pr67609.c: Add -fno-common option on hppa*-*-hpux*.
[official-gcc.git] / gcc / tree-ssa-pre.c
blob928b00ccfda3c3d4c44bfa79f4b79392690ae611
1 /* SSA-PRE for trees.
2 Copyright (C) 2001-2015 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 "backend.h"
26 #include "rtl.h"
27 #include "tree.h"
28 #include "gimple.h"
29 #include "predict.h"
30 #include "alloc-pool.h"
31 #include "tree-pass.h"
32 #include "ssa.h"
33 #include "cgraph.h"
34 #include "gimple-pretty-print.h"
35 #include "fold-const.h"
36 #include "cfganal.h"
37 #include "gimple-fold.h"
38 #include "tree-eh.h"
39 #include "gimplify.h"
40 #include "gimple-iterator.h"
41 #include "tree-cfg.h"
42 #include "tree-ssa-loop.h"
43 #include "tree-into-ssa.h"
44 #include "tree-dfa.h"
45 #include "tree-ssa.h"
46 #include "cfgloop.h"
47 #include "tree-ssa-sccvn.h"
48 #include "tree-scalar-evolution.h"
49 #include "params.h"
50 #include "dbgcnt.h"
51 #include "domwalk.h"
52 #include "tree-ssa-propagate.h"
53 #include "ipa-utils.h"
54 #include "tree-cfgcleanup.h"
56 /* TODO:
58 1. Avail sets can be shared by making an avail_find_leader that
59 walks up the dominator tree and looks in those avail sets.
60 This might affect code optimality, it's unclear right now.
61 2. Strength reduction can be performed by anticipating expressions
62 we can repair later on.
63 3. We can do back-substitution or smarter value numbering to catch
64 commutative expressions split up over multiple statements.
67 /* For ease of terminology, "expression node" in the below refers to
68 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
69 represent the actual statement containing the expressions we care about,
70 and we cache the value number by putting it in the expression. */
72 /* Basic algorithm
74 First we walk the statements to generate the AVAIL sets, the
75 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
76 generation of values/expressions by a given block. We use them
77 when computing the ANTIC sets. The AVAIL sets consist of
78 SSA_NAME's that represent values, so we know what values are
79 available in what blocks. AVAIL is a forward dataflow problem. In
80 SSA, values are never killed, so we don't need a kill set, or a
81 fixpoint iteration, in order to calculate the AVAIL sets. In
82 traditional parlance, AVAIL sets tell us the downsafety of the
83 expressions/values.
85 Next, we generate the ANTIC sets. These sets represent the
86 anticipatable expressions. ANTIC is a backwards dataflow
87 problem. An expression is anticipatable in a given block if it could
88 be generated in that block. This means that if we had to perform
89 an insertion in that block, of the value of that expression, we
90 could. Calculating the ANTIC sets requires phi translation of
91 expressions, because the flow goes backwards through phis. We must
92 iterate to a fixpoint of the ANTIC sets, because we have a kill
93 set. Even in SSA form, values are not live over the entire
94 function, only from their definition point onwards. So we have to
95 remove values from the ANTIC set once we go past the definition
96 point of the leaders that make them up.
97 compute_antic/compute_antic_aux performs this computation.
99 Third, we perform insertions to make partially redundant
100 expressions fully redundant.
102 An expression is partially redundant (excluding partial
103 anticipation) if:
105 1. It is AVAIL in some, but not all, of the predecessors of a
106 given block.
107 2. It is ANTIC in all the predecessors.
109 In order to make it fully redundant, we insert the expression into
110 the predecessors where it is not available, but is ANTIC.
112 For the partial anticipation case, we only perform insertion if it
113 is partially anticipated in some block, and fully available in all
114 of the predecessors.
116 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
117 performs these steps.
119 Fourth, we eliminate fully redundant expressions.
120 This is a simple statement walk that replaces redundant
121 calculations with the now available values. */
123 /* Representations of value numbers:
125 Value numbers are represented by a representative SSA_NAME. We
126 will create fake SSA_NAME's in situations where we need a
127 representative but do not have one (because it is a complex
128 expression). In order to facilitate storing the value numbers in
129 bitmaps, and keep the number of wasted SSA_NAME's down, we also
130 associate a value_id with each value number, and create full blown
131 ssa_name's only where we actually need them (IE in operands of
132 existing expressions).
134 Theoretically you could replace all the value_id's with
135 SSA_NAME_VERSION, but this would allocate a large number of
136 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
137 It would also require an additional indirection at each point we
138 use the value id. */
140 /* Representation of expressions on value numbers:
142 Expressions consisting of value numbers are represented the same
143 way as our VN internally represents them, with an additional
144 "pre_expr" wrapping around them in order to facilitate storing all
145 of the expressions in the same sets. */
147 /* Representation of sets:
149 The dataflow sets do not need to be sorted in any particular order
150 for the majority of their lifetime, are simply represented as two
151 bitmaps, one that keeps track of values present in the set, and one
152 that keeps track of expressions present in the set.
154 When we need them in topological order, we produce it on demand by
155 transforming the bitmap into an array and sorting it into topo
156 order. */
158 /* Type of expression, used to know which member of the PRE_EXPR union
159 is valid. */
161 enum pre_expr_kind
163 NAME,
164 NARY,
165 REFERENCE,
166 CONSTANT
169 union pre_expr_union
171 tree name;
172 tree constant;
173 vn_nary_op_t nary;
174 vn_reference_t reference;
177 typedef struct pre_expr_d : nofree_ptr_hash <pre_expr_d>
179 enum pre_expr_kind kind;
180 unsigned int id;
181 pre_expr_union u;
183 /* hash_table support. */
184 static inline hashval_t hash (const pre_expr_d *);
185 static inline int equal (const pre_expr_d *, const pre_expr_d *);
186 } *pre_expr;
188 #define PRE_EXPR_NAME(e) (e)->u.name
189 #define PRE_EXPR_NARY(e) (e)->u.nary
190 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
191 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
193 /* Compare E1 and E1 for equality. */
195 inline int
196 pre_expr_d::equal (const pre_expr_d *e1, const pre_expr_d *e2)
198 if (e1->kind != e2->kind)
199 return false;
201 switch (e1->kind)
203 case CONSTANT:
204 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
205 PRE_EXPR_CONSTANT (e2));
206 case NAME:
207 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
208 case NARY:
209 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
210 case REFERENCE:
211 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
212 PRE_EXPR_REFERENCE (e2));
213 default:
214 gcc_unreachable ();
218 /* Hash E. */
220 inline hashval_t
221 pre_expr_d::hash (const pre_expr_d *e)
223 switch (e->kind)
225 case CONSTANT:
226 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
227 case NAME:
228 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
229 case NARY:
230 return PRE_EXPR_NARY (e)->hashcode;
231 case REFERENCE:
232 return PRE_EXPR_REFERENCE (e)->hashcode;
233 default:
234 gcc_unreachable ();
238 /* Next global expression id number. */
239 static unsigned int next_expression_id;
241 /* Mapping from expression to id number we can use in bitmap sets. */
242 static vec<pre_expr> expressions;
243 static hash_table<pre_expr_d> *expression_to_id;
244 static vec<unsigned> name_to_id;
246 /* Allocate an expression id for EXPR. */
248 static inline unsigned int
249 alloc_expression_id (pre_expr expr)
251 struct pre_expr_d **slot;
252 /* Make sure we won't overflow. */
253 gcc_assert (next_expression_id + 1 > next_expression_id);
254 expr->id = next_expression_id++;
255 expressions.safe_push (expr);
256 if (expr->kind == NAME)
258 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
259 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
260 re-allocations by using vec::reserve upfront. */
261 unsigned old_len = name_to_id.length ();
262 name_to_id.reserve (num_ssa_names - old_len);
263 name_to_id.quick_grow_cleared (num_ssa_names);
264 gcc_assert (name_to_id[version] == 0);
265 name_to_id[version] = expr->id;
267 else
269 slot = expression_to_id->find_slot (expr, INSERT);
270 gcc_assert (!*slot);
271 *slot = expr;
273 return next_expression_id - 1;
276 /* Return the expression id for tree EXPR. */
278 static inline unsigned int
279 get_expression_id (const pre_expr expr)
281 return expr->id;
284 static inline unsigned int
285 lookup_expression_id (const pre_expr expr)
287 struct pre_expr_d **slot;
289 if (expr->kind == NAME)
291 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
292 if (name_to_id.length () <= version)
293 return 0;
294 return name_to_id[version];
296 else
298 slot = expression_to_id->find_slot (expr, NO_INSERT);
299 if (!slot)
300 return 0;
301 return ((pre_expr)*slot)->id;
305 /* Return the existing expression id for EXPR, or create one if one
306 does not exist yet. */
308 static inline unsigned int
309 get_or_alloc_expression_id (pre_expr expr)
311 unsigned int id = lookup_expression_id (expr);
312 if (id == 0)
313 return alloc_expression_id (expr);
314 return expr->id = id;
317 /* Return the expression that has expression id ID */
319 static inline pre_expr
320 expression_for_id (unsigned int id)
322 return expressions[id];
325 /* Free the expression id field in all of our expressions,
326 and then destroy the expressions array. */
328 static void
329 clear_expression_ids (void)
331 expressions.release ();
334 static object_allocator<pre_expr_d> pre_expr_pool ("pre_expr nodes");
336 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
338 static pre_expr
339 get_or_alloc_expr_for_name (tree name)
341 struct pre_expr_d expr;
342 pre_expr result;
343 unsigned int result_id;
345 expr.kind = NAME;
346 expr.id = 0;
347 PRE_EXPR_NAME (&expr) = name;
348 result_id = lookup_expression_id (&expr);
349 if (result_id != 0)
350 return expression_for_id (result_id);
352 result = pre_expr_pool.allocate ();
353 result->kind = NAME;
354 PRE_EXPR_NAME (result) = name;
355 alloc_expression_id (result);
356 return result;
359 /* An unordered bitmap set. One bitmap tracks values, the other,
360 expressions. */
361 typedef struct bitmap_set
363 bitmap_head expressions;
364 bitmap_head values;
365 } *bitmap_set_t;
367 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
368 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
370 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
371 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
373 /* Mapping from value id to expressions with that value_id. */
374 static vec<bitmap> value_expressions;
376 /* Sets that we need to keep track of. */
377 typedef struct bb_bitmap_sets
379 /* The EXP_GEN set, which represents expressions/values generated in
380 a basic block. */
381 bitmap_set_t exp_gen;
383 /* The PHI_GEN set, which represents PHI results generated in a
384 basic block. */
385 bitmap_set_t phi_gen;
387 /* The TMP_GEN set, which represents results/temporaries generated
388 in a basic block. IE the LHS of an expression. */
389 bitmap_set_t tmp_gen;
391 /* The AVAIL_OUT set, which represents which values are available in
392 a given basic block. */
393 bitmap_set_t avail_out;
395 /* The ANTIC_IN set, which represents which values are anticipatable
396 in a given basic block. */
397 bitmap_set_t antic_in;
399 /* The PA_IN set, which represents which values are
400 partially anticipatable in a given basic block. */
401 bitmap_set_t pa_in;
403 /* The NEW_SETS set, which is used during insertion to augment the
404 AVAIL_OUT set of blocks with the new insertions performed during
405 the current iteration. */
406 bitmap_set_t new_sets;
408 /* A cache for value_dies_in_block_x. */
409 bitmap expr_dies;
411 /* The live virtual operand on successor edges. */
412 tree vop_on_exit;
414 /* True if we have visited this block during ANTIC calculation. */
415 unsigned int visited : 1;
417 /* True when the block contains a call that might not return. */
418 unsigned int contains_may_not_return_call : 1;
419 } *bb_value_sets_t;
421 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
422 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
423 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
424 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
425 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
426 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
427 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
428 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
429 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
430 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
431 #define BB_LIVE_VOP_ON_EXIT(BB) ((bb_value_sets_t) ((BB)->aux))->vop_on_exit
434 /* Basic block list in postorder. */
435 static int *postorder;
436 static int postorder_num;
438 /* This structure is used to keep track of statistics on what
439 optimization PRE was able to perform. */
440 static struct
442 /* The number of RHS computations eliminated by PRE. */
443 int eliminations;
445 /* The number of new expressions/temporaries generated by PRE. */
446 int insertions;
448 /* The number of inserts found due to partial anticipation */
449 int pa_insert;
451 /* The number of new PHI nodes added by PRE. */
452 int phis;
453 } pre_stats;
455 static bool do_partial_partial;
456 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int);
457 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
458 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
459 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
460 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
461 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
462 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
463 unsigned int, bool);
464 static bitmap_set_t bitmap_set_new (void);
465 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
466 tree);
467 static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
468 static unsigned int get_expr_value_id (pre_expr);
470 /* We can add and remove elements and entries to and from sets
471 and hash tables, so we use alloc pools for them. */
473 static object_allocator<bitmap_set> bitmap_set_pool ("Bitmap sets");
474 static bitmap_obstack grand_bitmap_obstack;
476 /* Set of blocks with statements that have had their EH properties changed. */
477 static bitmap need_eh_cleanup;
479 /* Set of blocks with statements that have had their AB properties changed. */
480 static bitmap need_ab_cleanup;
482 /* A three tuple {e, pred, v} used to cache phi translations in the
483 phi_translate_table. */
485 typedef struct expr_pred_trans_d : free_ptr_hash<expr_pred_trans_d>
487 /* The expression. */
488 pre_expr e;
490 /* The predecessor block along which we translated the expression. */
491 basic_block pred;
493 /* The value that resulted from the translation. */
494 pre_expr v;
496 /* The hashcode for the expression, pred pair. This is cached for
497 speed reasons. */
498 hashval_t hashcode;
500 /* hash_table support. */
501 static inline hashval_t hash (const expr_pred_trans_d *);
502 static inline int equal (const expr_pred_trans_d *, const expr_pred_trans_d *);
503 } *expr_pred_trans_t;
504 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
506 inline hashval_t
507 expr_pred_trans_d::hash (const expr_pred_trans_d *e)
509 return e->hashcode;
512 inline int
513 expr_pred_trans_d::equal (const expr_pred_trans_d *ve1,
514 const expr_pred_trans_d *ve2)
516 basic_block b1 = ve1->pred;
517 basic_block b2 = ve2->pred;
519 /* If they are not translations for the same basic block, they can't
520 be equal. */
521 if (b1 != b2)
522 return false;
523 return pre_expr_d::equal (ve1->e, ve2->e);
526 /* The phi_translate_table caches phi translations for a given
527 expression and predecessor. */
528 static hash_table<expr_pred_trans_d> *phi_translate_table;
530 /* Add the tuple mapping from {expression E, basic block PRED} to
531 the phi translation table and return whether it pre-existed. */
533 static inline bool
534 phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred)
536 expr_pred_trans_t *slot;
537 expr_pred_trans_d tem;
538 hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e),
539 pred->index);
540 tem.e = e;
541 tem.pred = pred;
542 tem.hashcode = hash;
543 slot = phi_translate_table->find_slot_with_hash (&tem, hash, INSERT);
544 if (*slot)
546 *entry = *slot;
547 return true;
550 *entry = *slot = XNEW (struct expr_pred_trans_d);
551 (*entry)->e = e;
552 (*entry)->pred = pred;
553 (*entry)->hashcode = hash;
554 return false;
558 /* Add expression E to the expression set of value id V. */
560 static void
561 add_to_value (unsigned int v, pre_expr e)
563 bitmap set;
565 gcc_checking_assert (get_expr_value_id (e) == v);
567 if (v >= value_expressions.length ())
569 value_expressions.safe_grow_cleared (v + 1);
572 set = value_expressions[v];
573 if (!set)
575 set = BITMAP_ALLOC (&grand_bitmap_obstack);
576 value_expressions[v] = set;
579 bitmap_set_bit (set, get_or_alloc_expression_id (e));
582 /* Create a new bitmap set and return it. */
584 static bitmap_set_t
585 bitmap_set_new (void)
587 bitmap_set_t ret = bitmap_set_pool.allocate ();
588 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
589 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
590 return ret;
593 /* Return the value id for a PRE expression EXPR. */
595 static unsigned int
596 get_expr_value_id (pre_expr expr)
598 unsigned int id;
599 switch (expr->kind)
601 case CONSTANT:
602 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
603 break;
604 case NAME:
605 id = VN_INFO (PRE_EXPR_NAME (expr))->value_id;
606 break;
607 case NARY:
608 id = PRE_EXPR_NARY (expr)->value_id;
609 break;
610 case REFERENCE:
611 id = PRE_EXPR_REFERENCE (expr)->value_id;
612 break;
613 default:
614 gcc_unreachable ();
616 /* ??? We cannot assert that expr has a value-id (it can be 0), because
617 we assign value-ids only to expressions that have a result
618 in set_hashtable_value_ids. */
619 return id;
622 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
624 static tree
625 sccvn_valnum_from_value_id (unsigned int val)
627 bitmap_iterator bi;
628 unsigned int i;
629 bitmap exprset = value_expressions[val];
630 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
632 pre_expr vexpr = expression_for_id (i);
633 if (vexpr->kind == NAME)
634 return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
635 else if (vexpr->kind == CONSTANT)
636 return PRE_EXPR_CONSTANT (vexpr);
638 return NULL_TREE;
641 /* Remove an expression EXPR from a bitmapped set. */
643 static void
644 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
646 unsigned int val = get_expr_value_id (expr);
647 if (!value_id_constant_p (val))
649 bitmap_clear_bit (&set->values, val);
650 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
654 static void
655 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
656 unsigned int val, bool allow_constants)
658 if (allow_constants || !value_id_constant_p (val))
660 /* We specifically expect this and only this function to be able to
661 insert constants into a set. */
662 bitmap_set_bit (&set->values, val);
663 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
667 /* Insert an expression EXPR into a bitmapped set. */
669 static void
670 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
672 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
675 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
677 static void
678 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
680 bitmap_copy (&dest->expressions, &orig->expressions);
681 bitmap_copy (&dest->values, &orig->values);
685 /* Free memory used up by SET. */
686 static void
687 bitmap_set_free (bitmap_set_t set)
689 bitmap_clear (&set->expressions);
690 bitmap_clear (&set->values);
694 /* Generate an topological-ordered array of bitmap set SET. */
696 static vec<pre_expr>
697 sorted_array_from_bitmap_set (bitmap_set_t set)
699 unsigned int i, j;
700 bitmap_iterator bi, bj;
701 vec<pre_expr> result;
703 /* Pre-allocate enough space for the array. */
704 result.create (bitmap_count_bits (&set->expressions));
706 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
708 /* The number of expressions having a given value is usually
709 relatively small. Thus, rather than making a vector of all
710 the expressions and sorting it by value-id, we walk the values
711 and check in the reverse mapping that tells us what expressions
712 have a given value, to filter those in our set. As a result,
713 the expressions are inserted in value-id order, which means
714 topological order.
