2016-09-25 François Dumont <fdumont@gcc.gnu.org>
[official-gcc.git] / gcc / tree-ssa-pre.c
blob0c6f82093c41f6c6b75c062681b051862b4ff06c
1 /* Full and partial redundancy elimination and code hoisting on SSA GIMPLE.
2 Copyright (C) 2001-2016 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 void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
544 unsigned int, bool);
545 static bitmap_set_t bitmap_set_new (void);
546 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
547 tree);
548 static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
549 static unsigned int get_expr_value_id (pre_expr);
551 /* We can add and remove elements and entries to and from sets
552 and hash tables, so we use alloc pools for them. */
554 static object_allocator<bitmap_set> bitmap_set_pool ("Bitmap sets");
555 static bitmap_obstack grand_bitmap_obstack;
557 /* Set of blocks with statements that have had their EH properties changed. */
558 static bitmap need_eh_cleanup;
560 /* Set of blocks with statements that have had their AB properties changed. */
561 static bitmap need_ab_cleanup;
563 /* A three tuple {e, pred, v} used to cache phi translations in the
564 phi_translate_table. */
566 typedef struct expr_pred_trans_d : free_ptr_hash<expr_pred_trans_d>
568 /* The expression. */
569 pre_expr e;
571 /* The predecessor block along which we translated the expression. */
572 basic_block pred;
574 /* The value that resulted from the translation. */
575 pre_expr v;
577 /* The hashcode for the expression, pred pair. This is cached for
578 speed reasons. */
579 hashval_t hashcode;
581 /* hash_table support. */
582 static inline hashval_t hash (const expr_pred_trans_d *);
583 static inline int equal (const expr_pred_trans_d *, const expr_pred_trans_d *);
584 } *expr_pred_trans_t;
585 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
587 inline hashval_t
588 expr_pred_trans_d::hash (const expr_pred_trans_d *e)
590 return e->hashcode;
593 inline int
594 expr_pred_trans_d::equal (const expr_pred_trans_d *ve1,
595 const expr_pred_trans_d *ve2)
597 basic_block b1 = ve1->pred;
598 basic_block b2 = ve2->pred;
600 /* If they are not translations for the same basic block, they can't
601 be equal. */
602 if (b1 != b2)
603 return false;
604 return pre_expr_d::equal (ve1->e, ve2->e);
607 /* The phi_translate_table caches phi translations for a given
608 expression and predecessor. */
609 static hash_table<expr_pred_trans_d> *phi_translate_table;
611 /* Add the tuple mapping from {expression E, basic block PRED} to
612 the phi translation table and return whether it pre-existed. */
614 static inline bool
615 phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred)
617 expr_pred_trans_t *slot;
618 expr_pred_trans_d tem;
619 hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e),
620 pred->index);
621 tem.e = e;
622 tem.pred = pred;
623 tem.hashcode = hash;
624 slot = phi_translate_table->find_slot_with_hash (&tem, hash, INSERT);
625 if (*slot)
627 *entry = *slot;
628 return true;
631 *entry = *slot = XNEW (struct expr_pred_trans_d);
632 (*entry)->e = e;
633 (*entry)->pred = pred;
634 (*entry)->hashcode = hash;
635 return false;
639 /* Add expression E to the expression set of value id V. */
641 static void
642 add_to_value (unsigned int v, pre_expr e)
644 bitmap set;
646 gcc_checking_assert (get_expr_value_id (e) == v);
648 if (v >= value_expressions.length ())
650 value_expressions.safe_grow_cleared (v + 1);
653 set = value_expressions[v];
654 if (!set)
656 set = BITMAP_ALLOC (&grand_bitmap_obstack);
657 value_expressions[v] = set;
660 bitmap_set_bit (set, get_or_alloc_expression_id (e));
663 /* Create a new bitmap set and return it. */
665 static bitmap_set_t
666 bitmap_set_new (void)
668 bitmap_set_t ret = bitmap_set_pool.allocate ();
669 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
670 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
671 return ret;
674 /* Return the value id for a PRE expression EXPR. */
676 static unsigned int
677 get_expr_value_id (pre_expr expr)
679 unsigned int id;
680 switch (expr->kind)
682 case CONSTANT:
683 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
684 break;
685 case NAME:
686 id = VN_INFO (PRE_EXPR_NAME (expr))->value_id;
687 break;
688 case NARY:
689 id = PRE_EXPR_NARY (expr)->value_id;
690 break;
691 case REFERENCE:
692 id = PRE_EXPR_REFERENCE (expr)->value_id;
693 break;
694 default:
695 gcc_unreachable ();
697 /* ??? We cannot assert that expr has a value-id (it can be 0), because
698 we assign value-ids only to expressions that have a result
699 in set_hashtable_value_ids. */
700 return id;
703 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
705 static tree
706 sccvn_valnum_from_value_id (unsigned int val)
708 bitmap_iterator bi;
709 unsigned int i;
710 bitmap exprset = value_expressions[val];
711 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
713 pre_expr vexpr = expression_for_id (i);
714 if (vexpr->kind == NAME)
715 return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
716 else if (vexpr->kind == CONSTANT)
717 return PRE_EXPR_CONSTANT (vexpr);
719 return NULL_TREE;
722 /* Remove an expression EXPR from a bitmapped set. */
724 static void
725 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
727 unsigned int val = get_expr_value_id (expr);
728 if (!value_id_constant_p (val))
730 bitmap_clear_bit (&set->values, val);
731 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
735 static void
736 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
737 unsigned int val, bool allow_constants)
739 if (allow_constants || !value_id_constant_p (val))
741 /* We specifically expect this and only this function to be able to
742 insert constants into a set. */
743 bitmap_set_bit (&set->values, val);
744 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
748 /* Insert an expression EXPR into a bitmapped set. */
750 static void
751 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
753 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
756 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
758 static void
759 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
761 bitmap_copy (&dest->expressions, &orig->expressions);
762 bitmap_copy (&dest->values, &orig->values);
766 /* Free memory used up by SET. */
767 static void
768 bitmap_set_free (bitmap_set_t set)
770 bitmap_clear (&set->expressions);
771 bitmap_clear (&set->values);
775 /* Generate an topological-ordered array of bitmap set SET. */
777 static vec<pre_expr>
778 sorted_array_from_bitmap_set (bitmap_set_t set)
780 unsigned int i, j;
781 bitmap_iterator bi, bj;
782 vec<pre_expr> result;
784 /* Pre-allocate enough space for the array. */
785 result.create (bitmap_count_bits (&set->expressions));
787 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
789 /* The number of expressions having a given value is usually
790 relatively small. Thus, rather than making a vector of all
791 the expressions and sorting it by value-id, we walk the values
792 and check in the reverse mapping that tells us what expressions
793 have a given value, to filter those in our set. As a result,
794 the expressions are inserted in value-id order, which means
795 topological order.
797 If this is somehow a significant lose for some cases, we can
798 choose which set to walk based on the set size. */
799 bitmap exprset = value_expressions[i];
800 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj)
802 if (bitmap_bit_p (&set->expressions, j))
803 result.quick_push (expression_for_id (j));
807 return result;
810 /* Perform bitmapped set operation DEST &= ORIG. */
812 static void
813 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
815 bitmap_iterator bi;
816 unsigned int i;
818 if (dest != orig)
820 bitmap_head temp;
821 bitmap_initialize (&temp, &grand_bitmap_obstack);
823 bitmap_and_into (&dest->values, &orig->values);
824 bitmap_copy (&temp, &dest->expressions);
825 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
827 pre_expr expr = expression_for_id (i);
828 unsigned int value_id = get_expr_value_id (expr);
829 if (!bitmap_bit_p (&dest->values, value_id))
830 bitmap_clear_bit (&dest->expressions, i);
832 bitmap_clear (&temp);
836 /* Subtract all values and expressions contained in ORIG from DEST. */
838 static bitmap_set_t
839 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
841 bitmap_set_t result = bitmap_set_new ();
842 bitmap_iterator bi;
843 unsigned int i;
845 bitmap_and_compl (&result->expressions, &dest->expressions,
846 &orig->expressions);
848 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
850 pre_expr expr = expression_for_id (i);
851 unsigned int value_id = get_expr_value_id (expr);
852 bitmap_set_bit (&result->values, value_id);
855 return result;
858 /* Subtract all the values in bitmap set B from bitmap set A. */
860 static void
861 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
863 unsigned int i;
864 bitmap_iterator bi;
865 bitmap_head temp;
867 bitmap_initialize (&temp, &grand_bitmap_obstack);
869 bitmap_copy (&temp, &a->expressions);
870 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
872 pre_expr expr = expression_for_id (i);
873 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
874 bitmap_remove_from_set (a, expr);
876 bitmap_clear (&temp);
880 /* Return true if bitmapped set SET contains the value VALUE_ID. */
882 static bool
883 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
885 if (value_id_constant_p (value_id))
886 return true;
888 if (!set || bitmap_empty_p (&set->expressions))
889 return false;
891 return bitmap_bit_p (&set->values, value_id);
894 static inline bool
895 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
897 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
900 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
902 static void
903 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
904 const pre_expr expr)
906 bitmap exprset;
907 unsigned int i;
908 bitmap_iterator bi;
910 if (value_id_constant_p (lookfor))
911 return;
913 if (!bitmap_set_contains_value (set, lookfor))
914 return;
916 /* The number of expressions having a given value is usually
917 significantly less than the total number of expressions in SET.
