diagnostic-show-locus.c: remove unused field from class colorizer
[official-gcc.git] / gcc / tree-ssa-pre.c
bloba6a68cf837213696ea477f4a1c4b481fb7acf3e4
1 /* Full and partial redundancy elimination and code hoisting on SSA GIMPLE.
2 Copyright (C) 2001-2017 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"
55 #include "langhooks.h"
56 #include "alias.h"
58 /* Even though this file is called tree-ssa-pre.c, we actually
59 implement a bit more than just PRE here. All of them piggy-back
60 on GVN which is implemented in tree-ssa-sccvn.c.
62 1. Full Redundancy Elimination (FRE)
63 This is the elimination phase of GVN.
65 2. Partial Redundancy Elimination (PRE)
66 This is adds computation of AVAIL_OUT and ANTIC_IN and
67 doing expression insertion to form GVN-PRE.
69 3. Code hoisting
70 This optimization uses the ANTIC_IN sets computed for PRE
71 to move expressions further up than PRE would do, to make
72 multiple computations of the same value fully redundant.
73 This pass is explained below (after the explanation of the
74 basic algorithm for PRE).
77 /* TODO:
79 1. Avail sets can be shared by making an avail_find_leader that
80 walks up the dominator tree and looks in those avail sets.
81 This might affect code optimality, it's unclear right now.
82 Currently the AVAIL_OUT sets are the remaining quadraticness in
83 memory of GVN-PRE.
84 2. Strength reduction can be performed by anticipating expressions
85 we can repair later on.
86 3. We can do back-substitution or smarter value numbering to catch
87 commutative expressions split up over multiple statements.
90 /* For ease of terminology, "expression node" in the below refers to
91 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
92 represent the actual statement containing the expressions we care about,
93 and we cache the value number by putting it in the expression. */
95 /* Basic algorithm for Partial Redundancy Elimination:
97 First we walk the statements to generate the AVAIL sets, the
98 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
99 generation of values/expressions by a given block. We use them
100 when computing the ANTIC sets. The AVAIL sets consist of
101 SSA_NAME's that represent values, so we know what values are
102 available in what blocks. AVAIL is a forward dataflow problem. In
103 SSA, values are never killed, so we don't need a kill set, or a
104 fixpoint iteration, in order to calculate the AVAIL sets. In
105 traditional parlance, AVAIL sets tell us the downsafety of the
106 expressions/values.
108 Next, we generate the ANTIC sets. These sets represent the
109 anticipatable expressions. ANTIC is a backwards dataflow
110 problem. An expression is anticipatable in a given block if it could
111 be generated in that block. This means that if we had to perform
112 an insertion in that block, of the value of that expression, we
113 could. Calculating the ANTIC sets requires phi translation of
114 expressions, because the flow goes backwards through phis. We must
115 iterate to a fixpoint of the ANTIC sets, because we have a kill
116 set. Even in SSA form, values are not live over the entire
117 function, only from their definition point onwards. So we have to
118 remove values from the ANTIC set once we go past the definition
119 point of the leaders that make them up.
120 compute_antic/compute_antic_aux performs this computation.
122 Third, we perform insertions to make partially redundant
123 expressions fully redundant.
125 An expression is partially redundant (excluding partial
126 anticipation) if:
128 1. It is AVAIL in some, but not all, of the predecessors of a
129 given block.
130 2. It is ANTIC in all the predecessors.
132 In order to make it fully redundant, we insert the expression into
133 the predecessors where it is not available, but is ANTIC.
135 When optimizing for size, we only eliminate the partial redundancy
136 if we need to insert in only one predecessor. This avoids almost
137 completely the code size increase that PRE usually causes.
139 For the partial anticipation case, we only perform insertion if it
140 is partially anticipated in some block, and fully available in all
141 of the predecessors.
143 do_pre_regular_insertion/do_pre_partial_partial_insertion
144 performs these steps, driven by insert/insert_aux.
146 Fourth, we eliminate fully redundant expressions.
147 This is a simple statement walk that replaces redundant
148 calculations with the now available values. */
150 /* Basic algorithm for Code Hoisting:
152 Code hoisting is: Moving value computations up in the control flow
153 graph to make multiple copies redundant. Typically this is a size
154 optimization, but there are cases where it also is helpful for speed.
156 A simple code hoisting algorithm is implemented that piggy-backs on
157 the PRE infrastructure. For code hoisting, we have to know ANTIC_OUT
158 which is effectively ANTIC_IN - AVAIL_OUT. The latter two have to be
159 computed for PRE, and we can use them to perform a limited version of
160 code hoisting, too.
162 For the purpose of this implementation, a value is hoistable to a basic
163 block B if the following properties are met:
165 1. The value is in ANTIC_IN(B) -- the value will be computed on all
166 paths from B to function exit and it can be computed in B);
168 2. The value is not in AVAIL_OUT(B) -- there would be no need to
169 compute the value again and make it available twice;
171 3. All successors of B are dominated by B -- makes sure that inserting
172 a computation of the value in B will make the remaining
173 computations fully redundant;
175 4. At least one successor has the value in AVAIL_OUT -- to avoid
176 hoisting values up too far;
178 5. There are at least two successors of B -- hoisting in straight
179 line code is pointless.
181 The third condition is not strictly necessary, but it would complicate
182 the hoisting pass a lot. In fact, I don't know of any code hoisting
183 algorithm that does not have this requirement. Fortunately, experiments
184 have show that most candidate hoistable values are in regions that meet
185 this condition (e.g. diamond-shape regions).
187 The forth condition is necessary to avoid hoisting things up too far
188 away from the uses of the value. Nothing else limits the algorithm
189 from hoisting everything up as far as ANTIC_IN allows. Experiments
190 with SPEC and CSiBE have shown that hoisting up too far results in more
191 spilling, less benefits for code size, and worse benchmark scores.
192 Fortunately, in practice most of the interesting hoisting opportunities
193 are caught despite this limitation.
195 For hoistable values that meet all conditions, expressions are inserted
196 to make the calculation of the hoistable value fully redundant. We
197 perform code hoisting insertions after each round of PRE insertions,
198 because code hoisting never exposes new PRE opportunities, but PRE can
199 create new code hoisting opportunities.
201 The code hoisting algorithm is implemented in do_hoist_insert, driven
202 by insert/insert_aux. */
204 /* Representations of value numbers:
206 Value numbers are represented by a representative SSA_NAME. We
207 will create fake SSA_NAME's in situations where we need a
208 representative but do not have one (because it is a complex
209 expression). In order to facilitate storing the value numbers in
210 bitmaps, and keep the number of wasted SSA_NAME's down, we also
211 associate a value_id with each value number, and create full blown
212 ssa_name's only where we actually need them (IE in operands of
213 existing expressions).
215 Theoretically you could replace all the value_id's with
216 SSA_NAME_VERSION, but this would allocate a large number of
217 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
218 It would also require an additional indirection at each point we
219 use the value id. */
221 /* Representation of expressions on value numbers:
223 Expressions consisting of value numbers are represented the same
224 way as our VN internally represents them, with an additional
225 "pre_expr" wrapping around them in order to facilitate storing all
226 of the expressions in the same sets. */
228 /* Representation of sets:
230 The dataflow sets do not need to be sorted in any particular order
231 for the majority of their lifetime, are simply represented as two
232 bitmaps, one that keeps track of values present in the set, and one
233 that keeps track of expressions present in the set.
235 When we need them in topological order, we produce it on demand by
236 transforming the bitmap into an array and sorting it into topo
237 order. */
239 /* Type of expression, used to know which member of the PRE_EXPR union
240 is valid. */
242 enum pre_expr_kind
244 NAME,
245 NARY,
246 REFERENCE,
247 CONSTANT
250 union pre_expr_union
252 tree name;
253 tree constant;
254 vn_nary_op_t nary;
255 vn_reference_t reference;
258 typedef struct pre_expr_d : nofree_ptr_hash <pre_expr_d>
260 enum pre_expr_kind kind;
261 unsigned int id;
262 pre_expr_union u;
264 /* hash_table support. */
265 static inline hashval_t hash (const pre_expr_d *);
266 static inline int equal (const pre_expr_d *, const pre_expr_d *);
267 } *pre_expr;
269 #define PRE_EXPR_NAME(e) (e)->u.name
270 #define PRE_EXPR_NARY(e) (e)->u.nary
271 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
272 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
274 /* Compare E1 and E1 for equality. */
276 inline int
277 pre_expr_d::equal (const pre_expr_d *e1, const pre_expr_d *e2)
279 if (e1->kind != e2->kind)
280 return false;
282 switch (e1->kind)
284 case CONSTANT:
285 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
286 PRE_EXPR_CONSTANT (e2));
287 case NAME:
288 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
289 case NARY:
290 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
291 case REFERENCE:
292 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
293 PRE_EXPR_REFERENCE (e2));
294 default:
295 gcc_unreachable ();
299 /* Hash E. */
301 inline hashval_t
302 pre_expr_d::hash (const pre_expr_d *e)
304 switch (e->kind)
306 case CONSTANT:
307 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
308 case NAME:
309 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
310 case NARY:
311 return PRE_EXPR_NARY (e)->hashcode;
312 case REFERENCE:
313 return PRE_EXPR_REFERENCE (e)->hashcode;
314 default:
315 gcc_unreachable ();
319 /* Next global expression id number. */
320 static unsigned int next_expression_id;
322 /* Mapping from expression to id number we can use in bitmap sets. */
323 static vec<pre_expr> expressions;
324 static hash_table<pre_expr_d> *expression_to_id;
325 static vec<unsigned> name_to_id;
327 /* Allocate an expression id for EXPR. */
329 static inline unsigned int
330 alloc_expression_id (pre_expr expr)
332 struct pre_expr_d **slot;
333 /* Make sure we won't overflow. */
334 gcc_assert (next_expression_id + 1 > next_expression_id);
335 expr->id = next_expression_id++;
336 expressions.safe_push (expr);
337 if (expr->kind == NAME)
339 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
340 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
341 re-allocations by using vec::reserve upfront. */
342 unsigned old_len = name_to_id.length ();
343 name_to_id.reserve (num_ssa_names - old_len);
344 name_to_id.quick_grow_cleared (num_ssa_names);
345 gcc_assert (name_to_id[version] == 0);
346 name_to_id[version] = expr->id;
348 else
350 slot = expression_to_id->find_slot (expr, INSERT);
351 gcc_assert (!*slot);
352 *slot = expr;
354 return next_expression_id - 1;
357 /* Return the expression id for tree EXPR. */
359 static inline unsigned int
360 get_expression_id (const pre_expr expr)
362 return expr->id;
365 static inline unsigned int
366 lookup_expression_id (const pre_expr expr)
368 struct pre_expr_d **slot;
370 if (expr->kind == NAME)
372 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
373 if (name_to_id.length () <= version)
374 return 0;
375 return name_to_id[version];
377 else
379 slot = expression_to_id->find_slot (expr, NO_INSERT);
380 if (!slot)
381 return 0;
382 return ((pre_expr)*slot)->id;
386 /* Return the existing expression id for EXPR, or create one if one
387 does not exist yet. */
389 static inline unsigned int
390 get_or_alloc_expression_id (pre_expr expr)
392 unsigned int id = lookup_expression_id (expr);
393 if (id == 0)
394 return alloc_expression_id (expr);
395 return expr->id = id;
398 /* Return the expression that has expression id ID */
400 static inline pre_expr
401 expression_for_id (unsigned int id)
403 return expressions[id];
406 /* Free the expression id field in all of our expressions,
407 and then destroy the expressions array. */
409 static void
410 clear_expression_ids (void)
412 expressions.release ();
415 static object_allocator<pre_expr_d> pre_expr_pool ("pre_expr nodes");
417 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
419 static pre_expr
420 get_or_alloc_expr_for_name (tree name)
422 struct pre_expr_d expr;
423 pre_expr result;
424 unsigned int result_id;
426 expr.kind = NAME;
427 expr.id = 0;
428 PRE_EXPR_NAME (&expr) = name;
429 result_id = lookup_expression_id (&expr);
430 if (result_id != 0)
431 return expression_for_id (result_id);
433 result = pre_expr_pool.allocate ();
434 result->kind = NAME;
435 PRE_EXPR_NAME (result) = name;
436 alloc_expression_id (result);
437 return result;
440 /* An unordered bitmap set. One bitmap tracks values, the other,
441 expressions. */
442 typedef struct bitmap_set
444 bitmap_head expressions;
445 bitmap_head values;
446 } *bitmap_set_t;
448 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
449 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
451 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
452 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
454 /* Mapping from value id to expressions with that value_id. */
455 static vec<bitmap> value_expressions;
457 /* Sets that we need to keep track of. */
458 typedef struct bb_bitmap_sets
460 /* The EXP_GEN set, which represents expressions/values generated in
461 a basic block. */
462 bitmap_set_t exp_gen;
464 /* The PHI_GEN set, which represents PHI results generated in a
465 basic block. */
466 bitmap_set_t phi_gen;
468 /* The TMP_GEN set, which represents results/temporaries generated
469 in a basic block. IE the LHS of an expression. */
470 bitmap_set_t tmp_gen;
472 /* The AVAIL_OUT set, which represents which values are available in
473 a given basic block. */
474 bitmap_set_t avail_out;
476 /* The ANTIC_IN set, which represents which values are anticipatable
477 in a given basic block. */
478 bitmap_set_t antic_in;
480 /* The PA_IN set, which represents which values are
481 partially anticipatable in a given basic block. */
482 bitmap_set_t pa_in;
484 /* The NEW_SETS set, which is used during insertion to augment the
485 AVAIL_OUT set of blocks with the new insertions performed during
486 the current iteration. */
487 bitmap_set_t new_sets;
489 /* A cache for value_dies_in_block_x. */
490 bitmap expr_dies;
492 /* The live virtual operand on successor edges. */
493 tree vop_on_exit;
495 /* True if we have visited this block during ANTIC calculation. */
496 unsigned int visited : 1;
498 /* True when the block contains a call that might not return. */
499 unsigned int contains_may_not_return_call : 1;
500 } *bb_value_sets_t;
502 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
503 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
504 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
505 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
506 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
507 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
508 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
509 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
510 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
511 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
512 #define BB_LIVE_VOP_ON_EXIT(BB) ((bb_value_sets_t) ((BB)->aux))->vop_on_exit
515 /* This structure is used to keep track of statistics on what
516 optimization PRE was able to perform. */
517 static struct
519 /* The number of RHS computations eliminated by PRE. */
520 int eliminations;
522 /* The number of new expressions/temporaries generated by PRE. */
523 int insertions;
525 /* The number of inserts found due to partial anticipation */
526 int pa_insert;
528 /* The number of inserts made for code hoisting. */
529 int hoist_insert;
531 /* The number of new PHI nodes added by PRE. */
532 int phis;
533 } pre_stats;
535 static bool do_partial_partial;
536 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int);
537 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
538 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
539 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
540 static void bitmap_set_and (bitmap_set_t, bitmap_set_t);
541 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
542 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
543 static bitmap_set_t bitmap_set_new (void);
544 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
545 tree);
546 static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
547 static unsigned int get_expr_value_id (pre_expr);
549 /* We can add and remove elements and entries to and from sets
550 and hash tables, so we use alloc pools for them. */
552 static object_allocator<bitmap_set> bitmap_set_pool ("Bitmap sets");
553 static bitmap_obstack grand_bitmap_obstack;
555 /* Set of blocks with statements that have had their EH properties changed. */
556 static bitmap need_eh_cleanup;
558 /* Set of blocks with statements that have had their AB properties changed. */
559 static bitmap need_ab_cleanup;
561 /* A three tuple {e, pred, v} used to cache phi translations in the
562 phi_translate_table. */
564 typedef struct expr_pred_trans_d : free_ptr_hash<expr_pred_trans_d>
566 /* The expression. */
567 pre_expr e;
569 /* The predecessor block along which we translated the expression. */
570 basic_block pred;
572 /* The value that resulted from the translation. */
573 pre_expr v;
575 /* The hashcode for the expression, pred pair. This is cached for
576 speed reasons. */
577 hashval_t hashcode;
579 /* hash_table support. */
580 static inline hashval_t hash (const expr_pred_trans_d *);
581 static inline int equal (const expr_pred_trans_d *, const expr_pred_trans_d *);
582 } *expr_pred_trans_t;
583 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
585 inline hashval_t
586 expr_pred_trans_d::hash (const expr_pred_trans_d *e)
588 return e->hashcode;
591 inline int
592 expr_pred_trans_d::equal (const expr_pred_trans_d *ve1,
593 const expr_pred_trans_d *ve2)
595 basic_block b1 = ve1->pred;
596 basic_block b2 = ve2->pred;
598 /* If they are not translations for the same basic block, they can't
599 be equal. */
600 if (b1 != b2)
601 return false;
602 return pre_expr_d::equal (ve1->e, ve2->e);
605 /* The phi_translate_table caches phi translations for a given
606 expression and predecessor. */
607 static hash_table<expr_pred_trans_d> *phi_translate_table;
609 /* Add the tuple mapping from {expression E, basic block PRED} to
610 the phi translation table and return whether it pre-existed. */
612 static inline bool
613 phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred)
615 expr_pred_trans_t *slot;
616 expr_pred_trans_d tem;
617 hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e),
618 pred->index);
619 tem.e = e;
620 tem.pred = pred;
621 tem.hashcode = hash;
622 slot = phi_translate_table->find_slot_with_hash (&tem, hash, INSERT);
623 if (*slot)
625 *entry = *slot;
626 return true;
629 *entry = *slot = XNEW (struct expr_pred_trans_d);
630 (*entry)->e = e;
631 (*entry)->pred = pred;
632 (*entry)->hashcode = hash;
633 return false;
637 /* Add expression E to the expression set of value id V. */
639 static void
640 add_to_value (unsigned int v, pre_expr e)
642 bitmap set;
644 gcc_checking_assert (get_expr_value_id (e) == v);
646 if (v >= value_expressions.length ())
648 value_expressions.safe_grow_cleared (v + 1);
651 set = value_expressions[v];
652 if (!set)
654 set = BITMAP_ALLOC (&grand_bitmap_obstack);
655 value_expressions[v] = set;
658 bitmap_set_bit (set, get_or_alloc_expression_id (e));
661 /* Create a new bitmap set and return it. */
663 static bitmap_set_t
664 bitmap_set_new (void)
666 bitmap_set_t ret = bitmap_set_pool.allocate ();
667 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
668 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
669 return ret;
672 /* Return the value id for a PRE expression EXPR. */
674 static unsigned int
675 get_expr_value_id (pre_expr expr)
677 unsigned int id;
678 switch (expr->kind)
680 case CONSTANT:
681 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
682 break;
683 case NAME:
684 id = VN_INFO (PRE_EXPR_NAME (expr))->value_id;
685 break;
686 case NARY:
687 id = PRE_EXPR_NARY (expr)->value_id;
688 break;
689 case REFERENCE:
690 id = PRE_EXPR_REFERENCE (expr)->value_id;
691 break;
692 default:
693 gcc_unreachable ();
695 /* ??? We cannot assert that expr has a value-id (it can be 0), because
696 we assign value-ids only to expressions that have a result
697 in set_hashtable_value_ids. */
698 return id;
701 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
703 static tree
704 sccvn_valnum_from_value_id (unsigned int val)
706 bitmap_iterator bi;
707 unsigned int i;
708 bitmap exprset = value_expressions[val];
709 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
711 pre_expr vexpr = expression_for_id (i);
712 if (vexpr->kind == NAME)
713 return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
714 else if (vexpr->kind == CONSTANT)
715 return PRE_EXPR_CONSTANT (vexpr);
717 return NULL_TREE;
720 /* Remove an expression EXPR from a bitmapped set. */
722 static void
723 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
725 unsigned int val = get_expr_value_id (expr);
726 if (!value_id_constant_p (val))
728 bitmap_clear_bit (&set->values, val);
729 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
733 /* Insert an expression EXPR into a bitmapped set. */
735 static void
736 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
738 unsigned int val = get_expr_value_id (expr);
739 if (! value_id_constant_p (val))
741 /* Note this is the only function causing multiple expressions
742 for the same value to appear in a set. This is needed for
743 TMP_GEN, PHI_GEN and NEW_SETs. */
744 bitmap_set_bit (&set->values, val);
745 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
749 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
751 static void
752 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
754 bitmap_copy (&dest->expressions, &orig->expressions);
755 bitmap_copy (&dest->values, &orig->values);
759 /* Free memory used up by SET. */
760 static void
761 bitmap_set_free (bitmap_set_t set)
763 bitmap_clear (&set->expressions);
764 bitmap_clear (&set->values);
768 /* Generate an topological-ordered array of bitmap set SET. */
770 static vec<pre_expr>
771 sorted_array_from_bitmap_set (bitmap_set_t set)
773 unsigned int i, j;
774 bitmap_iterator bi, bj;
775 vec<pre_expr> result;
777 /* Pre-allocate enough space for the array. */
778 result.create (bitmap_count_bits (&set->expressions));
780 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
782 /* The number of expressions having a given value is usually
783 relatively small. Thus, rather than making a vector of all
784 the expressions and sorting it by value-id, we walk the values
785 and check in the reverse mapping that tells us what expressions
786 have a given value, to filter those in our set. As a result,
787 the expressions are inserted in value-id order, which means
788 topological order.