716 If this is somehow a significant lose for some cases, we can
717 choose which set to walk based on the set size. */
718 bitmap exprset = value_expressions[i];
719 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj)
721 if (bitmap_bit_p (&set->expressions, j))
722 result.quick_push (expression_for_id (j));
726 return result;
729 /* Perform bitmapped set operation DEST &= ORIG. */
731 static void
732 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
734 bitmap_iterator bi;
735 unsigned int i;
737 if (dest != orig)
739 bitmap_head temp;
740 bitmap_initialize (&temp, &grand_bitmap_obstack);
742 bitmap_and_into (&dest->values, &orig->values);
743 bitmap_copy (&temp, &dest->expressions);
744 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
746 pre_expr expr = expression_for_id (i);
747 unsigned int value_id = get_expr_value_id (expr);
748 if (!bitmap_bit_p (&dest->values, value_id))
749 bitmap_clear_bit (&dest->expressions, i);
751 bitmap_clear (&temp);
755 /* Subtract all values and expressions contained in ORIG from DEST. */
757 static bitmap_set_t
758 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
760 bitmap_set_t result = bitmap_set_new ();
761 bitmap_iterator bi;
762 unsigned int i;
764 bitmap_and_compl (&result->expressions, &dest->expressions,
765 &orig->expressions);
767 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
769 pre_expr expr = expression_for_id (i);
770 unsigned int value_id = get_expr_value_id (expr);
771 bitmap_set_bit (&result->values, value_id);
774 return result;
777 /* Subtract all the values in bitmap set B from bitmap set A. */
779 static void
780 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
782 unsigned int i;
783 bitmap_iterator bi;
784 bitmap_head temp;
786 bitmap_initialize (&temp, &grand_bitmap_obstack);
788 bitmap_copy (&temp, &a->expressions);
789 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
791 pre_expr expr = expression_for_id (i);
792 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
793 bitmap_remove_from_set (a, expr);
795 bitmap_clear (&temp);
799 /* Return true if bitmapped set SET contains the value VALUE_ID. */
801 static bool
802 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
804 if (value_id_constant_p (value_id))
805 return true;
807 if (!set || bitmap_empty_p (&set->expressions))
808 return false;
810 return bitmap_bit_p (&set->values, value_id);
813 static inline bool
814 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
816 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
819 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
821 static void
822 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
823 const pre_expr expr)
825 bitmap exprset;
826 unsigned int i;
827 bitmap_iterator bi;
829 if (value_id_constant_p (lookfor))
830 return;
832 if (!bitmap_set_contains_value (set, lookfor))
833 return;
835 /* The number of expressions having a given value is usually
836 significantly less than the total number of expressions in SET.
837 Thus, rather than check, for each expression in SET, whether it
838 has the value LOOKFOR, we walk the reverse mapping that tells us
839 what expressions have a given value, and see if any of those
840 expressions are in our set. For large testcases, this is about
841 5-10x faster than walking the bitmap. If this is somehow a
842 significant lose for some cases, we can choose which set to walk
843 based on the set size. */
844 exprset = value_expressions[lookfor];
845 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
847 if (bitmap_clear_bit (&set->expressions, i))
849 bitmap_set_bit (&set->expressions, get_expression_id (expr));
850 return;
854 gcc_unreachable ();
857 /* Return true if two bitmap sets are equal. */
859 static bool
860 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
862 return bitmap_equal_p (&a->values, &b->values);
865 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
866 and add it otherwise. */
868 static void
869 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
871 unsigned int val = get_expr_value_id (expr);
873 if (bitmap_set_contains_value (set, val))
874 bitmap_set_replace_value (set, val, expr);
875 else
876 bitmap_insert_into_set (set, expr);
879 /* Insert EXPR into SET if EXPR's value is not already present in
880 SET. */
882 static void
883 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
885 unsigned int val = get_expr_value_id (expr);
887 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
889 /* Constant values are always considered to be part of the set. */
890 if (value_id_constant_p (val))
891 return;
893 /* If the value membership changed, add the expression. */
894 if (bitmap_set_bit (&set->values, val))
895 bitmap_set_bit (&set->expressions, expr->id);
898 /* Print out EXPR to outfile. */
900 static void
901 print_pre_expr (FILE *outfile, const pre_expr expr)
903 switch (expr->kind)
905 case CONSTANT:
906 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
907 break;
908 case NAME:
909 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
910 break;
911 case NARY:
913 unsigned int i;
914 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
915 fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode));
916 for (i = 0; i < nary->length; i++)
918 print_generic_expr (outfile, nary->op[i], 0);
919 if (i != (unsigned) nary->length - 1)
920 fprintf (outfile, ",");
922 fprintf (outfile, "}");
924 break;
926 case REFERENCE:
928 vn_reference_op_t vro;
929 unsigned int i;
930 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
931 fprintf (outfile, "{");
932 for (i = 0;
933 ref->operands.iterate (i, &vro);
934 i++)
936 bool closebrace = false;
937 if (vro->opcode != SSA_NAME
938 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
940 fprintf (outfile, "%s", get_tree_code_name (vro->opcode));
941 if (vro->op0)
943 fprintf (outfile, "<");
944 closebrace = true;
947 if (vro->op0)
949 print_generic_expr (outfile, vro->op0, 0);
950 if (vro->op1)
952 fprintf (outfile, ",");
953 print_generic_expr (outfile, vro->op1, 0);
955 if (vro->op2)
957 fprintf (outfile, ",");
958 print_generic_expr (outfile, vro->op2, 0);
961 if (closebrace)
962 fprintf (outfile, ">");
963 if (i != ref->operands.length () - 1)
964 fprintf (outfile, ",");
966 fprintf (outfile, "}");
967 if (ref->vuse)
969 fprintf (outfile, "@");
970 print_generic_expr (outfile, ref->vuse, 0);
973 break;
976 void debug_pre_expr (pre_expr);
978 /* Like print_pre_expr but always prints to stderr. */
979 DEBUG_FUNCTION void
980 debug_pre_expr (pre_expr e)
982 print_pre_expr (stderr, e);
983 fprintf (stderr, "\n");
986 /* Print out SET to OUTFILE. */
988 static void
989 print_bitmap_set (FILE *outfile, bitmap_set_t set,
990 const char *setname, int blockindex)
992 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
993 if (set)
995 bool first = true;
996 unsigned i;
997 bitmap_iterator bi;
999 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1001 const pre_expr expr = expression_for_id (i);
1003 if (!first)
1004 fprintf (outfile, ", ");
1005 first = false;
1006 print_pre_expr (outfile, expr);
1008 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1011 fprintf (outfile, " }\n");
1014 void debug_bitmap_set (bitmap_set_t);
1016 DEBUG_FUNCTION void
1017 debug_bitmap_set (bitmap_set_t set)
1019 print_bitmap_set (stderr, set, "debug", 0);
1022 void debug_bitmap_sets_for (basic_block);
1024 DEBUG_FUNCTION void
1025 debug_bitmap_sets_for (basic_block bb)
1027 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1028 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1029 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1030 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1031 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1032 if (do_partial_partial)
1033 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1034 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1037 /* Print out the expressions that have VAL to OUTFILE. */
1039 static void
1040 print_value_expressions (FILE *outfile, unsigned int val)
1042 bitmap set = value_expressions[val];
1043 if (set)
1045 bitmap_set x;
1046 char s[10];
1047 sprintf (s, "%04d", val);
1048 x.expressions = *set;
1049 print_bitmap_set (outfile, &x, s, 0);
1054 DEBUG_FUNCTION void
1055 debug_value_expressions (unsigned int val)
1057 print_value_expressions (stderr, val);
1060 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1061 represent it. */
1063 static pre_expr
1064 get_or_alloc_expr_for_constant (tree constant)
1066 unsigned int result_id;
1067 unsigned int value_id;
1068 struct pre_expr_d expr;
1069 pre_expr newexpr;
1071 expr.kind = CONSTANT;
1072 PRE_EXPR_CONSTANT (&expr) = constant;
1073 result_id = lookup_expression_id (&expr);
1074 if (result_id != 0)
1075 return expression_for_id (result_id);
1077 newexpr = pre_expr_pool.allocate ();
1078 newexpr->kind = CONSTANT;
1079 PRE_EXPR_CONSTANT (newexpr) = constant;
1080 alloc_expression_id (newexpr);
1081 value_id = get_or_alloc_constant_value_id (constant);
1082 add_to_value (value_id, newexpr);
1083 return newexpr;
1086 /* Given a value id V, find the actual tree representing the constant
1087 value if there is one, and return it. Return NULL if we can't find
1088 a constant. */
1090 static tree
1091 get_constant_for_value_id (unsigned int v)
1093 if (value_id_constant_p (v))
1095 unsigned int i;
1096 bitmap_iterator bi;
1097 bitmap exprset = value_expressions[v];
1099 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1101 pre_expr expr = expression_for_id (i);
1102 if (expr->kind == CONSTANT)
1103 return PRE_EXPR_CONSTANT (expr);
1106 return NULL;
1109 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1110 Currently only supports constants and SSA_NAMES. */
1111 static pre_expr
1112 get_or_alloc_expr_for (tree t)
1114 if (TREE_CODE (t) == SSA_NAME)
1115 return get_or_alloc_expr_for_name (t);
1116 else if (is_gimple_min_invariant (t))
1117 return get_or_alloc_expr_for_constant (t);
1118 else
1120 /* More complex expressions can result from SCCVN expression
1121 simplification that inserts values for them. As they all
1122 do not have VOPs the get handled by the nary ops struct. */
1123 vn_nary_op_t result;
1124 unsigned int result_id;
1125 vn_nary_op_lookup (t, &result);
1126 if (result != NULL)
1128 pre_expr e = pre_expr_pool.allocate ();
1129 e->kind = NARY;
1130 PRE_EXPR_NARY (e) = result;
1131 result_id = lookup_expression_id (e);
1132 if (result_id != 0)
1134 pre_expr_pool.remove (e);
1135 e = expression_for_id (result_id);
1136 return e;
1138 alloc_expression_id (e);
1139 return e;
1142 return NULL;
1145 /* Return the folded version of T if T, when folded, is a gimple
1146 min_invariant. Otherwise, return T. */
1148 static pre_expr
1149 fully_constant_expression (pre_expr e)
1151 switch (e->kind)
1153 case CONSTANT:
1154 return e;
1155 case NARY:
1157 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1158 switch (TREE_CODE_CLASS (nary->opcode))
1160 case tcc_binary:
1161 case tcc_comparison:
1163 /* We have to go from trees to pre exprs to value ids to
1164 constants. */
1165 tree naryop0 = nary->op[0];
1166 tree naryop1 = nary->op[1];
1167 tree result;
1168 if (!is_gimple_min_invariant (naryop0))
1170 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1171 unsigned int vrep0 = get_expr_value_id (rep0);
1172 tree const0 = get_constant_for_value_id (vrep0);
1173 if (const0)
1174 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1176 if (!is_gimple_min_invariant (naryop1))
1178 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1179 unsigned int vrep1 = get_expr_value_id (rep1);
1180 tree const1 = get_constant_for_value_id (vrep1);
1181 if (const1)
1182 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1184 result = fold_binary (nary->opcode, nary->type,
1185 naryop0, naryop1);
1186 if (result && is_gimple_min_invariant (result))
1187 return get_or_alloc_expr_for_constant (result);
1188 /* We might have simplified the expression to a
1189 SSA_NAME for example from x_1 * 1. But we cannot
1190 insert a PHI for x_1 unconditionally as x_1 might
1191 not be available readily. */
1192 return e;
1194 case tcc_reference:
1195 if (nary->opcode != REALPART_EXPR
1196 && nary->opcode != IMAGPART_EXPR
1197 && nary->opcode != VIEW_CONVERT_EXPR)
1198 return e;
1199 /* Fallthrough. */
1200 case tcc_unary:
1202 /* We have to go from trees to pre exprs to value ids to
1203 constants. */
1204 tree naryop0 = nary->op[0];
1205 tree const0, result;
1206 if (is_gimple_min_invariant (naryop0))
1207 const0 = naryop0;
1208 else
1210 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1211 unsigned int vrep0 = get_expr_value_id (rep0);
1212 const0 = get_constant_for_value_id (vrep0);
1214 result = NULL;
1215 if (const0)
1217 tree type1 = TREE_TYPE (nary->op[0]);
1218 const0 = fold_convert (type1, const0);
1219 result = fold_unary (nary->opcode, nary->type, const0);
1221 if (result && is_gimple_min_invariant (result))
1222 return get_or_alloc_expr_for_constant (result);
1223 return e;
1225 default:
1226 return e;
1229 case REFERENCE:
1231 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1232 tree folded;
1233 if ((folded = fully_constant_vn_reference_p (ref)))
1234 return get_or_alloc_expr_for_constant (folded);
1235 return e;
1237 default:
1238 return e;
1240 return e;
1243 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1244 it has the value it would have in BLOCK. Set *SAME_VALID to true
1245 in case the new vuse doesn't change the value id of the OPERANDS. */
1247 static tree
1248 translate_vuse_through_block (vec<vn_reference_op_s> operands,
1249 alias_set_type set, tree type, tree vuse,
1250 basic_block phiblock,
1251 basic_block block, bool *same_valid)
1253 gimple *phi = SSA_NAME_DEF_STMT (vuse);
1254 ao_ref ref;
1255 edge e = NULL;
1256 bool use_oracle;
1258 *same_valid = true;
1260 if (gimple_bb (phi) != phiblock)
1261 return vuse;
1263 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1265 /* Use the alias-oracle to find either the PHI node in this block,
1266 the first VUSE used in this block that is equivalent to vuse or
1267 the first VUSE which definition in this block kills the value. */
1268 if (gimple_code (phi) == GIMPLE_PHI)
1269 e = find_edge (block, phiblock);
1270 else if (use_oracle)
1271 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1273 vuse = gimple_vuse (phi);
1274 phi = SSA_NAME_DEF_STMT (vuse);
1275 if (gimple_bb (phi) != phiblock)
1276 return vuse;
1277 if (gimple_code (phi) == GIMPLE_PHI)
1279 e = find_edge (block, phiblock);
1280 break;
1283 else
1284 return NULL_TREE;
1286 if (e)
1288 if (use_oracle)
1290 bitmap visited = NULL;
1291 unsigned int cnt;
1292 /* Try to find a vuse that dominates this phi node by skipping
1293 non-clobbering statements. */
1294 vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false,
1295 NULL, NULL);
1296 if (visited)
1297 BITMAP_FREE (visited);
1299 else
1300 vuse = NULL_TREE;
1301 if (!vuse)
1303 /* If we didn't find any, the value ID can't stay the same,
1304 but return the translated vuse. */
1305 *same_valid = false;
1306 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1308 /* ??? We would like to return vuse here as this is the canonical
1309 upmost vdef that this reference is associated with. But during
1310 insertion of the references into the hash tables we only ever
1311 directly insert with their direct gimple_vuse, hence returning
1312 something else would make us not find the other expression. */
1313 return PHI_ARG_DEF (phi, e->dest_idx);
1316 return NULL_TREE;
1319 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1320 SET2. This is used to avoid making a set consisting of the union
1321 of PA_IN and ANTIC_IN during insert. */
1323 static inline pre_expr
1324 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1326 pre_expr result;
1328 result = bitmap_find_leader (set1, val);
1329 if (!result && set2)
1330 result = bitmap_find_leader (set2, val);
1331 return result;
1334 /* Get the tree type for our PRE expression e. */
1336 static tree
1337 get_expr_type (const pre_expr e)
1339 switch (e->kind)
1341 case NAME:
1342 return TREE_TYPE (PRE_EXPR_NAME (e));
1343 case CONSTANT:
1344 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1345 case REFERENCE:
1346 return PRE_EXPR_REFERENCE (e)->type;
1347 case NARY:
1348 return PRE_EXPR_NARY (e)->type;
1350 gcc_unreachable ();
1353 /* Get a representative SSA_NAME for a given expression.
1354 Since all of our sub-expressions are treated as values, we require
1355 them to be SSA_NAME's for simplicity.
