_to_remove.safe_push (stmt);
4294 /* Visit COND_EXPRs and fold the comparison with the
4295 available value-numbers. */
4296 else if (gimple_code (stmt) == GIMPLE_COND)
4298 tree op0 = gimple_cond_lhs (stmt);
4299 tree op1 = gimple_cond_rhs (stmt);
4300 tree result;
4302 if (TREE_CODE (op0) == SSA_NAME)
4303 op0 = VN_INFO (op0)->valnum;
4304 if (TREE_CODE (op1) == SSA_NAME)
4305 op1 = VN_INFO (op1)->valnum;
4306 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4307 op0, op1);
4308 if (result && TREE_CODE (result) == INTEGER_CST)
4310 if (integer_zerop (result))
4311 gimple_cond_make_false (stmt);
4312 else
4313 gimple_cond_make_true (stmt);
4314 update_stmt (stmt);
4315 el_todo = TODO_cleanup_cfg;
4318 /* Visit indirect calls and turn them into direct calls if
4319 possible. */
4320 if (is_gimple_call (stmt))
4322 tree orig_fn = gimple_call_fn (stmt);
4323 tree fn;
4324 if (!orig_fn)
4325 continue;
4326 if (TREE_CODE (orig_fn) == SSA_NAME)
4327 fn = VN_INFO (orig_fn)->valnum;
4328 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4329 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4331 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4332 if (!gimple_call_addr_fndecl (fn))
4334 fn = ipa_intraprocedural_devirtualization (stmt);
4335 if (fn)
4336 fn = build_fold_addr_expr (fn);
4339 else
4340 continue;
4341 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4342 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4343 TREE_TYPE (fn)))
4345 bool can_make_abnormal_goto
4346 = stmt_can_make_abnormal_goto (stmt);
4347 bool was_noreturn = gimple_call_noreturn_p (stmt);
4349 if (dump_file && (dump_flags & TDF_DETAILS))
4351 fprintf (dump_file, "Replacing call target with ");
4352 print_generic_expr (dump_file, fn, 0);
4353 fprintf (dump_file, " in ");
4354 print_gimple_stmt (dump_file, stmt, 0, 0);
4357 gimple_call_set_fn (stmt, fn);
4358 el_to_update.safe_push (stmt);
4360 /* When changing a call into a noreturn call, cfg cleanup
4361 is needed to fix up the noreturn call. */
4362 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4363 el_todo |= TODO_cleanup_cfg;
4365 /* If we removed EH side-effects from the statement, clean
4366 its EH information. */
4367 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4369 bitmap_set_bit (need_eh_cleanup,
4370 gimple_bb (stmt)->index);
4371 if (dump_file && (dump_flags & TDF_DETAILS))
4372 fprintf (dump_file, " Removed EH side-effects.\n");
4375 /* Likewise for AB side-effects. */
4376 if (can_make_abnormal_goto
4377 && !stmt_can_make_abnormal_goto (stmt))
4379 bitmap_set_bit (need_ab_cleanup,
4380 gimple_bb (stmt)->index);
4381 if (dump_file && (dump_flags & TDF_DETAILS))
4382 fprintf (dump_file, " Removed AB side-effects.\n");
4385 /* Changing an indirect call to a direct call may
4386 have exposed different semantics. This may
4387 require an SSA update. */
4388 el_todo |= TODO_update_ssa_only_virtuals;
4394 /* Make no longer available leaders no longer available. */
4396 static void
4397 eliminate_leave_block (dom_walk_data *, basic_block)
4399 tree entry;
4400 while ((entry = el_avail_stack.pop ()) != NULL_TREE)
4401 el_avail[SSA_NAME_VERSION (VN_INFO (entry)->valnum)] = NULL_TREE;
4404 /* Eliminate fully redundant computations. */
4406 static unsigned int
4407 eliminate (void)
4409 struct dom_walk_data walk_data;
4410 gimple_stmt_iterator gsi;
4411 gimple stmt;
4412 unsigned i;
4414 need_eh_cleanup = BITMAP_ALLOC (NULL);
4415 need_ab_cleanup = BITMAP_ALLOC (NULL);
4417 el_to_remove.create (0);
4418 el_to_update.create (0);
4419 el_todo = 0;
4420 el_avail.create (0);
4421 el_avail_stack.create (0);
4423 walk_data.dom_direction = CDI_DOMINATORS;
4424 walk_data.initialize_block_local_data = NULL;
4425 walk_data.before_dom_children = eliminate_bb;
4426 walk_data.after_dom_children = eliminate_leave_block;
4427 walk_data.global_data = NULL;
4428 walk_data.block_local_data_size = 0;
4429 init_walk_dominator_tree (&walk_data);
4430 walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
4431 fini_walk_dominator_tree (&walk_data);
4433 el_avail.release ();