790 If this is somehow a significant lose for some cases, we can
791 choose which set to walk based on the set size. */
792 bitmap exprset = value_expressions[i];
793 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj)
795 if (bitmap_bit_p (&set->expressions, j))
796 result.quick_push (expression_for_id (j));
800 return result;
803 /* Perform bitmapped set operation DEST &= ORIG. */
805 static void
806 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
808 bitmap_iterator bi;
809 unsigned int i;
811 if (dest != orig)
813 bitmap_and_into (&dest->values, &orig->values);
815 unsigned int to_clear = -1U;
816 FOR_EACH_EXPR_ID_IN_SET (dest, i, bi)
818 if (to_clear != -1U)
820 bitmap_clear_bit (&dest->expressions, to_clear);
821 to_clear = -1U;
823 pre_expr expr = expression_for_id (i);
824 unsigned int value_id = get_expr_value_id (expr);
825 if (!bitmap_bit_p (&dest->values, value_id))
826 to_clear = i;
828 if (to_clear != -1U)
829 bitmap_clear_bit (&dest->expressions, to_clear);
833 /* Subtract all expressions contained in ORIG from DEST. */
835 static bitmap_set_t
836 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
838 bitmap_set_t result = bitmap_set_new ();
839 bitmap_iterator bi;
840 unsigned int i;
842 bitmap_and_compl (&result->expressions, &dest->expressions,
843 &orig->expressions);
845 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
847 pre_expr expr = expression_for_id (i);
848 unsigned int value_id = get_expr_value_id (expr);
849 bitmap_set_bit (&result->values, value_id);
852 return result;
855 /* Subtract all values in bitmap set B from bitmap set A. */
857 static void
858 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
860 unsigned int i;
861 bitmap_iterator bi;
862 pre_expr to_remove = NULL;
863 FOR_EACH_EXPR_ID_IN_SET (a, i, bi)
865 if (to_remove)
867 bitmap_remove_from_set (a, to_remove);
868 to_remove = NULL;
870 pre_expr expr = expression_for_id (i);
871 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
872 to_remove = expr;
874 if (to_remove)
875 bitmap_remove_from_set (a, to_remove);
879 /* Return true if bitmapped set SET contains the value VALUE_ID. */
881 static bool
882 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
884 if (value_id_constant_p (value_id))
885 return true;
887 if (!set || bitmap_empty_p (&set->expressions))
888 return false;
890 return bitmap_bit_p (&set->values, value_id);
893 static inline bool
894 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
896 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
899 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
901 static void
902 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
903 const pre_expr expr)
905 bitmap exprset;
906 unsigned int i;
907 bitmap_iterator bi;
909 if (value_id_constant_p (lookfor))
910 return;
912 if (!bitmap_set_contains_value (set, lookfor))
913 return;
915 /* The number of expressions having a given value is usually
916 significantly less than the total number of expressions in SET.
917 Thus, rather than check, for each expression in SET, whether it
918 has the value LOOKFOR, we walk the reverse mapping that tells us
919 what expressions have a given value, and see if any of those
920 expressions are in our set. For large testcases, this is about
921 5-10x faster than walking the bitmap. If this is somehow a
922 significant lose for some cases, we can choose which set to walk
923 based on the set size. */
924 exprset = value_expressions[lookfor];
925 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
927 if (bitmap_clear_bit (&set->expressions, i))
929 bitmap_set_bit (&set->expressions, get_expression_id (expr));
930 return;
934 gcc_unreachable ();
937 /* Return true if two bitmap sets are equal. */
939 static bool
940 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
942 return bitmap_equal_p (&a->values, &b->values);
945 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
946 and add it otherwise. */
948 static void
949 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
951 unsigned int val = get_expr_value_id (expr);
953 if (bitmap_set_contains_value (set, val))
954 bitmap_set_replace_value (set, val, expr);
955 else
956 bitmap_insert_into_set (set, expr);
959 /* Insert EXPR into SET if EXPR's value is not already present in
960 SET. */
962 static void
963 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
965 unsigned int val = get_expr_value_id (expr);
967 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
969 /* Constant values are always considered to be part of the set. */
970 if (value_id_constant_p (val))
971 return;
973 /* If the value membership changed, add the expression. */
974 if (bitmap_set_bit (&set->values, val))
975 bitmap_set_bit (&set->expressions, expr->id);
978 /* Print out EXPR to outfile. */
980 static void
981 print_pre_expr (FILE *outfile, const pre_expr expr)
983 if (! expr)
985 fprintf (outfile, "NULL");
986 return;
988 switch (expr->kind)
990 case CONSTANT:
991 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr));
992 break;
993 case NAME:
994 print_generic_expr (outfile, PRE_EXPR_NAME (expr));
995 break;
996 case NARY:
998 unsigned int i;
999 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1000 fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode));
1001 for (i = 0; i < nary->length; i++)
1003 print_generic_expr (outfile, nary->op[i]);
1004 if (i != (unsigned) nary->length - 1)
1005 fprintf (outfile, ",");
1007 fprintf (outfile, "}");
1009 break;
1011 case REFERENCE:
1013 vn_reference_op_t vro;
1014 unsigned int i;
1015 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1016 fprintf (outfile, "{");
1017 for (i = 0;
1018 ref->operands.iterate (i, &vro);
1019 i++)
1021 bool closebrace = false;
1022 if (vro->opcode != SSA_NAME
1023 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
1025 fprintf (outfile, "%s", get_tree_code_name (vro->opcode));
1026 if (vro->op0)
1028 fprintf (outfile, "<");
1029 closebrace = true;
1032 if (vro->op0)
1034 print_generic_expr (outfile, vro->op0);
1035 if (vro->op1)
1037 fprintf (outfile, ",");
1038 print_generic_expr (outfile, vro->op1);
1040 if (vro->op2)
1042 fprintf (outfile, ",");
1043 print_generic_expr (outfile, vro->op2);
1046 if (closebrace)
1047 fprintf (outfile, ">");
1048 if (i != ref->operands.length () - 1)
1049 fprintf (outfile, ",");
1051 fprintf (outfile, "}");
1052 if (ref->vuse)
1054 fprintf (outfile, "@");
1055 print_generic_expr (outfile, ref->vuse);
1058 break;
1061 void debug_pre_expr (pre_expr);
1063 /* Like print_pre_expr but always prints to stderr. */
1064 DEBUG_FUNCTION void
1065 debug_pre_expr (pre_expr e)
1067 print_pre_expr (stderr, e);
1068 fprintf (stderr, "\n");
1071 /* Print out SET to OUTFILE. */
1073 static void
1074 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1075 const char *setname, int blockindex)
1077 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1078 if (set)
1080 bool first = true;
1081 unsigned i;
1082 bitmap_iterator bi;
1084 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1086 const pre_expr expr = expression_for_id (i);
1088 if (!first)
1089 fprintf (outfile, ", ");
1090 first = false;
1091 print_pre_expr (outfile, expr);
1093 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1096 fprintf (outfile, " }\n");
1099 void debug_bitmap_set (bitmap_set_t);
1101 DEBUG_FUNCTION void
1102 debug_bitmap_set (bitmap_set_t set)
1104 print_bitmap_set (stderr, set, "debug", 0);
1107 void debug_bitmap_sets_for (basic_block);
1109 DEBUG_FUNCTION void
1110 debug_bitmap_sets_for (basic_block bb)
1112 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1113 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1114 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1115 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1116 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1117 if (do_partial_partial)
1118 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1119 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1122 /* Print out the expressions that have VAL to OUTFILE. */
1124 static void
1125 print_value_expressions (FILE *outfile, unsigned int val)
1127 bitmap set = value_expressions[val];
1128 if (set)
1130 bitmap_set x;
1131 char s[10];
1132 sprintf (s, "%04d", val);
1133 x.expressions = *set;
1134 print_bitmap_set (outfile, &x, s, 0);
1139 DEBUG_FUNCTION void
1140 debug_value_expressions (unsigned int val)
1142 print_value_expressions (stderr, val);
1145 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1146 represent it. */
1148 static pre_expr
1149 get_or_alloc_expr_for_constant (tree constant)
1151 unsigned int result_id;
1152 unsigned int value_id;
1153 struct pre_expr_d expr;
1154 pre_expr newexpr;
1156 expr.kind = CONSTANT;
1157 PRE_EXPR_CONSTANT (&expr) = constant;
1158 result_id = lookup_expression_id (&expr);
1159 if (result_id != 0)
1160 return expression_for_id (result_id);
1162 newexpr = pre_expr_pool.allocate ();
1163 newexpr->kind = CONSTANT;
1164 PRE_EXPR_CONSTANT (newexpr) = constant;
1165 alloc_expression_id (newexpr);
1166 value_id = get_or_alloc_constant_value_id (constant);
1167 add_to_value (value_id, newexpr);
1168 return newexpr;
1171 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1172 Currently only supports constants and SSA_NAMES. */
1173 static pre_expr
1174 get_or_alloc_expr_for (tree t)
1176 if (TREE_CODE (t) == SSA_NAME)
1177 return get_or_alloc_expr_for_name (t);
1178 else if (is_gimple_min_invariant (t))
1179 return get_or_alloc_expr_for_constant (t);
1180 gcc_unreachable ();
1183 /* Return the folded version of T if T, when folded, is a gimple
1184 min_invariant or an SSA name. Otherwise, return T. */
1186 static pre_expr
1187 fully_constant_expression (pre_expr e)
1189 switch (e->kind)
1191 case CONSTANT:
1192 return e;
1193 case NARY:
1195 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1196 tree res = vn_nary_simplify (nary);
1197 if (!res)
1198 return e;
1199 if (is_gimple_min_invariant (res))
1200 return get_or_alloc_expr_for_constant (res);
1201 if (TREE_CODE (res) == SSA_NAME)
1202 return get_or_alloc_expr_for_name (res);
1203 return e;
1205 case REFERENCE:
1207 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1208 tree folded;
1209 if ((folded = fully_constant_vn_reference_p (ref)))
1210 return get_or_alloc_expr_for_constant (folded);
1211 return e;
1213 default:
1214 return e;
1216 return e;
1219 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1220 it has the value it would have in BLOCK. Set *SAME_VALID to true
1221 in case the new vuse doesn't change the value id of the OPERANDS. */
1223 static tree
1224 translate_vuse_through_block (vec<vn_reference_op_s> operands,
1225 alias_set_type set, tree type, tree vuse,
1226 basic_block phiblock,
1227 basic_block block, bool *same_valid)
1229 gimple *phi = SSA_NAME_DEF_STMT (vuse);
1230 ao_ref ref;
1231 edge e = NULL;
1232 bool use_oracle;
1234 *same_valid = true;
1236 if (gimple_bb (phi) != phiblock)
1237 return vuse;
1239 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1241 /* Use the alias-oracle to find either the PHI node in this block,
1242 the first VUSE used in this block that is equivalent to vuse or
1243 the first VUSE which definition in this block kills the value. */
1244 if (gimple_code (phi) == GIMPLE_PHI)
1245 e = find_edge (block, phiblock);
1246 else if (use_oracle)
1247 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1249 vuse = gimple_vuse (phi);
1250 phi = SSA_NAME_DEF_STMT (vuse);
1251 if (gimple_bb (phi) != phiblock)
1252 return vuse;
1253 if (gimple_code (phi) == GIMPLE_PHI)
1255 e = find_edge (block, phiblock);
1256 break;
1259 else
1260 return NULL_TREE;
1262 if (e)
1264 if (use_oracle)
1266 bitmap visited = NULL;
1267 unsigned int cnt;
1268 /* Try to find a vuse that dominates this phi node by skipping
1269 non-clobbering statements. */
1270 vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false,
1271 NULL, NULL);
1272 if (visited)
1273 BITMAP_FREE (visited);
1275 else
1276 vuse = NULL_TREE;
1277 if (!vuse)
1279 /* If we didn't find any, the value ID can't stay the same,
1280 but return the translated vuse. */
1281 *same_valid = false;
1282 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1284 /* ??? We would like to return vuse here as this is the canonical
1285 upmost vdef that this reference is associated with. But during
1286 insertion of the references into the hash tables we only ever
1287 directly insert with their direct gimple_vuse, hence returning
1288 something else would make us not find the other expression. */
1289 return PHI_ARG_DEF (phi, e->dest_idx);
1292 return NULL_TREE;
1295 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1296 SET2 *or* SET3. This is used to avoid making a set consisting of the union
1297 of PA_IN and ANTIC_IN during insert and phi-translation. */
1299 static inline pre_expr
1300 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2,
1301 bitmap_set_t set3 = NULL)
1303 pre_expr result;
1305 result = bitmap_find_leader (set1, val);
1306 if (!result && set2)
1307 result = bitmap_find_leader (set2, val);
1308 if (!result && set3)
1309 result = bitmap_find_leader (set3, val);
1310 return result;
1313 /* Get the tree type for our PRE expression e. */
1315 static tree
1316 get_expr_type (const pre_expr e)
1318 switch (e->kind)
1320 case NAME:
1321 return TREE_TYPE (PRE_EXPR_NAME (e));
1322 case CONSTANT:
1323 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1324 case REFERENCE:
1325 return PRE_EXPR_REFERENCE (e)->type;
1326 case NARY:
1327 return PRE_EXPR_NARY (e)->type;
1329 gcc_unreachable ();
1332 /* Get a representative SSA_NAME for a given expression.
1333 Since all of our sub-expressions are treated as values, we require
1334 them to be SSA_NAME's for simplicity.
1335 Prior versions of GVNPRE used to use "value handles" here, so that
1336 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1337 either case, the operands are really values (IE we do not expect
1338 them to be usable without finding leaders). */
1340 static tree
1341 get_representative_for (const pre_expr e)
1343 tree name;
1344 unsigned int value_id = get_expr_value_id (e);
1346 switch (e->kind)
1348 case NAME:
1349 return VN_INFO (PRE_EXPR_NAME (e))->valnum;
1350 case CONSTANT:
1351 return PRE_EXPR_CONSTANT (e);
1352 case NARY:
1353 case REFERENCE:
1355 /* Go through all of the expressions representing this value
1356 and pick out an SSA_NAME. */
1357 unsigned int i;
1358 bitmap_iterator bi;
1359 bitmap exprs = value_expressions[value_id];
1360 EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1362 pre_expr rep = expression_for_id (i);
1363 if (rep->kind == NAME)
1364 return VN_INFO (PRE_EXPR_NAME (rep))->valnum;
1365 else if (rep->kind == CONSTANT)
1366 return PRE_EXPR_CONSTANT (rep);
1369 break;
1372 /* If we reached here we couldn't find an SSA_NAME. This can
1373 happen when we've discovered a value that has never appeared in
1374 the program as set to an SSA_NAME, as the result of phi translation.
1375 Create one here.
1376 ??? We should be able to re-use this when we insert the statement
1377 to compute it. */
1378 name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1379 VN_INFO_GET (name)->value_id = value_id;
1380 VN_INFO (name)->valnum = name;
1381 /* ??? For now mark this SSA name for release by SCCVN. */
1382 VN_INFO (name)->needs_insertion = true;
1383 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1384 if (dump_file && (dump_flags & TDF_DETAILS))
1386 fprintf (dump_file, "Created SSA_NAME representative ");
1387 print_generic_expr (dump_file, name);
1388 fprintf (dump_file, " for expression:");
1389 print_pre_expr (dump_file, e);
1390 fprintf (dump_file, " (%04d)\n", value_id);
1393 return name;
1398 static pre_expr
1399 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1400 basic_block pred, basic_block phiblock);
1402 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1403 the phis in PRED. Return NULL if we can't find a leader for each part
1404 of the translated expression. */
1406 static pre_expr
1407 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1408 basic_block pred, basic_block phiblock)
1410 switch (expr->kind)
1412 case NARY:
1414 unsigned int i;
1415 bool changed = false;
1416 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1417 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1418 sizeof_vn_nary_op (nary->length));
1419 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1421 for (i = 0; i < newnary->length; i++)
1423 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1424 continue;
1425 else
1427 pre_expr leader, result;
1428 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1429 leader = find_leader_in_sets (op_val_id, set1, set2);
1430 result = phi_translate (leader, set1, set2, pred, phiblock);
1431 if (result && result != leader)
1432 newnary->op[i] = get_representative_for (result);
1433 else if (!result)
1434 return NULL;
1436 changed |= newnary->op[i] != nary->op[i];
1439 if (changed)
1441 pre_expr constant;
1442 unsigned int new_val_id;
1444 PRE_EXPR_NARY (expr) = newnary;
1445 constant = fully_constant_expression (expr);
1446 PRE_EXPR_NARY (expr) = nary;
1447 if (constant != expr)
1449 /* For non-CONSTANTs we have to make sure we can eventually
1450 insert the expression. Which means we need to have a
1451 leader for it. */
1452 if (constant->kind != CONSTANT)
1454 /* Do not allow simplifications to non-constants over
1455 backedges as this will likely result in a loop PHI node
1456 to be inserted and increased register pressure.