1356 Prior versions of GVNPRE used to use "value handles" here, so that
1357 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1358 either case, the operands are really values (IE we do not expect
1359 them to be usable without finding leaders). */
1361 static tree
1362 get_representative_for (const pre_expr e)
1364 tree name;
1365 unsigned int value_id = get_expr_value_id (e);
1367 switch (e->kind)
1369 case NAME:
1370 return PRE_EXPR_NAME (e);
1371 case CONSTANT:
1372 return PRE_EXPR_CONSTANT (e);
1373 case NARY:
1374 case REFERENCE:
1376 /* Go through all of the expressions representing this value
1377 and pick out an SSA_NAME. */
1378 unsigned int i;
1379 bitmap_iterator bi;
1380 bitmap exprs = value_expressions[value_id];
1381 EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1383 pre_expr rep = expression_for_id (i);
1384 if (rep->kind == NAME)
1385 return PRE_EXPR_NAME (rep);
1386 else if (rep->kind == CONSTANT)
1387 return PRE_EXPR_CONSTANT (rep);
1390 break;
1393 /* If we reached here we couldn't find an SSA_NAME. This can
1394 happen when we've discovered a value that has never appeared in
1395 the program as set to an SSA_NAME, as the result of phi translation.
1396 Create one here.
1397 ??? We should be able to re-use this when we insert the statement
1398 to compute it. */
1399 name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1400 VN_INFO_GET (name)->value_id = value_id;
1401 VN_INFO (name)->valnum = name;
1402 /* ??? For now mark this SSA name for release by SCCVN. */
1403 VN_INFO (name)->needs_insertion = true;
1404 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1405 if (dump_file && (dump_flags & TDF_DETAILS))
1407 fprintf (dump_file, "Created SSA_NAME representative ");
1408 print_generic_expr (dump_file, name, 0);
1409 fprintf (dump_file, " for expression:");
1410 print_pre_expr (dump_file, e);
1411 fprintf (dump_file, " (%04d)\n", value_id);
1414 return name;
1419 static pre_expr
1420 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1421 basic_block pred, basic_block phiblock);
1423 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1424 the phis in PRED. Return NULL if we can't find a leader for each part
1425 of the translated expression. */
1427 static pre_expr
1428 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1429 basic_block pred, basic_block phiblock)
1431 switch (expr->kind)
1433 case NARY:
1435 unsigned int i;
1436 bool changed = false;
1437 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1438 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1439 sizeof_vn_nary_op (nary->length));
1440 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1442 for (i = 0; i < newnary->length; i++)
1444 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1445 continue;
1446 else
1448 pre_expr leader, result;
1449 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1450 leader = find_leader_in_sets (op_val_id, set1, set2);
1451 result = phi_translate (leader, set1, set2, pred, phiblock);
1452 if (result && result != leader)
1454 tree name = get_representative_for (result);
1455 if (!name)
1456 return NULL;
1457 newnary->op[i] = name;
1459 else if (!result)
1460 return NULL;
1462 changed |= newnary->op[i] != nary->op[i];
1465 if (changed)
1467 pre_expr constant;
1468 unsigned int new_val_id;
1470 tree result = vn_nary_op_lookup_pieces (newnary->length,
1471 newnary->opcode,
1472 newnary->type,
1473 &newnary->op[0],
1474 &nary);
1475 if (result && is_gimple_min_invariant (result))
1476 return get_or_alloc_expr_for_constant (result);
1478 expr = pre_expr_pool.allocate ();
1479 expr->kind = NARY;
1480 expr->id = 0;
1481 if (nary)
1483 PRE_EXPR_NARY (expr) = nary;
1484 constant = fully_constant_expression (expr);
1485 if (constant != expr)
1486 return constant;
1488 new_val_id = nary->value_id;
1489 get_or_alloc_expression_id (expr);
1491 else
1493 new_val_id = get_next_value_id ();
1494 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
1495 nary = vn_nary_op_insert_pieces (newnary->length,
1496 newnary->opcode,
1497 newnary->type,
1498 &newnary->op[0],
1499 result, new_val_id);
1500 PRE_EXPR_NARY (expr) = nary;
1501 constant = fully_constant_expression (expr);
1502 if (constant != expr)
1503 return constant;
1504 get_or_alloc_expression_id (expr);
1506 add_to_value (new_val_id, expr);
1508 return expr;
1510 break;
1512 case REFERENCE:
1514 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1515 vec<vn_reference_op_s> operands = ref->operands;
1516 tree vuse = ref->vuse;
1517 tree newvuse = vuse;
1518 vec<vn_reference_op_s> newoperands = vNULL;
1519 bool changed = false, same_valid = true;
1520 unsigned int i, n;
1521 vn_reference_op_t operand;
1522 vn_reference_t newref;
1524 for (i = 0; operands.iterate (i, &operand); i++)
1526 pre_expr opresult;
1527 pre_expr leader;
1528 tree op[3];
1529 tree type = operand->type;
1530 vn_reference_op_s newop = *operand;
1531 op[0] = operand->op0;
1532 op[1] = operand->op1;
1533 op[2] = operand->op2;
1534 for (n = 0; n < 3; ++n)
1536 unsigned int op_val_id;
1537 if (!op[n])
1538 continue;
1539 if (TREE_CODE (op[n]) != SSA_NAME)
1541 /* We can't possibly insert these. */
1542 if (n != 0
1543 && !is_gimple_min_invariant (op[n]))
1544 break;
1545 continue;
1547 op_val_id = VN_INFO (op[n])->value_id;
1548 leader = find_leader_in_sets (op_val_id, set1, set2);
1549 if (!leader)
1550 break;
1551 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1552 if (!opresult)
1553 break;
1554 if (opresult != leader)
1556 tree name = get_representative_for (opresult);
1557 if (!name)
1558 break;
1559 changed |= name != op[n];
1560 op[n] = name;
1563 if (n != 3)
1565 newoperands.release ();
1566 return NULL;
1568 if (!changed)
1569 continue;
1570 if (!newoperands.exists ())
1571 newoperands = operands.copy ();
1572 /* We may have changed from an SSA_NAME to a constant */
1573 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1574 newop.opcode = TREE_CODE (op[0]);
1575 newop.type = type;
1576 newop.op0 = op[0];
1577 newop.op1 = op[1];
1578 newop.op2 = op[2];
1579 newoperands[i] = newop;
1581 gcc_checking_assert (i == operands.length ());
1583 if (vuse)
1585 newvuse = translate_vuse_through_block (newoperands.exists ()
1586 ? newoperands : operands,
1587 ref->set, ref->type,
1588 vuse, phiblock, pred,
1589 &same_valid);
1590 if (newvuse == NULL_TREE)
1592 newoperands.release ();
1593 return NULL;
1597 if (changed || newvuse != vuse)
1599 unsigned int new_val_id;
1600 pre_expr constant;
1602 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1603 ref->type,
1604 newoperands.exists ()
1605 ? newoperands : operands,
1606 &newref, VN_WALK);
1607 if (result)
1608 newoperands.release ();
1610 /* We can always insert constants, so if we have a partial
1611 redundant constant load of another type try to translate it
1612 to a constant of appropriate type. */
1613 if (result && is_gimple_min_invariant (result))
1615 tree tem = result;
1616 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1618 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1619 if (tem && !is_gimple_min_invariant (tem))
1620 tem = NULL_TREE;
1622 if (tem)
1623 return get_or_alloc_expr_for_constant (tem);
1626 /* If we'd have to convert things we would need to validate
1627 if we can insert the translated expression. So fail
1628 here for now - we cannot insert an alias with a different
1629 type in the VN tables either, as that would assert. */
1630 if (result
1631 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1632 return NULL;
1633 else if (!result && newref
1634 && !useless_type_conversion_p (ref->type, newref->type))
1636 newoperands.release ();
1637 return NULL;
1640 expr = pre_expr_pool.allocate ();
1641 expr->kind = REFERENCE;
1642 expr->id = 0;
1644 if (newref)
1646 PRE_EXPR_REFERENCE (expr) = newref;
1647 constant = fully_constant_expression (expr);
1648 if (constant != expr)
1649 return constant;
1651 new_val_id = newref->value_id;
1652 get_or_alloc_expression_id (expr);
1654 else
1656 if (changed || !same_valid)
1658 new_val_id = get_next_value_id ();
1659 value_expressions.safe_grow_cleared
1660 (get_max_value_id () + 1);
1662 else
1663 new_val_id = ref->value_id;
1664 if (!newoperands.exists ())
1665 newoperands = operands.copy ();
1666 newref = vn_reference_insert_pieces (newvuse, ref->set,
1667 ref->type,
1668 newoperands,
1669 result, new_val_id);
1670 newoperands = vNULL;
1671 PRE_EXPR_REFERENCE (expr) = newref;
1672 constant = fully_constant_expression (expr);
1673 if (constant != expr)
1674 return constant;
1675 get_or_alloc_expression_id (expr);
1677 add_to_value (new_val_id, expr);
1679 newoperands.release ();
1680 return expr;
1682 break;
1684 case NAME:
1686 tree name = PRE_EXPR_NAME (expr);
1687 gimple *def_stmt = SSA_NAME_DEF_STMT (name);
1688 /* If the SSA name is defined by a PHI node in this block,
1689 translate it. */
1690 if (gimple_code (def_stmt) == GIMPLE_PHI
1691 && gimple_bb (def_stmt) == phiblock)
1693 edge e = find_edge (pred, gimple_bb (def_stmt));
1694 tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1696 /* Handle constant. */
1697 if (is_gimple_min_invariant (def))
1698 return get_or_alloc_expr_for_constant (def);
1700 return get_or_alloc_expr_for_name (def);
1702 /* Otherwise return it unchanged - it will get removed if its
1703 value is not available in PREDs AVAIL_OUT set of expressions
1704 by the subtraction of TMP_GEN. */
1705 return expr;
1708 default:
1709 gcc_unreachable ();
1713 /* Wrapper around phi_translate_1 providing caching functionality. */
1715 static pre_expr
1716 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1717 basic_block pred, basic_block phiblock)
1719 expr_pred_trans_t slot = NULL;
1720 pre_expr phitrans;
1722 if (!expr)
1723 return NULL;
1725 /* Constants contain no values that need translation. */
1726 if (expr->kind == CONSTANT)
1727 return expr;
1729 if (value_id_constant_p (get_expr_value_id (expr)))
1730 return expr;
1732 /* Don't add translations of NAMEs as those are cheap to translate. */
1733 if (expr->kind != NAME)
1735 if (phi_trans_add (&slot, expr, pred))
1736 return slot->v;
1737 /* Store NULL for the value we want to return in the case of
1738 recursing. */
1739 slot->v = NULL;
1742 /* Translate. */
1743 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1745 if (slot)
1747 if (phitrans)
1748 slot->v = phitrans;
1749 else
1750 /* Remove failed translations again, they cause insert
1751 iteration to not pick up new opportunities reliably. */
1752 phi_translate_table->remove_elt_with_hash (slot, slot->hashcode);
1755 return phitrans;
1759 /* For each expression in SET, translate the values through phi nodes
1760 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1761 expressions in DEST. */
1763 static void
1764 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1765 basic_block phiblock)
1767 vec<pre_expr> exprs;
1768 pre_expr expr;
1769 int i;
1771 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1773 bitmap_set_copy (dest, set);
1774 return;
1777 exprs = sorted_array_from_bitmap_set (set);
1778 FOR_EACH_VEC_ELT (exprs, i, expr)
1780 pre_expr translated;
1781 translated = phi_translate (expr, set, NULL, pred, phiblock);
1782 if (!translated)
1783 continue;
1785 /* We might end up with multiple expressions from SET being
1786 translated to the same value. In this case we do not want
1787 to retain the NARY or REFERENCE expression but prefer a NAME
1788 which would be the leader. */
1789 if (translated->kind == NAME)
1790 bitmap_value_replace_in_set (dest, translated);
1791 else
1792 bitmap_value_insert_into_set (dest, translated);
1794 exprs.release ();
1797 /* Find the leader for a value (i.e., the name representing that
1798 value) in a given set, and return it. Return NULL if no leader
1799 is found. */
1801 static pre_expr
1802 bitmap_find_leader (bitmap_set_t set, unsigned int val)
1804 if (value_id_constant_p (val))
1806 unsigned int i;
1807 bitmap_iterator bi;
1808 bitmap exprset = value_expressions[val];
1810 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1812 pre_expr expr = expression_for_id (i);
1813 if (expr->kind == CONSTANT)
1814 return expr;
1817 if (bitmap_set_contains_value (set, val))
1819 /* Rather than walk the entire bitmap of expressions, and see
1820 whether any of them has the value we are looking for, we look
1821 at the reverse mapping, which tells us the set of expressions
1822 that have a given value (IE value->expressions with that
1823 value) and see if any of those expressions are in our set.
1824 The number of expressions per value is usually significantly
1825 less than the number of expressions in the set. In fact, for
1826 large testcases, doing it this way is roughly 5-10x faster
1827 than walking the bitmap.
1828 If this is somehow a significant lose for some cases, we can
1829 choose which set to walk based on which set is smaller. */
1830 unsigned int i;
1831 bitmap_iterator bi;
1832 bitmap exprset = value_expressions[val];
1834 EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1835 return expression_for_id (i);
1837 return NULL;
1840 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1841 BLOCK by seeing if it is not killed in the block. Note that we are
1842 only determining whether there is a store that kills it. Because
1843 of the order in which clean iterates over values, we are guaranteed
1844 that altered operands will have caused us to be eliminated from the
1845 ANTIC_IN set already. */
1847 static bool
1848 value_dies_in_block_x (pre_expr expr, basic_block block)
1850 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1851 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1852 gimple *def;
1853 gimple_stmt_iterator gsi;
1854 unsigned id = get_expression_id (expr);
1855 bool res = false;
1856 ao_ref ref;
1858 if (!vuse)
1859 return false;
1861 /* Lookup a previously calculated result. */
1862 if (EXPR_DIES (block)
1863 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1864 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1866 /* A memory expression {e, VUSE} dies in the block if there is a
1867 statement that may clobber e. If, starting statement walk from the
1868 top of the basic block, a statement uses VUSE there can be no kill
1869 inbetween that use and the original statement that loaded {e, VUSE},
1870 so we can stop walking. */
1871 ref.base = NULL_TREE;
1872 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1874 tree def_vuse, def_vdef;
1875 def = gsi_stmt (gsi);
1876 def_vuse = gimple_vuse (def);
1877 def_vdef = gimple_vdef (def);
1879 /* Not a memory statement. */
1880 if (!def_vuse)
1881 continue;
1883 /* Not a may-def. */
1884 if (!def_vdef)
1886 /* A load with the same VUSE, we're done. */
1887 if (def_vuse == vuse)
1888 break;
1890 continue;
1893 /* Init ref only if we really need it. */
1894 if (ref.base == NULL_TREE
1895 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1896 refx->operands))
1898 res = true;
1899 break;
1901 /* If the statement may clobber expr, it dies. */
1902 if (stmt_may_clobber_ref_p_1 (def, &ref))
1904 res = true;
1905 break;
1909 /* Remember the result. */
1910 if (!EXPR_DIES (block))
1911 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1912 bitmap_set_bit (EXPR_DIES (block), id * 2);
1913 if (res)
1914 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1916 return res;
1920 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1921 contains its value-id. */
1923 static bool
1924 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1926 if (op && TREE_CODE (op) == SSA_NAME)
1928 unsigned int value_id = VN_INFO (op)->value_id;
1929 if (!(bitmap_set_contains_value (set1, value_id)
1930 || (set2 && bitmap_set_contains_value (set2, value_id))))
1931 return false;
1933 return true;
1936 /* Determine if the expression EXPR is valid in SET1 U SET2.
1937 ONLY SET2 CAN BE NULL.
1938 This means that we have a leader for each part of the expression
1939 (if it consists of values), or the expression is an SSA_NAME.
1940 For loads/calls, we also see if the vuse is killed in this block. */
1942 static bool
1943 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr)
1945 switch (expr->kind)
1947 case NAME:
1948 /* By construction all NAMEs are available. Non-available
1949 NAMEs are removed by subtracting TMP_GEN from the sets. */
1950 return true;
1951 case NARY:
1953 unsigned int i;
1954 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1955 for (i = 0; i < nary->length; i++)
1956 if (!op_valid_in_sets (set1, set2, nary->op[i]))
1957 return false;
1958 return true;
1960 break;
1961 case REFERENCE:
1963 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1964 vn_reference_op_t vro;
1965 unsigned int i;
1967 FOR_EACH_VEC_ELT (ref->operands, i, vro)
1969 if (!op_valid_in_sets (set1, set2, vro->op0)
1970 || !op_valid_in_sets (set1, set2, vro->op1)
1971 || !op_valid_in_sets (set1, set2, vro->op2))
1972 return false;
1974 return true;
1976 default:
1977 gcc_unreachable ();
1981 /* Clean the set of expressions that are no longer valid in SET1 or
1982 SET2. This means expressions that are made up of values we have no
1983 leaders for in SET1 or SET2. This version is used for partial
1984 anticipation, which means it is not valid in either ANTIC_IN or
1985 PA_IN. */
1987 static void
1988 dependent_clean (bitmap_set_t set1, bitmap_set_t set2)
1990 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
1991 pre_expr expr;
1992 int i;
1994 FOR_EACH_VEC_ELT (exprs, i, expr)
1996 if (!valid_in_sets (set1, set2, expr))
1997 bitmap_remove_from_set (set1, expr);
1999 exprs.release ();
2002 /* Clean the set of expressions that are no longer valid in SET. This
2003 means expressions that are made up of values we have no leaders for
2004 in SET. */
2006 static void
2007 clean (bitmap_set_t set)
2009 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set);
2010 pre_expr expr;
2011 int i;
2013 FOR_EACH_VEC_ELT (exprs, i, expr)
2015 if (!valid_in_sets (set, NULL, expr))
2016 bitmap_remove_from_set (set, expr);
2018 exprs.release ();
2021 /* Clean the set of expressions that are no longer valid in SET because
2022 they are clobbered in BLOCK or because they trap and may not be executed. */
2024 static void
2025 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2027 bitmap_iterator bi;
2028 unsigned i;
2030 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2032 pre_expr expr = expression_for_id (i);
2033 if (expr->kind == REFERENCE)
2035 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2036 if (ref->vuse)
2038 gimple *def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2039 if (!gimple_nop_p (def_stmt)
2040 && ((gimple_bb (def_stmt) != block
2041 && !dominated_by_p (CDI_DOMINATORS,
2042 block, gimple_bb (def_stmt)))
2043 || (gimple_bb (def_stmt) == block
2044 && value_dies_in_block_x (expr, block))))
2045 bitmap_remove_from_set (set, expr);
2048 else if (expr->kind == NARY)
2050 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2051 /* If the NARY may trap make sure the block does not contain
2052 a possible exit point.