918 Thus, rather than check, for each expression in SET, whether it
919 has the value LOOKFOR, we walk the reverse mapping that tells us
920 what expressions have a given value, and see if any of those
921 expressions are in our set. For large testcases, this is about
922 5-10x faster than walking the bitmap. If this is somehow a
923 significant lose for some cases, we can choose which set to walk
924 based on the set size. */
925 exprset = value_expressions[lookfor];
926 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
928 if (bitmap_clear_bit (&set->expressions, i))
930 bitmap_set_bit (&set->expressions, get_expression_id (expr));
931 return;
935 gcc_unreachable ();
938 /* Return true if two bitmap sets are equal. */
940 static bool
941 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
943 return bitmap_equal_p (&a->values, &b->values);
946 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
947 and add it otherwise. */
949 static void
950 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
952 unsigned int val = get_expr_value_id (expr);
954 if (bitmap_set_contains_value (set, val))
955 bitmap_set_replace_value (set, val, expr);
956 else
957 bitmap_insert_into_set (set, expr);
960 /* Insert EXPR into SET if EXPR's value is not already present in
961 SET. */
963 static void
964 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
966 unsigned int val = get_expr_value_id (expr);
968 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
970 /* Constant values are always considered to be part of the set. */
971 if (value_id_constant_p (val))
972 return;
974 /* If the value membership changed, add the expression. */
975 if (bitmap_set_bit (&set->values, val))
976 bitmap_set_bit (&set->expressions, expr->id);
979 /* Print out EXPR to outfile. */
981 static void
982 print_pre_expr (FILE *outfile, const pre_expr expr)
984 switch (expr->kind)
986 case CONSTANT:
987 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
988 break;
989 case NAME:
990 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
991 break;
992 case NARY:
994 unsigned int i;
995 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
996 fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode));
997 for (i = 0; i < nary->length; i++)
999 print_generic_expr (outfile, nary->op[i], 0);
1000 if (i != (unsigned) nary->length - 1)
1001 fprintf (outfile, ",");
1003 fprintf (outfile, "}");
1005 break;
1007 case REFERENCE:
1009 vn_reference_op_t vro;
1010 unsigned int i;
1011 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1012 fprintf (outfile, "{");
1013 for (i = 0;
1014 ref->operands.iterate (i, &vro);
1015 i++)
1017 bool closebrace = false;
1018 if (vro->opcode != SSA_NAME
1019 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
1021 fprintf (outfile, "%s", get_tree_code_name (vro->opcode));
1022 if (vro->op0)
1024 fprintf (outfile, "<");
1025 closebrace = true;
1028 if (vro->op0)
1030 print_generic_expr (outfile, vro->op0, 0);
1031 if (vro->op1)
1033 fprintf (outfile, ",");
1034 print_generic_expr (outfile, vro->op1, 0);
1036 if (vro->op2)
1038 fprintf (outfile, ",");
1039 print_generic_expr (outfile, vro->op2, 0);
1042 if (closebrace)
1043 fprintf (outfile, ">");
1044 if (i != ref->operands.length () - 1)
1045 fprintf (outfile, ",");
1047 fprintf (outfile, "}");
1048 if (ref->vuse)
1050 fprintf (outfile, "@");
1051 print_generic_expr (outfile, ref->vuse, 0);
1054 break;
1057 void debug_pre_expr (pre_expr);
1059 /* Like print_pre_expr but always prints to stderr. */
1060 DEBUG_FUNCTION void
1061 debug_pre_expr (pre_expr e)
1063 print_pre_expr (stderr, e);
1064 fprintf (stderr, "\n");
1067 /* Print out SET to OUTFILE. */
1069 static void
1070 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1071 const char *setname, int blockindex)
1073 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1074 if (set)
1076 bool first = true;
1077 unsigned i;
1078 bitmap_iterator bi;
1080 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1082 const pre_expr expr = expression_for_id (i);
1084 if (!first)
1085 fprintf (outfile, ", ");
1086 first = false;
1087 print_pre_expr (outfile, expr);
1089 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1092 fprintf (outfile, " }\n");
1095 void debug_bitmap_set (bitmap_set_t);
1097 DEBUG_FUNCTION void
1098 debug_bitmap_set (bitmap_set_t set)
1100 print_bitmap_set (stderr, set, "debug", 0);
1103 void debug_bitmap_sets_for (basic_block);
1105 DEBUG_FUNCTION void
1106 debug_bitmap_sets_for (basic_block bb)
1108 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1109 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1110 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1111 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1112 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1113 if (do_partial_partial)
1114 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1115 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1118 /* Print out the expressions that have VAL to OUTFILE. */
1120 static void
1121 print_value_expressions (FILE *outfile, unsigned int val)
1123 bitmap set = value_expressions[val];
1124 if (set)
1126 bitmap_set x;
1127 char s[10];
1128 sprintf (s, "%04d", val);
1129 x.expressions = *set;
1130 print_bitmap_set (outfile, &x, s, 0);
1135 DEBUG_FUNCTION void
1136 debug_value_expressions (unsigned int val)
1138 print_value_expressions (stderr, val);
1141 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1142 represent it. */
1144 static pre_expr
1145 get_or_alloc_expr_for_constant (tree constant)
1147 unsigned int result_id;
1148 unsigned int value_id;
1149 struct pre_expr_d expr;
1150 pre_expr newexpr;
1152 expr.kind = CONSTANT;
1153 PRE_EXPR_CONSTANT (&expr) = constant;
1154 result_id = lookup_expression_id (&expr);
1155 if (result_id != 0)
1156 return expression_for_id (result_id);
1158 newexpr = pre_expr_pool.allocate ();
1159 newexpr->kind = CONSTANT;
1160 PRE_EXPR_CONSTANT (newexpr) = constant;
1161 alloc_expression_id (newexpr);
1162 value_id = get_or_alloc_constant_value_id (constant);
1163 add_to_value (value_id, newexpr);
1164 return newexpr;
1167 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1168 Currently only supports constants and SSA_NAMES. */
1169 static pre_expr
1170 get_or_alloc_expr_for (tree t)
1172 if (TREE_CODE (t) == SSA_NAME)
1173 return get_or_alloc_expr_for_name (t);
1174 else if (is_gimple_min_invariant (t))
1175 return get_or_alloc_expr_for_constant (t);
1176 else
1178 /* More complex expressions can result from SCCVN expression
1179 simplification that inserts values for them. As they all
1180 do not have VOPs the get handled by the nary ops struct. */
1181 vn_nary_op_t result;
1182 unsigned int result_id;
1183 vn_nary_op_lookup (t, &result);
1184 if (result != NULL)
1186 pre_expr e = pre_expr_pool.allocate ();
1187 e->kind = NARY;
1188 PRE_EXPR_NARY (e) = result;
1189 result_id = lookup_expression_id (e);
1190 if (result_id != 0)
1192 pre_expr_pool.remove (e);
1193 e = expression_for_id (result_id);
1194 return e;
1196 alloc_expression_id (e);
1197 return e;
1200 return NULL;
1203 /* Return the folded version of T if T, when folded, is a gimple
1204 min_invariant or an SSA name. Otherwise, return T. */
1206 static pre_expr
1207 fully_constant_expression (pre_expr e)
1209 switch (e->kind)
1211 case CONSTANT:
1212 return e;
1213 case NARY:
1215 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1216 tree res = vn_nary_simplify (nary);
1217 if (!res)
1218 return e;
1219 if (is_gimple_min_invariant (res))
1220 return get_or_alloc_expr_for_constant (res);
1221 if (TREE_CODE (res) == SSA_NAME)
1222 return get_or_alloc_expr_for_name (res);
1223 return e;
1225 case REFERENCE:
1227 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1228 tree folded;
1229 if ((folded = fully_constant_vn_reference_p (ref)))
1230 return get_or_alloc_expr_for_constant (folded);
1231 return e;
1233 default:
1234 return e;
1236 return e;
1239 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1240 it has the value it would have in BLOCK. Set *SAME_VALID to true
1241 in case the new vuse doesn't change the value id of the OPERANDS. */
1243 static tree
1244 translate_vuse_through_block (vec<vn_reference_op_s> operands,
1245 alias_set_type set, tree type, tree vuse,
1246 basic_block phiblock,
1247 basic_block block, bool *same_valid)
1249 gimple *phi = SSA_NAME_DEF_STMT (vuse);
1250 ao_ref ref;
1251 edge e = NULL;
1252 bool use_oracle;
1254 *same_valid = true;
1256 if (gimple_bb (phi) != phiblock)
1257 return vuse;
1259 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1261 /* Use the alias-oracle to find either the PHI node in this block,
1262 the first VUSE used in this block that is equivalent to vuse or
1263 the first VUSE which definition in this block kills the value. */
1264 if (gimple_code (phi) == GIMPLE_PHI)
1265 e = find_edge (block, phiblock);
1266 else if (use_oracle)
1267 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1269 vuse = gimple_vuse (phi);
1270 phi = SSA_NAME_DEF_STMT (vuse);
1271 if (gimple_bb (phi) != phiblock)
1272 return vuse;
1273 if (gimple_code (phi) == GIMPLE_PHI)
1275 e = find_edge (block, phiblock);
1276 break;
1279 else
1280 return NULL_TREE;
1282 if (e)
1284 if (use_oracle)
1286 bitmap visited = NULL;
1287 unsigned int cnt;
1288 /* Try to find a vuse that dominates this phi node by skipping
1289 non-clobbering statements. */
1290 vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false,
1291 NULL, NULL);
1292 if (visited)
1293 BITMAP_FREE (visited);
1295 else
1296 vuse = NULL_TREE;
1297 if (!vuse)
1299 /* If we didn't find any, the value ID can't stay the same,
1300 but return the translated vuse. */
1301 *same_valid = false;
1302 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1304 /* ??? We would like to return vuse here as this is the canonical
1305 upmost vdef that this reference is associated with. But during
1306 insertion of the references into the hash tables we only ever
1307 directly insert with their direct gimple_vuse, hence returning
1308 something else would make us not find the other expression. */
1309 return PHI_ARG_DEF (phi, e->dest_idx);
1312 return NULL_TREE;
1315 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1316 SET2. This is used to avoid making a set consisting of the union
1317 of PA_IN and ANTIC_IN during insert. */
1319 static inline pre_expr
1320 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1322 pre_expr result;
1324 result = bitmap_find_leader (set1, val);
1325 if (!result && set2)
1326 result = bitmap_find_leader (set2, val);
1327 return result;
1330 /* Get the tree type for our PRE expression e. */
1332 static tree
1333 get_expr_type (const pre_expr e)
1335 switch (e->kind)
1337 case NAME:
1338 return TREE_TYPE (PRE_EXPR_NAME (e));
1339 case CONSTANT:
1340 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1341 case REFERENCE:
1342 return PRE_EXPR_REFERENCE (e)->type;
1343 case NARY:
1344 return PRE_EXPR_NARY (e)->type;
1346 gcc_unreachable ();
1349 /* Get a representative SSA_NAME for a given expression.
1350 Since all of our sub-expressions are treated as values, we require
1351 them to be SSA_NAME's for simplicity.
1352 Prior versions of GVNPRE used to use "value handles" here, so that
1353 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1354 either case, the operands are really values (IE we do not expect
1355 them to be usable without finding leaders). */
1357 static tree
1358 get_representative_for (const pre_expr e)
1360 tree name;
1361 unsigned int value_id = get_expr_value_id (e);
1363 switch (e->kind)
1365 case NAME:
1366 return VN_INFO (PRE_EXPR_NAME (e))->valnum;
1367 case CONSTANT:
1368 return PRE_EXPR_CONSTANT (e);
1369 case NARY:
1370 case REFERENCE:
1372 /* Go through all of the expressions representing this value
1373 and pick out an SSA_NAME. */
1374 unsigned int i;
1375 bitmap_iterator bi;
1376 bitmap exprs = value_expressions[value_id];
1377 EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1379 pre_expr rep = expression_for_id (i);
1380 if (rep->kind == NAME)
1381 return VN_INFO (PRE_EXPR_NAME (rep))->valnum;
1382 else if (rep->kind == CONSTANT)
1383 return PRE_EXPR_CONSTANT (rep);
1386 break;
1389 /* If we reached here we couldn't find an SSA_NAME. This can
1390 happen when we've discovered a value that has never appeared in
1391 the program as set to an SSA_NAME, as the result of phi translation.