1457 See PR77498 - this avoids doing predcoms work in
1458 a less efficient way. */
1459 if (find_edge (pred, phiblock)->flags & EDGE_DFS_BACK)
1461 else
1463 unsigned value_id = get_expr_value_id (constant);
1464 constant = find_leader_in_sets (value_id, set1, set2,
1465 AVAIL_OUT (pred));
1466 if (constant)
1467 return constant;
1470 else
1471 return constant;
1474 tree result = vn_nary_op_lookup_pieces (newnary->length,
1475 newnary->opcode,
1476 newnary->type,
1477 &newnary->op[0],
1478 &nary);
1479 if (result && is_gimple_min_invariant (result))
1480 return get_or_alloc_expr_for_constant (result);
1482 expr = pre_expr_pool.allocate ();
1483 expr->kind = NARY;
1484 expr->id = 0;
1485 if (nary)
1487 PRE_EXPR_NARY (expr) = nary;
1488 new_val_id = nary->value_id;
1489 get_or_alloc_expression_id (expr);
1490 /* When we end up re-using a value number make sure that
1491 doesn't have unrelated (which we can't check here)
1492 range or points-to info on it. */
1493 if (result
1494 && INTEGRAL_TYPE_P (TREE_TYPE (result))
1495 && SSA_NAME_RANGE_INFO (result)
1496 && ! SSA_NAME_IS_DEFAULT_DEF (result))
1498 if (! VN_INFO (result)->info.range_info)
1500 VN_INFO (result)->info.range_info
1501 = SSA_NAME_RANGE_INFO (result);
1502 VN_INFO (result)->range_info_anti_range_p
1503 = SSA_NAME_ANTI_RANGE_P (result);
1505 if (dump_file && (dump_flags & TDF_DETAILS))
1507 fprintf (dump_file, "clearing range info of ");
1508 print_generic_expr (dump_file, result);
1509 fprintf (dump_file, "\n");
1511 SSA_NAME_RANGE_INFO (result) = NULL;
1513 else if (result
1514 && POINTER_TYPE_P (TREE_TYPE (result))
1515 && SSA_NAME_PTR_INFO (result)
1516 && ! SSA_NAME_IS_DEFAULT_DEF (result))
1518 if (! VN_INFO (result)->info.ptr_info)
1519 VN_INFO (result)->info.ptr_info
1520 = SSA_NAME_PTR_INFO (result);
1521 if (dump_file && (dump_flags & TDF_DETAILS))
1523 fprintf (dump_file, "clearing points-to info of ");
1524 print_generic_expr (dump_file, result);
1525 fprintf (dump_file, "\n");
1527 SSA_NAME_PTR_INFO (result) = NULL;
1530 else
1532 new_val_id = get_next_value_id ();
1533 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
1534 nary = vn_nary_op_insert_pieces (newnary->length,
1535 newnary->opcode,
1536 newnary->type,
1537 &newnary->op[0],
1538 result, new_val_id);
1539 PRE_EXPR_NARY (expr) = nary;
1540 get_or_alloc_expression_id (expr);
1542 add_to_value (new_val_id, expr);
1544 return expr;
1546 break;
1548 case REFERENCE:
1550 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1551 vec<vn_reference_op_s> operands = ref->operands;
1552 tree vuse = ref->vuse;
1553 tree newvuse = vuse;
1554 vec<vn_reference_op_s> newoperands = vNULL;
1555 bool changed = false, same_valid = true;
1556 unsigned int i, n;
1557 vn_reference_op_t operand;
1558 vn_reference_t newref;
1560 for (i = 0; operands.iterate (i, &operand); i++)
1562 pre_expr opresult;
1563 pre_expr leader;
1564 tree op[3];
1565 tree type = operand->type;
1566 vn_reference_op_s newop = *operand;
1567 op[0] = operand->op0;
1568 op[1] = operand->op1;
1569 op[2] = operand->op2;
1570 for (n = 0; n < 3; ++n)
1572 unsigned int op_val_id;
1573 if (!op[n])
1574 continue;
1575 if (TREE_CODE (op[n]) != SSA_NAME)
1577 /* We can't possibly insert these. */
1578 if (n != 0
1579 && !is_gimple_min_invariant (op[n]))
1580 break;
1581 continue;
1583 op_val_id = VN_INFO (op[n])->value_id;
1584 leader = find_leader_in_sets (op_val_id, set1, set2);
1585 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1586 if (opresult && opresult != leader)
1588 tree name = get_representative_for (opresult);
1589 changed |= name != op[n];
1590 op[n] = name;
1592 else if (!opresult)
1593 break;
1595 if (n != 3)
1597 newoperands.release ();
1598 return NULL;
1600 if (!changed)
1601 continue;
1602 if (!newoperands.exists ())
1603 newoperands = operands.copy ();
1604 /* We may have changed from an SSA_NAME to a constant */
1605 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1606 newop.opcode = TREE_CODE (op[0]);
1607 newop.type = type;
1608 newop.op0 = op[0];
1609 newop.op1 = op[1];
1610 newop.op2 = op[2];
1611 newoperands[i] = newop;
1613 gcc_checking_assert (i == operands.length ());
1615 if (vuse)
1617 newvuse = translate_vuse_through_block (newoperands.exists ()
1618 ? newoperands : operands,
1619 ref->set, ref->type,
1620 vuse, phiblock, pred,
1621 &same_valid);
1622 if (newvuse == NULL_TREE)
1624 newoperands.release ();
1625 return NULL;
1629 if (changed || newvuse != vuse)
1631 unsigned int new_val_id;
1632 pre_expr constant;
1634 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1635 ref->type,
1636 newoperands.exists ()
1637 ? newoperands : operands,
1638 &newref, VN_WALK);
1639 if (result)
1640 newoperands.release ();
1642 /* We can always insert constants, so if we have a partial
1643 redundant constant load of another type try to translate it
1644 to a constant of appropriate type. */
1645 if (result && is_gimple_min_invariant (result))
1647 tree tem = result;
1648 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1650 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1651 if (tem && !is_gimple_min_invariant (tem))
1652 tem = NULL_TREE;
1654 if (tem)
1655 return get_or_alloc_expr_for_constant (tem);
1658 /* If we'd have to convert things we would need to validate
1659 if we can insert the translated expression. So fail
1660 here for now - we cannot insert an alias with a different
1661 type in the VN tables either, as that would assert. */
1662 if (result
1663 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1664 return NULL;
1665 else if (!result && newref
1666 && !useless_type_conversion_p (ref->type, newref->type))
1668 newoperands.release ();
1669 return NULL;
1672 expr = pre_expr_pool.allocate ();
1673 expr->kind = REFERENCE;
1674 expr->id = 0;
1676 if (newref)
1678 PRE_EXPR_REFERENCE (expr) = newref;
1679 constant = fully_constant_expression (expr);
1680 if (constant != expr)
1681 return constant;
1683 new_val_id = newref->value_id;
1684 get_or_alloc_expression_id (expr);
1686 else
1688 if (changed || !same_valid)
1690 new_val_id = get_next_value_id ();
1691 value_expressions.safe_grow_cleared
1692 (get_max_value_id () + 1);
1694 else
1695 new_val_id = ref->value_id;
1696 if (!newoperands.exists ())
1697 newoperands = operands.copy ();
1698 newref = vn_reference_insert_pieces (newvuse, ref->set,
1699 ref->type,
1700 newoperands,
1701 result, new_val_id);
1702 newoperands = vNULL;
1703 PRE_EXPR_REFERENCE (expr) = newref;
1704 constant = fully_constant_expression (expr);
1705 if (constant != expr)
1706 return constant;
1707 get_or_alloc_expression_id (expr);
1709 add_to_value (new_val_id, expr);
1711 newoperands.release ();
1712 return expr;
1714 break;
1716 case NAME:
1718 tree name = PRE_EXPR_NAME (expr);
1719 gimple *def_stmt = SSA_NAME_DEF_STMT (name);
1720 /* If the SSA name is defined by a PHI node in this block,
1721 translate it. */
1722 if (gimple_code (def_stmt) == GIMPLE_PHI
1723 && gimple_bb (def_stmt) == phiblock)
1725 edge e = find_edge (pred, gimple_bb (def_stmt));
1726 tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1728 /* Handle constant. */
1729 if (is_gimple_min_invariant (def))
1730 return get_or_alloc_expr_for_constant (def);
1732 return get_or_alloc_expr_for_name (def);
1734 /* Otherwise return it unchanged - it will get removed if its
1735 value is not available in PREDs AVAIL_OUT set of expressions
1736 by the subtraction of TMP_GEN. */
1737 return expr;
1740 default:
1741 gcc_unreachable ();
1745 /* Wrapper around phi_translate_1 providing caching functionality. */
1747 static pre_expr
1748 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1749 basic_block pred, basic_block phiblock)
1751 expr_pred_trans_t slot = NULL;
1752 pre_expr phitrans;
1754 if (!expr)
1755 return NULL;
1757 /* Constants contain no values that need translation. */
1758 if (expr->kind == CONSTANT)
1759 return expr;
1761 if (value_id_constant_p (get_expr_value_id (expr)))
1762 return expr;
1764 /* Don't add translations of NAMEs as those are cheap to translate. */
1765 if (expr->kind != NAME)
1767 if (phi_trans_add (&slot, expr, pred))
1768 return slot->v;
1769 /* Store NULL for the value we want to return in the case of
1770 recursing. */
1771 slot->v = NULL;
1774 /* Translate. */
1775 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1777 if (slot)
1779 if (phitrans)
1780 slot->v = phitrans;
1781 else
1782 /* Remove failed translations again, they cause insert
1783 iteration to not pick up new opportunities reliably. */
1784 phi_translate_table->remove_elt_with_hash (slot, slot->hashcode);
1787 return phitrans;
1791 /* For each expression in SET, translate the values through phi nodes
1792 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1793 expressions in DEST. */
1795 static void
1796 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1797 basic_block phiblock)
1799 vec<pre_expr> exprs;
1800 pre_expr expr;
1801 int i;
1803 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1805 bitmap_set_copy (dest, set);
1806 return;
1809 exprs = sorted_array_from_bitmap_set (set);
1810 FOR_EACH_VEC_ELT (exprs, i, expr)
1812 pre_expr translated;
1813 translated = phi_translate (expr, set, NULL, pred, phiblock);
1814 if (!translated)
1815 continue;
1817 /* We might end up with multiple expressions from SET being
1818 translated to the same value. In this case we do not want
1819 to retain the NARY or REFERENCE expression but prefer a NAME
1820 which would be the leader. */
1821 if (translated->kind == NAME)
1822 bitmap_value_replace_in_set (dest, translated);
1823 else
1824 bitmap_value_insert_into_set (dest, translated);
1826 exprs.release ();
1829 /* Find the leader for a value (i.e., the name representing that
1830 value) in a given set, and return it. Return NULL if no leader
1831 is found. */
1833 static pre_expr
1834 bitmap_find_leader (bitmap_set_t set, unsigned int val)
1836 if (value_id_constant_p (val))
1838 unsigned int i;
1839 bitmap_iterator bi;
1840 bitmap exprset = value_expressions[val];
1842 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1844 pre_expr expr = expression_for_id (i);
1845 if (expr->kind == CONSTANT)
1846 return expr;
1849 if (bitmap_set_contains_value (set, val))
1851 /* Rather than walk the entire bitmap of expressions, and see
1852 whether any of them has the value we are looking for, we look
1853 at the reverse mapping, which tells us the set of expressions
1854 that have a given value (IE value->expressions with that
1855 value) and see if any of those expressions are in our set.
1856 The number of expressions per value is usually significantly
1857 less than the number of expressions in the set. In fact, for
1858 large testcases, doing it this way is roughly 5-10x faster
1859 than walking the bitmap.
1860 If this is somehow a significant lose for some cases, we can
1861 choose which set to walk based on which set is smaller. */
1862 unsigned int i;
1863 bitmap_iterator bi;
1864 bitmap exprset = value_expressions[val];
1866 EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1867 return expression_for_id (i);
1869 return NULL;
1872 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1873 BLOCK by seeing if it is not killed in the block. Note that we are
1874 only determining whether there is a store that kills it. Because
1875 of the order in which clean iterates over values, we are guaranteed
1876 that altered operands will have caused us to be eliminated from the
1877 ANTIC_IN set already. */
1879 static bool
1880 value_dies_in_block_x (pre_expr expr, basic_block block)
1882 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1883 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1884 gimple *def;
1885 gimple_stmt_iterator gsi;
1886 unsigned id = get_expression_id (expr);
1887 bool res = false;
1888 ao_ref ref;
1890 if (!vuse)
1891 return false;
1893 /* Lookup a previously calculated result. */
1894 if (EXPR_DIES (block)
1895 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1896 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1898 /* A memory expression {e, VUSE} dies in the block if there is a
1899 statement that may clobber e. If, starting statement walk from the
1900 top of the basic block, a statement uses VUSE there can be no kill
1901 inbetween that use and the original statement that loaded {e, VUSE},
1902 so we can stop walking. */
1903 ref.base = NULL_TREE;
1904 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1906 tree def_vuse, def_vdef;
1907 def = gsi_stmt (gsi);
1908 def_vuse = gimple_vuse (def);
1909 def_vdef = gimple_vdef (def);
1911 /* Not a memory statement. */
1912 if (!def_vuse)
1913 continue;
1915 /* Not a may-def. */
1916 if (!def_vdef)
1918 /* A load with the same VUSE, we're done. */
1919 if (def_vuse == vuse)
1920 break;
1922 continue;
1925 /* Init ref only if we really need it. */
1926 if (ref.base == NULL_TREE
1927 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1928 refx->operands))
1930 res = true;
1931 break;
1933 /* If the statement may clobber expr, it dies. */
1934 if (stmt_may_clobber_ref_p_1 (def, &ref))
1936 res = true;
1937 break;
1941 /* Remember the result. */
1942 if (!EXPR_DIES (block))
1943 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1944 bitmap_set_bit (EXPR_DIES (block), id * 2);
1945 if (res)
1946 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1948 return res;
1952 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1953 contains its value-id. */
1955 static bool
1956 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1958 if (op && TREE_CODE (op) == SSA_NAME)
1960 unsigned int value_id = VN_INFO (op)->value_id;
1961 if (!(bitmap_set_contains_value (set1, value_id)
1962 || (set2 && bitmap_set_contains_value (set2, value_id))))
1963 return false;
1965 return true;
1968 /* Determine if the expression EXPR is valid in SET1 U SET2.
1969 ONLY SET2 CAN BE NULL.
1970 This means that we have a leader for each part of the expression
1971 (if it consists of values), or the expression is an SSA_NAME.
1972 For loads/calls, we also see if the vuse is killed in this block. */
1974 static bool
1975 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr)
1977 switch (expr->kind)
1979 case NAME:
1980 /* By construction all NAMEs are available. Non-available
1981 NAMEs are removed by subtracting TMP_GEN from the sets. */
1982 return true;
1983 case NARY:
1985 unsigned int i;
1986 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1987 for (i = 0; i < nary->length; i++)
1988 if (!op_valid_in_sets (set1, set2, nary->op[i]))
1989 return false;
1990 return true;
1992 break;
1993 case REFERENCE:
1995 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1996 vn_reference_op_t vro;
1997 unsigned int i;
1999 FOR_EACH_VEC_ELT (ref->operands, i, vro)
2001 if (!op_valid_in_sets (set1, set2, vro->op0)
2002 || !op_valid_in_sets (set1, set2, vro->op1)
2003 || !op_valid_in_sets (set1, set2, vro->op2))
2004 return false;
2006 return true;
2008 default:
2009 gcc_unreachable ();
2013 /* Clean the set of expressions that are no longer valid in SET1 or
2014 SET2. This means expressions that are made up of values we have no
2015 leaders for in SET1 or SET2. This version is used for partial
2016 anticipation, which means it is not valid in either ANTIC_IN or
2017 PA_IN. */
2019 static void
2020 dependent_clean (bitmap_set_t set1, bitmap_set_t set2)
2022 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
2023 pre_expr expr;
2024 int i;
2026 FOR_EACH_VEC_ELT (exprs, i, expr)
2028 if (!valid_in_sets (set1, set2, expr))
2029 bitmap_remove_from_set (set1, expr);
2031 exprs.release ();
2034 /* Clean the set of expressions that are no longer valid in SET. This
2035 means expressions that are made up of values we have no leaders for
2036 in SET. */
2038 static void
2039 clean (bitmap_set_t set)
2041 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set);
2042 pre_expr expr;
2043 int i;
2045 FOR_EACH_VEC_ELT (exprs, i, expr)
2047 if (!valid_in_sets (set, NULL, expr))
2048 bitmap_remove_from_set (set, expr);
2050 exprs.release ();
2053 /* Clean the set of expressions that are no longer valid in SET because
2054 they are clobbered in BLOCK or because they trap and may not be executed. */
2056 static void
2057 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2059 bitmap_iterator bi;
2060 unsigned i;
2061 pre_expr to_remove = NULL;
2063 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2065 /* Remove queued expr. */
2066 if (to_remove)
2068 bitmap_remove_from_set (set, to_remove);
2069 to_remove = NULL;
2072 pre_expr expr = expression_for_id (i);
2073 if (expr->kind == REFERENCE)
2075 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2076 if (ref->vuse)
2078 gimple *def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2079 if (!gimple_nop_p (def_stmt)
2080 && ((gimple_bb (def_stmt) != block
2081 && !dominated_by_p (CDI_DOMINATORS,
2082 block, gimple_bb (def_stmt)))
2083 || (gimple_bb (def_stmt) == block
2084 && value_dies_in_block_x (expr, block))))
2085 to_remove = expr;
2088 else if (expr->kind == NARY)
2090 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2091 /* If the NARY may trap make sure the block does not contain
2092 a possible exit point.
2093 ??? This is overly conservative if we translate AVAIL_OUT
2094 as the available expression might be after the exit point. */
2095 if (BB_MAY_NOTRETURN (block)
2096 && vn_nary_may_trap (nary))
2097 to_remove = expr;
2101 /* Remove queued expr. */
2102 if (to_remove)
2103 bitmap_remove_from_set (set, to_remove);
2106 static sbitmap has_abnormal_preds;
2108 /* Compute the ANTIC set for BLOCK.
2110 If succs(BLOCK) > 1 then
2111 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2112 else if succs(BLOCK) == 1 then
2113 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2115 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2117 Note that clean() is deferred until after the iteration. */
2119 static bool
2120 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2122 bool changed = false;
2123 bitmap_set_t S, old, ANTIC_OUT;
2124 bitmap_iterator bi;
2125 unsigned int bii;
2126 edge e;
2127 edge_iterator ei;
2128 bool was_visited = BB_VISITED (block);
2130 old = ANTIC_OUT = S = NULL;
2131 BB_VISITED (block) = 1;
2133 /* If any edges from predecessors are abnormal, antic_in is empty,
2134 so do nothing. */
2135 if (block_has_abnormal_pred_edge)
2136 goto maybe_dump_sets;
2138 old = ANTIC_IN (block);
2139 ANTIC_OUT = bitmap_set_new ();
2141 /* If the block has no successors, ANTIC_OUT is empty. */
2142 if (EDGE_COUNT (block->succs) == 0)
2144 /* If we have one successor, we could have some phi nodes to
2145 translate through. */
2146 else if (single_succ_p (block))
2148 basic_block succ_bb = single_succ (block);
2149 gcc_assert (BB_VISITED (succ_bb));
2150 phi_translate_set (ANTIC_OUT, ANTIC_IN (succ_bb), block, succ_bb);
2152 /* If we have multiple successors, we take the intersection of all of
2153 them. Note that in the case of loop exit phi nodes, we may have
2154 phis to translate through. */
2155 else
2157 size_t i;
2158 basic_block bprime, first = NULL;
2160 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2161 FOR_EACH_EDGE (e, ei, block->succs)
2163 if (!first
2164 && BB_VISITED (e->dest))
2165 first = e->dest;
2166 else if (BB_VISITED (e->dest))
2167 worklist.quick_push (e->dest);
2168 else
2170 /* Unvisited successors get their ANTIC_IN replaced by the
2171 maximal set to arrive at a maximum ANTIC_IN solution.