2053 ??? This is overly conservative if we translate AVAIL_OUT
2054 as the available expression might be after the exit point. */
2055 if (BB_MAY_NOTRETURN (block)
2056 && vn_nary_may_trap (nary))
2057 bitmap_remove_from_set (set, expr);
2062 static sbitmap has_abnormal_preds;
2064 /* List of blocks that may have changed during ANTIC computation and
2065 thus need to be iterated over. */
2067 static sbitmap changed_blocks;
2069 /* Compute the ANTIC set for BLOCK.
2071 If succs(BLOCK) > 1 then
2072 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2073 else if succs(BLOCK) == 1 then
2074 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2076 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2079 static bool
2080 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2082 bool changed = false;
2083 bitmap_set_t S, old, ANTIC_OUT;
2084 bitmap_iterator bi;
2085 unsigned int bii;
2086 edge e;
2087 edge_iterator ei;
2089 old = ANTIC_OUT = S = NULL;
2090 BB_VISITED (block) = 1;
2092 /* If any edges from predecessors are abnormal, antic_in is empty,
2093 so do nothing. */
2094 if (block_has_abnormal_pred_edge)
2095 goto maybe_dump_sets;
2097 old = ANTIC_IN (block);
2098 ANTIC_OUT = bitmap_set_new ();
2100 /* If the block has no successors, ANTIC_OUT is empty. */
2101 if (EDGE_COUNT (block->succs) == 0)
2103 /* If we have one successor, we could have some phi nodes to
2104 translate through. */
2105 else if (single_succ_p (block))
2107 basic_block succ_bb = single_succ (block);
2108 gcc_assert (BB_VISITED (succ_bb));
2109 phi_translate_set (ANTIC_OUT, ANTIC_IN (succ_bb), block, succ_bb);
2111 /* If we have multiple successors, we take the intersection of all of
2112 them. Note that in the case of loop exit phi nodes, we may have
2113 phis to translate through. */
2114 else
2116 size_t i;
2117 basic_block bprime, first = NULL;
2119 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2120 FOR_EACH_EDGE (e, ei, block->succs)
2122 if (!first
2123 && BB_VISITED (e->dest))
2124 first = e->dest;
2125 else if (BB_VISITED (e->dest))
2126 worklist.quick_push (e->dest);
2129 /* Of multiple successors we have to have visited one already
2130 which is guaranteed by iteration order. */
2131 gcc_assert (first != NULL);
2133 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2135 FOR_EACH_VEC_ELT (worklist, i, bprime)
2137 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2139 bitmap_set_t tmp = bitmap_set_new ();
2140 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2141 bitmap_set_and (ANTIC_OUT, tmp);
2142 bitmap_set_free (tmp);
2144 else
2145 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2149 /* Prune expressions that are clobbered in block and thus become
2150 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2151 prune_clobbered_mems (ANTIC_OUT, block);
2153 /* Generate ANTIC_OUT - TMP_GEN. */
2154 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2156 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2157 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2158 TMP_GEN (block));
2160 /* Then union in the ANTIC_OUT - TMP_GEN values,
2161 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2162 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2163 bitmap_value_insert_into_set (ANTIC_IN (block),
2164 expression_for_id (bii));
2166 clean (ANTIC_IN (block));
2168 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2170 changed = true;
2171 bitmap_set_bit (changed_blocks, block->index);
2172 FOR_EACH_EDGE (e, ei, block->preds)
2173 bitmap_set_bit (changed_blocks, e->src->index);
2175 else
2176 bitmap_clear_bit (changed_blocks, block->index);
2178 maybe_dump_sets:
2179 if (dump_file && (dump_flags & TDF_DETAILS))
2181 if (ANTIC_OUT)
2182 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2184 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2185 block->index);
2187 if (S)
2188 print_bitmap_set (dump_file, S, "S", block->index);
2190 if (old)
2191 bitmap_set_free (old);
2192 if (S)
2193 bitmap_set_free (S);
2194 if (ANTIC_OUT)
2195 bitmap_set_free (ANTIC_OUT);
2196 return changed;
2199 /* Compute PARTIAL_ANTIC for BLOCK.
2201 If succs(BLOCK) > 1 then
2202 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2203 in ANTIC_OUT for all succ(BLOCK)
2204 else if succs(BLOCK) == 1 then
2205 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2207 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2208 - ANTIC_IN[BLOCK])
2211 static bool
2212 compute_partial_antic_aux (basic_block block,
2213 bool block_has_abnormal_pred_edge)
2215 bool changed = false;
2216 bitmap_set_t old_PA_IN;
2217 bitmap_set_t PA_OUT;
2218 edge e;
2219 edge_iterator ei;
2220 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2222 old_PA_IN = PA_OUT = NULL;
2224 /* If any edges from predecessors are abnormal, antic_in is empty,
2225 so do nothing. */
2226 if (block_has_abnormal_pred_edge)
2227 goto maybe_dump_sets;
2229 /* If there are too many partially anticipatable values in the
2230 block, phi_translate_set can take an exponential time: stop
2231 before the translation starts. */
2232 if (max_pa
2233 && single_succ_p (block)
2234 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2235 goto maybe_dump_sets;
2237 old_PA_IN = PA_IN (block);
2238 PA_OUT = bitmap_set_new ();
2240 /* If the block has no successors, ANTIC_OUT is empty. */
2241 if (EDGE_COUNT (block->succs) == 0)
2243 /* If we have one successor, we could have some phi nodes to
2244 translate through. Note that we can't phi translate across DFS
2245 back edges in partial antic, because it uses a union operation on
2246 the successors. For recurrences like IV's, we will end up
2247 generating a new value in the set on each go around (i + 3 (VH.1)
2248 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2249 else if (single_succ_p (block))
2251 basic_block succ = single_succ (block);
2252 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2253 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2255 /* If we have multiple successors, we take the union of all of
2256 them. */
2257 else
2259 size_t i;
2260 basic_block bprime;
2262 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2263 FOR_EACH_EDGE (e, ei, block->succs)
2265 if (e->flags & EDGE_DFS_BACK)
2266 continue;
2267 worklist.quick_push (e->dest);
2269 if (worklist.length () > 0)
2271 FOR_EACH_VEC_ELT (worklist, i, bprime)
2273 unsigned int i;
2274 bitmap_iterator bi;
2276 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2277 bitmap_value_insert_into_set (PA_OUT,
2278 expression_for_id (i));
2279 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2281 bitmap_set_t pa_in = bitmap_set_new ();
2282 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2283 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2284 bitmap_value_insert_into_set (PA_OUT,
2285 expression_for_id (i));
2286 bitmap_set_free (pa_in);
2288 else
2289 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2290 bitmap_value_insert_into_set (PA_OUT,
2291 expression_for_id (i));
2296 /* Prune expressions that are clobbered in block and thus become
2297 invalid if translated from PA_OUT to PA_IN. */
2298 prune_clobbered_mems (PA_OUT, block);
2300 /* PA_IN starts with PA_OUT - TMP_GEN.
2301 Then we subtract things from ANTIC_IN. */
2302 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2304 /* For partial antic, we want to put back in the phi results, since
2305 we will properly avoid making them partially antic over backedges. */
2306 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2307 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2309 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2310 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2312 dependent_clean (PA_IN (block), ANTIC_IN (block));
2314 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2316 changed = true;
2317 bitmap_set_bit (changed_blocks, block->index);
2318 FOR_EACH_EDGE (e, ei, block->preds)
2319 bitmap_set_bit (changed_blocks, e->src->index);
2321 else
2322 bitmap_clear_bit (changed_blocks, block->index);
2324 maybe_dump_sets:
2325 if (dump_file && (dump_flags & TDF_DETAILS))
2327 if (PA_OUT)
2328 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2330 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2332 if (old_PA_IN)
2333 bitmap_set_free (old_PA_IN);
2334 if (PA_OUT)
2335 bitmap_set_free (PA_OUT);
2336 return changed;
2339 /* Compute ANTIC and partial ANTIC sets. */
2341 static void
2342 compute_antic (void)
2344 bool changed = true;
2345 int num_iterations = 0;
2346 basic_block block;
2347 int i;
2349 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2350 We pre-build the map of blocks with incoming abnormal edges here. */
2351 has_abnormal_preds = sbitmap_alloc (last_basic_block_for_fn (cfun));
2352 bitmap_clear (has_abnormal_preds);
2354 FOR_ALL_BB_FN (block, cfun)
2356 edge_iterator ei;
2357 edge e;
2359 FOR_EACH_EDGE (e, ei, block->preds)
2361 e->flags &= ~EDGE_DFS_BACK;
2362 if (e->flags & EDGE_ABNORMAL)
2364 bitmap_set_bit (has_abnormal_preds, block->index);
2365 break;
2369 BB_VISITED (block) = 0;
2371 /* While we are here, give empty ANTIC_IN sets to each block. */
2372 ANTIC_IN (block) = bitmap_set_new ();
2373 PA_IN (block) = bitmap_set_new ();
2376 /* At the exit block we anticipate nothing. */
2377 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun)) = 1;
2379 changed_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun) + 1);
2380 bitmap_ones (changed_blocks);
2381 while (changed)
2383 if (dump_file && (dump_flags & TDF_DETAILS))
2384 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2385 /* ??? We need to clear our PHI translation cache here as the
2386 ANTIC sets shrink and we restrict valid translations to
2387 those having operands with leaders in ANTIC. Same below
2388 for PA ANTIC computation. */
2389 num_iterations++;
2390 changed = false;
2391 for (i = postorder_num - 1; i >= 0; i--)
2393 if (bitmap_bit_p (changed_blocks, postorder[i]))
2395 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2396 changed |= compute_antic_aux (block,
2397 bitmap_bit_p (has_abnormal_preds,
2398 block->index));
2401 /* Theoretically possible, but *highly* unlikely. */
2402 gcc_checking_assert (num_iterations < 500);
2405 statistics_histogram_event (cfun, "compute_antic iterations",
2406 num_iterations);
2408 if (do_partial_partial)
2410 bitmap_ones (changed_blocks);
2411 mark_dfs_back_edges ();
2412 num_iterations = 0;
2413 changed = true;
2414 while (changed)
2416 if (dump_file && (dump_flags & TDF_DETAILS))
2417 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2418 num_iterations++;
2419 changed = false;
2420 for (i = postorder_num - 1 ; i >= 0; i--)
2422 if (bitmap_bit_p (changed_blocks, postorder[i]))
2424 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2425 changed
2426 |= compute_partial_antic_aux (block,
2427 bitmap_bit_p (has_abnormal_preds,
2428 block->index));
2431 /* Theoretically possible, but *highly* unlikely. */
2432 gcc_checking_assert (num_iterations < 500);
2434 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2435 num_iterations);
2437 sbitmap_free (has_abnormal_preds);
2438 sbitmap_free (changed_blocks);
2442 /* Inserted expressions are placed onto this worklist, which is used
2443 for performing quick dead code elimination of insertions we made
2444 that didn't turn out to be necessary. */
2445 static bitmap inserted_exprs;
2447 /* The actual worker for create_component_ref_by_pieces. */
2449 static tree
2450 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2451 unsigned int *operand, gimple_seq *stmts)
2453 vn_reference_op_t currop = &ref->operands[*operand];
2454 tree genop;
2455 ++*operand;
2456 switch (currop->opcode)
2458 case CALL_EXPR:
2459 gcc_unreachable ();
2461 case MEM_REF:
2463 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2464 stmts);
2465 if (!baseop)
2466 return NULL_TREE;
2467 tree offset = currop->op0;
2468 if (TREE_CODE (baseop) == ADDR_EXPR
2469 && handled_component_p (TREE_OPERAND (baseop, 0)))
2471 HOST_WIDE_INT off;
2472 tree base;
2473 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2474 &off);
2475 gcc_assert (base);
2476 offset = int_const_binop (PLUS_EXPR, offset,
2477 build_int_cst (TREE_TYPE (offset),
2478 off));
2479 baseop = build_fold_addr_expr (base);
2481 genop = build2 (MEM_REF, currop->type, baseop, offset);
2482 MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2483 MR_DEPENDENCE_BASE (genop) = currop->base;
2484 REF_REVERSE_STORAGE_ORDER (genop) = currop->reverse;
2485 return genop;
2488 case TARGET_MEM_REF:
2490 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2491 vn_reference_op_t nextop = &ref->operands[++*operand];
2492 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2493 stmts);
2494 if (!baseop)
2495 return NULL_TREE;
2496 if (currop->op0)
2498 genop0 = find_or_generate_expression (block, currop->op0, stmts);
2499 if (!genop0)
2500 return NULL_TREE;
2502 if (nextop->op0)
2504 genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2505 if (!genop1)
2506 return NULL_TREE;
2508 genop = build5 (TARGET_MEM_REF, currop->type,
2509 baseop, currop->op2, genop0, currop->op1, genop1);
2511 MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2512 MR_DEPENDENCE_BASE (genop) = currop->base;
2513 return genop;
2516 case ADDR_EXPR:
2517 if (currop->op0)
2519 gcc_assert (is_gimple_min_invariant (currop->op0));
2520 return currop->op0;
2522 /* Fallthrough. */
2523 case REALPART_EXPR:
2524 case IMAGPART_EXPR:
2525 case VIEW_CONVERT_EXPR:
2527 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2528 stmts);
2529 if (!genop0)
2530 return NULL_TREE;
2531 return fold_build1 (currop->opcode, currop->type, genop0);
2534 case WITH_SIZE_EXPR:
2536 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2537 stmts);
2538 if (!genop0)
2539 return NULL_TREE;
2540 tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2541 if (!genop1)
2542 return NULL_TREE;
2543 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2546 case BIT_FIELD_REF:
2548 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2549 stmts);
2550 if (!genop0)
2551 return NULL_TREE;
2552 tree op1 = currop->op0;
2553 tree op2 = currop->op1;
2554 tree t = build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2555 REF_REVERSE_STORAGE_ORDER (t) = currop->reverse;
2556 return fold (t);
2559 /* For array ref vn_reference_op's, operand 1 of the array ref
2560 is op0 of the reference op and operand 3 of the array ref is
2561 op1. */
2562 case ARRAY_RANGE_REF:
2563 case ARRAY_REF:
2565 tree genop0;
2566 tree genop1 = currop->op0;
2567 tree genop2 = currop->op1;
2568 tree genop3 = currop->op2;
2569 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2570 stmts);
2571 if (!genop0)
2572 return NULL_TREE;
2573 genop1 = find_or_generate_expression (block, genop1, stmts);
2574 if (!genop1)
2575 return NULL_TREE;
2576 if (genop2)
2578 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2579 /* Drop zero minimum index if redundant. */
2580 if (integer_zerop (genop2)
2581 && (!domain_type
2582 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2583 genop2 = NULL_TREE;
2584 else
2586 genop2 = find_or_generate_expression (block, genop2, stmts);
2587 if (!genop2)
2588 return NULL_TREE;
2591 if (genop3)
2593 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2594 /* We can't always put a size in units of the element alignment
2595 here as the element alignment may be not visible. See
2596 PR43783. Simply drop the element size for constant
2597 sizes. */
2598 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2599 genop3 = NULL_TREE;
2600 else
2602 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2603 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2604 genop3 = find_or_generate_expression (block, genop3, stmts);
2605 if (!genop3)
2606 return NULL_TREE;
2609 return build4 (currop->opcode, currop->type, genop0, genop1,
2610 genop2, genop3);
2612 case COMPONENT_REF:
2614 tree op0;
2615 tree op1;
2616 tree genop2 = currop->op1;
2617 op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2618 if (!op0)
2619 return NULL_TREE;
2620 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2621 op1 = currop->op0;
2622 if (genop2)
2624 genop2 = find_or_generate_expression (block, genop2, stmts);
2625 if (!genop2)
2626 return NULL_TREE;
2628 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2631 case SSA_NAME:
2633 genop = find_or_generate_expression (block, currop->op0, stmts);
2634 return genop;
2636 case STRING_CST:
2637 case INTEGER_CST:
2638 case COMPLEX_CST:
2639 case VECTOR_CST:
2640 case REAL_CST:
2641 case CONSTRUCTOR:
2642 case VAR_DECL:
2643 case PARM_DECL:
2644 case CONST_DECL:
2645 case RESULT_DECL:
2646 case FUNCTION_DECL:
2647 return currop->op0;
2649 default:
2650 gcc_unreachable ();
2654 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2655 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2656 trying to rename aggregates into ssa form directly, which is a no no.
2658 Thus, this routine doesn't create temporaries, it just builds a
2659 single access expression for the array, calling
2660 find_or_generate_expression to build the innermost pieces.
2662 This function is a subroutine of create_expression_by_pieces, and
2663 should not be called on it's own unless you really know what you
2664 are doing. */
2666 static tree
2667 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2668 gimple_seq *stmts)
2670 unsigned int op = 0;
2671 return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2674 /* Find a simple leader for an expression, or generate one using
2675 create_expression_by_pieces from a NARY expression for the value.