4434 el_avail_stack.release ();
4436 /* We cannot remove stmts during BB walk, especially not release SSA
4437 names there as this confuses the VN machinery. The stmts ending
4438 up in el_to_remove are either stores or simple copies. */
4439 FOR_EACH_VEC_ELT (el_to_remove, i, stmt)
4441 tree lhs = gimple_assign_lhs (stmt);
4442 tree rhs = gimple_assign_rhs1 (stmt);
4443 use_operand_p use_p;
4444 gimple use_stmt;
4446 /* If there is a single use only, propagate the equivalency
4447 instead of keeping the copy. */
4448 if (TREE_CODE (lhs) == SSA_NAME
4449 && TREE_CODE (rhs) == SSA_NAME
4450 && single_imm_use (lhs, &use_p, &use_stmt)
4451 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4453 SET_USE (use_p, rhs);
4454 update_stmt (use_stmt);
4455 if (inserted_exprs
4456 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4457 && TREE_CODE (rhs) == SSA_NAME)
4458 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4461 /* If this is a store or a now unused copy, remove it. */
4462 if (TREE_CODE (lhs) != SSA_NAME
4463 || has_zero_uses (lhs))
4465 basic_block bb = gimple_bb (stmt);
4466 gsi = gsi_for_stmt (stmt);
4467 unlink_stmt_vdef (stmt);
4468 if (gsi_remove (&gsi, true))
4469 bitmap_set_bit (need_eh_cleanup, bb->index);
4470 if (inserted_exprs
4471 && TREE_CODE (lhs) == SSA_NAME)
4472 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4473 release_defs (stmt);
4476 el_to_remove.release ();
4478 /* We cannot update call statements with virtual operands during
4479 SSA walk. This might remove them which in turn makes our
4480 VN lattice invalid. */
4481 FOR_EACH_VEC_ELT (el_to_update, i, stmt)
4482 update_stmt (stmt);
4483 el_to_update.release ();
4485 return el_todo;
4488 /* Perform CFG cleanups made necessary by elimination. */
4490 static unsigned
4491 fini_eliminate (void)
4493 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4494 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4496 if (do_eh_cleanup)
4497 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4499 if (do_ab_cleanup)
4500 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4502 BITMAP_FREE (need_eh_cleanup);
4503 BITMAP_FREE (need_ab_cleanup);
4505 if (do_eh_cleanup || do_ab_cleanup)
4506 return TODO_cleanup_cfg;
4507 return 0;
4510 /* Borrow a bit of tree-ssa-dce.c for the moment.
4511 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4512 this may be a bit faster, and we may want critical edges kept split. */
4514 /* If OP's defining statement has not already been determined to be necessary,
4515 mark that statement necessary. Return the stmt, if it is newly
4516 necessary. */
4518 static inline gimple
4519 mark_operand_necessary (tree op)
4521 gimple stmt;
4523 gcc_assert (op);
4525 if (TREE_CODE (op) != SSA_NAME)
4526 return NULL;
4528 stmt = SSA_NAME_DEF_STMT (op);
4529 gcc_assert (stmt);
4531 if (gimple_plf (stmt, NECESSARY)
4532 || gimple_nop_p (stmt))
4533 return NULL;
4535 gimple_set_plf (stmt, NECESSARY, true);
4536 return stmt;
4539 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4540 to insert PHI nodes sometimes, and because value numbering of casts isn't
4541 perfect, we sometimes end up inserting dead code. This simple DCE-like
4542 pass removes any insertions we made that weren't actually used. */
4544 static void
4545 remove_dead_inserted_code (void)
4547 bitmap worklist;
4548 unsigned i;
4549 bitmap_iterator bi;
4550 gimple t;
4552 worklist = BITMAP_ALLOC (NULL);
4553 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4555 t = SSA_NAME_DEF_STMT (ssa_name (i));
4556 if (gimple_plf (t, NECESSARY))
4557 bitmap_set_bit (worklist, i);
4559 while (!bitmap_empty_p (worklist))
4561 i = bitmap_first_set_bit (worklist);
4562 bitmap_clear_bit (worklist, i);
4563 t = SSA_NAME_DEF_STMT (ssa_name (i));
4565 /* PHI nodes are somewhat special in that each PHI alternative has
4566 data and control dependencies. All the statements feeding the
4567 PHI node's arguments are always necessary. */
4568 if (gimple_code (t) == GIMPLE_PHI)
4570 unsigned k;
4572 for (k = 0; k < gimple_phi_num_args (t); k++)
4574 tree arg = PHI_ARG_DEF (t, k);