1392 Create one here.
1393 ??? We should be able to re-use this when we insert the statement
1394 to compute it. */
1395 name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1396 VN_INFO_GET (name)->value_id = value_id;
1397 VN_INFO (name)->valnum = name;
1398 /* ??? For now mark this SSA name for release by SCCVN. */
1399 VN_INFO (name)->needs_insertion = true;
1400 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1401 if (dump_file && (dump_flags & TDF_DETAILS))
1403 fprintf (dump_file, "Created SSA_NAME representative ");
1404 print_generic_expr (dump_file, name, 0);
1405 fprintf (dump_file, " for expression:");
1406 print_pre_expr (dump_file, e);
1407 fprintf (dump_file, " (%04d)\n", value_id);
1410 return name;
1415 static pre_expr
1416 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1417 basic_block pred, basic_block phiblock);
1419 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1420 the phis in PRED. Return NULL if we can't find a leader for each part
1421 of the translated expression. */
1423 static pre_expr
1424 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1425 basic_block pred, basic_block phiblock)
1427 switch (expr->kind)
1429 case NARY:
1431 unsigned int i;
1432 bool changed = false;
1433 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1434 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1435 sizeof_vn_nary_op (nary->length));
1436 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1438 for (i = 0; i < newnary->length; i++)
1440 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1441 continue;
1442 else
1444 pre_expr leader, result;
1445 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1446 leader = find_leader_in_sets (op_val_id, set1, set2);
1447 result = phi_translate (leader, set1, set2, pred, phiblock);
1448 if (result && result != leader)
1449 newnary->op[i] = get_representative_for (result);
1450 else if (!result)
1451 return NULL;
1453 changed |= newnary->op[i] != nary->op[i];
1456 if (changed)
1458 pre_expr constant;
1459 unsigned int new_val_id;
1461 PRE_EXPR_NARY (expr) = newnary;
1462 constant = fully_constant_expression (expr);
1463 PRE_EXPR_NARY (expr) = nary;
1464 if (constant != expr)
1466 /* For non-CONSTANTs we have to make sure we can eventually
1467 insert the expression. Which means we need to have a
1468 leader for it. */
1469 if (constant->kind != CONSTANT)
1471 unsigned value_id = get_expr_value_id (constant);
1472 constant = find_leader_in_sets (value_id, set1, set2);
1473 if (constant)
1474 return constant;
1476 else
1477 return constant;
1480 tree result = vn_nary_op_lookup_pieces (newnary->length,
1481 newnary->opcode,
1482 newnary->type,
1483 &newnary->op[0],
1484 &nary);
1485 if (result && is_gimple_min_invariant (result))
1486 return get_or_alloc_expr_for_constant (result);
1488 expr = pre_expr_pool.allocate ();
1489 expr->kind = NARY;
1490 expr->id = 0;
1491 if (nary)
1493 PRE_EXPR_NARY (expr) = nary;
1494 new_val_id = nary->value_id;
1495 get_or_alloc_expression_id (expr);
1497 else
1499 new_val_id = get_next_value_id ();
1500 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
1501 nary = vn_nary_op_insert_pieces (newnary->length,
1502 newnary->opcode,
1503 newnary->type,
1504 &newnary->op[0],
1505 result, new_val_id);
1506 PRE_EXPR_NARY (expr) = nary;
1507 get_or_alloc_expression_id (expr);
1509 add_to_value (new_val_id, expr);
1511 return expr;
1513 break;
1515 case REFERENCE:
1517 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1518 vec<vn_reference_op_s> operands = ref->operands;
1519 tree vuse = ref->vuse;
1520 tree newvuse = vuse;
1521 vec<vn_reference_op_s> newoperands = vNULL;
1522 bool changed = false, same_valid = true;
1523 unsigned int i, n;
1524 vn_reference_op_t operand;
1525 vn_reference_t newref;
1527 for (i = 0; operands.iterate (i, &operand); i++)
1529 pre_expr opresult;
1530 pre_expr leader;
1531 tree op[3];
1532 tree type = operand->type;
1533 vn_reference_op_s newop = *operand;
1534 op[0] = operand->op0;
1535 op[1] = operand->op1;
1536 op[2] = operand->op2;
1537 for (n = 0; n < 3; ++n)
1539 unsigned int op_val_id;
1540 if (!op[n])
1541 continue;
1542 if (TREE_CODE (op[n]) != SSA_NAME)
1544 /* We can't possibly insert these. */
1545 if (n != 0
1546 && !is_gimple_min_invariant (op[n]))
1547 break;
1548 continue;
1550 op_val_id = VN_INFO (op[n])->value_id;
1551 leader = find_leader_in_sets (op_val_id, set1, set2);
1552 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1553 if (opresult && opresult != leader)
1555 tree name = get_representative_for (opresult);
1556 changed |= name != op[n];
1557 op[n] = name;
1559 else if (!opresult)
1560 break;
1562 if (n != 3)
1564 newoperands.release ();
1565 return NULL;
1567 if (!changed)
1568 continue;
1569 if (!newoperands.exists ())
1570 newoperands = operands.copy ();
1571 /* We may have changed from an SSA_NAME to a constant */
1572 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1573 newop.opcode = TREE_CODE (op[0]);
1574 newop.type = type;
1575 newop.op0 = op[0];
1576 newop.op1 = op[1];
1577 newop.op2 = op[2];
1578 newoperands[i] = newop;
1580 gcc_checking_assert (i == operands.length ());
1582 if (vuse)
1584 newvuse = translate_vuse_through_block (newoperands.exists ()
1585 ? newoperands : operands,
1586 ref->set, ref->type,
1587 vuse, phiblock, pred,
1588 &same_valid);
1589 if (newvuse == NULL_TREE)
1591 newoperands.release ();
1592 return NULL;
1596 if (changed || newvuse != vuse)
1598 unsigned int new_val_id;
1599 pre_expr constant;
1601 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1602 ref->type,
1603 newoperands.exists ()
1604 ? newoperands : operands,
1605 &newref, VN_WALK);
1606 if (result)
1607 newoperands.release ();
1609 /* We can always insert constants, so if we have a partial
1610 redundant constant load of another type try to translate it
1611 to a constant of appropriate type. */
1612 if (result && is_gimple_min_invariant (result))
1614 tree tem = result;
1615 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1617 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1618 if (tem && !is_gimple_min_invariant (tem))
1619 tem = NULL_TREE;
1621 if (tem)
1622 return get_or_alloc_expr_for_constant (tem);
1625 /* If we'd have to convert things we would need to validate
1626 if we can insert the translated expression. So fail
1627 here for now - we cannot insert an alias with a different
1628 type in the VN tables either, as that would assert. */
1629 if (result
1630 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1631 return NULL;
1632 else if (!result && newref
1633 && !useless_type_conversion_p (ref->type, newref->type))
1635 newoperands.release ();
1636 return NULL;
1639 expr = pre_expr_pool.allocate ();
1640 expr->kind = REFERENCE;
1641 expr->id = 0;
1643 if (newref)
1645 PRE_EXPR_REFERENCE (expr) = newref;
1646 constant = fully_constant_expression (expr);
1647 if (constant != expr)
1648 return constant;
1650 new_val_id = newref->value_id;
1651 get_or_alloc_expression_id (expr);
1653 else
1655 if (changed || !same_valid)
1657 new_val_id = get_next_value_id ();
1658 value_expressions.safe_grow_cleared
1659 (get_max_value_id () + 1);
1661 else
1662 new_val_id = ref->value_id;
1663 if (!newoperands.exists ())
1664 newoperands = operands.copy ();
1665 newref = vn_reference_insert_pieces (newvuse, ref->set,
1666 ref->type,
1667 newoperands,
1668 result, new_val_id);
1669 newoperands = vNULL;
1670 PRE_EXPR_REFERENCE (expr) = newref;
1671 constant = fully_constant_expression (expr);
1672 if (constant != expr)
1673 return constant;
1674 get_or_alloc_expression_id (expr);
1676 add_to_value (new_val_id, expr);
1678 newoperands.release ();
1679 return expr;
1681 break;
1683 case NAME:
1685 tree name = PRE_EXPR_NAME (expr);
1686 gimple *def_stmt = SSA_NAME_DEF_STMT (name);
1687 /* If the SSA name is defined by a PHI node in this block,
1688 translate it. */
1689 if (gimple_code (def_stmt) == GIMPLE_PHI
1690 && gimple_bb (def_stmt) == phiblock)
1692 edge e = find_edge (pred, gimple_bb (def_stmt));
1693 tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1695 /* Handle constant. */
1696 if (is_gimple_min_invariant (def))
1697 return get_or_alloc_expr_for_constant (def);
1699 return get_or_alloc_expr_for_name (def);
1701 /* Otherwise return it unchanged - it will get removed if its
1702 value is not available in PREDs AVAIL_OUT set of expressions
1703 by the subtraction of TMP_GEN. */
1704 return expr;
1707 default:
1708 gcc_unreachable ();
1712 /* Wrapper around phi_translate_1 providing caching functionality. */
1714 static pre_expr
1715 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1716 basic_block pred, basic_block phiblock)
1718 expr_pred_trans_t slot = NULL;
1719 pre_expr phitrans;
1721 if (!expr)
1722 return NULL;
1724 /* Constants contain no values that need translation. */
1725 if (expr->kind == CONSTANT)
1726 return expr;
1728 if (value_id_constant_p (get_expr_value_id (expr)))
1729 return expr;
1731 /* Don't add translations of NAMEs as those are cheap to translate. */
1732 if (expr->kind != NAME)
1734 if (phi_trans_add (&slot, expr, pred))
1735 return slot->v;
1736 /* Store NULL for the value we want to return in the case of
1737 recursing. */
1738 slot->v = NULL;
1741 /* Translate. */
1742 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1744 if (slot)
1746 if (phitrans)
1747 slot->v = phitrans;
1748 else
1749 /* Remove failed translations again, they cause insert
1750 iteration to not pick up new opportunities reliably. */
1751 phi_translate_table->remove_elt_with_hash (slot, slot->hashcode);
1754 return phitrans;
1758 /* For each expression in SET, translate the values through phi nodes
1759 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1760 expressions in DEST. */
1762 static void
1763 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1764 basic_block phiblock)
1766 vec<pre_expr> exprs;
1767 pre_expr expr;
1768 int i;
1770 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1772 bitmap_set_copy (dest, set);
1773 return;
1776 exprs = sorted_array_from_bitmap_set (set);
1777 FOR_EACH_VEC_ELT (exprs, i, expr)
1779 pre_expr translated;
1780 translated = phi_translate (expr, set, NULL, pred, phiblock);
1781 if (!translated)
1782 continue;
1784 /* We might end up with multiple expressions from SET being
1785 translated to the same value. In this case we do not want
1786 to retain the NARY or REFERENCE expression but prefer a NAME
1787 which would be the leader. */
1788 if (translated->kind == NAME)
1789 bitmap_value_replace_in_set (dest, translated);
1790 else
1791 bitmap_value_insert_into_set (dest, translated);
1793 exprs.release ();
1796 /* Find the leader for a value (i.e., the name representing that
1797 value) in a given set, and return it. Return NULL if no leader
1798 is found. */
1800 static pre_expr
1801 bitmap_find_leader (bitmap_set_t set, unsigned int val)
1803 if (value_id_constant_p (val))
1805 unsigned int i;
1806 bitmap_iterator bi;
1807 bitmap exprset = value_expressions[val];
1809 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1811 pre_expr expr = expression_for_id (i);
1812 if (expr->kind == CONSTANT)
1813 return expr;
1816 if (bitmap_set_contains_value (set, val))
1818 /* Rather than walk the entire bitmap of expressions, and see
1819 whether any of them has the value we are looking for, we look
1820 at the reverse mapping, which tells us the set of expressions
1821 that have a given value (IE value->expressions with that
1822 value) and see if any of those expressions are in our set.
1823 The number of expressions per value is usually significantly
1824 less than the number of expressions in the set. In fact, for
1825 large testcases, doing it this way is roughly 5-10x faster
1826 than walking the bitmap.
1827 If this is somehow a significant lose for some cases, we can
1828 choose which set to walk based on which set is smaller. */
1829 unsigned int i;
1830 bitmap_iterator bi;
1831 bitmap exprset = value_expressions[val];
1833 EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1834 return expression_for_id (i);
1836 return NULL;
1839 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1840 BLOCK by seeing if it is not killed in the block. Note that we are
1841 only determining whether there is a store that kills it. Because
1842 of the order in which clean iterates over values, we are guaranteed
1843 that altered operands will have caused us to be eliminated from the
1844 ANTIC_IN set already. */
1846 static bool
1847 value_dies_in_block_x (pre_expr expr, basic_block block)
1849 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1850 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1851 gimple *def;
1852 gimple_stmt_iterator gsi;
1853 unsigned id = get_expression_id (expr);
1854 bool res = false;
1855 ao_ref ref;
1857 if (!vuse)
1858 return false;
1860 /* Lookup a previously calculated result. */
1861 if (EXPR_DIES (block)
1862 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1863 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1865 /* A memory expression {e, VUSE} dies in the block if there is a
1866 statement that may clobber e. If, starting statement walk from the
1867 top of the basic block, a statement uses VUSE there can be no kill
1868 inbetween that use and the original statement that loaded {e, VUSE},
1869 so we can stop walking. */
1870 ref.base = NULL_TREE;
1871 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1873 tree def_vuse, def_vdef;
1874 def = gsi_stmt (gsi);
1875 def_vuse = gimple_vuse (def);
1876 def_vdef = gimple_vdef (def);
1878 /* Not a memory statement. */
1879 if (!def_vuse)
1880 continue;
1882 /* Not a may-def. */
1883 if (!def_vdef)
1885 /* A load with the same VUSE, we're done. */
1886 if (def_vuse == vuse)
1887 break;
1889 continue;
1892 /* Init ref only if we really need it. */
1893 if (ref.base == NULL_TREE
1894 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1895 refx->operands))
1897 res = true;
1898 break;
1900 /* If the statement may clobber expr, it dies. */
1901 if (stmt_may_clobber_ref_p_1 (def, &ref))
1903 res = true;
1904 break;
1908 /* Remember the result. */
1909 if (!EXPR_DIES (block))
1910 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1911 bitmap_set_bit (EXPR_DIES (block), id * 2);
1912 if (res)
1913 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1915 return res;
1919 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1920 contains its value-id. */
1922 static bool
1923 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1925 if (op && TREE_CODE (op) == SSA_NAME)
1927 unsigned int value_id = VN_INFO (op)->value_id;
1928 if (!(bitmap_set_contains_value (set1, value_id)
1929 || (set2 && bitmap_set_contains_value (set2, value_id))))
1930 return false;
1932 return true;
1935 /* Determine if the expression EXPR is valid in SET1 U SET2.