2676 BLOCK is the basic_block we are looking for leaders in.
2677 OP is the tree expression to find a leader for or generate.
2678 Returns the leader or NULL_TREE on failure. */
2680 static tree
2681 find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2683 pre_expr expr = get_or_alloc_expr_for (op);
2684 unsigned int lookfor = get_expr_value_id (expr);
2685 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2686 if (leader)
2688 if (leader->kind == NAME)
2689 return PRE_EXPR_NAME (leader);
2690 else if (leader->kind == CONSTANT)
2691 return PRE_EXPR_CONSTANT (leader);
2693 /* Defer. */
2694 return NULL_TREE;
2697 /* It must be a complex expression, so generate it recursively. Note
2698 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2699 where the insert algorithm fails to insert a required expression. */
2700 bitmap exprset = value_expressions[lookfor];
2701 bitmap_iterator bi;
2702 unsigned int i;
2703 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2705 pre_expr temp = expression_for_id (i);
2706 /* We cannot insert random REFERENCE expressions at arbitrary
2707 places. We can insert NARYs which eventually re-materializes
2708 its operand values. */
2709 if (temp->kind == NARY)
2710 return create_expression_by_pieces (block, temp, stmts,
2711 get_expr_type (expr));
2714 /* Defer. */
2715 return NULL_TREE;
2718 #define NECESSARY GF_PLF_1
2720 /* Create an expression in pieces, so that we can handle very complex
2721 expressions that may be ANTIC, but not necessary GIMPLE.
2722 BLOCK is the basic block the expression will be inserted into,
2723 EXPR is the expression to insert (in value form)
2724 STMTS is a statement list to append the necessary insertions into.
2726 This function will die if we hit some value that shouldn't be
2727 ANTIC but is (IE there is no leader for it, or its components).
2728 The function returns NULL_TREE in case a different antic expression
2729 has to be inserted first.
2730 This function may also generate expressions that are themselves
2731 partially or fully redundant. Those that are will be either made
2732 fully redundant during the next iteration of insert (for partially
2733 redundant ones), or eliminated by eliminate (for fully redundant
2734 ones). */
2736 static tree
2737 create_expression_by_pieces (basic_block block, pre_expr expr,
2738 gimple_seq *stmts, tree type)
2740 tree name;
2741 tree folded;
2742 gimple_seq forced_stmts = NULL;
2743 unsigned int value_id;
2744 gimple_stmt_iterator gsi;
2745 tree exprtype = type ? type : get_expr_type (expr);
2746 pre_expr nameexpr;
2747 gassign *newstmt;
2749 switch (expr->kind)
2751 /* We may hit the NAME/CONSTANT case if we have to convert types
2752 that value numbering saw through. */
2753 case NAME:
2754 folded = PRE_EXPR_NAME (expr);
2755 if (useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2756 return folded;
2757 break;
2758 case CONSTANT:
2760 folded = PRE_EXPR_CONSTANT (expr);
2761 tree tem = fold_convert (exprtype, folded);
2762 if (is_gimple_min_invariant (tem))
2763 return tem;
2764 break;
2766 case REFERENCE:
2767 if (PRE_EXPR_REFERENCE (expr)->operands[0].opcode == CALL_EXPR)
2769 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2770 unsigned int operand = 1;
2771 vn_reference_op_t currop = &ref->operands[0];
2772 tree sc = NULL_TREE;
2773 tree fn;
2774 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2775 fn = currop->op0;
2776 else
2777 fn = find_or_generate_expression (block, currop->op0, stmts);
2778 if (!fn)
2779 return NULL_TREE;
2780 if (currop->op1)
2782 sc = find_or_generate_expression (block, currop->op1, stmts);
2783 if (!sc)
2784 return NULL_TREE;
2786 auto_vec<tree> args (ref->operands.length () - 1);
2787 while (operand < ref->operands.length ())
2789 tree arg = create_component_ref_by_pieces_1 (block, ref,
2790 &operand, stmts);
2791 if (!arg)
2792 return NULL_TREE;
2793 args.quick_push (arg);
2795 gcall *call
2796 = gimple_build_call_vec ((TREE_CODE (fn) == FUNCTION_DECL
2797 ? build_fold_addr_expr (fn) : fn), args);
2798 gimple_call_set_with_bounds (call, currop->with_bounds);
2799 if (sc)
2800 gimple_call_set_chain (call, sc);
2801 tree forcedname = make_ssa_name (currop->type);
2802 gimple_call_set_lhs (call, forcedname);
2803 gimple_set_vuse (call, BB_LIVE_VOP_ON_EXIT (block));
2804 gimple_seq_add_stmt_without_update (&forced_stmts, call);
2805 folded = forcedname;
2807 else
2809 folded = create_component_ref_by_pieces (block,
2810 PRE_EXPR_REFERENCE (expr),
2811 stmts);
2812 if (!folded)
2813 return NULL_TREE;
2814 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2815 newstmt = gimple_build_assign (name, folded);
2816 gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2817 gimple_set_vuse (newstmt, BB_LIVE_VOP_ON_EXIT (block));
2818 folded = name;
2820 break;
2821 case NARY:
2823 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2824 tree *genop = XALLOCAVEC (tree, nary->length);
2825 unsigned i;
2826 for (i = 0; i < nary->length; ++i)
2828 genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2829 if (!genop[i])
2830 return NULL_TREE;
2831 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2832 may have conversions stripped. */
2833 if (nary->opcode == POINTER_PLUS_EXPR)
2835 if (i == 0)
2836 genop[i] = gimple_convert (&forced_stmts,
2837 nary->type, genop[i]);
2838 else if (i == 1)
2839 genop[i] = gimple_convert (&forced_stmts,
2840 sizetype, genop[i]);
2842 else
2843 genop[i] = gimple_convert (&forced_stmts,
2844 TREE_TYPE (nary->op[i]), genop[i]);
2846 if (nary->opcode == CONSTRUCTOR)
2848 vec<constructor_elt, va_gc> *elts = NULL;
2849 for (i = 0; i < nary->length; ++i)
2850 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2851 folded = build_constructor (nary->type, elts);
2852 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2853 newstmt = gimple_build_assign (name, folded);
2854 gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2855 folded = name;
2857 else
2859 switch (nary->length)
2861 case 1:
2862 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2863 genop[0]);
2864 break;
2865 case 2:
2866 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2867 genop[0], genop[1]);
2868 break;
2869 case 3:
2870 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2871 genop[0], genop[1], genop[2]);
2872 break;
2873 default:
2874 gcc_unreachable ();
2878 break;
2879 default:
2880 gcc_unreachable ();
2883 folded = gimple_convert (&forced_stmts, exprtype, folded);
2885 /* If there is nothing to insert, return the simplified result. */
2886 if (gimple_seq_empty_p (forced_stmts))
2887 return folded;
2888 /* If we simplified to a constant return it and discard eventually
2889 built stmts. */
2890 if (is_gimple_min_invariant (folded))
2892 gimple_seq_discard (forced_stmts);
2893 return folded;
2896 gcc_assert (TREE_CODE (folded) == SSA_NAME);
2898 /* If we have any intermediate expressions to the value sets, add them
2899 to the value sets and chain them in the instruction stream. */
2900 if (forced_stmts)
2902 gsi = gsi_start (forced_stmts);
2903 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2905 gimple *stmt = gsi_stmt (gsi);
2906 tree forcedname = gimple_get_lhs (stmt);
2907 pre_expr nameexpr;
2909 if (forcedname != folded)
2911 VN_INFO_GET (forcedname)->valnum = forcedname;
2912 VN_INFO (forcedname)->value_id = get_next_value_id ();
2913 nameexpr = get_or_alloc_expr_for_name (forcedname);
2914 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2915 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2916 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2919 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
2920 gimple_set_plf (stmt, NECESSARY, false);
2922 gimple_seq_add_seq (stmts, forced_stmts);
2925 name = folded;
2927 /* Fold the last statement. */
2928 gsi = gsi_last (*stmts);
2929 if (fold_stmt_inplace (&gsi))
2930 update_stmt (gsi_stmt (gsi));
2932 /* Add a value number to the temporary.
2933 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2934 we are creating the expression by pieces, and this particular piece of
2935 the expression may have been represented. There is no harm in replacing
2936 here. */
2937 value_id = get_expr_value_id (expr);
2938 VN_INFO_GET (name)->value_id = value_id;
2939 VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id);
2940 if (VN_INFO (name)->valnum == NULL_TREE)
2941 VN_INFO (name)->valnum = name;
2942 gcc_assert (VN_INFO (name)->valnum != NULL_TREE);
2943 nameexpr = get_or_alloc_expr_for_name (name);
2944 add_to_value (value_id, nameexpr);
2945 if (NEW_SETS (block))
2946 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2947 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2949 pre_stats.insertions++;
2950 if (dump_file && (dump_flags & TDF_DETAILS))
2952 fprintf (dump_file, "Inserted ");
2953 print_gimple_stmt (dump_file, gsi_stmt (gsi_last (*stmts)), 0, 0);
2954 fprintf (dump_file, " in predecessor %d (%04d)\n",
2955 block->index, value_id);
2958 return name;
2962 /* Insert the to-be-made-available values of expression EXPRNUM for each
2963 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
2964 merge the result with a phi node, given the same value number as
2965 NODE. Return true if we have inserted new stuff. */
2967 static bool
2968 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
2969 vec<pre_expr> avail)
2971 pre_expr expr = expression_for_id (exprnum);
2972 pre_expr newphi;
2973 unsigned int val = get_expr_value_id (expr);
2974 edge pred;
2975 bool insertions = false;
2976 bool nophi = false;
2977 basic_block bprime;
2978 pre_expr eprime;
2979 edge_iterator ei;
2980 tree type = get_expr_type (expr);
2981 tree temp;
2982 gphi *phi;
2984 /* Make sure we aren't creating an induction variable. */
2985 if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
2987 bool firstinsideloop = false;
2988 bool secondinsideloop = false;
2989 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
2990 EDGE_PRED (block, 0)->src);
2991 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
2992 EDGE_PRED (block, 1)->src);
2993 /* Induction variables only have one edge inside the loop. */
2994 if ((firstinsideloop ^ secondinsideloop)
2995 && expr->kind != REFERENCE)
2997 if (dump_file && (dump_flags & TDF_DETAILS))
2998 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
2999 nophi = true;
3003 /* Make the necessary insertions. */
3004 FOR_EACH_EDGE (pred, ei, block->preds)
3006 gimple_seq stmts = NULL;
3007 tree builtexpr;
3008 bprime = pred->src;
3009 eprime = avail[pred->dest_idx];
3010 builtexpr = create_expression_by_pieces (bprime, eprime,
3011 &stmts, type);
3012 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3013 if (!gimple_seq_empty_p (stmts))
3015 gsi_insert_seq_on_edge (pred, stmts);
3016 insertions = true;
3018 if (!builtexpr)
3020 /* We cannot insert a PHI node if we failed to insert
3021 on one edge. */
3022 nophi = true;
3023 continue;
3025 if (is_gimple_min_invariant (builtexpr))
3026 avail[pred->dest_idx] = get_or_alloc_expr_for_constant (builtexpr);
3027 else
3028 avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3030 /* If we didn't want a phi node, and we made insertions, we still have
3031 inserted new stuff, and thus return true. If we didn't want a phi node,
3032 and didn't make insertions, we haven't added anything new, so return
3033 false. */
3034 if (nophi && insertions)
3035 return true;
3036 else if (nophi && !insertions)
3037 return false;
3039 /* Now build a phi for the new variable. */
3040 temp = make_temp_ssa_name (type, NULL, "prephitmp");
3041 phi = create_phi_node (temp, block);
3043 gimple_set_plf (phi, NECESSARY, false);
3044 VN_INFO_GET (temp)->value_id = val;
3045 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3046 if (VN_INFO (temp)->valnum == NULL_TREE)
3047 VN_INFO (temp)->valnum = temp;
3048 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3049 FOR_EACH_EDGE (pred, ei, block->preds)
3051 pre_expr ae = avail[pred->dest_idx];
3052 gcc_assert (get_expr_type (ae) == type
3053 || useless_type_conversion_p (type, get_expr_type (ae)));
3054 if (ae->kind == CONSTANT)
3055 add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3056 pred, UNKNOWN_LOCATION);
3057 else
3058 add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3061 newphi = get_or_alloc_expr_for_name (temp);
3062 add_to_value (val, newphi);
3064 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3065 this insertion, since we test for the existence of this value in PHI_GEN
3066 before proceeding with the partial redundancy checks in insert_aux.
3068 The value may exist in AVAIL_OUT, in particular, it could be represented
3069 by the expression we are trying to eliminate, in which case we want the
3070 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3071 inserted there.
3073 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3074 this block, because if it did, it would have existed in our dominator's
3075 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3078 bitmap_insert_into_set (PHI_GEN (block), newphi);
3079 bitmap_value_replace_in_set (AVAIL_OUT (block),
3080 newphi);
3081 bitmap_insert_into_set (NEW_SETS (block),
3082 newphi);
3084 /* If we insert a PHI node for a conversion of another PHI node
3085 in the same basic-block try to preserve range information.
3086 This is important so that followup loop passes receive optimal
3087 number of iteration analysis results. See PR61743. */
3088 if (expr->kind == NARY
3089 && CONVERT_EXPR_CODE_P (expr->u.nary->opcode)
3090 && TREE_CODE (expr->u.nary->op[0]) == SSA_NAME
3091 && gimple_bb (SSA_NAME_DEF_STMT (expr->u.nary->op[0])) == block
3092 && INTEGRAL_TYPE_P (type)
3093 && INTEGRAL_TYPE_P (TREE_TYPE (expr->u.nary->op[0]))
3094 && (TYPE_PRECISION (type)
3095 >= TYPE_PRECISION (TREE_TYPE (expr->u.nary->op[0])))
3096 && SSA_NAME_RANGE_INFO (expr->u.nary->op[0]))
3098 wide_int min, max;
3099 if (get_range_info (expr->u.nary->op[0], &min, &max) == VR_RANGE
3100 && !wi::neg_p (min, SIGNED)
3101 && !wi::neg_p (max, SIGNED))
3102 /* Just handle extension and sign-changes of all-positive ranges. */
3103 set_range_info (temp,
3104 SSA_NAME_RANGE_TYPE (expr->u.nary->op[0]),
3105 wide_int_storage::from (min, TYPE_PRECISION (type),
3106 TYPE_SIGN (type)),
3107 wide_int_storage::from (max, TYPE_PRECISION (type),
3108 TYPE_SIGN (type)));
3111 if (dump_file && (dump_flags & TDF_DETAILS))
3113 fprintf (dump_file, "Created phi ");
3114 print_gimple_stmt (dump_file, phi, 0, 0);
3115 fprintf (dump_file, " in block %d (%04d)\n", block->index, val);
3117 pre_stats.phis++;
3118 return true;
3123 /* Perform insertion of partially redundant values.
3124 For BLOCK, do the following:
3125 1. Propagate the NEW_SETS of the dominator into the current block.
3126 If the block has multiple predecessors,
3127 2a. Iterate over the ANTIC expressions for the block to see if
3128 any of them are partially redundant.
3129 2b. If so, insert them into the necessary predecessors to make
3130 the expression fully redundant.
3131 2c. Insert a new PHI merging the values of the predecessors.
3132 2d. Insert the new PHI, and the new expressions, into the
3133 NEW_SETS set.
3134 3. Recursively call ourselves on the dominator children of BLOCK.
3136 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3137 do_regular_insertion and do_partial_insertion.
3141 static bool
3142 do_regular_insertion (basic_block block, basic_block dom)
3144 bool new_stuff = false;
3145 vec<pre_expr> exprs;
3146 pre_expr expr;
3147 auto_vec<pre_expr> avail;
3148 int i;
3150 exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3151 avail.safe_grow (EDGE_COUNT (block->preds));
3153 FOR_EACH_VEC_ELT (exprs, i, expr)
3155 if (expr->kind == NARY
3156 || expr->kind == REFERENCE)
3158 unsigned int val;
3159 bool by_some = false;
3160 bool cant_insert = false;
3161 bool all_same = true;
3162 pre_expr first_s = NULL;
3163 edge pred;
3164 basic_block bprime;
3165 pre_expr eprime = NULL;
3166 edge_iterator ei;
3167 pre_expr edoubleprime = NULL;
3168 bool do_insertion = false;
3170 val = get_expr_value_id (expr);
3171 if (bitmap_set_contains_value (PHI_GEN (block), val))
3172 continue;
3173 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3175 if (dump_file && (dump_flags & TDF_DETAILS))
3177 fprintf (dump_file, "Found fully redundant value: ");
3178 print_pre_expr (dump_file, expr);
3179 fprintf (dump_file, "\n");
3181 continue;
3184 FOR_EACH_EDGE (pred, ei, block->preds)
3186 unsigned int vprime;
3188 /* We should never run insertion for the exit block
3189 and so not come across fake pred edges. */
3190 gcc_assert (!(pred->flags & EDGE_FAKE));
3191 bprime = pred->src;
3192 /* We are looking at ANTIC_OUT of bprime. */
3193 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3194 bprime, block);
3196 /* eprime will generally only be NULL if the
3197 value of the expression, translated
3198 through the PHI for this predecessor, is
3199 undefined. If that is the case, we can't
3200 make the expression fully redundant,
3201 because its value is undefined along a
3202 predecessor path. We can thus break out
3203 early because it doesn't matter what the
3204 rest of the results are. */
3205 if (eprime == NULL)
3207 avail[pred->dest_idx] = NULL;
3208 cant_insert = true;
3209 break;
3212 eprime = fully_constant_expression (eprime);
3213 vprime = get_expr_value_id (eprime);
3214 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3215 vprime);
3216 if (edoubleprime == NULL)
3218 avail[pred->dest_idx] = eprime;
3219 all_same = false;
3221 else
3223 avail[pred->dest_idx] = edoubleprime;
3224 by_some = true;
3225 /* We want to perform insertions to remove a redundancy on
3226 a path in the CFG we want to optimize for speed. */
3227 if (optimize_edge_for_speed_p (pred))
3228 do_insertion = true;
3229 if (first_s == NULL)
3230 first_s = edoubleprime;
3231 else if (!pre_expr_d::equal (first_s, edoubleprime))
3232 all_same = false;
3235 /* If we can insert it, it's not the same value
3236 already existing along every predecessor, and
3237 it's defined by some predecessor, it is
3238 partially redundant. */
3239 if (!cant_insert && !all_same && by_some)
3241 if (!do_insertion)
3243 if (dump_file && (dump_flags & TDF_DETAILS))
3245 fprintf (dump_file, "Skipping partial redundancy for "
3246 "expression ");
3247 print_pre_expr (dump_file, expr);
3248 fprintf (dump_file, " (%04d), no redundancy on to be "
3249 "optimized for speed edge\n", val);
3252 else if (dbg_cnt (treepre_insert))
3254 if (dump_file && (dump_flags & TDF_DETAILS))
3256 fprintf (dump_file, "Found partial redundancy for "
3257 "expression ");
3258 print_pre_expr (dump_file, expr);
3259 fprintf (dump_file, " (%04d)\n",
3260 get_expr_value_id (expr));
3262 if (insert_into_preds_of_block (block,
3263 get_expression_id (expr),
3264 avail))
3265 new_stuff = true;
3268 /* If all edges produce the same value and that value is
3269 an invariant, then the PHI has the same value on all
3270 edges. Note this. */
3271 else if (!cant_insert && all_same)
3273 gcc_assert (edoubleprime->kind == CONSTANT
3274 || edoubleprime->kind == NAME);
3276 tree temp = make_temp_ssa_name (get_expr_type (expr),
3277 NULL, "pretmp");
3278 gassign *assign
3279 = gimple_build_assign (temp,
3280 edoubleprime->kind == CONSTANT ?