2172 We can ignore them in the intersection operation and thus
2173 need not explicitely represent that maximum solution. */
2174 if (dump_file && (dump_flags & TDF_DETAILS))
2175 fprintf (dump_file, "ANTIC_IN is MAX on %d->%d\n",
2176 e->src->index, e->dest->index);
2180 /* Of multiple successors we have to have visited one already
2181 which is guaranteed by iteration order. */
2182 gcc_assert (first != NULL);
2184 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2186 FOR_EACH_VEC_ELT (worklist, i, bprime)
2188 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2190 bitmap_set_t tmp = bitmap_set_new ();
2191 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2192 bitmap_set_and (ANTIC_OUT, tmp);
2193 bitmap_set_free (tmp);
2195 else
2196 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2200 /* Prune expressions that are clobbered in block and thus become
2201 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2202 prune_clobbered_mems (ANTIC_OUT, block);
2204 /* Generate ANTIC_OUT - TMP_GEN. */
2205 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2207 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2208 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2209 TMP_GEN (block));
2211 /* Then union in the ANTIC_OUT - TMP_GEN values,
2212 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2213 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2214 bitmap_value_insert_into_set (ANTIC_IN (block),
2215 expression_for_id (bii));
2217 /* clean (ANTIC_IN (block)) is defered to after the iteration converged
2218 because it can cause non-convergence, see for example PR81181. */
2220 if (!was_visited || !bitmap_set_equal (old, ANTIC_IN (block)))
2221 changed = true;
2223 maybe_dump_sets:
2224 if (dump_file && (dump_flags & TDF_DETAILS))
2226 if (ANTIC_OUT)
2227 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2229 if (changed)
2230 fprintf (dump_file, "[changed] ");
2231 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2232 block->index);
2234 if (S)
2235 print_bitmap_set (dump_file, S, "S", block->index);
2237 if (old)
2238 bitmap_set_free (old);
2239 if (S)
2240 bitmap_set_free (S);
2241 if (ANTIC_OUT)
2242 bitmap_set_free (ANTIC_OUT);
2243 return changed;
2246 /* Compute PARTIAL_ANTIC for BLOCK.
2248 If succs(BLOCK) > 1 then
2249 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2250 in ANTIC_OUT for all succ(BLOCK)
2251 else if succs(BLOCK) == 1 then
2252 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2254 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2255 - ANTIC_IN[BLOCK])
2258 static void
2259 compute_partial_antic_aux (basic_block block,
2260 bool block_has_abnormal_pred_edge)
2262 bitmap_set_t old_PA_IN;
2263 bitmap_set_t PA_OUT;
2264 edge e;
2265 edge_iterator ei;
2266 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2268 old_PA_IN = PA_OUT = NULL;
2270 /* If any edges from predecessors are abnormal, antic_in is empty,
2271 so do nothing. */
2272 if (block_has_abnormal_pred_edge)
2273 goto maybe_dump_sets;
2275 /* If there are too many partially anticipatable values in the
2276 block, phi_translate_set can take an exponential time: stop
2277 before the translation starts. */
2278 if (max_pa
2279 && single_succ_p (block)
2280 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2281 goto maybe_dump_sets;
2283 old_PA_IN = PA_IN (block);
2284 PA_OUT = bitmap_set_new ();
2286 /* If the block has no successors, ANTIC_OUT is empty. */
2287 if (EDGE_COUNT (block->succs) == 0)
2289 /* If we have one successor, we could have some phi nodes to
2290 translate through. Note that we can't phi translate across DFS
2291 back edges in partial antic, because it uses a union operation on
2292 the successors. For recurrences like IV's, we will end up
2293 generating a new value in the set on each go around (i + 3 (VH.1)
2294 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2295 else if (single_succ_p (block))
2297 basic_block succ = single_succ (block);
2298 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2299 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2301 /* If we have multiple successors, we take the union of all of
2302 them. */
2303 else
2305 size_t i;
2306 basic_block bprime;
2308 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2309 FOR_EACH_EDGE (e, ei, block->succs)
2311 if (e->flags & EDGE_DFS_BACK)
2312 continue;
2313 worklist.quick_push (e->dest);
2315 if (worklist.length () > 0)
2317 FOR_EACH_VEC_ELT (worklist, i, bprime)
2319 unsigned int i;
2320 bitmap_iterator bi;
2322 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2323 bitmap_value_insert_into_set (PA_OUT,
2324 expression_for_id (i));
2325 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2327 bitmap_set_t pa_in = bitmap_set_new ();
2328 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2329 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2330 bitmap_value_insert_into_set (PA_OUT,
2331 expression_for_id (i));
2332 bitmap_set_free (pa_in);
2334 else
2335 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2336 bitmap_value_insert_into_set (PA_OUT,
2337 expression_for_id (i));
2342 /* Prune expressions that are clobbered in block and thus become
2343 invalid if translated from PA_OUT to PA_IN. */
2344 prune_clobbered_mems (PA_OUT, block);
2346 /* PA_IN starts with PA_OUT - TMP_GEN.
2347 Then we subtract things from ANTIC_IN. */
2348 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2350 /* For partial antic, we want to put back in the phi results, since
2351 we will properly avoid making them partially antic over backedges. */
2352 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2353 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2355 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2356 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2358 dependent_clean (PA_IN (block), ANTIC_IN (block));
2360 maybe_dump_sets:
2361 if (dump_file && (dump_flags & TDF_DETAILS))
2363 if (PA_OUT)
2364 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2366 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2368 if (old_PA_IN)
2369 bitmap_set_free (old_PA_IN);
2370 if (PA_OUT)
2371 bitmap_set_free (PA_OUT);
2374 /* Compute ANTIC and partial ANTIC sets. */
2376 static void
2377 compute_antic (void)
2379 bool changed = true;
2380 int num_iterations = 0;
2381 basic_block block;
2382 int i;
2383 edge_iterator ei;
2384 edge e;
2386 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2387 We pre-build the map of blocks with incoming abnormal edges here. */
2388 has_abnormal_preds = sbitmap_alloc (last_basic_block_for_fn (cfun));
2389 bitmap_clear (has_abnormal_preds);
2391 FOR_ALL_BB_FN (block, cfun)
2393 BB_VISITED (block) = 0;
2395 FOR_EACH_EDGE (e, ei, block->preds)
2396 if (e->flags & EDGE_ABNORMAL)
2398 bitmap_set_bit (has_abnormal_preds, block->index);
2400 /* We also anticipate nothing. */
2401 BB_VISITED (block) = 1;
2402 break;
2405 /* While we are here, give empty ANTIC_IN sets to each block. */
2406 ANTIC_IN (block) = bitmap_set_new ();
2407 if (do_partial_partial)
2408 PA_IN (block) = bitmap_set_new ();
2411 /* At the exit block we anticipate nothing. */
2412 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun)) = 1;
2414 /* For ANTIC computation we need a postorder that also guarantees that
2415 a block with a single successor is visited after its successor.
2416 RPO on the inverted CFG has this property. */
2417 auto_vec<int, 20> postorder;
2418 inverted_post_order_compute (&postorder);
2420 auto_sbitmap worklist (last_basic_block_for_fn (cfun) + 1);
2421 bitmap_clear (worklist);
2422 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
2423 bitmap_set_bit (worklist, e->src->index);
2424 while (changed)
2426 if (dump_file && (dump_flags & TDF_DETAILS))
2427 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2428 /* ??? We need to clear our PHI translation cache here as the
2429 ANTIC sets shrink and we restrict valid translations to
2430 those having operands with leaders in ANTIC. Same below
2431 for PA ANTIC computation. */
2432 num_iterations++;
2433 changed = false;
2434 for (i = postorder.length () - 1; i >= 0; i--)
2436 if (bitmap_bit_p (worklist, postorder[i]))
2438 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2439 bitmap_clear_bit (worklist, block->index);
2440 if (compute_antic_aux (block,
2441 bitmap_bit_p (has_abnormal_preds,
2442 block->index)))
2444 FOR_EACH_EDGE (e, ei, block->preds)
2445 bitmap_set_bit (worklist, e->src->index);
2446 changed = true;
2450 /* Theoretically possible, but *highly* unlikely. */
2451 gcc_checking_assert (num_iterations < 500);
2454 /* We have to clean after the dataflow problem converged as cleaning
2455 can cause non-convergence because it is based on expressions
2456 rather than values. */
2457 FOR_EACH_BB_FN (block, cfun)
2458 clean (ANTIC_IN (block));
2460 statistics_histogram_event (cfun, "compute_antic iterations",
2461 num_iterations);
2463 if (do_partial_partial)
2465 /* For partial antic we ignore backedges and thus we do not need
2466 to perform any iteration when we process blocks in postorder. */
2467 int postorder_num
2468 = pre_and_rev_post_order_compute (NULL, postorder.address (), false);
2469 for (i = postorder_num - 1 ; i >= 0; i--)
2471 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2472 compute_partial_antic_aux (block,
2473 bitmap_bit_p (has_abnormal_preds,
2474 block->index));
2478 sbitmap_free (has_abnormal_preds);
2482 /* Inserted expressions are placed onto this worklist, which is used
2483 for performing quick dead code elimination of insertions we made
2484 that didn't turn out to be necessary. */
2485 static bitmap inserted_exprs;
2487 /* The actual worker for create_component_ref_by_pieces. */
2489 static tree
2490 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2491 unsigned int *operand, gimple_seq *stmts)
2493 vn_reference_op_t currop = &ref->operands[*operand];
2494 tree genop;
2495 ++*operand;
2496 switch (currop->opcode)
2498 case CALL_EXPR:
2499 gcc_unreachable ();
2501 case MEM_REF:
2503 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2504 stmts);
2505 if (!baseop)
2506 return NULL_TREE;
2507 tree offset = currop->op0;
2508 if (TREE_CODE (baseop) == ADDR_EXPR
2509 && handled_component_p (TREE_OPERAND (baseop, 0)))
2511 HOST_WIDE_INT off;
2512 tree base;
2513 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2514 &off);
2515 gcc_assert (base);
2516 offset = int_const_binop (PLUS_EXPR, offset,
2517 build_int_cst (TREE_TYPE (offset),
2518 off));
2519 baseop = build_fold_addr_expr (base);
2521 genop = build2 (MEM_REF, currop->type, baseop, offset);
2522 MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2523 MR_DEPENDENCE_BASE (genop) = currop->base;
2524 REF_REVERSE_STORAGE_ORDER (genop) = currop->reverse;
2525 return genop;
2528 case TARGET_MEM_REF:
2530 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2531 vn_reference_op_t nextop = &ref->operands[++*operand];
2532 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2533 stmts);
2534 if (!baseop)
2535 return NULL_TREE;
2536 if (currop->op0)
2538 genop0 = find_or_generate_expression (block, currop->op0, stmts);
2539 if (!genop0)
2540 return NULL_TREE;
2542 if (nextop->op0)
2544 genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2545 if (!genop1)
2546 return NULL_TREE;
2548 genop = build5 (TARGET_MEM_REF, currop->type,
2549 baseop, currop->op2, genop0, currop->op1, genop1);
2551 MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2552 MR_DEPENDENCE_BASE (genop) = currop->base;
2553 return genop;
2556 case ADDR_EXPR:
2557 if (currop->op0)
2559 gcc_assert (is_gimple_min_invariant (currop->op0));
2560 return currop->op0;
2562 /* Fallthrough. */
2563 case REALPART_EXPR:
2564 case IMAGPART_EXPR:
2565 case VIEW_CONVERT_EXPR:
2567 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2568 stmts);
2569 if (!genop0)
2570 return NULL_TREE;
2571 return fold_build1 (currop->opcode, currop->type, genop0);
2574 case WITH_SIZE_EXPR:
2576 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2577 stmts);
2578 if (!genop0)
2579 return NULL_TREE;
2580 tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2581 if (!genop1)
2582 return NULL_TREE;
2583 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2586 case BIT_FIELD_REF:
2588 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2589 stmts);
2590 if (!genop0)
2591 return NULL_TREE;
2592 tree op1 = currop->op0;
2593 tree op2 = currop->op1;
2594 tree t = build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2595 REF_REVERSE_STORAGE_ORDER (t) = currop->reverse;
2596 return fold (t);
2599 /* For array ref vn_reference_op's, operand 1 of the array ref
2600 is op0 of the reference op and operand 3 of the array ref is
2601 op1. */
2602 case ARRAY_RANGE_REF:
2603 case ARRAY_REF:
2605 tree genop0;
2606 tree genop1 = currop->op0;
2607 tree genop2 = currop->op1;
2608 tree genop3 = currop->op2;
2609 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2610 stmts);
2611 if (!genop0)
2612 return NULL_TREE;
2613 genop1 = find_or_generate_expression (block, genop1, stmts);
2614 if (!genop1)
2615 return NULL_TREE;
2616 if (genop2)
2618 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2619 /* Drop zero minimum index if redundant. */
2620 if (integer_zerop (genop2)
2621 && (!domain_type
2622 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2623 genop2 = NULL_TREE;
2624 else
2626 genop2 = find_or_generate_expression (block, genop2, stmts);
2627 if (!genop2)
2628 return NULL_TREE;
2631 if (genop3)
2633 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2634 /* We can't always put a size in units of the element alignment
2635 here as the element alignment may be not visible. See
2636 PR43783. Simply drop the element size for constant
2637 sizes. */
2638 if (TREE_CODE (genop3) == INTEGER_CST
2639 && TREE_CODE (TYPE_SIZE_UNIT (elmt_type)) == INTEGER_CST
2640 && wi::eq_p (wi::to_offset (TYPE_SIZE_UNIT (elmt_type)),
2641 (wi::to_offset (genop3)
2642 * vn_ref_op_align_unit (currop))))
2643 genop3 = NULL_TREE;
2644 else
2646 genop3 = find_or_generate_expression (block, genop3, stmts);
2647 if (!genop3)
2648 return NULL_TREE;
2651 return build4 (currop->opcode, currop->type, genop0, genop1,
2652 genop2, genop3);
2654 case COMPONENT_REF:
2656 tree op0;
2657 tree op1;
2658 tree genop2 = currop->op1;
2659 op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2660 if (!op0)
2661 return NULL_TREE;
2662 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2663 op1 = currop->op0;
2664 if (genop2)
2666 genop2 = find_or_generate_expression (block, genop2, stmts);
2667 if (!genop2)
2668 return NULL_TREE;
2670 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2673 case SSA_NAME:
2675 genop = find_or_generate_expression (block, currop->op0, stmts);
2676 return genop;
2678 case STRING_CST:
2679 case INTEGER_CST:
2680 case COMPLEX_CST:
2681 case VECTOR_CST:
2682 case REAL_CST:
2683 case CONSTRUCTOR:
2684 case VAR_DECL:
2685 case PARM_DECL:
2686 case CONST_DECL:
2687 case RESULT_DECL:
2688 case FUNCTION_DECL:
2689 return currop->op0;
2691 default:
2692 gcc_unreachable ();
2696 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2697 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2698 trying to rename aggregates into ssa form directly, which is a no no.
2700 Thus, this routine doesn't create temporaries, it just builds a
2701 single access expression for the array, calling
2702 find_or_generate_expression to build the innermost pieces.
2704 This function is a subroutine of create_expression_by_pieces, and
2705 should not be called on it's own unless you really know what you
2706 are doing. */
2708 static tree
2709 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2710 gimple_seq *stmts)
2712 unsigned int op = 0;
2713 return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2716 /* Find a simple leader for an expression, or generate one using
2717 create_expression_by_pieces from a NARY expression for the value.
2718 BLOCK is the basic_block we are looking for leaders in.
2719 OP is the tree expression to find a leader for or generate.
2720 Returns the leader or NULL_TREE on failure. */
2722 static tree
2723 find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2725 pre_expr expr = get_or_alloc_expr_for (op);
2726 unsigned int lookfor = get_expr_value_id (expr);
2727 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2728 if (leader)
2730 if (leader->kind == NAME)
2731 return PRE_EXPR_NAME (leader);
2732 else if (leader->kind == CONSTANT)
2733 return PRE_EXPR_CONSTANT (leader);
2735 /* Defer. */
2736 return NULL_TREE;
2739 /* It must be a complex expression, so generate it recursively. Note
2740 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2741 where the insert algorithm fails to insert a required expression. */
2742 bitmap exprset = value_expressions[lookfor];
2743 bitmap_iterator bi;
2744 unsigned int i;
2745 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2747 pre_expr temp = expression_for_id (i);
2748 /* We cannot insert random REFERENCE expressions at arbitrary
2749 places. We can insert NARYs which eventually re-materializes
2750 its operand values. */
2751 if (temp->kind == NARY)
2752 return create_expression_by_pieces (block, temp, stmts,
2753 get_expr_type (expr));
2756 /* Defer. */
2757 return NULL_TREE;
2760 #define NECESSARY GF_PLF_1
2762 /* Create an expression in pieces, so that we can handle very complex
2763 expressions that may be ANTIC, but not necessary GIMPLE.
2764 BLOCK is the basic block the expression will be inserted into,
2765 EXPR is the expression to insert (in value form)
2766 STMTS is a statement list to append the necessary insertions into.
2768 This function will die if we hit some value that shouldn't be
2769 ANTIC but is (IE there is no leader for it, or its components).
2770 The function returns NULL_TREE in case a different antic expression
2771 has to be inserted first.
2772 This function may also generate expressions that are themselves
2773 partially or fully redundant. Those that are will be either made
2774 fully redundant during the next iteration of insert (for partially
2775 redundant ones), or eliminated by eliminate (for fully redundant
2776 ones). */
2778 static tree
2779 create_expression_by_pieces (basic_block block, pre_expr expr,
2780 gimple_seq *stmts, tree type)
2782 tree name;
2783 tree folded;
2784 gimple_seq forced_stmts = NULL;
2785 unsigned int value_id;
2786 gimple_stmt_iterator gsi;
2787 tree exprtype = type ? type : get_expr_type (expr);
2788 pre_expr nameexpr;
2789 gassign *newstmt;
2791 switch (expr->kind)
2793 /* We may hit the NAME/CONSTANT case if we have to convert types
2794 that value numbering saw through. */
2795 case NAME:
2796 folded = PRE_EXPR_NAME (expr);
2797 if (useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2798 return folded;
2799 break;
2800 case CONSTANT:
2802 folded = PRE_EXPR_CONSTANT (expr);
2803 tree tem = fold_convert (exprtype, folded);
2804 if (is_gimple_min_invariant (tem))
2805 return tem;
2806 break;
2808 case REFERENCE:
2809 if (PRE_EXPR_REFERENCE (expr)->operands[0].opcode == CALL_EXPR)
2811 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2812 unsigned int operand = 1;
2813 vn_reference_op_t currop = &ref->operands[0];
2814 tree sc = NULL_TREE;
2815 tree fn;
2816 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2817 fn = currop->op0;
2818 else
2819 fn = find_or_generate_expression (block, currop->op0, stmts);
2820 if (!fn)
2821 return NULL_TREE;
2822 if (currop->op1)
2824 sc = find_or_generate_expression (block, currop->op1, stmts);
2825 if (!sc)
2826 return NULL_TREE;
2828 auto_vec<tree> args (ref->operands.length () - 1);
2829 while (operand < ref->operands.length ())
2831 tree arg = create_component_ref_by_pieces_1 (block, ref,
2832 &operand, stmts);
2833 if (!arg)
2834 return NULL_TREE;
2835 args.quick_push (arg);
2837 gcall *call
2838 = gimple_build_call_vec ((TREE_CODE (fn) == FUNCTION_DECL
2839 ? build_fold_addr_expr (fn) : fn), args);
2840 gimple_call_set_with_bounds (call, currop->with_bounds);
2841 if (sc)
2842 gimple_call_set_chain (call, sc);
2843 tree forcedname = make_ssa_name (currop->type);
2844 gimple_call_set_lhs (call, forcedname);
2845 gimple_set_vuse (call, BB_LIVE_VOP_ON_EXIT (block));
2846 gimple_seq_add_stmt_without_update (&forced_stmts, call);
2847 folded = forcedname;
2849 else
2851 folded = create_component_ref_by_pieces (block,
2852 PRE_EXPR_REFERENCE (expr),
2853 stmts);
2854 if (!folded)
2855 return NULL_TREE;
2856 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2857 newstmt = gimple_build_assign (name, folded);
2858 gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2859 gimple_set_vuse (newstmt, BB_LIVE_VOP_ON_EXIT (block));
2860 folded = name;
2862 break;
2863 case NARY:
2865 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2866 tree *genop = XALLOCAVEC (tree, nary->length);
2867 unsigned i;
2868 for (i = 0; i < nary->length; ++i)
2870 genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2871 if (!genop[i])
2872 return NULL_TREE;
2873 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2874 may have conversions stripped. */
2875 if (nary->opcode == POINTER_PLUS_EXPR)
2877 if (i == 0)
2878 genop[i] = gimple_convert (&forced_stmts,
2879 nary->type, genop[i]);
2880 else if (i == 1)
2881 genop[i] = gimple_convert (&forced_stmts,
2882 sizetype, genop[i]);
2884 else
2885 genop[i] = gimple_convert (&forced_stmts,
2886 TREE_TYPE (nary->op[i]), genop[i]);
2888 if (nary->opcode == CONSTRUCTOR)
2890 vec<constructor_elt, va_gc> *elts = NULL;
2891 for (i = 0; i < nary->length; ++i)
2892 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2893 folded = build_constructor (nary->type, elts);
2894 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2895 newstmt = gimple_build_assign (name, folded);
2896 gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2897 folded = name;
2899 else
2901 switch (nary->length)
2903 case 1:
2904 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2905 genop[0]);
2906 break;
2907 case 2:
2908 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2909 genop[0], genop[1]);
2910 break;
2911 case 3:
2912 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2913 genop[0], genop[1], genop[2]);
2914 break;
2915 default:
2916 gcc_unreachable ();
2920 break;
2921 default:
2922 gcc_unreachable ();
2925 folded = gimple_convert (&forced_stmts, exprtype, folded);
2927 /* If there is nothing to insert, return the simplified result. */
2928 if (gimple_seq_empty_p (forced_stmts))
2929 return folded;
2930 /* If we simplified to a constant return it and discard eventually
2931 built stmts. */
2932 if (is_gimple_min_invariant (folded))
2934 gimple_seq_discard (forced_stmts);
2935 return folded;
2937 /* Likewise if we simplified to sth not queued for insertion. */
2938 bool found = false;
2939 gsi = gsi_last (forced_stmts);
2940 for (; !gsi_end_p (gsi); gsi_prev (&gsi))
2942 gimple *stmt = gsi_stmt (gsi);
2943 tree forcedname = gimple_get_lhs (stmt);
2944 if (forcedname == folded)
2946 found = true;
2947 break;
2950 if (! found)
2952 gimple_seq_discard (forced_stmts);
2953 return folded;
2955 gcc_assert (TREE_CODE (folded) == SSA_NAME);
2957 /* If we have any intermediate expressions to the value sets, add them
2958 to the value sets and chain them in the instruction stream. */
2959 if (forced_stmts)
2961 gsi = gsi_start (forced_stmts);
2962 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2964 gimple *stmt = gsi_stmt (gsi);
2965 tree forcedname = gimple_get_lhs (stmt);
2966 pre_expr nameexpr;
2968 if (forcedname != folded)
2970 VN_INFO_GET (forcedname)->valnum = forcedname;
2971 VN_INFO (forcedname)->value_id = get_next_value_id ();
2972 nameexpr = get_or_alloc_expr_for_name (forcedname);
2973 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2974 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2975 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2978 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
2979 gimple_set_plf (stmt, NECESSARY, false);
2981 gimple_seq_add_seq (stmts, forced_stmts);
2984 name = folded;
2986 /* Fold the last statement. */
2987 gsi = gsi_last (*stmts);
2988 if (fold_stmt_inplace (&gsi))
2989 update_stmt (gsi_stmt (gsi));
2991 /* Add a value number to the temporary.