1936 ONLY SET2 CAN BE NULL.
1937 This means that we have a leader for each part of the expression
1938 (if it consists of values), or the expression is an SSA_NAME.
1939 For loads/calls, we also see if the vuse is killed in this block. */
1941 static bool
1942 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr)
1944 switch (expr->kind)
1946 case NAME:
1947 /* By construction all NAMEs are available. Non-available
1948 NAMEs are removed by subtracting TMP_GEN from the sets. */
1949 return true;
1950 case NARY:
1952 unsigned int i;
1953 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1954 for (i = 0; i < nary->length; i++)
1955 if (!op_valid_in_sets (set1, set2, nary->op[i]))
1956 return false;
1957 return true;
1959 break;
1960 case REFERENCE:
1962 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1963 vn_reference_op_t vro;
1964 unsigned int i;
1966 FOR_EACH_VEC_ELT (ref->operands, i, vro)
1968 if (!op_valid_in_sets (set1, set2, vro->op0)
1969 || !op_valid_in_sets (set1, set2, vro->op1)
1970 || !op_valid_in_sets (set1, set2, vro->op2))
1971 return false;
1973 return true;
1975 default:
1976 gcc_unreachable ();
1980 /* Clean the set of expressions that are no longer valid in SET1 or
1981 SET2. This means expressions that are made up of values we have no
1982 leaders for in SET1 or SET2. This version is used for partial
1983 anticipation, which means it is not valid in either ANTIC_IN or
1984 PA_IN. */
1986 static void
1987 dependent_clean (bitmap_set_t set1, bitmap_set_t set2)
1989 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
1990 pre_expr expr;
1991 int i;
1993 FOR_EACH_VEC_ELT (exprs, i, expr)
1995 if (!valid_in_sets (set1, set2, expr))
1996 bitmap_remove_from_set (set1, expr);
1998 exprs.release ();
2001 /* Clean the set of expressions that are no longer valid in SET. This
2002 means expressions that are made up of values we have no leaders for
2003 in SET. */
2005 static void
2006 clean (bitmap_set_t set)
2008 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set);
2009 pre_expr expr;
2010 int i;
2012 FOR_EACH_VEC_ELT (exprs, i, expr)
2014 if (!valid_in_sets (set, NULL, expr))
2015 bitmap_remove_from_set (set, expr);
2017 exprs.release ();
2020 /* Clean the set of expressions that are no longer valid in SET because
2021 they are clobbered in BLOCK or because they trap and may not be executed. */
2023 static void
2024 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2026 bitmap_iterator bi;
2027 unsigned i;
2029 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2031 pre_expr expr = expression_for_id (i);
2032 if (expr->kind == REFERENCE)
2034 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2035 if (ref->vuse)
2037 gimple *def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2038 if (!gimple_nop_p (def_stmt)
2039 && ((gimple_bb (def_stmt) != block
2040 && !dominated_by_p (CDI_DOMINATORS,
2041 block, gimple_bb (def_stmt)))
2042 || (gimple_bb (def_stmt) == block
2043 && value_dies_in_block_x (expr, block))))
2044 bitmap_remove_from_set (set, expr);
2047 else if (expr->kind == NARY)
2049 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2050 /* If the NARY may trap make sure the block does not contain
2051 a possible exit point.
2052 ??? This is overly conservative if we translate AVAIL_OUT
2053 as the available expression might be after the exit point. */
2054 if (BB_MAY_NOTRETURN (block)
2055 && vn_nary_may_trap (nary))
2056 bitmap_remove_from_set (set, expr);
2061 static sbitmap has_abnormal_preds;
2063 /* Compute the ANTIC set for BLOCK.
2065 If succs(BLOCK) > 1 then
2066 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2067 else if succs(BLOCK) == 1 then
2068 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2070 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2073 static bool
2074 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2076 bool changed = false;
2077 bitmap_set_t S, old, ANTIC_OUT;
2078 bitmap_iterator bi;
2079 unsigned int bii;
2080 edge e;
2081 edge_iterator ei;
2082 bool was_visited = BB_VISITED (block);
2084 old = ANTIC_OUT = S = NULL;
2085 BB_VISITED (block) = 1;
2087 /* If any edges from predecessors are abnormal, antic_in is empty,
2088 so do nothing. */
2089 if (block_has_abnormal_pred_edge)
2090 goto maybe_dump_sets;
2092 old = ANTIC_IN (block);
2093 ANTIC_OUT = bitmap_set_new ();
2095 /* If the block has no successors, ANTIC_OUT is empty. */
2096 if (EDGE_COUNT (block->succs) == 0)
2098 /* If we have one successor, we could have some phi nodes to
2099 translate through. */
2100 else if (single_succ_p (block))
2102 basic_block succ_bb = single_succ (block);
2103 gcc_assert (BB_VISITED (succ_bb));
2104 phi_translate_set (ANTIC_OUT, ANTIC_IN (succ_bb), block, succ_bb);
2106 /* If we have multiple successors, we take the intersection of all of
2107 them. Note that in the case of loop exit phi nodes, we may have
2108 phis to translate through. */
2109 else
2111 size_t i;
2112 basic_block bprime, first = NULL;
2114 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2115 FOR_EACH_EDGE (e, ei, block->succs)
2117 if (!first
2118 && BB_VISITED (e->dest))
2119 first = e->dest;
2120 else if (BB_VISITED (e->dest))
2121 worklist.quick_push (e->dest);
2122 else
2124 /* Unvisited successors get their ANTIC_IN replaced by the
2125 maximal set to arrive at a maximum ANTIC_IN solution.
2126 We can ignore them in the intersection operation and thus
2127 need not explicitely represent that maximum solution. */
2128 if (dump_file && (dump_flags & TDF_DETAILS))
2129 fprintf (dump_file, "ANTIC_IN is MAX on %d->%d\n",
2130 e->src->index, e->dest->index);
2134 /* Of multiple successors we have to have visited one already
2135 which is guaranteed by iteration order. */
2136 gcc_assert (first != NULL);
2138 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2140 FOR_EACH_VEC_ELT (worklist, i, bprime)
2142 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2144 bitmap_set_t tmp = bitmap_set_new ();
2145 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2146 bitmap_set_and (ANTIC_OUT, tmp);
2147 bitmap_set_free (tmp);
2149 else
2150 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2154 /* Prune expressions that are clobbered in block and thus become
2155 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2156 prune_clobbered_mems (ANTIC_OUT, block);
2158 /* Generate ANTIC_OUT - TMP_GEN. */
2159 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2161 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2162 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2163 TMP_GEN (block));
2165 /* Then union in the ANTIC_OUT - TMP_GEN values,
2166 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2167 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2168 bitmap_value_insert_into_set (ANTIC_IN (block),
2169 expression_for_id (bii));
2171 clean (ANTIC_IN (block));
2173 if (!was_visited || !bitmap_set_equal (old, ANTIC_IN (block)))
2174 changed = true;
2176 maybe_dump_sets:
2177 if (dump_file && (dump_flags & TDF_DETAILS))
2179 if (ANTIC_OUT)
2180 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2182 if (changed)
2183 fprintf (dump_file, "[changed] ");
2184 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2185 block->index);
2187 if (S)
2188 print_bitmap_set (dump_file, S, "S", block->index);
2190 if (old)
2191 bitmap_set_free (old);
2192 if (S)
2193 bitmap_set_free (S);
2194 if (ANTIC_OUT)
2195 bitmap_set_free (ANTIC_OUT);
2196 return changed;
2199 /* Compute PARTIAL_ANTIC for BLOCK.
2201 If succs(BLOCK) > 1 then
2202 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2203 in ANTIC_OUT for all succ(BLOCK)
2204 else if succs(BLOCK) == 1 then
2205 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2207 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2208 - ANTIC_IN[BLOCK])
2211 static void
2212 compute_partial_antic_aux (basic_block block,
2213 bool block_has_abnormal_pred_edge)
2215 bitmap_set_t old_PA_IN;
2216 bitmap_set_t PA_OUT;
2217 edge e;
2218 edge_iterator ei;
2219 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2221 old_PA_IN = PA_OUT = NULL;
2223 /* If any edges from predecessors are abnormal, antic_in is empty,
2224 so do nothing. */
2225 if (block_has_abnormal_pred_edge)
2226 goto maybe_dump_sets;
2228 /* If there are too many partially anticipatable values in the
2229 block, phi_translate_set can take an exponential time: stop
2230 before the translation starts. */
2231 if (max_pa
2232 && single_succ_p (block)
2233 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2234 goto maybe_dump_sets;
2236 old_PA_IN = PA_IN (block);
2237 PA_OUT = bitmap_set_new ();
2239 /* If the block has no successors, ANTIC_OUT is empty. */
2240 if (EDGE_COUNT (block->succs) == 0)
2242 /* If we have one successor, we could have some phi nodes to
2243 translate through. Note that we can't phi translate across DFS
2244 back edges in partial antic, because it uses a union operation on
2245 the successors. For recurrences like IV's, we will end up
2246 generating a new value in the set on each go around (i + 3 (VH.1)
2247 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2248 else if (single_succ_p (block))
2250 basic_block succ = single_succ (block);
2251 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2252 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2254 /* If we have multiple successors, we take the union of all of
2255 them. */
2256 else
2258 size_t i;
2259 basic_block bprime;
2261 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2262 FOR_EACH_EDGE (e, ei, block->succs)
2264 if (e->flags & EDGE_DFS_BACK)
2265 continue;
2266 worklist.quick_push (e->dest);
2268 if (worklist.length () > 0)
2270 FOR_EACH_VEC_ELT (worklist, i, bprime)
2272 unsigned int i;
2273 bitmap_iterator bi;
2275 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2276 bitmap_value_insert_into_set (PA_OUT,
2277 expression_for_id (i));
2278 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2280 bitmap_set_t pa_in = bitmap_set_new ();
2281 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2282 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2283 bitmap_value_insert_into_set (PA_OUT,
2284 expression_for_id (i));
2285 bitmap_set_free (pa_in);
2287 else
2288 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2289 bitmap_value_insert_into_set (PA_OUT,
2290 expression_for_id (i));
2295 /* Prune expressions that are clobbered in block and thus become
2296 invalid if translated from PA_OUT to PA_IN. */
2297 prune_clobbered_mems (PA_OUT, block);
2299 /* PA_IN starts with PA_OUT - TMP_GEN.
2300 Then we subtract things from ANTIC_IN. */
2301 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2303 /* For partial antic, we want to put back in the phi results, since
2304 we will properly avoid making them partially antic over backedges. */
2305 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2306 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2308 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2309 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2311 dependent_clean (PA_IN (block), ANTIC_IN (block));
2313 maybe_dump_sets:
2314 if (dump_file && (dump_flags & TDF_DETAILS))
2316 if (PA_OUT)
2317 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2319 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2321 if (old_PA_IN)
2322 bitmap_set_free (old_PA_IN);
2323 if (PA_OUT)
2324 bitmap_set_free (PA_OUT);
2327 /* Compute ANTIC and partial ANTIC sets. */
2329 static void
2330 compute_antic (void)
2332 bool changed = true;
2333 int num_iterations = 0;
2334 basic_block block;
2335 int i;
2336 edge_iterator ei;
2337 edge e;
2339 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2340 We pre-build the map of blocks with incoming abnormal edges here. */
2341 has_abnormal_preds = sbitmap_alloc (last_basic_block_for_fn (cfun));
2342 bitmap_clear (has_abnormal_preds);
2344 FOR_ALL_BB_FN (block, cfun)
2346 BB_VISITED (block) = 0;
2348 FOR_EACH_EDGE (e, ei, block->preds)
2349 if (e->flags & EDGE_ABNORMAL)
2351 bitmap_set_bit (has_abnormal_preds, block->index);
2353 /* We also anticipate nothing. */
2354 BB_VISITED (block) = 1;
2355 break;
2358 /* While we are here, give empty ANTIC_IN sets to each block. */
2359 ANTIC_IN (block) = bitmap_set_new ();
2360 if (do_partial_partial)
2361 PA_IN (block) = bitmap_set_new ();
2364 /* At the exit block we anticipate nothing. */
2365 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun)) = 1;
2367 /* For ANTIC computation we need a postorder that also guarantees that
2368 a block with a single successor is visited after its successor.