3281 PRE_EXPR_CONSTANT (edoubleprime) :
3282 PRE_EXPR_NAME (edoubleprime));
3283 gimple_stmt_iterator gsi = gsi_after_labels (block);
3284 gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3286 gimple_set_plf (assign, NECESSARY, false);
3287 VN_INFO_GET (temp)->value_id = val;
3288 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3289 if (VN_INFO (temp)->valnum == NULL_TREE)
3290 VN_INFO (temp)->valnum = temp;
3291 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3292 pre_expr newe = get_or_alloc_expr_for_name (temp);
3293 add_to_value (val, newe);
3294 bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3295 bitmap_insert_into_set (NEW_SETS (block), newe);
3300 exprs.release ();
3301 return new_stuff;
3305 /* Perform insertion for partially anticipatable expressions. There
3306 is only one case we will perform insertion for these. This case is
3307 if the expression is partially anticipatable, and fully available.
3308 In this case, we know that putting it earlier will enable us to
3309 remove the later computation. */
3312 static bool
3313 do_partial_partial_insertion (basic_block block, basic_block dom)
3315 bool new_stuff = false;
3316 vec<pre_expr> exprs;
3317 pre_expr expr;
3318 auto_vec<pre_expr> avail;
3319 int i;
3321 exprs = sorted_array_from_bitmap_set (PA_IN (block));
3322 avail.safe_grow (EDGE_COUNT (block->preds));
3324 FOR_EACH_VEC_ELT (exprs, i, expr)
3326 if (expr->kind == NARY
3327 || expr->kind == REFERENCE)
3329 unsigned int val;
3330 bool by_all = true;
3331 bool cant_insert = false;
3332 edge pred;
3333 basic_block bprime;
3334 pre_expr eprime = NULL;
3335 edge_iterator ei;
3337 val = get_expr_value_id (expr);
3338 if (bitmap_set_contains_value (PHI_GEN (block), val))
3339 continue;
3340 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3341 continue;
3343 FOR_EACH_EDGE (pred, ei, block->preds)
3345 unsigned int vprime;
3346 pre_expr edoubleprime;
3348 /* We should never run insertion for the exit block
3349 and so not come across fake pred edges. */
3350 gcc_assert (!(pred->flags & EDGE_FAKE));
3351 bprime = pred->src;
3352 eprime = phi_translate (expr, ANTIC_IN (block),
3353 PA_IN (block),
3354 bprime, block);
3356 /* eprime will generally only be NULL if the
3357 value of the expression, translated
3358 through the PHI for this predecessor, is
3359 undefined. If that is the case, we can't
3360 make the expression fully redundant,
3361 because its value is undefined along a
3362 predecessor path. We can thus break out
3363 early because it doesn't matter what the
3364 rest of the results are. */
3365 if (eprime == NULL)
3367 avail[pred->dest_idx] = NULL;
3368 cant_insert = true;
3369 break;
3372 eprime = fully_constant_expression (eprime);
3373 vprime = get_expr_value_id (eprime);
3374 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3375 avail[pred->dest_idx] = edoubleprime;
3376 if (edoubleprime == NULL)
3378 by_all = false;
3379 break;
3383 /* If we can insert it, it's not the same value
3384 already existing along every predecessor, and
3385 it's defined by some predecessor, it is
3386 partially redundant. */
3387 if (!cant_insert && by_all)
3389 edge succ;
3390 bool do_insertion = false;
3392 /* Insert only if we can remove a later expression on a path
3393 that we want to optimize for speed.
3394 The phi node that we will be inserting in BLOCK is not free,
3395 and inserting it for the sake of !optimize_for_speed successor
3396 may cause regressions on the speed path. */
3397 FOR_EACH_EDGE (succ, ei, block->succs)
3399 if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3400 || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3402 if (optimize_edge_for_speed_p (succ))
3403 do_insertion = true;
3407 if (!do_insertion)
3409 if (dump_file && (dump_flags & TDF_DETAILS))
3411 fprintf (dump_file, "Skipping partial partial redundancy "
3412 "for expression ");
3413 print_pre_expr (dump_file, expr);
3414 fprintf (dump_file, " (%04d), not (partially) anticipated "
3415 "on any to be optimized for speed edges\n", val);
3418 else if (dbg_cnt (treepre_insert))
3420 pre_stats.pa_insert++;
3421 if (dump_file && (dump_flags & TDF_DETAILS))
3423 fprintf (dump_file, "Found partial partial redundancy "
3424 "for expression ");
3425 print_pre_expr (dump_file, expr);
3426 fprintf (dump_file, " (%04d)\n",
3427 get_expr_value_id (expr));
3429 if (insert_into_preds_of_block (block,
3430 get_expression_id (expr),
3431 avail))
3432 new_stuff = true;
3438 exprs.release ();
3439 return new_stuff;
3442 static bool
3443 insert_aux (basic_block block)
3445 basic_block son;
3446 bool new_stuff = false;
3448 if (block)
3450 basic_block dom;
3451 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3452 if (dom)
3454 unsigned i;
3455 bitmap_iterator bi;
3456 bitmap_set_t newset = NEW_SETS (dom);
3457 if (newset)
3459 /* Note that we need to value_replace both NEW_SETS, and
3460 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3461 represented by some non-simple expression here that we want
3462 to replace it with. */
3463 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3465 pre_expr expr = expression_for_id (i);
3466 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3467 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3470 if (!single_pred_p (block))
3472 new_stuff |= do_regular_insertion (block, dom);
3473 if (do_partial_partial)
3474 new_stuff |= do_partial_partial_insertion (block, dom);
3478 for (son = first_dom_son (CDI_DOMINATORS, block);
3479 son;
3480 son = next_dom_son (CDI_DOMINATORS, son))
3482 new_stuff |= insert_aux (son);
3485 return new_stuff;
3488 /* Perform insertion of partially redundant values. */
3490 static void
3491 insert (void)
3493 bool new_stuff = true;
3494 basic_block bb;
3495 int num_iterations = 0;
3497 FOR_ALL_BB_FN (bb, cfun)
3498 NEW_SETS (bb) = bitmap_set_new ();
3500 while (new_stuff)
3502 num_iterations++;
3503 if (dump_file && dump_flags & TDF_DETAILS)
3504 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3505 new_stuff = insert_aux (ENTRY_BLOCK_PTR_FOR_FN (cfun));
3507 /* Clear the NEW sets before the next iteration. We have already
3508 fully propagated its contents. */
3509 if (new_stuff)
3510 FOR_ALL_BB_FN (bb, cfun)
3511 bitmap_set_free (NEW_SETS (bb));
3513 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3517 /* Compute the AVAIL set for all basic blocks.
3519 This function performs value numbering of the statements in each basic
3520 block. The AVAIL sets are built from information we glean while doing
3521 this value numbering, since the AVAIL sets contain only one entry per
3522 value.
3524 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3525 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3527 static void
3528 compute_avail (void)
3531 basic_block block, son;
3532 basic_block *worklist;
3533 size_t sp = 0;
3534 unsigned i;
3536 /* We pretend that default definitions are defined in the entry block.
3537 This includes function arguments and the static chain decl. */
3538 for (i = 1; i < num_ssa_names; ++i)
3540 tree name = ssa_name (i);
3541 pre_expr e;
3542 if (!name
3543 || !SSA_NAME_IS_DEFAULT_DEF (name)
3544 || has_zero_uses (name)
3545 || virtual_operand_p (name))
3546 continue;
3548 e = get_or_alloc_expr_for_name (name);
3549 add_to_value (get_expr_value_id (e), e);
3550 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)), e);
3551 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3555 if (dump_file && (dump_flags & TDF_DETAILS))
3557 print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3558 "tmp_gen", ENTRY_BLOCK);
3559 print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3560 "avail_out", ENTRY_BLOCK);
3563 /* Allocate the worklist. */
3564 worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
3566 /* Seed the algorithm by putting the dominator children of the entry
3567 block on the worklist. */
3568 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR_FOR_FN (cfun));
3569 son;
3570 son = next_dom_son (CDI_DOMINATORS, son))
3571 worklist[sp++] = son;
3573 BB_LIVE_VOP_ON_EXIT (ENTRY_BLOCK_PTR_FOR_FN (cfun))
3574 = ssa_default_def (cfun, gimple_vop (cfun));
3576 /* Loop until the worklist is empty. */
3577 while (sp)
3579 gimple *stmt;
3580 basic_block dom;
3582 /* Pick a block from the worklist. */
3583 block = worklist[--sp];
3585 /* Initially, the set of available values in BLOCK is that of
3586 its immediate dominator. */
3587 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3588 if (dom)
3590 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3591 BB_LIVE_VOP_ON_EXIT (block) = BB_LIVE_VOP_ON_EXIT (dom);
3594 /* Generate values for PHI nodes. */
3595 for (gphi_iterator gsi = gsi_start_phis (block); !gsi_end_p (gsi);
3596 gsi_next (&gsi))
3598 tree result = gimple_phi_result (gsi.phi ());
3600 /* We have no need for virtual phis, as they don't represent
3601 actual computations. */
3602 if (virtual_operand_p (result))
3604 BB_LIVE_VOP_ON_EXIT (block) = result;
3605 continue;
3608 pre_expr e = get_or_alloc_expr_for_name (result);
3609 add_to_value (get_expr_value_id (e), e);
3610 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3611 bitmap_insert_into_set (PHI_GEN (block), e);
3614 BB_MAY_NOTRETURN (block) = 0;
3616 /* Now compute value numbers and populate value sets with all
3617 the expressions computed in BLOCK. */
3618 for (gimple_stmt_iterator gsi = gsi_start_bb (block); !gsi_end_p (gsi);
3619 gsi_next (&gsi))
3621 ssa_op_iter iter;
3622 tree op;
3624 stmt = gsi_stmt (gsi);
3626 /* Cache whether the basic-block has any non-visible side-effect
3627 or control flow.
3628 If this isn't a call or it is the last stmt in the
3629 basic-block then the CFG represents things correctly. */
3630 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3632 /* Non-looping const functions always return normally.
3633 Otherwise the call might not return or have side-effects
3634 that forbids hoisting possibly trapping expressions
3635 before it. */
3636 int flags = gimple_call_flags (stmt);
3637 if (!(flags & ECF_CONST)
3638 || (flags & ECF_LOOPING_CONST_OR_PURE))
3639 BB_MAY_NOTRETURN (block) = 1;
3642 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3644 pre_expr e = get_or_alloc_expr_for_name (op);
3646 add_to_value (get_expr_value_id (e), e);
3647 bitmap_insert_into_set (TMP_GEN (block), e);
3648 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3651 if (gimple_vdef (stmt))
3652 BB_LIVE_VOP_ON_EXIT (block) = gimple_vdef (stmt);
3654 if (gimple_has_side_effects (stmt)
3655 || stmt_could_throw_p (stmt)
3656 || is_gimple_debug (stmt))
3657 continue;
3659 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3661 if (ssa_undefined_value_p (op))
3662 continue;
3663 pre_expr e = get_or_alloc_expr_for_name (op);
3664 bitmap_value_insert_into_set (EXP_GEN (block), e);
3667 switch (gimple_code (stmt))
3669 case GIMPLE_RETURN:
3670 continue;
3672 case GIMPLE_CALL:
3674 vn_reference_t ref;
3675 vn_reference_s ref1;
3676 pre_expr result = NULL;
3678 /* We can value number only calls to real functions. */
3679 if (gimple_call_internal_p (stmt))
3680 continue;
3682 vn_reference_lookup_call (as_a <gcall *> (stmt), &ref, &ref1);
3683 if (!ref)
3684 continue;
3686 /* If the value of the call is not invalidated in
3687 this block until it is computed, add the expression
3688 to EXP_GEN. */
3689 if (!gimple_vuse (stmt)
3690 || gimple_code
3691 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
3692 || gimple_bb (SSA_NAME_DEF_STMT
3693 (gimple_vuse (stmt))) != block)
3695 result = pre_expr_pool.allocate ();
3696 result->kind = REFERENCE;
3697 result->id = 0;
3698 PRE_EXPR_REFERENCE (result) = ref;
3700 get_or_alloc_expression_id (result);
3701 add_to_value (get_expr_value_id (result), result);
3702 bitmap_value_insert_into_set (EXP_GEN (block), result);
3704 continue;
3707 case GIMPLE_ASSIGN:
3709 pre_expr result = NULL;
3710 switch (vn_get_stmt_kind (stmt))
3712 case VN_NARY:
3714 enum tree_code code = gimple_assign_rhs_code (stmt);
3715 vn_nary_op_t nary;
3717 /* COND_EXPR and VEC_COND_EXPR are awkward in
3718 that they contain an embedded complex expression.