2992 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2993 we are creating the expression by pieces, and this particular piece of
2994 the expression may have been represented. There is no harm in replacing
2995 here. */
2996 value_id = get_expr_value_id (expr);
2997 VN_INFO_GET (name)->value_id = value_id;
2998 VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id);
2999 if (VN_INFO (name)->valnum == NULL_TREE)
3000 VN_INFO (name)->valnum = name;
3001 gcc_assert (VN_INFO (name)->valnum != NULL_TREE);
3002 nameexpr = get_or_alloc_expr_for_name (name);
3003 add_to_value (value_id, nameexpr);
3004 if (NEW_SETS (block))
3005 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3006 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3008 pre_stats.insertions++;
3009 if (dump_file && (dump_flags & TDF_DETAILS))
3011 fprintf (dump_file, "Inserted ");
3012 print_gimple_stmt (dump_file, gsi_stmt (gsi_last (*stmts)), 0);
3013 fprintf (dump_file, " in predecessor %d (%04d)\n",
3014 block->index, value_id);
3017 return name;
3021 /* Insert the to-be-made-available values of expression EXPRNUM for each
3022 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3023 merge the result with a phi node, given the same value number as
3024 NODE. Return true if we have inserted new stuff. */
3026 static bool
3027 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3028 vec<pre_expr> avail)
3030 pre_expr expr = expression_for_id (exprnum);
3031 pre_expr newphi;
3032 unsigned int val = get_expr_value_id (expr);
3033 edge pred;
3034 bool insertions = false;
3035 bool nophi = false;
3036 basic_block bprime;
3037 pre_expr eprime;
3038 edge_iterator ei;
3039 tree type = get_expr_type (expr);
3040 tree temp;
3041 gphi *phi;
3043 /* Make sure we aren't creating an induction variable. */
3044 if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
3046 bool firstinsideloop = false;
3047 bool secondinsideloop = false;
3048 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3049 EDGE_PRED (block, 0)->src);
3050 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3051 EDGE_PRED (block, 1)->src);
3052 /* Induction variables only have one edge inside the loop. */
3053 if ((firstinsideloop ^ secondinsideloop)
3054 && expr->kind != REFERENCE)
3056 if (dump_file && (dump_flags & TDF_DETAILS))
3057 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3058 nophi = true;
3062 /* Make the necessary insertions. */
3063 FOR_EACH_EDGE (pred, ei, block->preds)
3065 gimple_seq stmts = NULL;
3066 tree builtexpr;
3067 bprime = pred->src;
3068 eprime = avail[pred->dest_idx];
3069 builtexpr = create_expression_by_pieces (bprime, eprime,
3070 &stmts, type);
3071 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3072 if (!gimple_seq_empty_p (stmts))
3074 gsi_insert_seq_on_edge (pred, stmts);
3075 insertions = true;
3077 if (!builtexpr)
3079 /* We cannot insert a PHI node if we failed to insert
3080 on one edge. */
3081 nophi = true;
3082 continue;
3084 if (is_gimple_min_invariant (builtexpr))
3085 avail[pred->dest_idx] = get_or_alloc_expr_for_constant (builtexpr);
3086 else
3087 avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3089 /* If we didn't want a phi node, and we made insertions, we still have
3090 inserted new stuff, and thus return true. If we didn't want a phi node,
3091 and didn't make insertions, we haven't added anything new, so return
3092 false. */
3093 if (nophi && insertions)
3094 return true;
3095 else if (nophi && !insertions)
3096 return false;
3098 /* Now build a phi for the new variable. */
3099 temp = make_temp_ssa_name (type, NULL, "prephitmp");
3100 phi = create_phi_node (temp, block);
3102 gimple_set_plf (phi, NECESSARY, false);
3103 VN_INFO_GET (temp)->value_id = val;
3104 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3105 if (VN_INFO (temp)->valnum == NULL_TREE)
3106 VN_INFO (temp)->valnum = temp;
3107 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3108 FOR_EACH_EDGE (pred, ei, block->preds)
3110 pre_expr ae = avail[pred->dest_idx];
3111 gcc_assert (get_expr_type (ae) == type
3112 || useless_type_conversion_p (type, get_expr_type (ae)));
3113 if (ae->kind == CONSTANT)
3114 add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3115 pred, UNKNOWN_LOCATION);
3116 else
3117 add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3120 newphi = get_or_alloc_expr_for_name (temp);
3121 add_to_value (val, newphi);
3123 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3124 this insertion, since we test for the existence of this value in PHI_GEN
3125 before proceeding with the partial redundancy checks in insert_aux.
3127 The value may exist in AVAIL_OUT, in particular, it could be represented
3128 by the expression we are trying to eliminate, in which case we want the
3129 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3130 inserted there.
3132 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3133 this block, because if it did, it would have existed in our dominator's
3134 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3137 bitmap_insert_into_set (PHI_GEN (block), newphi);
3138 bitmap_value_replace_in_set (AVAIL_OUT (block),
3139 newphi);
3140 bitmap_insert_into_set (NEW_SETS (block),
3141 newphi);
3143 /* If we insert a PHI node for a conversion of another PHI node
3144 in the same basic-block try to preserve range information.
3145 This is important so that followup loop passes receive optimal
3146 number of iteration analysis results. See PR61743. */
3147 if (expr->kind == NARY
3148 && CONVERT_EXPR_CODE_P (expr->u.nary->opcode)
3149 && TREE_CODE (expr->u.nary->op[0]) == SSA_NAME
3150 && gimple_bb (SSA_NAME_DEF_STMT (expr->u.nary->op[0])) == block
3151 && INTEGRAL_TYPE_P (type)
3152 && INTEGRAL_TYPE_P (TREE_TYPE (expr->u.nary->op[0]))
3153 && (TYPE_PRECISION (type)
3154 >= TYPE_PRECISION (TREE_TYPE (expr->u.nary->op[0])))
3155 && SSA_NAME_RANGE_INFO (expr->u.nary->op[0]))
3157 wide_int min, max;
3158 if (get_range_info (expr->u.nary->op[0], &min, &max) == VR_RANGE
3159 && !wi::neg_p (min, SIGNED)
3160 && !wi::neg_p (max, SIGNED))
3161 /* Just handle extension and sign-changes of all-positive ranges. */
3162 set_range_info (temp,
3163 SSA_NAME_RANGE_TYPE (expr->u.nary->op[0]),
3164 wide_int_storage::from (min, TYPE_PRECISION (type),
3165 TYPE_SIGN (type)),
3166 wide_int_storage::from (max, TYPE_PRECISION (type),
3167 TYPE_SIGN (type)));
3170 if (dump_file && (dump_flags & TDF_DETAILS))
3172 fprintf (dump_file, "Created phi ");
3173 print_gimple_stmt (dump_file, phi, 0);
3174 fprintf (dump_file, " in block %d (%04d)\n", block->index, val);
3176 pre_stats.phis++;
3177 return true;
3182 /* Perform insertion of partially redundant or hoistable values.
3183 For BLOCK, do the following:
3184 1. Propagate the NEW_SETS of the dominator into the current block.
3185 If the block has multiple predecessors,
3186 2a. Iterate over the ANTIC expressions for the block to see if
3187 any of them are partially redundant.
3188 2b. If so, insert them into the necessary predecessors to make
3189 the expression fully redundant.
3190 2c. Insert a new PHI merging the values of the predecessors.
3191 2d. Insert the new PHI, and the new expressions, into the
3192 NEW_SETS set.
3193 If the block has multiple successors,
3194 3a. Iterate over the ANTIC values for the block to see if
3195 any of them are good candidates for hoisting.
3196 3b. If so, insert expressions computing the values in BLOCK,
3197 and add the new expressions into the NEW_SETS set.
3198 4. Recursively call ourselves on the dominator children of BLOCK.
3200 Steps 1, 2a, and 4 are done by insert_aux. 2b, 2c and 2d are done by
3201 do_pre_regular_insertion and do_partial_insertion. 3a and 3b are
3202 done in do_hoist_insertion.
3205 static bool
3206 do_pre_regular_insertion (basic_block block, basic_block dom)
3208 bool new_stuff = false;
3209 vec<pre_expr> exprs;
3210 pre_expr expr;
3211 auto_vec<pre_expr> avail;
3212 int i;
3214 exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3215 avail.safe_grow (EDGE_COUNT (block->preds));
3217 FOR_EACH_VEC_ELT (exprs, i, expr)
3219 if (expr->kind == NARY
3220 || expr->kind == REFERENCE)
3222 unsigned int val;
3223 bool by_some = false;
3224 bool cant_insert = false;
3225 bool all_same = true;
3226 pre_expr first_s = NULL;
3227 edge pred;
3228 basic_block bprime;
3229 pre_expr eprime = NULL;
3230 edge_iterator ei;
3231 pre_expr edoubleprime = NULL;
3232 bool do_insertion = false;
3234 val = get_expr_value_id (expr);
3235 if (bitmap_set_contains_value (PHI_GEN (block), val))
3236 continue;
3237 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3239 if (dump_file && (dump_flags & TDF_DETAILS))
3241 fprintf (dump_file, "Found fully redundant value: ");
3242 print_pre_expr (dump_file, expr);
3243 fprintf (dump_file, "\n");
3245 continue;
3248 FOR_EACH_EDGE (pred, ei, block->preds)
3250 unsigned int vprime;
3252 /* We should never run insertion for the exit block
3253 and so not come across fake pred edges. */
3254 gcc_assert (!(pred->flags & EDGE_FAKE));
3255 bprime = pred->src;
3256 /* We are looking at ANTIC_OUT of bprime. */
3257 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3258 bprime, block);
3260 /* eprime will generally only be NULL if the
3261 value of the expression, translated
3262 through the PHI for this predecessor, is
3263 undefined. If that is the case, we can't
3264 make the expression fully redundant,
3265 because its value is undefined along a
3266 predecessor path. We can thus break out
3267 early because it doesn't matter what the
3268 rest of the results are. */
3269 if (eprime == NULL)
3271 avail[pred->dest_idx] = NULL;
3272 cant_insert = true;
3273 break;
3276 vprime = get_expr_value_id (eprime);
3277 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3278 vprime);
3279 if (edoubleprime == NULL)
3281 avail[pred->dest_idx] = eprime;
3282 all_same = false;
3284 else
3286 avail[pred->dest_idx] = edoubleprime;
3287 by_some = true;
3288 /* We want to perform insertions to remove a redundancy on
3289 a path in the CFG we want to optimize for speed. */
3290 if (optimize_edge_for_speed_p (pred))
3291 do_insertion = true;
3292 if (first_s == NULL)
3293 first_s = edoubleprime;
3294 else if (!pre_expr_d::equal (first_s, edoubleprime))
3295 all_same = false;
3298 /* If we can insert it, it's not the same value
3299 already existing along every predecessor, and
3300 it's defined by some predecessor, it is
3301 partially redundant. */
3302 if (!cant_insert && !all_same && by_some)
3304 if (!do_insertion)
3306 if (dump_file && (dump_flags & TDF_DETAILS))
3308 fprintf (dump_file, "Skipping partial redundancy for "
3309 "expression ");
3310 print_pre_expr (dump_file, expr);
3311 fprintf (dump_file, " (%04d), no redundancy on to be "
3312 "optimized for speed edge\n", val);
3315 else if (dbg_cnt (treepre_insert))
3317 if (dump_file && (dump_flags & TDF_DETAILS))
3319 fprintf (dump_file, "Found partial redundancy for "
3320 "expression ");
3321 print_pre_expr (dump_file, expr);
3322 fprintf (dump_file, " (%04d)\n",
3323 get_expr_value_id (expr));
3325 if (insert_into_preds_of_block (block,
3326 get_expression_id (expr),
3327 avail))
3328 new_stuff = true;
3331 /* If all edges produce the same value and that value is
3332 an invariant, then the PHI has the same value on all
3333 edges. Note this. */
3334 else if (!cant_insert && all_same)
3336 gcc_assert (edoubleprime->kind == CONSTANT
3337 || edoubleprime->kind == NAME);
3339 tree temp = make_temp_ssa_name (get_expr_type (expr),
3340 NULL, "pretmp");
3341 gassign *assign
3342 = gimple_build_assign (temp,
3343 edoubleprime->kind == CONSTANT ?
3344 PRE_EXPR_CONSTANT (edoubleprime) :
3345 PRE_EXPR_NAME (edoubleprime));
3346 gimple_stmt_iterator gsi = gsi_after_labels (block);
3347 gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3349 gimple_set_plf (assign, NECESSARY, false);
3350 VN_INFO_GET (temp)->value_id = val;
3351 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3352 if (VN_INFO (temp)->valnum == NULL_TREE)
3353 VN_INFO (temp)->valnum = temp;
3354 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3355 pre_expr newe = get_or_alloc_expr_for_name (temp);
3356 add_to_value (val, newe);
3357 bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3358 bitmap_insert_into_set (NEW_SETS (block), newe);
3363 exprs.release ();
3364 return new_stuff;
3368 /* Perform insertion for partially anticipatable expressions. There
3369 is only one case we will perform insertion for these. This case is
3370 if the expression is partially anticipatable, and fully available.
3371 In this case, we know that putting it earlier will enable us to
3372 remove the later computation. */
3374 static bool
3375 do_pre_partial_partial_insertion (basic_block block, basic_block dom)
3377 bool new_stuff = false;
3378 vec<pre_expr> exprs;
3379 pre_expr expr;
3380 auto_vec<pre_expr> avail;
3381 int i;
3383 exprs = sorted_array_from_bitmap_set (PA_IN (block));
3384 avail.safe_grow (EDGE_COUNT (block->preds));
3386 FOR_EACH_VEC_ELT (exprs, i, expr)
3388 if (expr->kind == NARY
3389 || expr->kind == REFERENCE)
3391 unsigned int val;
3392 bool by_all = true;
3393 bool cant_insert = false;
3394 edge pred;
3395 basic_block bprime;
3396 pre_expr eprime = NULL;
3397 edge_iterator ei;
3399 val = get_expr_value_id (expr);
3400 if (bitmap_set_contains_value (PHI_GEN (block), val))
3401 continue;
3402 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3403 continue;
3405 FOR_EACH_EDGE (pred, ei, block->preds)
3407 unsigned int vprime;
3408 pre_expr edoubleprime;
3410 /* We should never run insertion for the exit block
3411 and so not come across fake pred edges. */
3412 gcc_assert (!(pred->flags & EDGE_FAKE));
3413 bprime = pred->src;
3414 eprime = phi_translate (expr, ANTIC_IN (block),
3415 PA_IN (block),
3416 bprime, block);
3418 /* eprime will generally only be NULL if the
3419 value of the expression, translated
3420 through the PHI for this predecessor, is
3421 undefined. If that is the case, we can't
3422 make the expression fully redundant,
3423 because its value is undefined along a
3424 predecessor path. We can thus break out
3425 early because it doesn't matter what the
3426 rest of the results are. */
3427 if (eprime == NULL)
3429 avail[pred->dest_idx] = NULL;
3430 cant_insert = true;
3431 break;
3434 vprime = get_expr_value_id (eprime);
3435 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3436 avail[pred->dest_idx] = edoubleprime;
3437 if (edoubleprime == NULL)
3439 by_all = false;
3440 break;
3444 /* If we can insert it, it's not the same value
3445 already existing along every predecessor, and
3446 it's defined by some predecessor, it is
3447 partially redundant. */
3448 if (!cant_insert && by_all)
3450 edge succ;
3451 bool do_insertion = false;
3453 /* Insert only if we can remove a later expression on a path
3454 that we want to optimize for speed.
3455 The phi node that we will be inserting in BLOCK is not free,
3456 and inserting it for the sake of !optimize_for_speed successor
3457 may cause regressions on the speed path. */
3458 FOR_EACH_EDGE (succ, ei, block->succs)
3460 if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3461 || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3463 if (optimize_edge_for_speed_p (succ))
3464 do_insertion = true;
3468 if (!do_insertion)
3470 if (dump_file && (dump_flags & TDF_DETAILS))
3472 fprintf (dump_file, "Skipping partial partial redundancy "
3473 "for expression ");
3474 print_pre_expr (dump_file, expr);
3475 fprintf (dump_file, " (%04d), not (partially) anticipated "
3476 "on any to be optimized for speed edges\n", val);
3479 else if (dbg_cnt (treepre_insert))
3481 pre_stats.pa_insert++;
3482 if (dump_file && (dump_flags & TDF_DETAILS))
3484 fprintf (dump_file, "Found partial partial redundancy "
3485 "for expression ");
3486 print_pre_expr (dump_file, expr);
3487 fprintf (dump_file, " (%04d)\n",
3488 get_expr_value_id (expr));
3490 if (insert_into_preds_of_block (block,
3491 get_expression_id (expr),
3492 avail))
3493 new_stuff = true;
3499 exprs.release ();
3500 return new_stuff;
3503 /* Insert expressions in BLOCK to compute hoistable values up.