2369 RPO on the inverted CFG has this property. */
2370 int *postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
2371 int postorder_num = inverted_post_order_compute (postorder);
2373 auto_sbitmap worklist (last_basic_block_for_fn (cfun) + 1);
2374 bitmap_ones (worklist);
2375 while (changed)
2377 if (dump_file && (dump_flags & TDF_DETAILS))
2378 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2379 /* ??? We need to clear our PHI translation cache here as the
2380 ANTIC sets shrink and we restrict valid translations to
2381 those having operands with leaders in ANTIC. Same below
2382 for PA ANTIC computation. */
2383 num_iterations++;
2384 changed = false;
2385 for (i = postorder_num - 1; i >= 0; i--)
2387 if (bitmap_bit_p (worklist, postorder[i]))
2389 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2390 bitmap_clear_bit (worklist, block->index);
2391 if (compute_antic_aux (block,
2392 bitmap_bit_p (has_abnormal_preds,
2393 block->index)))
2395 FOR_EACH_EDGE (e, ei, block->preds)
2396 bitmap_set_bit (worklist, e->src->index);
2397 changed = true;
2401 /* Theoretically possible, but *highly* unlikely. */
2402 gcc_checking_assert (num_iterations < 500);
2405 statistics_histogram_event (cfun, "compute_antic iterations",
2406 num_iterations);
2408 if (do_partial_partial)
2410 /* For partial antic we ignore backedges and thus we do not need
2411 to perform any iteration when we process blocks in postorder. */
2412 postorder_num = pre_and_rev_post_order_compute (NULL, postorder, false);
2413 for (i = postorder_num - 1 ; i >= 0; i--)
2415 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2416 compute_partial_antic_aux (block,
2417 bitmap_bit_p (has_abnormal_preds,
2418 block->index));
2422 sbitmap_free (has_abnormal_preds);
2423 free (postorder);
2427 /* Inserted expressions are placed onto this worklist, which is used
2428 for performing quick dead code elimination of insertions we made
2429 that didn't turn out to be necessary. */
2430 static bitmap inserted_exprs;
2432 /* The actual worker for create_component_ref_by_pieces. */
2434 static tree
2435 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2436 unsigned int *operand, gimple_seq *stmts)
2438 vn_reference_op_t currop = &ref->operands[*operand];
2439 tree genop;
2440 ++*operand;
2441 switch (currop->opcode)
2443 case CALL_EXPR:
2444 gcc_unreachable ();
2446 case MEM_REF:
2448 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2449 stmts);
2450 if (!baseop)
2451 return NULL_TREE;
2452 tree offset = currop->op0;
2453 if (TREE_CODE (baseop) == ADDR_EXPR
2454 && handled_component_p (TREE_OPERAND (baseop, 0)))
2456 HOST_WIDE_INT off;
2457 tree base;
2458 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2459 &off);
2460 gcc_assert (base);
2461 offset = int_const_binop (PLUS_EXPR, offset,
2462 build_int_cst (TREE_TYPE (offset),
2463 off));
2464 baseop = build_fold_addr_expr (base);
2466 genop = build2 (MEM_REF, currop->type, baseop, offset);
2467 MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2468 MR_DEPENDENCE_BASE (genop) = currop->base;
2469 REF_REVERSE_STORAGE_ORDER (genop) = currop->reverse;
2470 return genop;
2473 case TARGET_MEM_REF:
2475 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2476 vn_reference_op_t nextop = &ref->operands[++*operand];
2477 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2478 stmts);
2479 if (!baseop)
2480 return NULL_TREE;
2481 if (currop->op0)
2483 genop0 = find_or_generate_expression (block, currop->op0, stmts);
2484 if (!genop0)
2485 return NULL_TREE;
2487 if (nextop->op0)
2489 genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2490 if (!genop1)
2491 return NULL_TREE;
2493 genop = build5 (TARGET_MEM_REF, currop->type,
2494 baseop, currop->op2, genop0, currop->op1, genop1);
2496 MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2497 MR_DEPENDENCE_BASE (genop) = currop->base;
2498 return genop;
2501 case ADDR_EXPR:
2502 if (currop->op0)
2504 gcc_assert (is_gimple_min_invariant (currop->op0));
2505 return currop->op0;
2507 /* Fallthrough. */
2508 case REALPART_EXPR:
2509 case IMAGPART_EXPR:
2510 case VIEW_CONVERT_EXPR:
2512 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2513 stmts);
2514 if (!genop0)
2515 return NULL_TREE;
2516 return fold_build1 (currop->opcode, currop->type, genop0);
2519 case WITH_SIZE_EXPR:
2521 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2522 stmts);
2523 if (!genop0)
2524 return NULL_TREE;
2525 tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2526 if (!genop1)
2527 return NULL_TREE;
2528 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2531 case BIT_FIELD_REF:
2533 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2534 stmts);
2535 if (!genop0)
2536 return NULL_TREE;
2537 tree op1 = currop->op0;
2538 tree op2 = currop->op1;
2539 tree t = build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2540 REF_REVERSE_STORAGE_ORDER (t) = currop->reverse;
2541 return fold (t);
2544 /* For array ref vn_reference_op's, operand 1 of the array ref
2545 is op0 of the reference op and operand 3 of the array ref is
2546 op1. */
2547 case ARRAY_RANGE_REF:
2548 case ARRAY_REF:
2550 tree genop0;
2551 tree genop1 = currop->op0;
2552 tree genop2 = currop->op1;
2553 tree genop3 = currop->op2;
2554 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2555 stmts);
2556 if (!genop0)
2557 return NULL_TREE;
2558 genop1 = find_or_generate_expression (block, genop1, stmts);
2559 if (!genop1)
2560 return NULL_TREE;
2561 if (genop2)
2563 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2564 /* Drop zero minimum index if redundant. */
2565 if (integer_zerop (genop2)
2566 && (!domain_type
2567 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2568 genop2 = NULL_TREE;
2569 else
2571 genop2 = find_or_generate_expression (block, genop2, stmts);
2572 if (!genop2)
2573 return NULL_TREE;
2576 if (genop3)
2578 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2579 /* We can't always put a size in units of the element alignment
2580 here as the element alignment may be not visible. See
2581 PR43783. Simply drop the element size for constant
2582 sizes. */
2583 if (TREE_CODE (genop3) == INTEGER_CST
2584 && TREE_CODE (TYPE_SIZE_UNIT (elmt_type)) == INTEGER_CST
2585 && wi::eq_p (wi::to_offset (TYPE_SIZE_UNIT (elmt_type)),
2586 (wi::to_offset (genop3)
2587 * vn_ref_op_align_unit (currop))))
2588 genop3 = NULL_TREE;
2589 else
2591 genop3 = find_or_generate_expression (block, genop3, stmts);
2592 if (!genop3)
2593 return NULL_TREE;
2596 return build4 (currop->opcode, currop->type, genop0, genop1,
2597 genop2, genop3);
2599 case COMPONENT_REF:
2601 tree op0;
2602 tree op1;
2603 tree genop2 = currop->op1;
2604 op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2605 if (!op0)
2606 return NULL_TREE;
2607 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2608 op1 = currop->op0;
2609 if (genop2)
2611 genop2 = find_or_generate_expression (block, genop2, stmts);
2612 if (!genop2)
2613 return NULL_TREE;
2615 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2618 case SSA_NAME:
2620 genop = find_or_generate_expression (block, currop->op0, stmts);
2621 return genop;
2623 case STRING_CST:
2624 case INTEGER_CST:
2625 case COMPLEX_CST:
2626 case VECTOR_CST:
2627 case REAL_CST:
2628 case CONSTRUCTOR:
2629 case VAR_DECL:
2630 case PARM_DECL:
2631 case CONST_DECL:
2632 case RESULT_DECL:
2633 case FUNCTION_DECL:
2634 return currop->op0;
2636 default:
2637 gcc_unreachable ();
2641 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2642 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2643 trying to rename aggregates into ssa form directly, which is a no no.
2645 Thus, this routine doesn't create temporaries, it just builds a
2646 single access expression for the array, calling
2647 find_or_generate_expression to build the innermost pieces.
2649 This function is a subroutine of create_expression_by_pieces, and
2650 should not be called on it's own unless you really know what you
2651 are doing. */
2653 static tree
2654 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2655 gimple_seq *stmts)
2657 unsigned int op = 0;
2658 return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2661 /* Find a simple leader for an expression, or generate one using
2662 create_expression_by_pieces from a NARY expression for the value.
2663 BLOCK is the basic_block we are looking for leaders in.
2664 OP is the tree expression to find a leader for or generate.
2665 Returns the leader or NULL_TREE on failure. */
2667 static tree
2668 find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2670 pre_expr expr = get_or_alloc_expr_for (op);
2671 unsigned int lookfor = get_expr_value_id (expr);
2672 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2673 if (leader)
2675 if (leader->kind == NAME)
2676 return PRE_EXPR_NAME (leader);
2677 else if (leader->kind == CONSTANT)
2678 return PRE_EXPR_CONSTANT (leader);
2680 /* Defer. */
2681 return NULL_TREE;
2684 /* It must be a complex expression, so generate it recursively. Note
2685 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2686 where the insert algorithm fails to insert a required expression. */
2687 bitmap exprset = value_expressions[lookfor];
2688 bitmap_iterator bi;
2689 unsigned int i;
2690 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2692 pre_expr temp = expression_for_id (i);
2693 /* We cannot insert random REFERENCE expressions at arbitrary
2694 places. We can insert NARYs which eventually re-materializes
2695 its operand values. */
2696 if (temp->kind == NARY)
2697 return create_expression_by_pieces (block, temp, stmts,
2698 get_expr_type (expr));
2701 /* Defer. */
2702 return NULL_TREE;
2705 #define NECESSARY GF_PLF_1
2707 /* Create an expression in pieces, so that we can handle very complex
2708 expressions that may be ANTIC, but not necessary GIMPLE.
2709 BLOCK is the basic block the expression will be inserted into,
2710 EXPR is the expression to insert (in value form)
2711 STMTS is a statement list to append the necessary insertions into.
2713 This function will die if we hit some value that shouldn't be
2714 ANTIC but is (IE there is no leader for it, or its components).
2715 The function returns NULL_TREE in case a different antic expression
2716 has to be inserted first.