4575 if (TREE_CODE (arg) == SSA_NAME)
4577 gimple n = mark_operand_necessary (arg);
4578 if (n)
4579 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4583 else
4585 /* Propagate through the operands. Examine all the USE, VUSE and
4586 VDEF operands in this statement. Mark all the statements
4587 which feed this statement's uses as necessary. */
4588 ssa_op_iter iter;
4589 tree use;
4591 /* The operands of VDEF expressions are also needed as they
4592 represent potential definitions that may reach this
4593 statement (VDEF operands allow us to follow def-def
4594 links). */
4596 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4598 gimple n = mark_operand_necessary (use);
4599 if (n)
4600 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4605 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4607 t = SSA_NAME_DEF_STMT (ssa_name (i));
4608 if (!gimple_plf (t, NECESSARY))
4610 gimple_stmt_iterator gsi;
4612 if (dump_file && (dump_flags & TDF_DETAILS))
4614 fprintf (dump_file, "Removing unnecessary insertion:");
4615 print_gimple_stmt (dump_file, t, 0, 0);
4618 gsi = gsi_for_stmt (t);
4619 if (gimple_code (t) == GIMPLE_PHI)
4620 remove_phi_node (&gsi, true);
4621 else
4623 gsi_remove (&gsi, true);
4624 release_defs (t);
4628 BITMAP_FREE (worklist);
4632 /* Initialize data structures used by PRE. */
4634 static void
4635 init_pre (void)
4637 basic_block bb;
4639 next_expression_id = 1;
4640 expressions.create (0);
4641 expressions.safe_push (NULL);
4642 value_expressions.create (get_max_value_id () + 1);
4643 value_expressions.safe_grow_cleared (get_max_value_id() + 1);
4644 name_to_id.create (0);
4646 inserted_exprs = BITMAP_ALLOC (NULL);
4648 connect_infinite_loops_to_exit ();
4649 memset (&pre_stats, 0, sizeof (pre_stats));
4651 postorder = XNEWVEC (int, n_basic_blocks);
4652 postorder_num = inverted_post_order_compute (postorder);
4654 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4656 calculate_dominance_info (CDI_POST_DOMINATORS);
4657 calculate_dominance_info (CDI_DOMINATORS);
4659 bitmap_obstack_initialize (&grand_bitmap_obstack);
4660 phi_translate_table.create (5110);
4661 expression_to_id.create (num_ssa_names * 3);
4662 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4663 sizeof (struct bitmap_set), 30);
4664 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4665 sizeof (struct pre_expr_d), 30);
4666 FOR_ALL_BB (bb)
4668 EXP_GEN (bb) = bitmap_set_new ();
4669 PHI_GEN (bb) = bitmap_set_new ();
4670 TMP_GEN (bb) = bitmap_set_new ();
4671 AVAIL_OUT (bb) = bitmap_set_new ();
4676 /* Deallocate data structures used by PRE. */
4678 static void
4679 fini_pre ()
4681 free (postorder);
4682 value_expressions.release ();
4683 BITMAP_FREE (inserted_exprs);
4684 bitmap_obstack_release (&grand_bitmap_obstack);
4685 free_alloc_pool (bitmap_set_pool);
4686 free_alloc_pool (pre_expr_pool);
4687 phi_translate_table.dispose ();
4688 expression_to_id.dispose ();
4689 name_to_id.release ();
4691 free_aux_for_blocks ();
4693 free_dominance_info (CDI_POST_DOMINATORS);
4696 /* Gate and execute functions for PRE. */
4698 static unsigned int
4699 do_pre (void)
4701 unsigned int todo = 0;
4703 do_partial_partial =
4704 flag_tree_partial_pre && optimize_function_for_speed_p (cfun);
4706 /* This has to happen before SCCVN runs because
4707 loop_optimizer_init may create new phis, etc. */
4708 loop_optimizer_init (LOOPS_NORMAL);
4710 if (!run_scc_vn (VN_WALK))
4712 loop_optimizer_finalize ();
4713 return 0;
4716 init_pre ();
4717 scev_initialize ();
4719 /* Collect and value number expressions computed in each basic block. */
4720 compute_avail ();
4722 /* Insert can get quite slow on an incredibly large number of basic
4723 blocks due to some quadratic behavior. Until this behavior is
4724 fixed, don't run it when he have an incredibly large number of
4725 bb's. If we aren't going to run insert, there is no point in
4726 computing ANTIC, either, even though it's plenty fast. */
4727 if (n_basic_blocks < 4000)
4729 compute_antic ();
4730 insert ();
4733 /* Make sure to remove fake edges before committing our inserts.