3504 Return TRUE if something was inserted, otherwise return FALSE.
3505 The caller has to make sure that BLOCK has at least two successors. */
3507 static bool
3508 do_hoist_insertion (basic_block block)
3510 edge e;
3511 edge_iterator ei;
3512 bool new_stuff = false;
3513 unsigned i;
3514 gimple_stmt_iterator last;
3516 /* At least two successors, or else... */
3517 gcc_assert (EDGE_COUNT (block->succs) >= 2);
3519 /* Check that all successors of BLOCK are dominated by block.
3520 We could use dominated_by_p() for this, but actually there is a much
3521 quicker check: any successor that is dominated by BLOCK can't have
3522 more than one predecessor edge. */
3523 FOR_EACH_EDGE (e, ei, block->succs)
3524 if (! single_pred_p (e->dest))
3525 return false;
3527 /* Determine the insertion point. If we cannot safely insert before
3528 the last stmt if we'd have to, bail out. */
3529 last = gsi_last_bb (block);
3530 if (!gsi_end_p (last)
3531 && !is_ctrl_stmt (gsi_stmt (last))
3532 && stmt_ends_bb_p (gsi_stmt (last)))
3533 return false;
3535 /* Compute the set of hoistable expressions from ANTIC_IN. First compute
3536 hoistable values. */
3537 bitmap_set hoistable_set;
3539 /* A hoistable value must be in ANTIC_IN(block)
3540 but not in AVAIL_OUT(BLOCK). */
3541 bitmap_initialize (&hoistable_set.values, &grand_bitmap_obstack);
3542 bitmap_and_compl (&hoistable_set.values,
3543 &ANTIC_IN (block)->values, &AVAIL_OUT (block)->values);
3545 /* Short-cut for a common case: hoistable_set is empty. */
3546 if (bitmap_empty_p (&hoistable_set.values))
3547 return false;
3549 /* Compute which of the hoistable values is in AVAIL_OUT of
3550 at least one of the successors of BLOCK. */
3551 bitmap_head availout_in_some;
3552 bitmap_initialize (&availout_in_some, &grand_bitmap_obstack);
3553 FOR_EACH_EDGE (e, ei, block->succs)
3554 /* Do not consider expressions solely because their availability
3555 on loop exits. They'd be ANTIC-IN throughout the whole loop
3556 and thus effectively hoisted across loops by combination of
3557 PRE and hoisting. */
3558 if (! loop_exit_edge_p (block->loop_father, e))
3559 bitmap_ior_and_into (&availout_in_some, &hoistable_set.values,
3560 &AVAIL_OUT (e->dest)->values);
3561 bitmap_clear (&hoistable_set.values);
3563 /* Short-cut for a common case: availout_in_some is empty. */
3564 if (bitmap_empty_p (&availout_in_some))
3565 return false;
3567 /* Hack hoitable_set in-place so we can use sorted_array_from_bitmap_set. */
3568 hoistable_set.values = availout_in_some;
3569 hoistable_set.expressions = ANTIC_IN (block)->expressions;
3571 /* Now finally construct the topological-ordered expression set. */
3572 vec<pre_expr> exprs = sorted_array_from_bitmap_set (&hoistable_set);
3574 bitmap_clear (&hoistable_set.values);
3576 /* If there are candidate values for hoisting, insert expressions
3577 strategically to make the hoistable expressions fully redundant. */
3578 pre_expr expr;
3579 FOR_EACH_VEC_ELT (exprs, i, expr)
3581 /* While we try to sort expressions topologically above the
3582 sorting doesn't work out perfectly. Catch expressions we
3583 already inserted. */
3584 unsigned int value_id = get_expr_value_id (expr);
3585 if (bitmap_set_contains_value (AVAIL_OUT (block), value_id))
3587 if (dump_file && (dump_flags & TDF_DETAILS))
3589 fprintf (dump_file,
3590 "Already inserted expression for ");
3591 print_pre_expr (dump_file, expr);
3592 fprintf (dump_file, " (%04d)\n", value_id);
3594 continue;
3597 /* OK, we should hoist this value. Perform the transformation. */
3598 pre_stats.hoist_insert++;
3599 if (dump_file && (dump_flags & TDF_DETAILS))
3601 fprintf (dump_file,
3602 "Inserting expression in block %d for code hoisting: ",
3603 block->index);
3604 print_pre_expr (dump_file, expr);
3605 fprintf (dump_file, " (%04d)\n", value_id);
3608 gimple_seq stmts = NULL;
3609 tree res = create_expression_by_pieces (block, expr, &stmts,
3610 get_expr_type (expr));
3612 /* Do not return true if expression creation ultimately
3613 did not insert any statements. */
3614 if (gimple_seq_empty_p (stmts))
3615 res = NULL_TREE;
3616 else
3618 if (gsi_end_p (last) || is_ctrl_stmt (gsi_stmt (last)))
3619 gsi_insert_seq_before (&last, stmts, GSI_SAME_STMT);
3620 else
3621 gsi_insert_seq_after (&last, stmts, GSI_NEW_STMT);
3624 /* Make sure to not return true if expression creation ultimately
3625 failed but also make sure to insert any stmts produced as they
3626 are tracked in inserted_exprs. */
3627 if (! res)
3628 continue;
3630 new_stuff = true;
3633 exprs.release ();
3635 return new_stuff;
3638 /* Do a dominator walk on the control flow graph, and insert computations
3639 of values as necessary for PRE and hoisting. */
3641 static bool
3642 insert_aux (basic_block block, bool do_pre, bool do_hoist)
3644 basic_block son;
3645 bool new_stuff = false;
3647 if (block)
3649 basic_block dom;
3650 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3651 if (dom)
3653 unsigned i;
3654 bitmap_iterator bi;
3655 bitmap_set_t newset;
3657 /* First, update the AVAIL_OUT set with anything we may have
3658 inserted higher up in the dominator tree. */
3659 newset = NEW_SETS (dom);
3660 if (newset)
3662 /* Note that we need to value_replace both NEW_SETS, and
3663 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3664 represented by some non-simple expression here that we want
3665 to replace it with. */
3666 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3668 pre_expr expr = expression_for_id (i);
3669 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3670 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3674 /* Insert expressions for partial redundancies. */
3675 if (do_pre && !single_pred_p (block))
3677 new_stuff |= do_pre_regular_insertion (block, dom);
3678 if (do_partial_partial)
3679 new_stuff |= do_pre_partial_partial_insertion (block, dom);
3682 /* Insert expressions for hoisting. */
3683 if (do_hoist && EDGE_COUNT (block->succs) >= 2)
3684 new_stuff |= do_hoist_insertion (block);
3687 for (son = first_dom_son (CDI_DOMINATORS, block);
3688 son;
3689 son = next_dom_son (CDI_DOMINATORS, son))
3691 new_stuff |= insert_aux (son, do_pre, do_hoist);
3694 return new_stuff;
3697 /* Perform insertion of partially redundant and hoistable values. */
3699 static void
3700 insert (void)
3702 bool new_stuff = true;
3703 basic_block bb;
3704 int num_iterations = 0;
3706 FOR_ALL_BB_FN (bb, cfun)
3707 NEW_SETS (bb) = bitmap_set_new ();
3709 while (new_stuff)
3711 num_iterations++;
3712 if (dump_file && dump_flags & TDF_DETAILS)
3713 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3714 new_stuff = insert_aux (ENTRY_BLOCK_PTR_FOR_FN (cfun), flag_tree_pre,
3715 flag_code_hoisting);
3717 /* Clear the NEW sets before the next iteration. We have already
3718 fully propagated its contents. */
3719 if (new_stuff)
3720 FOR_ALL_BB_FN (bb, cfun)
3721 bitmap_set_free (NEW_SETS (bb));
3723 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3727 /* Compute the AVAIL set for all basic blocks.
3729 This function performs value numbering of the statements in each basic
3730 block. The AVAIL sets are built from information we glean while doing
3731 this value numbering, since the AVAIL sets contain only one entry per
3732 value.
3734 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3735 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3737 static void
3738 compute_avail (void)
3741 basic_block block, son;
3742 basic_block *worklist;
3743 size_t sp = 0;
3744 unsigned i;
3745 tree name;
3747 /* We pretend that default definitions are defined in the entry block.
3748 This includes function arguments and the static chain decl. */
3749 FOR_EACH_SSA_NAME (i, name, cfun)
3751 pre_expr e;
3752 if (!SSA_NAME_IS_DEFAULT_DEF (name)
3753 || has_zero_uses (name)
3754 || virtual_operand_p (name))
3755 continue;
3757 e = get_or_alloc_expr_for_name (name);
3758 add_to_value (get_expr_value_id (e), e);
3759 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)), e);
3760 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3764 if (dump_file && (dump_flags & TDF_DETAILS))
3766 print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3767 "tmp_gen", ENTRY_BLOCK);
3768 print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3769 "avail_out", ENTRY_BLOCK);
3772 /* Allocate the worklist. */
3773 worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
3775 /* Seed the algorithm by putting the dominator children of the entry
3776 block on the worklist. */
3777 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR_FOR_FN (cfun));
3778 son;
3779 son = next_dom_son (CDI_DOMINATORS, son))
3780 worklist[sp++] = son;
3782 BB_LIVE_VOP_ON_EXIT (ENTRY_BLOCK_PTR_FOR_FN (cfun))
3783 = ssa_default_def (cfun, gimple_vop (cfun));
3785 /* Loop until the worklist is empty. */
3786 while (sp)
3788 gimple *stmt;
3789 basic_block dom;
3791 /* Pick a block from the worklist. */
3792 block = worklist[--sp];
3794 /* Initially, the set of available values in BLOCK is that of
3795 its immediate dominator. */
3796 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3797 if (dom)
3799 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3800 BB_LIVE_VOP_ON_EXIT (block) = BB_LIVE_VOP_ON_EXIT (dom);
3803 /* Generate values for PHI nodes. */
3804 for (gphi_iterator gsi = gsi_start_phis (block); !gsi_end_p (gsi);
3805 gsi_next (&gsi))
3807 tree result = gimple_phi_result (gsi.phi ());
3809 /* We have no need for virtual phis, as they don't represent
3810 actual computations. */
3811 if (virtual_operand_p (result))
3813 BB_LIVE_VOP_ON_EXIT (block) = result;
3814 continue;
3817 pre_expr e = get_or_alloc_expr_for_name (result);
3818 add_to_value (get_expr_value_id (e), e);
3819 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3820 bitmap_insert_into_set (PHI_GEN (block), e);
3823 BB_MAY_NOTRETURN (block) = 0;
3825 /* Now compute value numbers and populate value sets with all
3826 the expressions computed in BLOCK. */
3827 for (gimple_stmt_iterator gsi = gsi_start_bb (block); !gsi_end_p (gsi);
3828 gsi_next (&gsi))
3830 ssa_op_iter iter;
3831 tree op;
3833 stmt = gsi_stmt (gsi);
3835 /* Cache whether the basic-block has any non-visible side-effect
3836 or control flow.
3837 If this isn't a call or it is the last stmt in the
3838 basic-block then the CFG represents things correctly. */
3839 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3841 /* Non-looping const functions always return normally.
3842 Otherwise the call might not return or have side-effects
3843 that forbids hoisting possibly trapping expressions
3844 before it. */
3845 int flags = gimple_call_flags (stmt);
3846 if (!(flags & ECF_CONST)
3847 || (flags & ECF_LOOPING_CONST_OR_PURE))
3848 BB_MAY_NOTRETURN (block) = 1;
3851 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3853 pre_expr e = get_or_alloc_expr_for_name (op);
3855 add_to_value (get_expr_value_id (e), e);
3856 bitmap_insert_into_set (TMP_GEN (block), e);
3857 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3860 if (gimple_vdef (stmt))
3861 BB_LIVE_VOP_ON_EXIT (block) = gimple_vdef (stmt);
3863 if (gimple_has_side_effects (stmt)
3864 || stmt_could_throw_p (stmt)
3865 || is_gimple_debug (stmt))
3866 continue;
3868 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3870 if (ssa_undefined_value_p (op))
3871 continue;
3872 pre_expr e = get_or_alloc_expr_for_name (op);
3873 bitmap_value_insert_into_set (EXP_GEN (block), e);
3876 switch (gimple_code (stmt))
3878 case GIMPLE_RETURN:
3879 continue;
3881 case GIMPLE_CALL:
3883 vn_reference_t ref;
3884 vn_reference_s ref1;
3885 pre_expr result = NULL;
3887 /* We can value number only calls to real functions. */
3888 if (gimple_call_internal_p (stmt))
3889 continue;
3891 vn_reference_lookup_call (as_a <gcall *> (stmt), &ref, &ref1);
3892 if (!ref)
3893 continue;
3895 /* If the value of the call is not invalidated in
3896 this block until it is computed, add the expression
3897 to EXP_GEN. */
3898 if (!gimple_vuse (stmt)
3899 || gimple_code
3900 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
3901 || gimple_bb (SSA_NAME_DEF_STMT
3902 (gimple_vuse (stmt))) != block)
3904 result = pre_expr_pool.allocate ();
3905 result->kind = REFERENCE;
3906 result->id = 0;
3907 PRE_EXPR_REFERENCE (result) = ref;
3909 get_or_alloc_expression_id (result);
3910 add_to_value (get_expr_value_id (result), result);
3911 bitmap_value_insert_into_set (EXP_GEN (block), result);
3913 continue;
3916 case GIMPLE_ASSIGN:
3918 pre_expr result = NULL;
3919 switch (vn_get_stmt_kind (stmt))
3921 case VN_NARY:
3923 enum tree_code code = gimple_assign_rhs_code (stmt);
3924 vn_nary_op_t nary;
3926 /* COND_EXPR and VEC_COND_EXPR are awkward in
3927 that they contain an embedded complex expression.