2717 This function may also generate expressions that are themselves
2718 partially or fully redundant. Those that are will be either made
2719 fully redundant during the next iteration of insert (for partially
2720 redundant ones), or eliminated by eliminate (for fully redundant
2721 ones). */
2723 static tree
2724 create_expression_by_pieces (basic_block block, pre_expr expr,
2725 gimple_seq *stmts, tree type)
2727 tree name;
2728 tree folded;
2729 gimple_seq forced_stmts = NULL;
2730 unsigned int value_id;
2731 gimple_stmt_iterator gsi;
2732 tree exprtype = type ? type : get_expr_type (expr);
2733 pre_expr nameexpr;
2734 gassign *newstmt;
2736 switch (expr->kind)
2738 /* We may hit the NAME/CONSTANT case if we have to convert types
2739 that value numbering saw through. */
2740 case NAME:
2741 folded = PRE_EXPR_NAME (expr);
2742 if (useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2743 return folded;
2744 break;
2745 case CONSTANT:
2747 folded = PRE_EXPR_CONSTANT (expr);
2748 tree tem = fold_convert (exprtype, folded);
2749 if (is_gimple_min_invariant (tem))
2750 return tem;
2751 break;
2753 case REFERENCE:
2754 if (PRE_EXPR_REFERENCE (expr)->operands[0].opcode == CALL_EXPR)
2756 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2757 unsigned int operand = 1;
2758 vn_reference_op_t currop = &ref->operands[0];
2759 tree sc = NULL_TREE;
2760 tree fn;
2761 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2762 fn = currop->op0;
2763 else
2764 fn = find_or_generate_expression (block, currop->op0, stmts);
2765 if (!fn)
2766 return NULL_TREE;
2767 if (currop->op1)
2769 sc = find_or_generate_expression (block, currop->op1, stmts);
2770 if (!sc)
2771 return NULL_TREE;
2773 auto_vec<tree> args (ref->operands.length () - 1);
2774 while (operand < ref->operands.length ())
2776 tree arg = create_component_ref_by_pieces_1 (block, ref,
2777 &operand, stmts);
2778 if (!arg)
2779 return NULL_TREE;
2780 args.quick_push (arg);
2782 gcall *call
2783 = gimple_build_call_vec ((TREE_CODE (fn) == FUNCTION_DECL
2784 ? build_fold_addr_expr (fn) : fn), args);
2785 gimple_call_set_with_bounds (call, currop->with_bounds);
2786 if (sc)
2787 gimple_call_set_chain (call, sc);
2788 tree forcedname = make_ssa_name (currop->type);
2789 gimple_call_set_lhs (call, forcedname);
2790 gimple_set_vuse (call, BB_LIVE_VOP_ON_EXIT (block));
2791 gimple_seq_add_stmt_without_update (&forced_stmts, call);
2792 folded = forcedname;
2794 else
2796 folded = create_component_ref_by_pieces (block,
2797 PRE_EXPR_REFERENCE (expr),
2798 stmts);
2799 if (!folded)
2800 return NULL_TREE;
2801 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2802 newstmt = gimple_build_assign (name, folded);
2803 gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2804 gimple_set_vuse (newstmt, BB_LIVE_VOP_ON_EXIT (block));
2805 folded = name;
2807 break;
2808 case NARY:
2810 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2811 tree *genop = XALLOCAVEC (tree, nary->length);
2812 unsigned i;
2813 for (i = 0; i < nary->length; ++i)
2815 genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2816 if (!genop[i])
2817 return NULL_TREE;
2818 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2819 may have conversions stripped. */
2820 if (nary->opcode == POINTER_PLUS_EXPR)
2822 if (i == 0)
2823 genop[i] = gimple_convert (&forced_stmts,
2824 nary->type, genop[i]);
2825 else if (i == 1)
2826 genop[i] = gimple_convert (&forced_stmts,
2827 sizetype, genop[i]);
2829 else
2830 genop[i] = gimple_convert (&forced_stmts,
2831 TREE_TYPE (nary->op[i]), genop[i]);
2833 if (nary->opcode == CONSTRUCTOR)
2835 vec<constructor_elt, va_gc> *elts = NULL;
2836 for (i = 0; i < nary->length; ++i)
2837 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2838 folded = build_constructor (nary->type, elts);
2839 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2840 newstmt = gimple_build_assign (name, folded);
2841 gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2842 folded = name;
2844 else
2846 switch (nary->length)
2848 case 1:
2849 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2850 genop[0]);
2851 break;
2852 case 2:
2853 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2854 genop[0], genop[1]);
2855 break;
2856 case 3:
2857 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2858 genop[0], genop[1], genop[2]);
2859 break;
2860 default:
2861 gcc_unreachable ();
2865 break;
2866 default:
2867 gcc_unreachable ();
2870 folded = gimple_convert (&forced_stmts, exprtype, folded);
2872 /* If there is nothing to insert, return the simplified result. */
2873 if (gimple_seq_empty_p (forced_stmts))
2874 return folded;
2875 /* If we simplified to a constant return it and discard eventually
2876 built stmts. */
2877 if (is_gimple_min_invariant (folded))
2879 gimple_seq_discard (forced_stmts);
2880 return folded;
2882 /* Likewise if we simplified to sth not queued for insertion. */
2883 bool found = false;
2884 gsi = gsi_last (forced_stmts);
2885 for (; !gsi_end_p (gsi); gsi_prev (&gsi))
2887 gimple *stmt = gsi_stmt (gsi);
2888 tree forcedname = gimple_get_lhs (stmt);
2889 if (forcedname == folded)
2891 found = true;
2892 break;
2895 if (! found)
2897 gimple_seq_discard (forced_stmts);
2898 return folded;
2900 gcc_assert (TREE_CODE (folded) == SSA_NAME);
2902 /* If we have any intermediate expressions to the value sets, add them
2903 to the value sets and chain them in the instruction stream. */
2904 if (forced_stmts)
2906 gsi = gsi_start (forced_stmts);
2907 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2909 gimple *stmt = gsi_stmt (gsi);
2910 tree forcedname = gimple_get_lhs (stmt);
2911 pre_expr nameexpr;
2913 if (forcedname != folded)
2915 VN_INFO_GET (forcedname)->valnum = forcedname;
2916 VN_INFO (forcedname)->value_id = get_next_value_id ();
2917 nameexpr = get_or_alloc_expr_for_name (forcedname);
2918 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2919 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2920 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2923 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
2924 gimple_set_plf (stmt, NECESSARY, false);
2926 gimple_seq_add_seq (stmts, forced_stmts);
2929 name = folded;
2931 /* Fold the last statement. */
2932 gsi = gsi_last (*stmts);
2933 if (fold_stmt_inplace (&gsi))
2934 update_stmt (gsi_stmt (gsi));
2936 /* Add a value number to the temporary.
2937 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2938 we are creating the expression by pieces, and this particular piece of
2939 the expression may have been represented. There is no harm in replacing
2940 here. */
2941 value_id = get_expr_value_id (expr);
2942 VN_INFO_GET (name)->value_id = value_id;
2943 VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id);
2944 if (VN_INFO (name)->valnum == NULL_TREE)
2945 VN_INFO (name)->valnum = name;
2946 gcc_assert (VN_INFO (name)->valnum != NULL_TREE);
2947 nameexpr = get_or_alloc_expr_for_name (name);
2948 add_to_value (value_id, nameexpr);
2949 if (NEW_SETS (block))
2950 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2951 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2953 pre_stats.insertions++;
2954 if (dump_file && (dump_flags & TDF_DETAILS))
2956 fprintf (dump_file, "Inserted ");
2957 print_gimple_stmt (dump_file, gsi_stmt (gsi_last (*stmts)), 0, 0);
2958 fprintf (dump_file, " in predecessor %d (%04d)\n",
2959 block->index, value_id);
2962 return name;
2966 /* Insert the to-be-made-available values of expression EXPRNUM for each
2967 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
2968 merge the result with a phi node, given the same value number as
2969 NODE. Return true if we have inserted new stuff. */
2971 static bool
2972 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
2973 vec<pre_expr> avail)
2975 pre_expr expr = expression_for_id (exprnum);
2976 pre_expr newphi;
2977 unsigned int val = get_expr_value_id (expr);
2978 edge pred;
2979 bool insertions = false;
2980 bool nophi = false;
2981 basic_block bprime;
2982 pre_expr eprime;
2983 edge_iterator ei;
2984 tree type = get_expr_type (expr);
2985 tree temp;
2986 gphi *phi;
2988 /* Make sure we aren't creating an induction variable. */
2989 if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
2991 bool firstinsideloop = false;
2992 bool secondinsideloop = false;
2993 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
2994 EDGE_PRED (block, 0)->src);
2995 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
2996 EDGE_PRED (block, 1)->src);
2997 /* Induction variables only have one edge inside the loop. */
2998 if ((firstinsideloop ^ secondinsideloop)
2999 && expr->kind != REFERENCE)
3001 if (dump_file && (dump_flags & TDF_DETAILS))
3002 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3003 nophi = true;
3007 /* Make the necessary insertions. */
3008 FOR_EACH_EDGE (pred, ei, block->preds)
3010 gimple_seq stmts = NULL;
3011 tree builtexpr;
3012 bprime = pred->src;
3013 eprime = avail[pred->dest_idx];
3014 builtexpr = create_expression_by_pieces (bprime, eprime,
3015 &stmts, type);
3016 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3017 if (!gimple_seq_empty_p (stmts))
3019 gsi_insert_seq_on_edge (pred, stmts);
3020 insertions = true;
3022 if (!builtexpr)
3024 /* We cannot insert a PHI node if we failed to insert
3025 on one edge. */
3026 nophi = true;
3027 continue;
3029 if (is_gimple_min_invariant (builtexpr))
3030 avail[pred->dest_idx] = get_or_alloc_expr_for_constant (builtexpr);
3031 else
3032 avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3034 /* If we didn't want a phi node, and we made insertions, we still have
3035 inserted new stuff, and thus return true. If we didn't want a phi node,
3036 and didn't make insertions, we haven't added anything new, so return
3037 false. */
3038 if (nophi && insertions)
3039 return true;
3040 else if (nophi && !insertions)
3041 return false;
3043 /* Now build a phi for the new variable. */
3044 temp = make_temp_ssa_name (type, NULL, "prephitmp");
3045 phi = create_phi_node (temp, block);
3047 gimple_set_plf (phi, NECESSARY, false);
3048 VN_INFO_GET (temp)->value_id = val;
3049 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3050 if (VN_INFO (temp)->valnum == NULL_TREE)
3051 VN_INFO (temp)->valnum = temp;
3052 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3053 FOR_EACH_EDGE (pred, ei, block->preds)
3055 pre_expr ae = avail[pred->dest_idx];
3056 gcc_assert (get_expr_type (ae) == type
3057 || useless_type_conversion_p (type, get_expr_type (ae)));
3058 if (ae->kind == CONSTANT)
3059 add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3060 pred, UNKNOWN_LOCATION);
3061 else
3062 add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3065 newphi = get_or_alloc_expr_for_name (temp);
3066 add_to_value (val, newphi);
3068 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3069 this insertion, since we test for the existence of this value in PHI_GEN
3070 before proceeding with the partial redundancy checks in insert_aux.
3072 The value may exist in AVAIL_OUT, in particular, it could be represented
3073 by the expression we are trying to eliminate, in which case we want the
3074 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3075 inserted there.
3077 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3078 this block, because if it did, it would have existed in our dominator's
3079 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3082 bitmap_insert_into_set (PHI_GEN (block), newphi);
3083 bitmap_value_replace_in_set (AVAIL_OUT (block),
3084 newphi);
3085 bitmap_insert_into_set (NEW_SETS (block),
3086 newphi);
3088 /* If we insert a PHI node for a conversion of another PHI node
3089 in the same basic-block try to preserve range information.
3090 This is important so that followup loop passes receive optimal
3091 number of iteration analysis results. See PR61743. */
3092 if (expr->kind == NARY
3093 && CONVERT_EXPR_CODE_P (expr->u.nary->opcode)
3094 && TREE_CODE (expr->u.nary->op[0]) == SSA_NAME
3095 && gimple_bb (SSA_NAME_DEF_STMT (expr->u.nary->op[0])) == block
3096 && INTEGRAL_TYPE_P (type)
3097 && INTEGRAL_TYPE_P (TREE_TYPE (expr->u.nary->op[0]))
3098 && (TYPE_PRECISION (type)
3099 >= TYPE_PRECISION (TREE_TYPE (expr->u.nary->op[0])))
3100 && SSA_NAME_RANGE_INFO (expr->u.nary->op[0]))
3102 wide_int min, max;
3103 if (get_range_info (expr->u.nary->op[0], &min, &max) == VR_RANGE
3104 && !wi::neg_p (min, SIGNED)
3105 && !wi::neg_p (max, SIGNED))
3106 /* Just handle extension and sign-changes of all-positive ranges. */
3107 set_range_info (temp,
3108 SSA_NAME_RANGE_TYPE (expr->u.nary->op[0]),
3109 wide_int_storage::from (min, TYPE_PRECISION (type),
3110 TYPE_SIGN (type)),
3111 wide_int_storage::from (max, TYPE_PRECISION (type),
3112 TYPE_SIGN (type)));
3115 if (dump_file && (dump_flags & TDF_DETAILS))
3117 fprintf (dump_file, "Created phi ");
3118 print_gimple_stmt (dump_file, phi, 0, 0);
3119 fprintf (dump_file, " in block %d (%04d)\n", block->index, val);
3121 pre_stats.phis++;
3122 return true;
3127 /* Perform insertion of partially redundant or hoistable values.
3128 For BLOCK, do the following:
3129 1. Propagate the NEW_SETS of the dominator into the current block.
3130 If the block has multiple predecessors,
3131 2a. Iterate over the ANTIC expressions for the block to see if
3132 any of them are partially redundant.
3133 2b. If so, insert them into the necessary predecessors to make
3134 the expression fully redundant.
3135 2c. Insert a new PHI merging the values of the predecessors.