3719 Don't even try to shove those through PRE. */
3720 if (code == COND_EXPR
3721 || code == VEC_COND_EXPR)
3722 continue;
3724 vn_nary_op_lookup_stmt (stmt, &nary);
3725 if (!nary)
3726 continue;
3728 /* If the NARY traps and there was a preceding
3729 point in the block that might not return avoid
3730 adding the nary to EXP_GEN. */
3731 if (BB_MAY_NOTRETURN (block)
3732 && vn_nary_may_trap (nary))
3733 continue;
3735 result = pre_expr_pool.allocate ();
3736 result->kind = NARY;
3737 result->id = 0;
3738 PRE_EXPR_NARY (result) = nary;
3739 break;
3742 case VN_REFERENCE:
3744 vn_reference_t ref;
3745 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3746 gimple_vuse (stmt),
3747 VN_WALK, &ref);
3748 if (!ref)
3749 continue;
3751 /* If the value of the reference is not invalidated in
3752 this block until it is computed, add the expression
3753 to EXP_GEN. */
3754 if (gimple_vuse (stmt))
3756 gimple *def_stmt;
3757 bool ok = true;
3758 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
3759 while (!gimple_nop_p (def_stmt)
3760 && gimple_code (def_stmt) != GIMPLE_PHI
3761 && gimple_bb (def_stmt) == block)
3763 if (stmt_may_clobber_ref_p
3764 (def_stmt, gimple_assign_rhs1 (stmt)))
3766 ok = false;
3767 break;
3769 def_stmt
3770 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
3772 if (!ok)
3773 continue;
3776 result = pre_expr_pool.allocate ();
3777 result->kind = REFERENCE;
3778 result->id = 0;
3779 PRE_EXPR_REFERENCE (result) = ref;
3780 break;
3783 default:
3784 continue;
3787 get_or_alloc_expression_id (result);
3788 add_to_value (get_expr_value_id (result), result);
3789 bitmap_value_insert_into_set (EXP_GEN (block), result);
3790 continue;
3792 default:
3793 break;
3797 if (dump_file && (dump_flags & TDF_DETAILS))
3799 print_bitmap_set (dump_file, EXP_GEN (block),
3800 "exp_gen", block->index);
3801 print_bitmap_set (dump_file, PHI_GEN (block),
3802 "phi_gen", block->index);
3803 print_bitmap_set (dump_file, TMP_GEN (block),
3804 "tmp_gen", block->index);
3805 print_bitmap_set (dump_file, AVAIL_OUT (block),
3806 "avail_out", block->index);
3809 /* Put the dominator children of BLOCK on the worklist of blocks
3810 to compute available sets for. */
3811 for (son = first_dom_son (CDI_DOMINATORS, block);
3812 son;
3813 son = next_dom_son (CDI_DOMINATORS, son))
3814 worklist[sp++] = son;
3817 free (worklist);
3821 /* Local state for the eliminate domwalk. */
3822 static vec<gimple *> el_to_remove;
3823 static vec<gimple *> el_to_fixup;
3824 static unsigned int el_todo;
3825 static vec<tree> el_avail;
3826 static vec<tree> el_avail_stack;
3828 /* Return a leader for OP that is available at the current point of the
3829 eliminate domwalk. */
3831 static tree
3832 eliminate_avail (tree op)
3834 tree valnum = VN_INFO (op)->valnum;
3835 if (TREE_CODE (valnum) == SSA_NAME)
3837 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
3838 return valnum;
3839 if (el_avail.length () > SSA_NAME_VERSION (valnum))
3840 return el_avail[SSA_NAME_VERSION (valnum)];
3842 else if (is_gimple_min_invariant (valnum))
3843 return valnum;
3844 return NULL_TREE;
3847 /* At the current point of the eliminate domwalk make OP available. */
3849 static void
3850 eliminate_push_avail (tree op)
3852 tree valnum = VN_INFO (op)->valnum;
3853 if (TREE_CODE (valnum) == SSA_NAME)
3855 if (el_avail.length () <= SSA_NAME_VERSION (valnum))
3856 el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1);
3857 tree pushop = op;
3858 if (el_avail[SSA_NAME_VERSION (valnum)])
3859 pushop = el_avail[SSA_NAME_VERSION (valnum)];
3860 el_avail_stack.safe_push (pushop);
3861 el_avail[SSA_NAME_VERSION (valnum)] = op;
3865 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
3866 the leader for the expression if insertion was successful. */
3868 static tree
3869 eliminate_insert (gimple_stmt_iterator *gsi, tree val)
3871 gimple *stmt = gimple_seq_first_stmt (VN_INFO (val)->expr);
3872 if (!is_gimple_assign (stmt)
3873 || (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
3874 && gimple_assign_rhs_code (stmt) != VIEW_CONVERT_EXPR))
3875 return NULL_TREE;
3877 tree op = gimple_assign_rhs1 (stmt);
3878 if (gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR)
3879 op = TREE_OPERAND (op, 0);
3880 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op;
3881 if (!leader)
3882 return NULL_TREE;
3884 gimple_seq stmts = NULL;
3885 tree res = gimple_build (&stmts, gimple_assign_rhs_code (stmt),
3886 TREE_TYPE (val), leader);
3887 gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
3888 VN_INFO_GET (res)->valnum = val;
3890 if (TREE_CODE (leader) == SSA_NAME)
3891 gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true);
3893 pre_stats.insertions++;
3894 if (dump_file && (dump_flags & TDF_DETAILS))
3896 fprintf (dump_file, "Inserted ");
3897 print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (res), 0, 0);
3900 return res;
3903 class eliminate_dom_walker : public dom_walker
3905 public:
3906 eliminate_dom_walker (cdi_direction direction, bool do_pre_)
3907 : dom_walker (direction), do_pre (do_pre_) {}
3909 virtual edge before_dom_children (basic_block);
3910 virtual void after_dom_children (basic_block);
3912 bool do_pre;
3915 /* Perform elimination for the basic-block B during the domwalk. */
3917 edge
3918 eliminate_dom_walker::before_dom_children (basic_block b)
3920 /* Mark new bb. */
3921 el_avail_stack.safe_push (NULL_TREE);
3923 /* ??? If we do nothing for unreachable blocks then this will confuse
3924 tailmerging. Eventually we can reduce its reliance on SCCVN now
3925 that we fully copy/constant-propagate (most) things. */
3927 for (gphi_iterator gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
3929 gphi *phi = gsi.phi ();
3930 tree res = PHI_RESULT (phi);
3932 if (virtual_operand_p (res))
3934 gsi_next (&gsi);
3935 continue;
3938 tree sprime = eliminate_avail (res);
3939 if (sprime
3940 && sprime != res)
3942 if (dump_file && (dump_flags & TDF_DETAILS))
3944 fprintf (dump_file, "Replaced redundant PHI node defining ");
3945 print_generic_expr (dump_file, res, 0);
3946 fprintf (dump_file, " with ");
3947 print_generic_expr (dump_file, sprime, 0);
3948 fprintf (dump_file, "\n");
3951 /* If we inserted this PHI node ourself, it's not an elimination. */
3952 if (inserted_exprs
3953 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
3954 pre_stats.phis--;
3955 else
3956 pre_stats.eliminations++;
3958 /* If we will propagate into all uses don't bother to do
3959 anything. */
3960 if (may_propagate_copy (res, sprime))
3962 /* Mark the PHI for removal. */
3963 el_to_remove.safe_push (phi);
3964 gsi_next (&gsi);
3965 continue;
3968 remove_phi_node (&gsi, false);
3970 if (inserted_exprs
3971 && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
3972 && TREE_CODE (sprime) == SSA_NAME)
3973 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
3975 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
3976 sprime = fold_convert (TREE_TYPE (res), sprime);
3977 gimple *stmt = gimple_build_assign (res, sprime);
3978 /* ??? It cannot yet be necessary (DOM walk). */
3979 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
3981 gimple_stmt_iterator gsi2 = gsi_after_labels (b);
3982 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
3983 continue;
3986 eliminate_push_avail (res);
3987 gsi_next (&gsi);
3990 for (gimple_stmt_iterator gsi = gsi_start_bb (b);
3991 !gsi_end_p (gsi);
3992 gsi_next (&gsi))
3994 tree sprime = NULL_TREE;
3995 gimple *stmt = gsi_stmt (gsi);
3996 tree lhs = gimple_get_lhs (stmt);
3997 if (lhs && TREE_CODE (lhs) == SSA_NAME
3998 && !gimple_has_volatile_ops (stmt)
3999 /* See PR43491. Do not replace a global register variable when
4000 it is a the RHS of an assignment. Do replace local register
4001 variables since gcc does not guarantee a local variable will
4002 be allocated in register.
4003 ??? The fix isn't effective here. This should instead
4004 be ensured by not value-numbering them the same but treating
4005 them like volatiles? */
4006 && !(gimple_assign_single_p (stmt)
4007 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL
4008 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt))
4009 && is_global_var (gimple_assign_rhs1 (stmt)))))
4011 sprime = eliminate_avail (lhs);
4012 if (!sprime)
4014 /* If there is no existing usable leader but SCCVN thinks
4015 it has an expression it wants to use as replacement,
4016 insert that. */
4017 tree val = VN_INFO (lhs)->valnum;
4018 if (val != VN_TOP
4019 && TREE_CODE (val) == SSA_NAME
4020 && VN_INFO (val)->needs_insertion
4021 && VN_INFO (val)->expr != NULL
4022 && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE)
4023 eliminate_push_avail (sprime);
4026 /* If this now constitutes a copy duplicate points-to
4027 and range info appropriately. This is especially
4028 important for inserted code. See tree-ssa-copy.c
4029 for similar code. */
4030 if (sprime
4031 && TREE_CODE (sprime) == SSA_NAME)
4033 basic_block sprime_b = gimple_bb (SSA_NAME_DEF_STMT (sprime));
4034 if (POINTER_TYPE_P (TREE_TYPE (lhs))
4035 && SSA_NAME_PTR_INFO (lhs)
4036 && !SSA_NAME_PTR_INFO (sprime))
4038 duplicate_ssa_name_ptr_info (sprime,
4039 SSA_NAME_PTR_INFO (lhs));
4040 if (b != sprime_b)
4041 mark_ptr_info_alignment_unknown
4042 (SSA_NAME_PTR_INFO (sprime));
4044 else if (!POINTER_TYPE_P (TREE_TYPE (lhs))
4045 && SSA_NAME_RANGE_INFO (lhs)
4046 && !SSA_NAME_RANGE_INFO (sprime)
4047 && b == sprime_b)
4048 duplicate_ssa_name_range_info (sprime,
4049 SSA_NAME_RANGE_TYPE (lhs),
4050 SSA_NAME_RANGE_INFO (lhs));
4053 /* Inhibit the use of an inserted PHI on a loop header when
4054 the address of the memory reference is a simple induction
4055 variable. In other cases the vectorizer won't do anything
4056 anyway (either it's loop invariant or a complicated
4057 expression). */
4058 if (sprime
4059 && TREE_CODE (sprime) == SSA_NAME
4060 && do_pre
4061 && flag_tree_loop_vectorize
4062 && loop_outer (b->loop_father)
4063 && has_zero_uses (sprime)
4064 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))
4065 && gimple_assign_load_p (stmt))
4067 gimple *def_stmt = SSA_NAME_DEF_STMT (sprime);
4068 basic_block def_bb = gimple_bb (def_stmt);
4069 if (gimple_code (def_stmt) == GIMPLE_PHI
4070 && def_bb->loop_father->header == def_bb)
4072 loop_p loop = def_bb->loop_father;
4073 ssa_op_iter iter;
4074 tree op;
4075 bool found = false;
4076 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4078 affine_iv iv;
4079 def_bb = gimple_bb (SSA_NAME_DEF_STMT (op));
4080 if (def_bb
4081 && flow_bb_inside_loop_p (loop, def_bb)
4082 && simple_iv (loop, loop, op, &iv, true))
4084 found = true;
4085 break;
4088 if (found)
4090 if (dump_file && (dump_flags & TDF_DETAILS))
4092 fprintf (dump_file, "Not replacing ");
4093 print_gimple_expr (dump_file, stmt, 0, 0);
4094 fprintf (dump_file, " with ");
4095 print_generic_expr (dump_file, sprime, 0);
4096 fprintf (dump_file, " which would add a loop"
4097 " carried dependence to loop %d\n",
4098 loop->num);
4100 /* Don't keep sprime available. */
4101 sprime = NULL_TREE;
4106 if (sprime)
4108 /* If we can propagate the value computed for LHS into
4109 all uses don't bother doing anything with this stmt. */
4110 if (may_propagate_copy (lhs, sprime))
4112 /* Mark it for removal. */
4113 el_to_remove.safe_push (stmt);
4115 /* ??? Don't count copy/constant propagations. */
4116 if (gimple_assign_single_p (stmt)
4117 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4118 || gimple_assign_rhs1 (stmt) == sprime))
4119 continue;
4121 if (dump_file && (dump_flags & TDF_DETAILS))
4123 fprintf (dump_file, "Replaced ");
4124 print_gimple_expr (dump_file, stmt, 0, 0);
4125 fprintf (dump_file, " with ");
4126 print_generic_expr (dump_file, sprime, 0);
4127 fprintf (dump_file, " in all uses of ");
4128 print_gimple_stmt (dump_file, stmt, 0, 0);
4131 pre_stats.eliminations++;
4132 continue;
4135 /* If this is an assignment from our leader (which
4136 happens in the case the value-number is a constant)
4137 then there is nothing to do. */
4138 if (gimple_assign_single_p (stmt)
4139 && sprime == gimple_assign_rhs1 (stmt))
4140 continue;
4142 /* Else replace its RHS. */
4143 bool can_make_abnormal_goto
4144 = is_gimple_call (stmt)
4145 && stmt_can_make_abnormal_goto (stmt);
4147 if (dump_file && (dump_flags & TDF_DETAILS))
4149 fprintf (dump_file, "Replaced ");
4150 print_gimple_expr (dump_file, stmt, 0, 0);
4151 fprintf (dump_file, " with ");
4152 print_generic_expr (dump_file, sprime, 0);
4153 fprintf (dump_file, " in ");
4154 print_gimple_stmt (dump_file, stmt, 0, 0);
4157 if (TREE_CODE (sprime) == SSA_NAME)
4158 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4159 NECESSARY, true);
4161 pre_stats.eliminations++;
4162 gimple *orig_stmt = stmt;
4163 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4164 TREE_TYPE (sprime)))
4165 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4166 tree vdef = gimple_vdef (stmt);
4167 tree vuse = gimple_vuse (stmt);
4168 propagate_tree_value_into_stmt (&gsi, sprime);
4169 stmt = gsi_stmt (gsi);
4170 update_stmt (stmt);
4171 if (vdef != gimple_vdef (stmt))
4172 VN_INFO (vdef)->valnum = vuse;
4174 /* If we removed EH side-effects from the statement, clean
4175 its EH information. */
4176 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4178 bitmap_set_bit (need_eh_cleanup,
4179 gimple_bb (stmt)->index);
4180 if (dump_file && (dump_flags & TDF_DETAILS))
4181 fprintf (dump_file, " Removed EH side-effects.\n");
4184 /* Likewise for AB side-effects. */
4185 if (can_make_abnormal_goto
4186 && !stmt_can_make_abnormal_goto (stmt))
4188 bitmap_set_bit (need_ab_cleanup,
4189 gimple_bb (stmt)->index);
4190 if (dump_file && (dump_flags & TDF_DETAILS))
4191 fprintf (dump_file, " Removed AB side-effects.\n");
4194 continue;
4198 /* If the statement is a scalar store, see if the expression
4199 has the same value number as its rhs. If so, the store is
4200 dead. */
4201 if (gimple_assign_single_p (stmt)
4202 && !gimple_has_volatile_ops (stmt)
4203 && !is_gimple_reg (gimple_assign_lhs (stmt))
4204 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4205 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4207 tree val;
4208 tree rhs = gimple_assign_rhs1 (stmt);
4209 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4210 gimple_vuse (stmt), VN_WALK, NULL);
4211 if (TREE_CODE (rhs) == SSA_NAME)
4212 rhs = VN_INFO (rhs)->valnum;
4213 if (val
4214 && operand_equal_p (val, rhs, 0))
4216 if (dump_file && (dump_flags & TDF_DETAILS))
4218 fprintf (dump_file, "Deleted redundant store ");
4219 print_gimple_stmt (dump_file, stmt, 0, 0);
4222 /* Queue stmt for removal. */
4223 el_to_remove.safe_push (stmt);
4224 continue;
4228 /* If this is a control statement value numbering left edges
4229 unexecuted on force the condition in a way consistent with
4230 that. */
4231 if (gcond *cond = dyn_cast <gcond *> (stmt))
4233 if ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE)
4234 ^ (EDGE_SUCC (b, 1)->flags & EDGE_EXECUTABLE))
4236 if (dump_file && (dump_flags & TDF_DETAILS))
4238 fprintf (dump_file, "Removing unexecutable edge from ");
4239 print_gimple_stmt (dump_file, stmt, 0, 0);
4241 if (((EDGE_SUCC (b, 0)->flags & EDGE_TRUE_VALUE) != 0)
4242 == ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE) != 0))
4243 gimple_cond_make_true (cond);
4244 else
4245 gimple_cond_make_false (cond);
4246 update_stmt (cond);
4247 el_todo |= TODO_cleanup_cfg;
4248 continue;
4252 bool can_make_abnormal_goto = stmt_can_make_abnormal_goto (stmt);
4253 bool was_noreturn = (is_gimple_call (stmt)
4254 && gimple_call_noreturn_p (stmt));
4255 tree vdef = gimple_vdef (stmt);
4256 tree vuse = gimple_vuse (stmt);
4258 /* If we didn't replace the whole stmt (or propagate the result
4259 into all uses), replace all uses on this stmt with their
4260 leaders. */
4261 use_operand_p use_p;
4262 ssa_op_iter iter;
4263 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
4265 tree use = USE_FROM_PTR (use_p);
4266 /* ??? The call code above leaves stmt operands un-updated. */
4267 if (TREE_CODE (use) != SSA_NAME)
4268 continue;
4269 tree sprime = eliminate_avail (use);
4270 if (sprime && sprime != use
4271 && may_propagate_copy (use, sprime)
4272 /* We substitute into debug stmts to avoid excessive
4273 debug temporaries created by removed stmts, but we need
4274 to avoid doing so for inserted sprimes as we never want
4275 to create debug temporaries for them. */
4276 && (!inserted_exprs
4277 || TREE_CODE (sprime) != SSA_NAME
4278 || !is_gimple_debug (stmt)
4279 || !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))))
4281 propagate_value (use_p, sprime);
4282 gimple_set_modified (stmt, true);
4283 if (TREE_CODE (sprime) == SSA_NAME
4284 && !is_gimple_debug (stmt))
4285 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4286 NECESSARY, true);
4290 /* Visit indirect calls and turn them into direct calls if
4291 possible using the devirtualization machinery. */
4292 if (gcall *call_stmt = dyn_cast <gcall *> (stmt))
4294 tree fn = gimple_call_fn (call_stmt);
4295 if (fn
4296 && flag_devirtualize
4297 && virtual_method_call_p (fn))
4299 tree otr_type = obj_type_ref_class (fn);
4300 tree instance;
4301 ipa_polymorphic_call_context context (current_function_decl, fn, stmt, &instance);
4302 bool final;
4304 context.get_dynamic_type (instance, OBJ_TYPE_REF_OBJECT (fn), otr_type, stmt);
4306 vec <cgraph_node *>targets
4307 = possible_polymorphic_call_targets (obj_type_ref_class (fn),
4308 tree_to_uhwi
4309 (OBJ_TYPE_REF_TOKEN (fn)),
4310 context,
4311 &final);
4312 if (dump_file)
4313 dump_possible_polymorphic_call_targets (dump_file,
4314 obj_type_ref_class (fn),
4315 tree_to_uhwi
4316 (OBJ_TYPE_REF_TOKEN (fn)),
4317 context);
4318 if (final && targets.length () <= 1 && dbg_cnt (devirt))
4320 tree fn;
4321 if (targets.length () == 1)
4322 fn = targets[0]->decl;
4323 else
4324 fn = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
4325 if (dump_enabled_p ())
4327 location_t loc = gimple_location_safe (stmt);
4328 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc,
4329 "converting indirect call to "
4330 "function %s\n",
4331 cgraph_node::get (fn)->name ());
4333 gimple_call_set_fndecl (call_stmt, fn);
4334 maybe_remove_unused_call_args (cfun, call_stmt);
4335 gimple_set_modified (stmt, true);
4340 if (gimple_modified_p (stmt))
4342 /* If a formerly non-invariant ADDR_EXPR is turned into an
4343 invariant one it was on a separate stmt. */
4344 if (gimple_assign_single_p (stmt)
4345 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR)
4346 recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
4347 gimple *old_stmt = stmt;
4348 if (is_gimple_call (stmt))
4350 /* ??? Only fold calls inplace for now, this may create new
4351 SSA names which in turn will confuse free_scc_vn SSA name
4352 release code. */
4353 fold_stmt_inplace (&gsi);
4354 /* When changing a call into a noreturn call, cfg cleanup
4355 is needed to fix up the noreturn call. */
4356 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4357 el_to_fixup.safe_push (stmt);
4359 else
4361 fold_stmt (&gsi);
4362 stmt = gsi_stmt (gsi);
4363 if ((gimple_code (stmt) == GIMPLE_COND
4364 && (gimple_cond_true_p (as_a <gcond *> (stmt))
4365 || gimple_cond_false_p (as_a <gcond *> (stmt))))
4366 || (gimple_code (stmt) == GIMPLE_SWITCH
4367 && TREE_CODE (gimple_switch_index (
4368 as_a <gswitch *> (stmt)))
4369 == INTEGER_CST))
4370 el_todo |= TODO_cleanup_cfg;
4372 /* If we removed EH side-effects from the statement, clean
4373 its EH information. */
4374 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
4376 bitmap_set_bit (need_eh_cleanup,
4377 gimple_bb (stmt)->index);
4378 if (dump_file && (dump_flags & TDF_DETAILS))
4379 fprintf (dump_file, " Removed EH side-effects.\n");
4381 /* Likewise for AB side-effects. */
4382 if (can_make_abnormal_goto
4383 && !stmt_can_make_abnormal_goto (stmt))
4385 bitmap_set_bit (need_ab_cleanup,
4386 gimple_bb (stmt)->index);
4387 if (dump_file && (dump_flags & TDF_DETAILS))
4388 fprintf (dump_file, " Removed AB side-effects.\n");
4390 update_stmt (stmt);
4391 if (vdef != gimple_vdef (stmt))
4392 VN_INFO (vdef)->valnum = vuse;
4395 /* Make new values available - for fully redundant LHS we
4396 continue with the next stmt above and skip this. */
4397 def_operand_p defp;
4398 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_DEF)
4399 eliminate_push_avail (DEF_FROM_PTR (defp));
4402 /* Replace destination PHI arguments. */
4403 edge_iterator ei;
4404 edge e;
4405 FOR_EACH_EDGE (e, ei, b->succs)
4407 for (gphi_iterator gsi = gsi_start_phis (e->dest);
4408 !gsi_end_p (gsi);
4409 gsi_next (&gsi))
4411 gphi *phi = gsi.phi ();
4412 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
4413 tree arg = USE_FROM_PTR (use_p);
4414 if (TREE_CODE (arg) != SSA_NAME
4415 || virtual_operand_p (arg))
4416 continue;
4417 tree sprime = eliminate_avail (arg);
4418 if (sprime && may_propagate_copy (arg, sprime))
4420 propagate_value (use_p, sprime);
4421 if (TREE_CODE (sprime) == SSA_NAME)
4422 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4426 return NULL;
4429 /* Make no longer available leaders no longer available. */
4431 void
4432 eliminate_dom_walker::after_dom_children (basic_block)
4434 tree entry;
4435 while ((entry = el_avail_stack.pop ()) != NULL_TREE)
4437 tree valnum = VN_INFO (entry)->valnum;
4438 tree old = el_avail[SSA_NAME_VERSION (valnum)];
4439 if (old == entry)
4440 el_avail[SSA_NAME_VERSION (valnum)] = NULL_TREE;
4441 else
4442 el_avail[SSA_NAME_VERSION (valnum)] = entry;
4446 /* Eliminate fully redundant computations. */
4448 static unsigned int
4449 eliminate (bool do_pre)
4451 gimple_stmt_iterator gsi;
4452 gimple *stmt;
4454 need_eh_cleanup = BITMAP_ALLOC (NULL);
4455 need_ab_cleanup = BITMAP_ALLOC (NULL);
4457 el_to_remove.create (0);
4458 el_to_fixup.create (0);
4459 el_todo = 0;
4460 el_avail.create (num_ssa_names);
4461 el_avail_stack.create (0);
4463 eliminate_dom_walker (CDI_DOMINATORS,
4464 do_pre).walk (cfun->cfg->x_entry_block_ptr);
4466 el_avail.release ();
4467 el_avail_stack.release ();
4469 /* We cannot remove stmts during BB walk, especially not release SSA
4470 names there as this confuses the VN machinery. The stmts ending
4471 up in el_to_remove are either stores or simple copies.