3928 Don't even try to shove those through PRE. */
3929 if (code == COND_EXPR
3930 || code == VEC_COND_EXPR)
3931 continue;
3933 vn_nary_op_lookup_stmt (stmt, &nary);
3934 if (!nary)
3935 continue;
3937 /* If the NARY traps and there was a preceding
3938 point in the block that might not return avoid
3939 adding the nary to EXP_GEN. */
3940 if (BB_MAY_NOTRETURN (block)
3941 && vn_nary_may_trap (nary))
3942 continue;
3944 result = pre_expr_pool.allocate ();
3945 result->kind = NARY;
3946 result->id = 0;
3947 PRE_EXPR_NARY (result) = nary;
3948 break;
3951 case VN_REFERENCE:
3953 tree rhs1 = gimple_assign_rhs1 (stmt);
3954 alias_set_type set = get_alias_set (rhs1);
3955 vec<vn_reference_op_s> operands
3956 = vn_reference_operands_for_lookup (rhs1);
3957 vn_reference_t ref;
3958 vn_reference_lookup_pieces (gimple_vuse (stmt), set,
3959 TREE_TYPE (rhs1),
3960 operands, &ref, VN_WALK);
3961 if (!ref)
3963 operands.release ();
3964 continue;
3967 /* If the value of the reference is not invalidated in
3968 this block until it is computed, add the expression
3969 to EXP_GEN. */
3970 if (gimple_vuse (stmt))
3972 gimple *def_stmt;
3973 bool ok = true;
3974 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
3975 while (!gimple_nop_p (def_stmt)
3976 && gimple_code (def_stmt) != GIMPLE_PHI
3977 && gimple_bb (def_stmt) == block)
3979 if (stmt_may_clobber_ref_p
3980 (def_stmt, gimple_assign_rhs1 (stmt)))
3982 ok = false;
3983 break;
3985 def_stmt
3986 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
3988 if (!ok)
3990 operands.release ();
3991 continue;
3995 /* If the load was value-numbered to another
3996 load make sure we do not use its expression
3997 for insertion if it wouldn't be a valid
3998 replacement. */
3999 /* At the momemt we have a testcase
4000 for hoist insertion of aligned vs. misaligned
4001 variants in gcc.dg/torture/pr65270-1.c thus
4002 with just alignment to be considered we can
4003 simply replace the expression in the hashtable
4004 with the most conservative one. */
4005 vn_reference_op_t ref1 = &ref->operands.last ();
4006 while (ref1->opcode != TARGET_MEM_REF
4007 && ref1->opcode != MEM_REF
4008 && ref1 != &ref->operands[0])
4009 --ref1;
4010 vn_reference_op_t ref2 = &operands.last ();
4011 while (ref2->opcode != TARGET_MEM_REF
4012 && ref2->opcode != MEM_REF
4013 && ref2 != &operands[0])
4014 --ref2;
4015 if ((ref1->opcode == TARGET_MEM_REF
4016 || ref1->opcode == MEM_REF)
4017 && (TYPE_ALIGN (ref1->type)
4018 > TYPE_ALIGN (ref2->type)))
4019 ref1->type
4020 = build_aligned_type (ref1->type,
4021 TYPE_ALIGN (ref2->type));
4022 /* TBAA behavior is an obvious part so make sure
4023 that the hashtable one covers this as well
4024 by adjusting the ref alias set and its base. */
4025 if (ref->set == set
4026 || alias_set_subset_of (set, ref->set))
4028 else if (alias_set_subset_of (ref->set, set))
4030 ref->set = set;
4031 if (ref1->opcode == MEM_REF)
4032 ref1->op0 = wide_int_to_tree (TREE_TYPE (ref2->op0),
4033 ref1->op0);
4034 else
4035 ref1->op2 = wide_int_to_tree (TREE_TYPE (ref2->op2),
4036 ref1->op2);
4038 else
4040 ref->set = 0;
4041 if (ref1->opcode == MEM_REF)
4042 ref1->op0 = wide_int_to_tree (ptr_type_node,
4043 ref1->op0);
4044 else
4045 ref1->op2 = wide_int_to_tree (ptr_type_node,
4046 ref1->op2);
4048 operands.release ();
4050 result = pre_expr_pool.allocate ();
4051 result->kind = REFERENCE;
4052 result->id = 0;
4053 PRE_EXPR_REFERENCE (result) = ref;
4054 break;
4057 default:
4058 continue;
4061 get_or_alloc_expression_id (result);
4062 add_to_value (get_expr_value_id (result), result);
4063 bitmap_value_insert_into_set (EXP_GEN (block), result);
4064 continue;
4066 default:
4067 break;
4071 if (dump_file && (dump_flags & TDF_DETAILS))
4073 print_bitmap_set (dump_file, EXP_GEN (block),
4074 "exp_gen", block->index);
4075 print_bitmap_set (dump_file, PHI_GEN (block),
4076 "phi_gen", block->index);
4077 print_bitmap_set (dump_file, TMP_GEN (block),
4078 "tmp_gen", block->index);
4079 print_bitmap_set (dump_file, AVAIL_OUT (block),
4080 "avail_out", block->index);
4083 /* Put the dominator children of BLOCK on the worklist of blocks
4084 to compute available sets for. */
4085 for (son = first_dom_son (CDI_DOMINATORS, block);
4086 son;
4087 son = next_dom_son (CDI_DOMINATORS, son))
4088 worklist[sp++] = son;
4091 free (worklist);
4095 /* Local state for the eliminate domwalk. */
4096 static vec<gimple *> el_to_remove;
4097 static vec<gimple *> el_to_fixup;
4098 static unsigned int el_todo;
4099 static vec<tree> el_avail;
4100 static vec<tree> el_avail_stack;
4102 /* Return a leader for OP that is available at the current point of the
4103 eliminate domwalk. */
4105 static tree
4106 eliminate_avail (tree op)
4108 tree valnum = VN_INFO (op)->valnum;
4109 if (TREE_CODE (valnum) == SSA_NAME)
4111 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
4112 return valnum;
4113 if (el_avail.length () > SSA_NAME_VERSION (valnum))
4114 return el_avail[SSA_NAME_VERSION (valnum)];
4116 else if (is_gimple_min_invariant (valnum))
4117 return valnum;
4118 return NULL_TREE;
4121 /* At the current point of the eliminate domwalk make OP available. */
4123 static void
4124 eliminate_push_avail (tree op)
4126 tree valnum = VN_INFO (op)->valnum;
4127 if (TREE_CODE (valnum) == SSA_NAME)
4129 if (el_avail.length () <= SSA_NAME_VERSION (valnum))
4130 el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1);
4131 tree pushop = op;
4132 if (el_avail[SSA_NAME_VERSION (valnum)])
4133 pushop = el_avail[SSA_NAME_VERSION (valnum)];
4134 el_avail_stack.safe_push (pushop);
4135 el_avail[SSA_NAME_VERSION (valnum)] = op;
4139 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
4140 the leader for the expression if insertion was successful. */
4142 static tree
4143 eliminate_insert (gimple_stmt_iterator *gsi, tree val)
4145 /* We can insert a sequence with a single assignment only. */
4146 gimple_seq stmts = VN_INFO (val)->expr;
4147 if (!gimple_seq_singleton_p (stmts))
4148 return NULL_TREE;
4149 gassign *stmt = dyn_cast <gassign *> (gimple_seq_first_stmt (stmts));
4150 if (!stmt
4151 || (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
4152 && gimple_assign_rhs_code (stmt) != VIEW_CONVERT_EXPR
4153 && gimple_assign_rhs_code (stmt) != BIT_FIELD_REF
4154 && (gimple_assign_rhs_code (stmt) != BIT_AND_EXPR
4155 || TREE_CODE (gimple_assign_rhs2 (stmt)) != INTEGER_CST)))
4156 return NULL_TREE;
4158 tree op = gimple_assign_rhs1 (stmt);
4159 if (gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR
4160 || gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
4161 op = TREE_OPERAND (op, 0);
4162 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op;
4163 if (!leader)
4164 return NULL_TREE;
4166 tree res;
4167 stmts = NULL;
4168 if (gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
4169 res = gimple_build (&stmts, BIT_FIELD_REF,
4170 TREE_TYPE (val), leader,
4171 TREE_OPERAND (gimple_assign_rhs1 (stmt), 1),
4172 TREE_OPERAND (gimple_assign_rhs1 (stmt), 2));
4173 else if (gimple_assign_rhs_code (stmt) == BIT_AND_EXPR)
4174 res = gimple_build (&stmts, BIT_AND_EXPR,
4175 TREE_TYPE (val), leader, gimple_assign_rhs2 (stmt));
4176 else
4177 res = gimple_build (&stmts, gimple_assign_rhs_code (stmt),
4178 TREE_TYPE (val), leader);
4179 if (TREE_CODE (res) != SSA_NAME
4180 || SSA_NAME_IS_DEFAULT_DEF (res)
4181 || gimple_bb (SSA_NAME_DEF_STMT (res)))
4183 gimple_seq_discard (stmts);
4185 /* During propagation we have to treat SSA info conservatively
4186 and thus we can end up simplifying the inserted expression
4187 at elimination time to sth not defined in stmts. */
4188 /* But then this is a redundancy we failed to detect. Which means
4189 res now has two values. That doesn't play well with how
4190 we track availability here, so give up. */
4191 if (dump_file && (dump_flags & TDF_DETAILS))
4193 if (TREE_CODE (res) == SSA_NAME)
4194 res = eliminate_avail (res);
4195 if (res)
4197 fprintf (dump_file, "Failed to insert expression for value ");
4198 print_generic_expr (dump_file, val);
4199 fprintf (dump_file, " which is really fully redundant to ");
4200 print_generic_expr (dump_file, res);
4201 fprintf (dump_file, "\n");
4205 return NULL_TREE;
4207 else
4209 gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
4210 VN_INFO_GET (res)->valnum = val;
4212 if (TREE_CODE (leader) == SSA_NAME)
4213 gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true);
4216 pre_stats.insertions++;
4217 if (dump_file && (dump_flags & TDF_DETAILS))
4219 fprintf (dump_file, "Inserted ");
4220 print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (res), 0);
4223 return res;
4226 class eliminate_dom_walker : public dom_walker
4228 public:
4229 eliminate_dom_walker (cdi_direction direction, bool do_pre_)
4230 : dom_walker (direction), do_pre (do_pre_) {}
4232 virtual edge before_dom_children (basic_block);
4233 virtual void after_dom_children (basic_block);
4235 bool do_pre;
4238 /* Perform elimination for the basic-block B during the domwalk. */
4240 edge
4241 eliminate_dom_walker::before_dom_children (basic_block b)
4243 /* Mark new bb. */
4244 el_avail_stack.safe_push (NULL_TREE);
4246 /* Skip unreachable blocks marked unreachable during the SCCVN domwalk. */
4247 edge_iterator ei;
4248 edge e;
4249 FOR_EACH_EDGE (e, ei, b->preds)
4250 if (e->flags & EDGE_EXECUTABLE)
4251 break;
4252 if (! e)
4253 return NULL;
4255 for (gphi_iterator gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4257 gphi *phi = gsi.phi ();
4258 tree res = PHI_RESULT (phi);
4260 if (virtual_operand_p (res))
4262 gsi_next (&gsi);
4263 continue;
4266 tree sprime = eliminate_avail (res);
4267 if (sprime
4268 && sprime != res)
4270 if (dump_file && (dump_flags & TDF_DETAILS))
4272 fprintf (dump_file, "Replaced redundant PHI node defining ");
4273 print_generic_expr (dump_file, res);
4274 fprintf (dump_file, " with ");
4275 print_generic_expr (dump_file, sprime);
4276 fprintf (dump_file, "\n");
4279 /* If we inserted this PHI node ourself, it's not an elimination. */
4280 if (inserted_exprs
4281 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4282 pre_stats.phis--;
4283 else
4284 pre_stats.eliminations++;
4286 /* If we will propagate into all uses don't bother to do
4287 anything. */
4288 if (may_propagate_copy (res, sprime))
4290 /* Mark the PHI for removal. */
4291 el_to_remove.safe_push (phi);
4292 gsi_next (&gsi);
4293 continue;
4296 remove_phi_node (&gsi, false);
4298 if (inserted_exprs
4299 && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4300 && TREE_CODE (sprime) == SSA_NAME)
4301 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4303 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4304 sprime = fold_convert (TREE_TYPE (res), sprime);
4305 gimple *stmt = gimple_build_assign (res, sprime);
4306 /* ??? It cannot yet be necessary (DOM walk). */
4307 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4309 gimple_stmt_iterator gsi2 = gsi_after_labels (b);
4310 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4311 continue;
4314 eliminate_push_avail (res);
4315 gsi_next (&gsi);
4318 for (gimple_stmt_iterator gsi = gsi_start_bb (b);
4319 !gsi_end_p (gsi);
4320 gsi_next (&gsi))
4322 tree sprime = NULL_TREE;
4323 gimple *stmt = gsi_stmt (gsi);
4324 tree lhs = gimple_get_lhs (stmt);
4325 if (lhs && TREE_CODE (lhs) == SSA_NAME
4326 && !gimple_has_volatile_ops (stmt)
4327 /* See PR43491. Do not replace a global register variable when
4328 it is a the RHS of an assignment. Do replace local register
4329 variables since gcc does not guarantee a local variable will
4330 be allocated in register.
4331 ??? The fix isn't effective here. This should instead
4332 be ensured by not value-numbering them the same but treating
4333 them like volatiles? */
4334 && !(gimple_assign_single_p (stmt)
4335 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL
4336 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt))
4337 && is_global_var (gimple_assign_rhs1 (stmt)))))
4339 sprime = eliminate_avail (lhs);
4340 if (!sprime)
4342 /* If there is no existing usable leader but SCCVN thinks
4343 it has an expression it wants to use as replacement,
4344 insert that. */
4345 tree val = VN_INFO (lhs)->valnum;
4346 if (val != VN_TOP
4347 && TREE_CODE (val) == SSA_NAME
4348 && VN_INFO (val)->needs_insertion
4349 && VN_INFO (val)->expr != NULL
4350 && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE)
4351 eliminate_push_avail (sprime);
4354 /* If this now constitutes a copy duplicate points-to
4355 and range info appropriately. This is especially
4356 important for inserted code. See tree-ssa-copy.c
4357 for similar code. */
4358 if (sprime
4359 && TREE_CODE (sprime) == SSA_NAME)
4361 basic_block sprime_b = gimple_bb (SSA_NAME_DEF_STMT (sprime));
4362 if (POINTER_TYPE_P (TREE_TYPE (lhs))
4363 && VN_INFO_PTR_INFO (lhs)
4364 && ! VN_INFO_PTR_INFO (sprime))
4366 duplicate_ssa_name_ptr_info (sprime,
4367 VN_INFO_PTR_INFO (lhs));
4368 if (b != sprime_b)
4369 mark_ptr_info_alignment_unknown
4370 (SSA_NAME_PTR_INFO (sprime));
4372 else if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))
4373 && VN_INFO_RANGE_INFO (lhs)
4374 && ! VN_INFO_RANGE_INFO (sprime)
4375 && b == sprime_b)
4376 duplicate_ssa_name_range_info (sprime,
4377 VN_INFO_RANGE_TYPE (lhs),
4378 VN_INFO_RANGE_INFO (lhs));
4381 /* Inhibit the use of an inserted PHI on a loop header when
4382 the address of the memory reference is a simple induction
4383 variable. In other cases the vectorizer won't do anything
4384 anyway (either it's loop invariant or a complicated
4385 expression). */
4386 if (sprime
4387 && TREE_CODE (sprime) == SSA_NAME
4388 && do_pre
4389 && (flag_tree_loop_vectorize || flag_tree_parallelize_loops > 1)
4390 && loop_outer (b->loop_father)
4391 && has_zero_uses (sprime)
4392 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))
4393 && gimple_assign_load_p (stmt))
4395 gimple *def_stmt = SSA_NAME_DEF_STMT (sprime);
4396 basic_block def_bb = gimple_bb (def_stmt);
4397 if (gimple_code (def_stmt) == GIMPLE_PHI
4398 && def_bb->loop_father->header == def_bb)
4400 loop_p loop = def_bb->loop_father;
4401 ssa_op_iter iter;
4402 tree op;
4403 bool found = false;
4404 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4406 affine_iv iv;
4407 def_bb = gimple_bb (SSA_NAME_DEF_STMT (op));
4408 if (def_bb
4409 && flow_bb_inside_loop_p (loop, def_bb)
4410 && simple_iv (loop, loop, op, &iv, true))
4412 found = true;
4413 break;
4416 if (found)
4418 if (dump_file && (dump_flags & TDF_DETAILS))
4420 fprintf (dump_file, "Not replacing ");
4421 print_gimple_expr (dump_file, stmt, 0);
4422 fprintf (dump_file, " with ");
4423 print_generic_expr (dump_file, sprime);
4424 fprintf (dump_file, " which would add a loop"
4425 " carried dependence to loop %d\n",
4426 loop->num);
4428 /* Don't keep sprime available. */
4429 sprime = NULL_TREE;
4434 if (sprime)
4436 /* If we can propagate the value computed for LHS into
4437 all uses don't bother doing anything with this stmt. */
4438 if (may_propagate_copy (lhs, sprime))
4440 /* Mark it for removal. */
4441 el_to_remove.safe_push (stmt);
4443 /* ??? Don't count copy/constant propagations. */
4444 if (gimple_assign_single_p (stmt)
4445 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4446 || gimple_assign_rhs1 (stmt) == sprime))
4447 continue;
4449 if (dump_file && (dump_flags & TDF_DETAILS))
4451 fprintf (dump_file, "Replaced ");
4452 print_gimple_expr (dump_file, stmt, 0);
4453 fprintf (dump_file, " with ");
4454 print_generic_expr (dump_file, sprime);
4455 fprintf (dump_file, " in all uses of ");
4456 print_gimple_stmt (dump_file, stmt, 0);
4459 pre_stats.eliminations++;
4460 continue;
4463 /* If this is an assignment from our leader (which
4464 happens in the case the value-number is a constant)
4465 then there is nothing to do. */
4466 if (gimple_assign_single_p (stmt)
4467 && sprime == gimple_assign_rhs1 (stmt))
4468 continue;
4470 /* Else replace its RHS. */
4471 bool can_make_abnormal_goto
4472 = is_gimple_call (stmt)
4473 && stmt_can_make_abnormal_goto (stmt);
4475 if (dump_file && (dump_flags & TDF_DETAILS))
4477 fprintf (dump_file, "Replaced ");
4478 print_gimple_expr (dump_file, stmt, 0);
4479 fprintf (dump_file, " with ");
4480 print_generic_expr (dump_file, sprime);
4481 fprintf (dump_file, " in ");
4482 print_gimple_stmt (dump_file, stmt, 0);
4485 if (TREE_CODE (sprime) == SSA_NAME)
4486 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4487 NECESSARY, true);
4489 pre_stats.eliminations++;
4490 gimple *orig_stmt = stmt;
4491 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4492 TREE_TYPE (sprime)))
4493 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4494 tree vdef = gimple_vdef (stmt);
4495 tree vuse = gimple_vuse (stmt);
4496 propagate_tree_value_into_stmt (&gsi, sprime);
4497 stmt = gsi_stmt (gsi);
4498 update_stmt (stmt);
4499 if (vdef != gimple_vdef (stmt))
4500 VN_INFO (vdef)->valnum = vuse;
4502 /* If we removed EH side-effects from the statement, clean
4503 its EH information. */
4504 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4506 bitmap_set_bit (need_eh_cleanup,
4507 gimple_bb (stmt)->index);
4508 if (dump_file && (dump_flags & TDF_DETAILS))
4509 fprintf (dump_file, " Removed EH side-effects.\n");
4512 /* Likewise for AB side-effects. */
4513 if (can_make_abnormal_goto
4514 && !stmt_can_make_abnormal_goto (stmt))
4516 bitmap_set_bit (need_ab_cleanup,
4517 gimple_bb (stmt)->index);
4518 if (dump_file && (dump_flags & TDF_DETAILS))
4519 fprintf (dump_file, " Removed AB side-effects.\n");
4522 continue;
4526 /* If the statement is a scalar store, see if the expression
4527 has the same value number as its rhs. If so, the store is
4528 dead. */
4529 if (gimple_assign_single_p (stmt)
4530 && !gimple_has_volatile_ops (stmt)
4531 && !is_gimple_reg (gimple_assign_lhs (stmt))
4532 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4533 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4535 tree val;
4536 tree rhs = gimple_assign_rhs1 (stmt);
4537 vn_reference_t vnresult;
4538 val = vn_reference_lookup (lhs, gimple_vuse (stmt), VN_WALKREWRITE,
4539 &vnresult, false);
4540 if (TREE_CODE (rhs) == SSA_NAME)
4541 rhs = VN_INFO (rhs)->valnum;
4542 if (val
4543 && operand_equal_p (val, rhs, 0))
4545 /* We can only remove the later store if the former aliases
4546 at least all accesses the later one does or if the store
4547 was to readonly memory storing the same value. */
4548 alias_set_type set = get_alias_set (lhs);
4549 if (! vnresult
4550 || vnresult->set == set
4551 || alias_set_subset_of (set, vnresult->set))
4553 if (dump_file && (dump_flags & TDF_DETAILS))
4555 fprintf (dump_file, "Deleted redundant store ");
4556 print_gimple_stmt (dump_file, stmt, 0);
4559 /* Queue stmt for removal. */
4560 el_to_remove.safe_push (stmt);
4561 continue;
4566 /* If this is a control statement value numbering left edges
4567 unexecuted on force the condition in a way consistent with
4568 that. */
4569 if (gcond *cond = dyn_cast <gcond *> (stmt))
4571 if ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE)
4572 ^ (EDGE_SUCC (b, 1)->flags & EDGE_EXECUTABLE))
4574 if (dump_file && (dump_flags & TDF_DETAILS))
4576 fprintf (dump_file, "Removing unexecutable edge from ");
4577 print_gimple_stmt (dump_file, stmt, 0);
4579 if (((EDGE_SUCC (b, 0)->flags & EDGE_TRUE_VALUE) != 0)
4580 == ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE) != 0))
4581 gimple_cond_make_true (cond);
4582 else
4583 gimple_cond_make_false (cond);
4584 update_stmt (cond);
4585 el_todo |= TODO_cleanup_cfg;
4586 continue;
4590 bool can_make_abnormal_goto = stmt_can_make_abnormal_goto (stmt);
4591 bool was_noreturn = (is_gimple_call (stmt)
4592 && gimple_call_noreturn_p (stmt));
4593 tree vdef = gimple_vdef (stmt);
4594 tree vuse = gimple_vuse (stmt);
4596 /* If we didn't replace the whole stmt (or propagate the result
4597 into all uses), replace all uses on this stmt with their