3136 2d. Insert the new PHI, and the new expressions, into the
3137 NEW_SETS set.
3138 If the block has multiple successors,
3139 3a. Iterate over the ANTIC values for the block to see if
3140 any of them are good candidates for hoisting.
3141 3b. If so, insert expressions computing the values in BLOCK,
3142 and add the new expressions into the NEW_SETS set.
3143 4. Recursively call ourselves on the dominator children of BLOCK.
3145 Steps 1, 2a, and 4 are done by insert_aux. 2b, 2c and 2d are done by
3146 do_pre_regular_insertion and do_partial_insertion. 3a and 3b are
3147 done in do_hoist_insertion.
3150 static bool
3151 do_pre_regular_insertion (basic_block block, basic_block dom)
3153 bool new_stuff = false;
3154 vec<pre_expr> exprs;
3155 pre_expr expr;
3156 auto_vec<pre_expr> avail;
3157 int i;
3159 exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3160 avail.safe_grow (EDGE_COUNT (block->preds));
3162 FOR_EACH_VEC_ELT (exprs, i, expr)
3164 if (expr->kind == NARY
3165 || expr->kind == REFERENCE)
3167 unsigned int val;
3168 bool by_some = false;
3169 bool cant_insert = false;
3170 bool all_same = true;
3171 pre_expr first_s = NULL;
3172 edge pred;
3173 basic_block bprime;
3174 pre_expr eprime = NULL;
3175 edge_iterator ei;
3176 pre_expr edoubleprime = NULL;
3177 bool do_insertion = false;
3179 val = get_expr_value_id (expr);
3180 if (bitmap_set_contains_value (PHI_GEN (block), val))
3181 continue;
3182 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3184 if (dump_file && (dump_flags & TDF_DETAILS))
3186 fprintf (dump_file, "Found fully redundant value: ");
3187 print_pre_expr (dump_file, expr);
3188 fprintf (dump_file, "\n");
3190 continue;
3193 FOR_EACH_EDGE (pred, ei, block->preds)
3195 unsigned int vprime;
3197 /* We should never run insertion for the exit block
3198 and so not come across fake pred edges. */
3199 gcc_assert (!(pred->flags & EDGE_FAKE));
3200 bprime = pred->src;
3201 /* We are looking at ANTIC_OUT of bprime. */
3202 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3203 bprime, block);
3205 /* eprime will generally only be NULL if the
3206 value of the expression, translated
3207 through the PHI for this predecessor, is
3208 undefined. If that is the case, we can't
3209 make the expression fully redundant,
3210 because its value is undefined along a
3211 predecessor path. We can thus break out
3212 early because it doesn't matter what the
3213 rest of the results are. */
3214 if (eprime == NULL)
3216 avail[pred->dest_idx] = NULL;
3217 cant_insert = true;
3218 break;
3221 vprime = get_expr_value_id (eprime);
3222 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3223 vprime);
3224 if (edoubleprime == NULL)
3226 avail[pred->dest_idx] = eprime;
3227 all_same = false;
3229 else
3231 avail[pred->dest_idx] = edoubleprime;
3232 by_some = true;
3233 /* We want to perform insertions to remove a redundancy on
3234 a path in the CFG we want to optimize for speed. */
3235 if (optimize_edge_for_speed_p (pred))
3236 do_insertion = true;
3237 if (first_s == NULL)
3238 first_s = edoubleprime;
3239 else if (!pre_expr_d::equal (first_s, edoubleprime))
3240 all_same = false;
3243 /* If we can insert it, it's not the same value
3244 already existing along every predecessor, and
3245 it's defined by some predecessor, it is
3246 partially redundant. */
3247 if (!cant_insert && !all_same && by_some)
3249 if (!do_insertion)
3251 if (dump_file && (dump_flags & TDF_DETAILS))
3253 fprintf (dump_file, "Skipping partial redundancy for "
3254 "expression ");
3255 print_pre_expr (dump_file, expr);
3256 fprintf (dump_file, " (%04d), no redundancy on to be "
3257 "optimized for speed edge\n", val);
3260 else if (dbg_cnt (treepre_insert))
3262 if (dump_file && (dump_flags & TDF_DETAILS))
3264 fprintf (dump_file, "Found partial redundancy for "
3265 "expression ");
3266 print_pre_expr (dump_file, expr);
3267 fprintf (dump_file, " (%04d)\n",
3268 get_expr_value_id (expr));
3270 if (insert_into_preds_of_block (block,
3271 get_expression_id (expr),
3272 avail))
3273 new_stuff = true;
3276 /* If all edges produce the same value and that value is
3277 an invariant, then the PHI has the same value on all
3278 edges. Note this. */
3279 else if (!cant_insert && all_same)
3281 gcc_assert (edoubleprime->kind == CONSTANT
3282 || edoubleprime->kind == NAME);
3284 tree temp = make_temp_ssa_name (get_expr_type (expr),
3285 NULL, "pretmp");
3286 gassign *assign
3287 = gimple_build_assign (temp,
3288 edoubleprime->kind == CONSTANT ?
3289 PRE_EXPR_CONSTANT (edoubleprime) :
3290 PRE_EXPR_NAME (edoubleprime));
3291 gimple_stmt_iterator gsi = gsi_after_labels (block);
3292 gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3294 gimple_set_plf (assign, NECESSARY, false);
3295 VN_INFO_GET (temp)->value_id = val;
3296 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3297 if (VN_INFO (temp)->valnum == NULL_TREE)
3298 VN_INFO (temp)->valnum = temp;
3299 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3300 pre_expr newe = get_or_alloc_expr_for_name (temp);
3301 add_to_value (val, newe);
3302 bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3303 bitmap_insert_into_set (NEW_SETS (block), newe);
3308 exprs.release ();
3309 return new_stuff;
3313 /* Perform insertion for partially anticipatable expressions. There
3314 is only one case we will perform insertion for these. This case is
3315 if the expression is partially anticipatable, and fully available.
3316 In this case, we know that putting it earlier will enable us to
3317 remove the later computation. */
3319 static bool
3320 do_pre_partial_partial_insertion (basic_block block, basic_block dom)
3322 bool new_stuff = false;
3323 vec<pre_expr> exprs;
3324 pre_expr expr;
3325 auto_vec<pre_expr> avail;
3326 int i;
3328 exprs = sorted_array_from_bitmap_set (PA_IN (block));
3329 avail.safe_grow (EDGE_COUNT (block->preds));
3331 FOR_EACH_VEC_ELT (exprs, i, expr)
3333 if (expr->kind == NARY
3334 || expr->kind == REFERENCE)
3336 unsigned int val;
3337 bool by_all = true;
3338 bool cant_insert = false;
3339 edge pred;
3340 basic_block bprime;
3341 pre_expr eprime = NULL;
3342 edge_iterator ei;
3344 val = get_expr_value_id (expr);
3345 if (bitmap_set_contains_value (PHI_GEN (block), val))
3346 continue;
3347 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3348 continue;
3350 FOR_EACH_EDGE (pred, ei, block->preds)
3352 unsigned int vprime;
3353 pre_expr edoubleprime;
3355 /* We should never run insertion for the exit block
3356 and so not come across fake pred edges. */
3357 gcc_assert (!(pred->flags & EDGE_FAKE));
3358 bprime = pred->src;
3359 eprime = phi_translate (expr, ANTIC_IN (block),
3360 PA_IN (block),
3361 bprime, block);
3363 /* eprime will generally only be NULL if the
3364 value of the expression, translated
3365 through the PHI for this predecessor, is
3366 undefined. If that is the case, we can't
3367 make the expression fully redundant,
3368 because its value is undefined along a
3369 predecessor path. We can thus break out
3370 early because it doesn't matter what the
3371 rest of the results are. */
3372 if (eprime == NULL)
3374 avail[pred->dest_idx] = NULL;
3375 cant_insert = true;
3376 break;
3379 vprime = get_expr_value_id (eprime);
3380 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3381 avail[pred->dest_idx] = edoubleprime;
3382 if (edoubleprime == NULL)
3384 by_all = false;
3385 break;
3389 /* If we can insert it, it's not the same value
3390 already existing along every predecessor, and
3391 it's defined by some predecessor, it is
3392 partially redundant. */
3393 if (!cant_insert && by_all)
3395 edge succ;
3396 bool do_insertion = false;
3398 /* Insert only if we can remove a later expression on a path
3399 that we want to optimize for speed.
3400 The phi node that we will be inserting in BLOCK is not free,
3401 and inserting it for the sake of !optimize_for_speed successor
3402 may cause regressions on the speed path. */
3403 FOR_EACH_EDGE (succ, ei, block->succs)
3405 if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3406 || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3408 if (optimize_edge_for_speed_p (succ))
3409 do_insertion = true;
3413 if (!do_insertion)
3415 if (dump_file && (dump_flags & TDF_DETAILS))
3417 fprintf (dump_file, "Skipping partial partial redundancy "
3418 "for expression ");
3419 print_pre_expr (dump_file, expr);
3420 fprintf (dump_file, " (%04d), not (partially) anticipated "
3421 "on any to be optimized for speed edges\n", val);
3424 else if (dbg_cnt (treepre_insert))
3426 pre_stats.pa_insert++;
3427 if (dump_file && (dump_flags & TDF_DETAILS))
3429 fprintf (dump_file, "Found partial partial redundancy "
3430 "for expression ");
3431 print_pre_expr (dump_file, expr);
3432 fprintf (dump_file, " (%04d)\n",
3433 get_expr_value_id (expr));
3435 if (insert_into_preds_of_block (block,
3436 get_expression_id (expr),
3437 avail))
3438 new_stuff = true;
3444 exprs.release ();
3445 return new_stuff;
3448 /* Insert expressions in BLOCK to compute hoistable values up.
3449 Return TRUE if something was inserted, otherwise return FALSE.
3450 The caller has to make sure that BLOCK has at least two successors. */
3452 static bool
3453 do_hoist_insertion (basic_block block)
3455 edge e;
3456 edge_iterator ei;
3457 bool new_stuff = false;
3458 unsigned i;
3459 gimple_stmt_iterator last;
3461 /* At least two successors, or else... */
3462 gcc_assert (EDGE_COUNT (block->succs) >= 2);
3464 /* Check that all successors of BLOCK are dominated by block.