4472 Remove stmts in reverse order to make debug stmt creation possible. */
4473 while (!el_to_remove.is_empty ())
4475 stmt = el_to_remove.pop ();
4477 if (dump_file && (dump_flags & TDF_DETAILS))
4479 fprintf (dump_file, "Removing dead stmt ");
4480 print_gimple_stmt (dump_file, stmt, 0, 0);
4483 tree lhs;
4484 if (gimple_code (stmt) == GIMPLE_PHI)
4485 lhs = gimple_phi_result (stmt);
4486 else
4487 lhs = gimple_get_lhs (stmt);
4489 if (inserted_exprs
4490 && TREE_CODE (lhs) == SSA_NAME)
4491 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4493 gsi = gsi_for_stmt (stmt);
4494 if (gimple_code (stmt) == GIMPLE_PHI)
4495 remove_phi_node (&gsi, true);
4496 else
4498 basic_block bb = gimple_bb (stmt);
4499 unlink_stmt_vdef (stmt);
4500 if (gsi_remove (&gsi, true))
4501 bitmap_set_bit (need_eh_cleanup, bb->index);
4502 if (is_gimple_call (stmt) && stmt_can_make_abnormal_goto (stmt))
4503 bitmap_set_bit (need_ab_cleanup, bb->index);
4504 release_defs (stmt);
4507 /* Removing a stmt may expose a forwarder block. */
4508 el_todo |= TODO_cleanup_cfg;
4510 el_to_remove.release ();
4512 /* Fixup stmts that became noreturn calls. This may require splitting
4513 blocks and thus isn't possible during the dominator walk. Do this
4514 in reverse order so we don't inadvertedly remove a stmt we want to
4515 fixup by visiting a dominating now noreturn call first. */
4516 while (!el_to_fixup.is_empty ())
4518 stmt = el_to_fixup.pop ();
4520 if (dump_file && (dump_flags & TDF_DETAILS))
4522 fprintf (dump_file, "Fixing up noreturn call ");
4523 print_gimple_stmt (dump_file, stmt, 0, 0);
4526 if (fixup_noreturn_call (stmt))
4527 el_todo |= TODO_cleanup_cfg;
4529 el_to_fixup.release ();
4531 return el_todo;
4534 /* Perform CFG cleanups made necessary by elimination. */
4536 static unsigned
4537 fini_eliminate (void)
4539 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4540 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4542 if (do_eh_cleanup)
4543 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4545 if (do_ab_cleanup)
4546 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4548 BITMAP_FREE (need_eh_cleanup);
4549 BITMAP_FREE (need_ab_cleanup);
4551 if (do_eh_cleanup || do_ab_cleanup)
4552 return TODO_cleanup_cfg;
4553 return 0;
4556 /* Borrow a bit of tree-ssa-dce.c for the moment.
4557 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4558 this may be a bit faster, and we may want critical edges kept split. */
4560 /* If OP's defining statement has not already been determined to be necessary,
4561 mark that statement necessary. Return the stmt, if it is newly
4562 necessary. */
4564 static inline gimple *
4565 mark_operand_necessary (tree op)
4567 gimple *stmt;
4569 gcc_assert (op);
4571 if (TREE_CODE (op) != SSA_NAME)
4572 return NULL;
4574 stmt = SSA_NAME_DEF_STMT (op);
4575 gcc_assert (stmt);
4577 if (gimple_plf (stmt, NECESSARY)
4578 || gimple_nop_p (stmt))
4579 return NULL;
4581 gimple_set_plf (stmt, NECESSARY, true);
4582 return stmt;
4585 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4586 to insert PHI nodes sometimes, and because value numbering of casts isn't
4587 perfect, we sometimes end up inserting dead code. This simple DCE-like
4588 pass removes any insertions we made that weren't actually used. */
4590 static void
4591 remove_dead_inserted_code (void)
4593 bitmap worklist;
4594 unsigned i;
4595 bitmap_iterator bi;
4596 gimple *t;
4598 worklist = BITMAP_ALLOC (NULL);
4599 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4601 t = SSA_NAME_DEF_STMT (ssa_name (i));
4602 if (gimple_plf (t, NECESSARY))
4603 bitmap_set_bit (worklist, i);
4605 while (!bitmap_empty_p (worklist))
4607 i = bitmap_first_set_bit (worklist);
4608 bitmap_clear_bit (worklist, i);
4609 t = SSA_NAME_DEF_STMT (ssa_name (i));
4611 /* PHI nodes are somewhat special in that each PHI alternative has
4612 data and control dependencies. All the statements feeding the
4613 PHI node's arguments are always necessary. */
4614 if (gimple_code (t) == GIMPLE_PHI)
4616 unsigned k;
4618 for (k = 0; k < gimple_phi_num_args (t); k++)
4620 tree arg = PHI_ARG_DEF (t, k);
4621 if (TREE_CODE (arg) == SSA_NAME)
4623 gimple *n = mark_operand_necessary (arg);
4624 if (n)
4625 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4629 else
4631 /* Propagate through the operands. Examine all the USE, VUSE and
4632 VDEF operands in this statement. Mark all the statements
4633 which feed this statement's uses as necessary. */
4634 ssa_op_iter iter;
4635 tree use;
4637 /* The operands of VDEF expressions are also needed as they
4638 represent potential definitions that may reach this
4639 statement (VDEF operands allow us to follow def-def
4640 links). */
4642 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4644 gimple *n = mark_operand_necessary (use);
4645 if (n)
4646 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4651 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4653 t = SSA_NAME_DEF_STMT (ssa_name (i));
4654 if (!gimple_plf (t, NECESSARY))
4656 gimple_stmt_iterator gsi;
4658 if (dump_file && (dump_flags & TDF_DETAILS))
4660 fprintf (dump_file, "Removing unnecessary insertion:");
4661 print_gimple_stmt (dump_file, t, 0, 0);
4664 gsi = gsi_for_stmt (t);
4665 if (gimple_code (t) == GIMPLE_PHI)
4666 remove_phi_node (&gsi, true);
4667 else
4669 gsi_remove (&gsi, true);
4670 release_defs (t);
4674 BITMAP_FREE (worklist);
4678 /* Initialize data structures used by PRE. */
4680 static void
4681 init_pre (void)
4683 basic_block bb;
4685 next_expression_id = 1;
4686 expressions.create (0);
4687 expressions.safe_push (NULL);
4688 value_expressions.create (get_max_value_id () + 1);
4689 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
4690 name_to_id.create (0);
4692 inserted_exprs = BITMAP_ALLOC (NULL);
4694 connect_infinite_loops_to_exit ();
4695 memset (&pre_stats, 0, sizeof (pre_stats));
4697 postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
4698 postorder_num = inverted_post_order_compute (postorder);
4700 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4702 calculate_dominance_info (CDI_POST_DOMINATORS);
4703 calculate_dominance_info (CDI_DOMINATORS);
4705 bitmap_obstack_initialize (&grand_bitmap_obstack);
4706 phi_translate_table = new hash_table<expr_pred_trans_d> (5110);
4707 expression_to_id = new hash_table<pre_expr_d> (num_ssa_names * 3);
4708 FOR_ALL_BB_FN (bb, cfun)
4710 EXP_GEN (bb) = bitmap_set_new ();
4711 PHI_GEN (bb) = bitmap_set_new ();
4712 TMP_GEN (bb) = bitmap_set_new ();
4713 AVAIL_OUT (bb) = bitmap_set_new ();
4718 /* Deallocate data structures used by PRE. */
4720 static void
4721 fini_pre ()
4723 free (postorder);
4724 value_expressions.release ();
4725 BITMAP_FREE (inserted_exprs);
4726 bitmap_obstack_release (&grand_bitmap_obstack);
4727 bitmap_set_pool.release ();
4728 pre_expr_pool.release ();
4729 delete phi_translate_table;
4730 phi_translate_table = NULL;
4731 delete expression_to_id;
4732 expression_to_id = NULL;
4733 name_to_id.release ();
4735 free_aux_for_blocks ();
4737 free_dominance_info (CDI_POST_DOMINATORS);
4740 namespace {
4742 const pass_data pass_data_pre =
4744 GIMPLE_PASS, /* type */
4745 "pre", /* name */
4746 OPTGROUP_NONE, /* optinfo_flags */
4747 TV_TREE_PRE, /* tv_id */
4748 /* PROP_no_crit_edges is ensured by placing pass_split_crit_edges before
4749 pass_pre. */
4750 ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */
4751 0, /* properties_provided */
4752 PROP_no_crit_edges, /* properties_destroyed */
4753 TODO_rebuild_alias, /* todo_flags_start */
4754 0, /* todo_flags_finish */
4757 class pass_pre : public gimple_opt_pass
4759 public:
4760 pass_pre (gcc::context *ctxt)
4761 : gimple_opt_pass (pass_data_pre, ctxt)
4764 /* opt_pass methods: */
4765 virtual bool gate (function *) { return flag_tree_pre != 0; }
4766 virtual unsigned int execute (function *);
4768 }; // class pass_pre
4770 unsigned int
4771 pass_pre::execute (function *fun)
4773 unsigned int todo = 0;
4775 do_partial_partial =
4776 flag_tree_partial_pre && optimize_function_for_speed_p (fun);
4778 /* This has to happen before SCCVN runs because
4779 loop_optimizer_init may create new phis, etc. */
4780 loop_optimizer_init (LOOPS_NORMAL);
4782 if (!run_scc_vn (VN_WALK))
4784 loop_optimizer_finalize ();
4785 return 0;
4788 init_pre ();
4789 scev_initialize ();
4791 /* Collect and value number expressions computed in each basic block. */
4792 compute_avail ();
4794 /* Insert can get quite slow on an incredibly large number of basic
4795 blocks due to some quadratic behavior. Until this behavior is
4796 fixed, don't run it when he have an incredibly large number of
4797 bb's. If we aren't going to run insert, there is no point in
4798 computing ANTIC, either, even though it's plenty fast. */
4799 if (n_basic_blocks_for_fn (fun) < 4000)
4801 compute_antic ();
4802 insert ();
4805 /* Make sure to remove fake edges before committing our inserts.
4806 This makes sure we don't end up with extra critical edges that
4807 we would need to split. */
4808 remove_fake_exit_edges ();
4809 gsi_commit_edge_inserts ();
4811 /* Eliminate folds statements which might (should not...) end up
4812 not keeping virtual operands up-to-date. */
4813 gcc_assert (!need_ssa_update_p (fun));
4815 /* Remove all the redundant expressions. */
4816 todo |= eliminate (true);
4818 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
4819 statistics_counter_event (fun, "PA inserted", pre_stats.pa_insert);
4820 statistics_counter_event (fun, "New PHIs", pre_stats.phis);
4821 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
4823 clear_expression_ids ();
4824 remove_dead_inserted_code ();
4826 scev_finalize ();
4827 fini_pre ();
4828 todo |= fini_eliminate ();
4829 loop_optimizer_finalize ();
4831 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4832 case we can merge the block with the remaining predecessor of the block.
4833 It should either:
4834 - call merge_blocks after each tail merge iteration
4835 - call merge_blocks after all tail merge iterations
4836 - mark TODO_cleanup_cfg when necessary
4837 - share the cfg cleanup with fini_pre. */
4838 todo |= tail_merge_optimize (todo);
4840 free_scc_vn ();
4842 /* Tail merging invalidates the virtual SSA web, together with
4843 cfg-cleanup opportunities exposed by PRE this will wreck the
4844 SSA updating machinery. So make sure to run update-ssa
4845 manually, before eventually scheduling cfg-cleanup as part of
4846 the todo. */
4847 update_ssa (TODO_update_ssa_only_virtuals);
4849 return todo;
4852 } // anon namespace
4854 gimple_opt_pass *
4855 make_pass_pre (gcc::context *ctxt)
4857 return new pass_pre (ctxt);
4860 namespace {
4862 const pass_data pass_data_fre =
4864 GIMPLE_PASS, /* type */
4865 "fre", /* name */
4866 OPTGROUP_NONE, /* optinfo_flags */
4867 TV_TREE_FRE, /* tv_id */
4868 ( PROP_cfg | PROP_ssa ), /* properties_required */
4869 0, /* properties_provided */
4870 0, /* properties_destroyed */
4871 0, /* todo_flags_start */
4872 0, /* todo_flags_finish */
4875 class pass_fre : public gimple_opt_pass
4877 public:
4878 pass_fre (gcc::context *ctxt)
4879 : gimple_opt_pass (pass_data_fre, ctxt)
4882 /* opt_pass methods: */
4883 opt_pass * clone () { return new pass_fre (m_ctxt); }
4884 virtual bool gate (function *) { return flag_tree_fre != 0; }
4885 virtual unsigned int execute (function *);
4887 }; // class pass_fre
4889 unsigned int
4890 pass_fre::execute (function *fun)
4892 unsigned int todo = 0;
4894 if (!run_scc_vn (VN_WALKREWRITE))
4895 return 0;
4897 memset (&pre_stats, 0, sizeof (pre_stats));
4899 /* Remove all the redundant expressions. */
4900 todo |= eliminate (false);
4902 todo |= fini_eliminate ();
4904 free_scc_vn ();
4906 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
4907 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
4909 return todo;
4912 } // anon namespace
4914 gimple_opt_pass *
4915 make_pass_fre (gcc::context *ctxt)
4917 return new pass_fre (ctxt);