4598 leaders. */
4599 use_operand_p use_p;
4600 ssa_op_iter iter;
4601 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
4603 tree use = USE_FROM_PTR (use_p);
4604 /* ??? The call code above leaves stmt operands un-updated. */
4605 if (TREE_CODE (use) != SSA_NAME)
4606 continue;
4607 tree sprime = eliminate_avail (use);
4608 if (sprime && sprime != use
4609 && may_propagate_copy (use, sprime)
4610 /* We substitute into debug stmts to avoid excessive
4611 debug temporaries created by removed stmts, but we need
4612 to avoid doing so for inserted sprimes as we never want
4613 to create debug temporaries for them. */
4614 && (!inserted_exprs
4615 || TREE_CODE (sprime) != SSA_NAME
4616 || !is_gimple_debug (stmt)
4617 || !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))))
4619 propagate_value (use_p, sprime);
4620 gimple_set_modified (stmt, true);
4621 if (TREE_CODE (sprime) == SSA_NAME
4622 && !is_gimple_debug (stmt))
4623 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4624 NECESSARY, true);
4628 /* Visit indirect calls and turn them into direct calls if
4629 possible using the devirtualization machinery. */
4630 if (gcall *call_stmt = dyn_cast <gcall *> (stmt))
4632 tree fn = gimple_call_fn (call_stmt);
4633 if (fn
4634 && flag_devirtualize
4635 && virtual_method_call_p (fn))
4637 tree otr_type = obj_type_ref_class (fn);
4638 tree instance;
4639 ipa_polymorphic_call_context context (current_function_decl, fn, stmt, &instance);
4640 bool final;
4642 context.get_dynamic_type (instance, OBJ_TYPE_REF_OBJECT (fn), otr_type, stmt);
4644 vec <cgraph_node *>targets
4645 = possible_polymorphic_call_targets (obj_type_ref_class (fn),
4646 tree_to_uhwi
4647 (OBJ_TYPE_REF_TOKEN (fn)),
4648 context,
4649 &final);
4650 if (dump_file)
4651 dump_possible_polymorphic_call_targets (dump_file,
4652 obj_type_ref_class (fn),
4653 tree_to_uhwi
4654 (OBJ_TYPE_REF_TOKEN (fn)),
4655 context);
4656 if (final && targets.length () <= 1 && dbg_cnt (devirt))
4658 tree fn;
4659 if (targets.length () == 1)
4660 fn = targets[0]->decl;
4661 else
4662 fn = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
4663 if (dump_enabled_p ())
4665 location_t loc = gimple_location_safe (stmt);
4666 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc,
4667 "converting indirect call to "
4668 "function %s\n",
4669 lang_hooks.decl_printable_name (fn, 2));
4671 gimple_call_set_fndecl (call_stmt, fn);
4672 /* If changing the call to __builtin_unreachable
4673 or similar noreturn function, adjust gimple_call_fntype
4674 too. */
4675 if (gimple_call_noreturn_p (call_stmt)
4676 && VOID_TYPE_P (TREE_TYPE (TREE_TYPE (fn)))
4677 && TYPE_ARG_TYPES (TREE_TYPE (fn))
4678 && (TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fn)))
4679 == void_type_node))
4680 gimple_call_set_fntype (call_stmt, TREE_TYPE (fn));
4681 maybe_remove_unused_call_args (cfun, call_stmt);
4682 gimple_set_modified (stmt, true);
4687 if (gimple_modified_p (stmt))
4689 /* If a formerly non-invariant ADDR_EXPR is turned into an
4690 invariant one it was on a separate stmt. */
4691 if (gimple_assign_single_p (stmt)
4692 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR)
4693 recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
4694 gimple *old_stmt = stmt;
4695 gimple_stmt_iterator prev = gsi;
4696 gsi_prev (&prev);
4697 if (fold_stmt (&gsi))
4699 /* fold_stmt may have created new stmts inbetween
4700 the previous stmt and the folded stmt. Mark
4701 all defs created there as varying to not confuse
4702 the SCCVN machinery as we're using that even during
4703 elimination. */
4704 if (gsi_end_p (prev))
4705 prev = gsi_start_bb (b);
4706 else
4707 gsi_next (&prev);
4708 if (gsi_stmt (prev) != gsi_stmt (gsi))
4711 tree def;
4712 ssa_op_iter dit;
4713 FOR_EACH_SSA_TREE_OPERAND (def, gsi_stmt (prev),
4714 dit, SSA_OP_ALL_DEFS)
4715 /* As existing DEFs may move between stmts
4716 we have to guard VN_INFO_GET. */
4717 if (! has_VN_INFO (def))
4718 VN_INFO_GET (def)->valnum = def;
4719 if (gsi_stmt (prev) == gsi_stmt (gsi))
4720 break;
4721 gsi_next (&prev);
4723 while (1);
4725 stmt = gsi_stmt (gsi);
4726 /* When changing a call into a noreturn call, cfg cleanup
4727 is needed to fix up the noreturn call. */
4728 if (!was_noreturn
4729 && is_gimple_call (stmt) && gimple_call_noreturn_p (stmt))
4730 el_to_fixup.safe_push (stmt);
4731 /* When changing a condition or switch into one we know what
4732 edge will be executed, schedule a cfg cleanup. */
4733 if ((gimple_code (stmt) == GIMPLE_COND
4734 && (gimple_cond_true_p (as_a <gcond *> (stmt))
4735 || gimple_cond_false_p (as_a <gcond *> (stmt))))
4736 || (gimple_code (stmt) == GIMPLE_SWITCH
4737 && TREE_CODE (gimple_switch_index
4738 (as_a <gswitch *> (stmt))) == INTEGER_CST))
4739 el_todo |= TODO_cleanup_cfg;
4740 /* If we removed EH side-effects from the statement, clean
4741 its EH information. */
4742 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
4744 bitmap_set_bit (need_eh_cleanup,
4745 gimple_bb (stmt)->index);
4746 if (dump_file && (dump_flags & TDF_DETAILS))
4747 fprintf (dump_file, " Removed EH side-effects.\n");
4749 /* Likewise for AB side-effects. */
4750 if (can_make_abnormal_goto
4751 && !stmt_can_make_abnormal_goto (stmt))
4753 bitmap_set_bit (need_ab_cleanup,
4754 gimple_bb (stmt)->index);
4755 if (dump_file && (dump_flags & TDF_DETAILS))
4756 fprintf (dump_file, " Removed AB side-effects.\n");
4758 update_stmt (stmt);
4759 if (vdef != gimple_vdef (stmt))
4760 VN_INFO (vdef)->valnum = vuse;
4763 /* Make new values available - for fully redundant LHS we
4764 continue with the next stmt above and skip this. */
4765 def_operand_p defp;
4766 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_DEF)
4767 eliminate_push_avail (DEF_FROM_PTR (defp));
4770 /* Replace destination PHI arguments. */
4771 FOR_EACH_EDGE (e, ei, b->succs)
4772 if (e->flags & EDGE_EXECUTABLE)
4773 for (gphi_iterator gsi = gsi_start_phis (e->dest);
4774 !gsi_end_p (gsi);
4775 gsi_next (&gsi))
4777 gphi *phi = gsi.phi ();
4778 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
4779 tree arg = USE_FROM_PTR (use_p);
4780 if (TREE_CODE (arg) != SSA_NAME
4781 || virtual_operand_p (arg))
4782 continue;
4783 tree sprime = eliminate_avail (arg);
4784 if (sprime && may_propagate_copy (arg, sprime))
4786 propagate_value (use_p, sprime);
4787 if (TREE_CODE (sprime) == SSA_NAME)
4788 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4791 return NULL;
4794 /* Make no longer available leaders no longer available. */
4796 void
4797 eliminate_dom_walker::after_dom_children (basic_block)
4799 tree entry;
4800 while ((entry = el_avail_stack.pop ()) != NULL_TREE)
4802 tree valnum = VN_INFO (entry)->valnum;
4803 tree old = el_avail[SSA_NAME_VERSION (valnum)];
4804 if (old == entry)
4805 el_avail[SSA_NAME_VERSION (valnum)] = NULL_TREE;
4806 else
4807 el_avail[SSA_NAME_VERSION (valnum)] = entry;
4811 /* Eliminate fully redundant computations. */
4813 static unsigned int
4814 eliminate (bool do_pre)
4816 need_eh_cleanup = BITMAP_ALLOC (NULL);
4817 need_ab_cleanup = BITMAP_ALLOC (NULL);
4819 el_to_remove.create (0);
4820 el_to_fixup.create (0);
4821 el_todo = 0;
4822 el_avail.create (num_ssa_names);
4823 el_avail_stack.create (0);
4825 eliminate_dom_walker (CDI_DOMINATORS,
4826 do_pre).walk (cfun->cfg->x_entry_block_ptr);
4828 el_avail.release ();
4829 el_avail_stack.release ();
4831 return el_todo;
4834 /* Perform CFG cleanups made necessary by elimination. */
4836 static unsigned
4837 fini_eliminate (void)
4839 gimple_stmt_iterator gsi;
4840 gimple *stmt;
4841 unsigned todo = 0;
4843 /* We cannot remove stmts during BB walk, especially not release SSA
4844 names there as this confuses the VN machinery. The stmts ending
4845 up in el_to_remove are either stores or simple copies.
4846 Remove stmts in reverse order to make debug stmt creation possible. */
4847 while (!el_to_remove.is_empty ())
4849 stmt = el_to_remove.pop ();
4851 tree lhs;
4852 if (gimple_code (stmt) == GIMPLE_PHI)
4853 lhs = gimple_phi_result (stmt);
4854 else
4855 lhs = gimple_get_lhs (stmt);
4857 if (inserted_exprs
4858 && TREE_CODE (lhs) == SSA_NAME
4859 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs)))
4860 continue;
4862 if (dump_file && (dump_flags & TDF_DETAILS))
4864 fprintf (dump_file, "Removing dead stmt ");
4865 print_gimple_stmt (dump_file, stmt, 0, 0);
4868 gsi = gsi_for_stmt (stmt);
4869 if (gimple_code (stmt) == GIMPLE_PHI)
4870 remove_phi_node (&gsi, true);
4871 else
4873 basic_block bb = gimple_bb (stmt);
4874 unlink_stmt_vdef (stmt);
4875 if (gsi_remove (&gsi, true))
4876 bitmap_set_bit (need_eh_cleanup, bb->index);
4877 if (is_gimple_call (stmt) && stmt_can_make_abnormal_goto (stmt))
4878 bitmap_set_bit (need_ab_cleanup, bb->index);
4879 release_defs (stmt);
4882 /* Removing a stmt may expose a forwarder block. */
4883 todo |= TODO_cleanup_cfg;
4885 el_to_remove.release ();
4887 /* Fixup stmts that became noreturn calls. This may require splitting
4888 blocks and thus isn't possible during the dominator walk. Do this
4889 in reverse order so we don't inadvertedly remove a stmt we want to
4890 fixup by visiting a dominating now noreturn call first. */
4891 while (!el_to_fixup.is_empty ())
4893 stmt = el_to_fixup.pop ();
4895 if (dump_file && (dump_flags & TDF_DETAILS))
4897 fprintf (dump_file, "Fixing up noreturn call ");
4898 print_gimple_stmt (dump_file, stmt, 0);
4901 if (fixup_noreturn_call (stmt))
4902 todo |= TODO_cleanup_cfg;
4904 el_to_fixup.release ();
4906 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4907 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4909 if (do_eh_cleanup)
4910 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4912 if (do_ab_cleanup)
4913 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4915 BITMAP_FREE (need_eh_cleanup);
4916 BITMAP_FREE (need_ab_cleanup);
4918 if (do_eh_cleanup || do_ab_cleanup)
4919 todo |= TODO_cleanup_cfg;
4920 return todo;
4923 /* Borrow a bit of tree-ssa-dce.c for the moment.
4924 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4925 this may be a bit faster, and we may want critical edges kept split. */
4927 /* If OP's defining statement has not already been determined to be necessary,
4928 mark that statement necessary. Return the stmt, if it is newly
4929 necessary. */
4931 static inline gimple *
4932 mark_operand_necessary (tree op)
4934 gimple *stmt;
4936 gcc_assert (op);
4938 if (TREE_CODE (op) != SSA_NAME)
4939 return NULL;
4941 stmt = SSA_NAME_DEF_STMT (op);
4942 gcc_assert (stmt);
4944 if (gimple_plf (stmt, NECESSARY)
4945 || gimple_nop_p (stmt))
4946 return NULL;
4948 gimple_set_plf (stmt, NECESSARY, true);
4949 return stmt;
4952 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4953 to insert PHI nodes sometimes, and because value numbering of casts isn't
4954 perfect, we sometimes end up inserting dead code. This simple DCE-like
4955 pass removes any insertions we made that weren't actually used. */
4957 static void
4958 remove_dead_inserted_code (void)
4960 unsigned i;
4961 bitmap_iterator bi;
4962 gimple *t;
4964 auto_bitmap worklist;
4965 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4967 t = SSA_NAME_DEF_STMT (ssa_name (i));
4968 if (gimple_plf (t, NECESSARY))
4969 bitmap_set_bit (worklist, i);
4971 while (!bitmap_empty_p (worklist))
4973 i = bitmap_first_set_bit (worklist);
4974 bitmap_clear_bit (worklist, i);
4975 t = SSA_NAME_DEF_STMT (ssa_name (i));
4977 /* PHI nodes are somewhat special in that each PHI alternative has
4978 data and control dependencies. All the statements feeding the
4979 PHI node's arguments are always necessary. */
4980 if (gimple_code (t) == GIMPLE_PHI)
4982 unsigned k;
4984 for (k = 0; k < gimple_phi_num_args (t); k++)
4986 tree arg = PHI_ARG_DEF (t, k);
4987 if (TREE_CODE (arg) == SSA_NAME)
4989 gimple *n = mark_operand_necessary (arg);
4990 if (n)
4991 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4995 else
4997 /* Propagate through the operands. Examine all the USE, VUSE and
4998 VDEF operands in this statement. Mark all the statements
4999 which feed this statement's uses as necessary. */
5000 ssa_op_iter iter;
5001 tree use;
5003 /* The operands of VDEF expressions are also needed as they
5004 represent potential definitions that may reach this
5005 statement (VDEF operands allow us to follow def-def
5006 links). */
5008 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
5010 gimple *n = mark_operand_necessary (use);
5011 if (n)
5012 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
5017 unsigned int to_clear = -1U;
5018 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
5020 if (to_clear != -1U)
5022 bitmap_clear_bit (inserted_exprs, to_clear);
5023 to_clear = -1U;
5025 t = SSA_NAME_DEF_STMT (ssa_name (i));
5026 if (!gimple_plf (t, NECESSARY))
5028 gimple_stmt_iterator gsi;
5030 if (dump_file && (dump_flags & TDF_DETAILS))
5032 fprintf (dump_file, "Removing unnecessary insertion:");
5033 print_gimple_stmt (dump_file, t, 0);
5036 gsi = gsi_for_stmt (t);
5037 if (gimple_code (t) == GIMPLE_PHI)
5038 remove_phi_node (&gsi, true);
5039 else
5041 gsi_remove (&gsi, true);
5042 release_defs (t);
5045 else
5046 /* eliminate_fini will skip stmts marked for removal if we
5047 already removed it and uses inserted_exprs for this, so
5048 clear those we didn't end up removing. */
5049 to_clear = i;
5051 if (to_clear != -1U)
5052 bitmap_clear_bit (inserted_exprs, to_clear);
5056 /* Initialize data structures used by PRE. */
5058 static void
5059 init_pre (void)
5061 basic_block bb;
5063 next_expression_id = 1;
5064 expressions.create (0);
5065 expressions.safe_push (NULL);
5066 value_expressions.create (get_max_value_id () + 1);
5067 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
5068 name_to_id.create (0);
5070 inserted_exprs = BITMAP_ALLOC (NULL);
5072 connect_infinite_loops_to_exit ();
5073 memset (&pre_stats, 0, sizeof (pre_stats));
5075 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
5077 calculate_dominance_info (CDI_DOMINATORS);
5079 bitmap_obstack_initialize (&grand_bitmap_obstack);
5080 phi_translate_table = new hash_table<expr_pred_trans_d> (5110);
5081 expression_to_id = new hash_table<pre_expr_d> (num_ssa_names * 3);
5082 FOR_ALL_BB_FN (bb, cfun)
5084 EXP_GEN (bb) = bitmap_set_new ();
5085 PHI_GEN (bb) = bitmap_set_new ();
5086 TMP_GEN (bb) = bitmap_set_new ();
5087 AVAIL_OUT (bb) = bitmap_set_new ();
5092 /* Deallocate data structures used by PRE. */
5094 static void
5095 fini_pre ()
5097 value_expressions.release ();
5098 BITMAP_FREE (inserted_exprs);
5099 bitmap_obstack_release (&grand_bitmap_obstack);
5100 bitmap_set_pool.release ();
5101 pre_expr_pool.release ();
5102 delete phi_translate_table;
5103 phi_translate_table = NULL;
5104 delete expression_to_id;
5105 expression_to_id = NULL;
5106 name_to_id.release ();
5108 free_aux_for_blocks ();
5111 namespace {
5113 const pass_data pass_data_pre =
5115 GIMPLE_PASS, /* type */
5116 "pre", /* name */
5117 OPTGROUP_NONE, /* optinfo_flags */
5118 TV_TREE_PRE, /* tv_id */
5119 /* PROP_no_crit_edges is ensured by placing pass_split_crit_edges before
5120 pass_pre. */
5121 ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */
5122 0, /* properties_provided */
5123 PROP_no_crit_edges, /* properties_destroyed */
5124 TODO_rebuild_alias, /* todo_flags_start */
5125 0, /* todo_flags_finish */
5128 class pass_pre : public gimple_opt_pass
5130 public:
5131 pass_pre (gcc::context *ctxt)
5132 : gimple_opt_pass (pass_data_pre, ctxt)
5135 /* opt_pass methods: */
5136 virtual bool gate (function *)
5137 { return flag_tree_pre != 0 || flag_code_hoisting != 0; }
5138 virtual unsigned int execute (function *);
5140 }; // class pass_pre
5142 unsigned int
5143 pass_pre::execute (function *fun)
5145 unsigned int todo = 0;
5147 do_partial_partial =
5148 flag_tree_partial_pre && optimize_function_for_speed_p (fun);
5150 /* This has to happen before SCCVN runs because
5151 loop_optimizer_init may create new phis, etc. */
5152 loop_optimizer_init (LOOPS_NORMAL);
5154 run_scc_vn (VN_WALK);
5156 init_pre ();
5157 scev_initialize ();
5159 /* Collect and value number expressions computed in each basic block. */
5160 compute_avail ();
5162 /* Insert can get quite slow on an incredibly large number of basic
5163 blocks due to some quadratic behavior. Until this behavior is
5164 fixed, don't run it when he have an incredibly large number of
5165 bb's. If we aren't going to run insert, there is no point in
5166 computing ANTIC, either, even though it's plenty fast. */
5167 if (n_basic_blocks_for_fn (fun) < 4000)
5169 compute_antic ();
5170 insert ();
5173 /* Make sure to remove fake edges before committing our inserts.
5174 This makes sure we don't end up with extra critical edges that
5175 we would need to split. */
5176 remove_fake_exit_edges ();
5177 gsi_commit_edge_inserts ();
5179 /* Eliminate folds statements which might (should not...) end up
5180 not keeping virtual operands up-to-date. */
5181 gcc_assert (!need_ssa_update_p (fun));
5183 /* Remove all the redundant expressions. */
5184 todo |= eliminate (true);
5186 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
5187 statistics_counter_event (fun, "PA inserted", pre_stats.pa_insert);
5188 statistics_counter_event (fun, "HOIST inserted", pre_stats.hoist_insert);
5189 statistics_counter_event (fun, "New PHIs", pre_stats.phis);
5190 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
5192 clear_expression_ids ();
5193 remove_dead_inserted_code ();
5195 scev_finalize ();
5196 todo |= fini_eliminate ();
5197 fini_pre ();
5198 loop_optimizer_finalize ();
5200 /* Restore SSA info before tail-merging as that resets it as well. */
5201 scc_vn_restore_ssa_info ();
5203 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
5204 case we can merge the block with the remaining predecessor of the block.
5205 It should either:
5206 - call merge_blocks after each tail merge iteration
5207 - call merge_blocks after all tail merge iterations
5208 - mark TODO_cleanup_cfg when necessary
5209 - share the cfg cleanup with fini_pre. */
5210 todo |= tail_merge_optimize (todo);
5212 free_scc_vn ();
5214 /* Tail merging invalidates the virtual SSA web, together with
5215 cfg-cleanup opportunities exposed by PRE this will wreck the
5216 SSA updating machinery. So make sure to run update-ssa
5217 manually, before eventually scheduling cfg-cleanup as part of
5218 the todo. */
5219 update_ssa (TODO_update_ssa_only_virtuals);
5221 return todo;
5224 } // anon namespace
5226 gimple_opt_pass *
5227 make_pass_pre (gcc::context *ctxt)
5229 return new pass_pre (ctxt);
5232 namespace {
5234 const pass_data pass_data_fre =
5236 GIMPLE_PASS, /* type */
5237 "fre", /* name */
5238 OPTGROUP_NONE, /* optinfo_flags */
5239 TV_TREE_FRE, /* tv_id */
5240 ( PROP_cfg | PROP_ssa ), /* properties_required */
5241 0, /* properties_provided */
5242 0, /* properties_destroyed */
5243 0, /* todo_flags_start */
5244 0, /* todo_flags_finish */
5247 class pass_fre : public gimple_opt_pass
5249 public:
5250 pass_fre (gcc::context *ctxt)
5251 : gimple_opt_pass (pass_data_fre, ctxt)
5254 /* opt_pass methods: */
5255 opt_pass * clone () { return new pass_fre (m_ctxt); }
5256 virtual bool gate (function *) { return flag_tree_fre != 0; }
5257 virtual unsigned int execute (function *);
5259 }; // class pass_fre
5261 unsigned int
5262 pass_fre::execute (function *fun)
5264 unsigned int todo = 0;
5266 run_scc_vn (VN_WALKREWRITE);
5268 memset (&pre_stats, 0, sizeof (pre_stats));
5270 /* Remove all the redundant expressions. */
5271 todo |= eliminate (false);
5273 todo |= fini_eliminate ();
5275 scc_vn_restore_ssa_info ();
5276 free_scc_vn ();
5278 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
5279 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
5281 return todo;
5284 } // anon namespace
5286 gimple_opt_pass *
5287 make_pass_fre (gcc::context *ctxt)
5289 return new pass_fre (ctxt);