4734 This makes sure we don't end up with extra critical edges that
4735 we would need to split. */
4736 remove_fake_exit_edges ();
4737 gsi_commit_edge_inserts ();
4739 /* Remove all the redundant expressions. */
4740 todo |= eliminate ();
4742 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4743 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4744 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4745 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4747 clear_expression_ids ();
4748 remove_dead_inserted_code ();
4749 todo |= TODO_verify_flow;
4751 scev_finalize ();
4752 fini_pre ();
4753 todo |= fini_eliminate ();
4754 loop_optimizer_finalize ();
4756 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4757 case we can merge the block with the remaining predecessor of the block.
4758 It should either:
4759 - call merge_blocks after each tail merge iteration
4760 - call merge_blocks after all tail merge iterations
4761 - mark TODO_cleanup_cfg when necessary
4762 - share the cfg cleanup with fini_pre. */
4763 todo |= tail_merge_optimize (todo);
4765 free_scc_vn ();
4767 /* Tail merging invalidates the virtual SSA web, together with
4768 cfg-cleanup opportunities exposed by PRE this will wreck the
4769 SSA updating machinery. So make sure to run update-ssa
4770 manually, before eventually scheduling cfg-cleanup as part of
4771 the todo. */
4772 update_ssa (TODO_update_ssa_only_virtuals);
4774 return todo;
4777 static bool
4778 gate_pre (void)
4780 return flag_tree_pre != 0;
4783 struct gimple_opt_pass pass_pre =
4786 GIMPLE_PASS,
4787 "pre", /* name */
4788 OPTGROUP_NONE, /* optinfo_flags */
4789 gate_pre, /* gate */
4790 do_pre, /* execute */
4791 NULL, /* sub */
4792 NULL, /* next */
4793 0, /* static_pass_number */
4794 TV_TREE_PRE, /* tv_id */
4795 PROP_no_crit_edges | PROP_cfg
4796 | PROP_ssa, /* properties_required */
4797 0, /* properties_provided */
4798 0, /* properties_destroyed */
4799 TODO_rebuild_alias, /* todo_flags_start */
4800 TODO_verify_ssa /* todo_flags_finish */
4805 /* Gate and execute functions for FRE. */
4807 static unsigned int
4808 execute_fre (void)
4810 unsigned int todo = 0;
4812 if (!run_scc_vn (VN_WALKREWRITE))
4813 return 0;
4815 memset (&pre_stats, 0, sizeof (pre_stats));
4817 /* Remove all the redundant expressions. */
4818 todo |= eliminate ();
4820 todo |= fini_eliminate ();
4822 free_scc_vn ();
4824 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4825 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4827 return todo;
4830 static bool
4831 gate_fre (void)
4833 return flag_tree_fre != 0;
4836 struct gimple_opt_pass pass_fre =
4839 GIMPLE_PASS,
4840 "fre", /* name */
4841 OPTGROUP_NONE, /* optinfo_flags */
4842 gate_fre, /* gate */
4843 execute_fre, /* execute */
4844 NULL, /* sub */
4845 NULL, /* next */
4846 0, /* static_pass_number */
4847 TV_TREE_FRE, /* tv_id */
4848 PROP_cfg | PROP_ssa, /* properties_required */
4849 0, /* properties_provided */
4850 0, /* properties_destroyed */
4851 0, /* todo_flags_start */
4852 TODO_verify_ssa /* todo_flags_finish */