3465 We could use dominated_by_p() for this, but actually there is a much
3466 quicker check: any successor that is dominated by BLOCK can't have
3467 more than one predecessor edge. */
3468 FOR_EACH_EDGE (e, ei, block->succs)
3469 if (! single_pred_p (e->dest))
3470 return false;
3472 /* Determine the insertion point. If we cannot safely insert before
3473 the last stmt if we'd have to, bail out. */
3474 last = gsi_last_bb (block);
3475 if (!gsi_end_p (last)
3476 && !is_ctrl_stmt (gsi_stmt (last))
3477 && stmt_ends_bb_p (gsi_stmt (last)))
3478 return false;
3480 /* Compute the set of hoistable expressions from ANTIC_IN. First compute
3481 hoistable values. */
3482 bitmap_set hoistable_set;
3484 /* A hoistable value must be in ANTIC_IN(block)
3485 but not in AVAIL_OUT(BLOCK). */
3486 bitmap_initialize (&hoistable_set.values, &grand_bitmap_obstack);
3487 bitmap_and_compl (&hoistable_set.values,
3488 &ANTIC_IN (block)->values, &AVAIL_OUT (block)->values);
3490 /* Short-cut for a common case: hoistable_set is empty. */
3491 if (bitmap_empty_p (&hoistable_set.values))
3492 return false;
3494 /* Compute which of the hoistable values is in AVAIL_OUT of
3495 at least one of the successors of BLOCK. */
3496 bitmap_head availout_in_some;
3497 bitmap_initialize (&availout_in_some, &grand_bitmap_obstack);
3498 FOR_EACH_EDGE (e, ei, block->succs)
3499 /* Do not consider expressions solely because their availability
3500 on loop exits. They'd be ANTIC-IN throughout the whole loop
3501 and thus effectively hoisted across loops by combination of
3502 PRE and hoisting. */
3503 if (! loop_exit_edge_p (block->loop_father, e))
3504 bitmap_ior_and_into (&availout_in_some, &hoistable_set.values,
3505 &AVAIL_OUT (e->dest)->values);
3506 bitmap_clear (&hoistable_set.values);
3508 /* Short-cut for a common case: availout_in_some is empty. */
3509 if (bitmap_empty_p (&availout_in_some))
3510 return false;
3512 /* Hack hoitable_set in-place so we can use sorted_array_from_bitmap_set. */
3513 hoistable_set.values = availout_in_some;
3514 hoistable_set.expressions = ANTIC_IN (block)->expressions;
3516 /* Now finally construct the topological-ordered expression set. */
3517 vec<pre_expr> exprs = sorted_array_from_bitmap_set (&hoistable_set);
3519 bitmap_clear (&hoistable_set.values);
3521 /* If there are candidate values for hoisting, insert expressions
3522 strategically to make the hoistable expressions fully redundant. */
3523 pre_expr expr;
3524 FOR_EACH_VEC_ELT (exprs, i, expr)
3526 /* While we try to sort expressions topologically above the
3527 sorting doesn't work out perfectly. Catch expressions we
3528 already inserted. */
3529 unsigned int value_id = get_expr_value_id (expr);
3530 if (bitmap_set_contains_value (AVAIL_OUT (block), value_id))
3532 if (dump_file && (dump_flags & TDF_DETAILS))
3534 fprintf (dump_file,
3535 "Already inserted expression for ");
3536 print_pre_expr (dump_file, expr);
3537 fprintf (dump_file, " (%04d)\n", value_id);
3539 continue;
3542 /* OK, we should hoist this value. Perform the transformation. */
3543 pre_stats.hoist_insert++;
3544 if (dump_file && (dump_flags & TDF_DETAILS))
3546 fprintf (dump_file,
3547 "Inserting expression in block %d for code hoisting: ",
3548 block->index);
3549 print_pre_expr (dump_file, expr);
3550 fprintf (dump_file, " (%04d)\n", value_id);
3553 gimple_seq stmts = NULL;
3554 tree res = create_expression_by_pieces (block, expr, &stmts,
3555 get_expr_type (expr));
3557 /* Do not return true if expression creation ultimately
3558 did not insert any statements. */
3559 if (gimple_seq_empty_p (stmts))
3560 res = NULL_TREE;
3561 else
3563 if (gsi_end_p (last) || is_ctrl_stmt (gsi_stmt (last)))
3564 gsi_insert_seq_before (&last, stmts, GSI_SAME_STMT);
3565 else
3566 gsi_insert_seq_after (&last, stmts, GSI_NEW_STMT);
3569 /* Make sure to not return true if expression creation ultimately
3570 failed but also make sure to insert any stmts produced as they
3571 are tracked in inserted_exprs. */
3572 if (! res)
3573 continue;
3575 new_stuff = true;
3578 exprs.release ();
3580 return new_stuff;
3583 /* Do a dominator walk on the control flow graph, and insert computations
3584 of values as necessary for PRE and hoisting. */
3586 static bool
3587 insert_aux (basic_block block, bool do_pre, bool do_hoist)
3589 basic_block son;
3590 bool new_stuff = false;
3592 if (block)
3594 basic_block dom;
3595 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3596 if (dom)
3598 unsigned i;
3599 bitmap_iterator bi;
3600 bitmap_set_t newset;
3602 /* First, update the AVAIL_OUT set with anything we may have
3603 inserted higher up in the dominator tree. */
3604 newset = NEW_SETS (dom);
3605 if (newset)
3607 /* Note that we need to value_replace both NEW_SETS, and
3608 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3609 represented by some non-simple expression here that we want
3610 to replace it with. */
3611 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3613 pre_expr expr = expression_for_id (i);
3614 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3615 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3619 /* Insert expressions for partial redundancies. */
3620 if (do_pre && !single_pred_p (block))
3622 new_stuff |= do_pre_regular_insertion (block, dom);
3623 if (do_partial_partial)
3624 new_stuff |= do_pre_partial_partial_insertion (block, dom);
3627 /* Insert expressions for hoisting. */
3628 if (do_hoist && EDGE_COUNT (block->succs) >= 2)
3629 new_stuff |= do_hoist_insertion (block);
3632 for (son = first_dom_son (CDI_DOMINATORS, block);
3633 son;
3634 son = next_dom_son (CDI_DOMINATORS, son))
3636 new_stuff |= insert_aux (son, do_pre, do_hoist);
3639 return new_stuff;
3642 /* Perform insertion of partially redundant and hoistable values. */
3644 static void
3645 insert (void)
3647 bool new_stuff = true;
3648 basic_block bb;
3649 int num_iterations = 0;
3651 FOR_ALL_BB_FN (bb, cfun)
3652 NEW_SETS (bb) = bitmap_set_new ();
3654 while (new_stuff)
3656 num_iterations++;
3657 if (dump_file && dump_flags & TDF_DETAILS)
3658 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3659 new_stuff = insert_aux (ENTRY_BLOCK_PTR_FOR_FN (cfun), flag_tree_pre,
3660 flag_code_hoisting);
3662 /* Clear the NEW sets before the next iteration. We have already
3663 fully propagated its contents. */
3664 if (new_stuff)
3665 FOR_ALL_BB_FN (bb, cfun)
3666 bitmap_set_free (NEW_SETS (bb));
3668 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3672 /* Compute the AVAIL set for all basic blocks.
3674 This function performs value numbering of the statements in each basic
3675 block. The AVAIL sets are built from information we glean while doing
3676 this value numbering, since the AVAIL sets contain only one entry per
3677 value.
3679 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3680 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3682 static void
3683 compute_avail (void)
3686 basic_block block, son;
3687 basic_block *worklist;
3688 size_t sp = 0;
3689 unsigned i;
3690 tree name;
3692 /* We pretend that default definitions are defined in the entry block.
3693 This includes function arguments and the static chain decl. */
3694 FOR_EACH_SSA_NAME (i, name, cfun)
3696 pre_expr e;
3697 if (!SSA_NAME_IS_DEFAULT_DEF (name)
3698 || has_zero_uses (name)
3699 || virtual_operand_p (name))
3700 continue;
3702 e = get_or_alloc_expr_for_name (name);
3703 add_to_value (get_expr_value_id (e), e);
3704 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)), e);
3705 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3709 if (dump_file && (dump_flags & TDF_DETAILS))
3711 print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3712 "tmp_gen", ENTRY_BLOCK);
3713 print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3714 "avail_out", ENTRY_BLOCK);
3717 /* Allocate the worklist. */
3718 worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
3720 /* Seed the algorithm by putting the dominator children of the entry
3721 block on the worklist. */
3722 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR_FOR_FN (cfun));
3723 son;
3724 son = next_dom_son (CDI_DOMINATORS, son))
3725 worklist[sp++] = son;
3727 BB_LIVE_VOP_ON_EXIT (ENTRY_BLOCK_PTR_FOR_FN (cfun))
3728 = ssa_default_def (cfun, gimple_vop (cfun));
3730 /* Loop until the worklist is empty. */
3731 while (sp)
3733 gimple *stmt;
3734 basic_block dom;
3736 /* Pick a block from the worklist. */
3737 block = worklist[--sp];
3739 /* Initially, the set of available values in BLOCK is that of
3740 its immediate dominator. */
3741 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3742 if (dom)
3744 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3745 BB_LIVE_VOP_ON_EXIT (block) = BB_LIVE_VOP_ON_EXIT (dom);
3748 /* Generate values for PHI nodes. */
3749 for (gphi_iterator gsi = gsi_start_phis (block); !gsi_end_p (gsi);
3750 gsi_next (&gsi))
3752 tree result = gimple_phi_result (gsi.phi ());
3754 /* We have no need for virtual phis, as they don't represent
3755 actual computations. */
3756 if (virtual_operand_p (result))
3758 BB_LIVE_VOP_ON_EXIT (block) = result;
3759 continue;
3762 pre_expr e = get_or_alloc_expr_for_name (result);
3763 add_to_value (get_expr_value_id (e), e);
3764 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3765 bitmap_insert_into_set (PHI_GEN (block), e);
3768 BB_MAY_NOTRETURN (block) = 0;
3770 /* Now compute value numbers and populate value sets with all
3771 the expressions computed in BLOCK. */
3772 for (gimple_stmt_iterator gsi = gsi_start_bb (block); !gsi_end_p (gsi);
3773 gsi_next (&gsi))
3775 ssa_op_iter iter;
3776 tree op;
3778 stmt = gsi_stmt (gsi);
3780 /* Cache whether the basic-block has any non-visible side-effect
3781 or control flow.
3782 If this isn't a call or it is the last stmt in the
3783 basic-block then the CFG represents things correctly. */
3784 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3786 /* Non-looping const functions always return normally.
3787 Otherwise the call might not return or have side-effects
3788 that forbids hoisting possibly trapping expressions
3789 before it. */
3790 int flags = gimple_call_flags (stmt);
3791 if (!(flags & ECF_CONST)
3792 || (flags & ECF_LOOPING_CONST_OR_PURE))
3793 BB_MAY_NOTRETURN (block) = 1;
3796 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3798 pre_expr e = get_or_alloc_expr_for_name (op);
3800 add_to_value (get_expr_value_id (e), e);
3801 bitmap_insert_into_set (TMP_GEN (block), e);
3802 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3805 if (gimple_vdef (stmt))
3806 BB_LIVE_VOP_ON_EXIT (block) = gimple_vdef (stmt);
3808 if (gimple_has_side_effects (stmt)
3809 || stmt_could_throw_p (stmt)
3810 || is_gimple_debug (stmt))
3811 continue;
3813 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3815 if (ssa_undefined_value_p (op))
3816 continue;
3817 pre_expr e = get_or_alloc_expr_for_name (op);
3818 bitmap_value_insert_into_set (EXP_GEN (block), e);
3821 switch (gimple_code (stmt))
3823 case GIMPLE_RETURN:
3824 continue;
3826 case GIMPLE_CALL:
3828 vn_reference_t ref;
3829 vn_reference_s ref1;
3830 pre_expr result = NULL;
3832 /* We can value number only calls to real functions. */
3833 if (gimple_call_internal_p (stmt))
3834 continue;
3836 vn_reference_lookup_call (as_a <gcall *> (stmt), &ref, &ref1);
3837 if (!ref)
3838 continue;
3840 /* If the value of the call is not invalidated in
3841 this block until it is computed, add the expression
3842 to EXP_GEN. */
3843 if (!gimple_vuse (stmt)
3844 || gimple_code
3845 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
3846 || gimple_bb (SSA_NAME_DEF_STMT
3847 (gimple_vuse (stmt))) != block)
3849 result = pre_expr_pool.allocate ();
3850 result->kind = REFERENCE;
3851 result->id = 0;
3852 PRE_EXPR_REFERENCE (result) = ref;
3854 get_or_alloc_expression_id (result);
3855 add_to_value (get_expr_value_id (result), result);
3856 bitmap_value_insert_into_set (EXP_GEN (block), result);
3858 continue;
3861 case GIMPLE_ASSIGN:
3863 pre_expr result = NULL;
3864 switch (vn_get_stmt_kind (stmt))
3866 case VN_NARY:
3868 enum tree_code code = gimple_assign_rhs_code (stmt);
3869 vn_nary_op_t nary;
3871 /* COND_EXPR and VEC_COND_EXPR are awkward in
3872 that they contain an embedded complex expression.
3873 Don't even try to shove those through PRE. */
3874 if (code == COND_EXPR
3875 || code == VEC_COND_EXPR)
3876 continue;
3878 vn_nary_op_lookup_stmt (stmt, &nary);
3879 if (!nary)
3880 continue;
3882 /* If the NARY traps and there was a preceding
3883 point in the block that might not return avoid
3884 adding the nary to EXP_GEN. */
3885 if (BB_MAY_NOTRETURN (block)
3886 && vn_nary_may_trap (nary))
3887 continue;
3889 result = pre_expr_pool.allocate ();
3890 result->kind = NARY;
3891 result->id = 0;
3892 PRE_EXPR_NARY (result) = nary;
3893 break;
3896 case VN_REFERENCE:
3898 tree rhs1 = gimple_assign_rhs1 (stmt);
3899 alias_set_type set = get_alias_set (rhs1);
3900 vec<vn_reference_op_s> operands
3901 = vn_reference_operands_for_lookup (rhs1);
3902 vn_reference_t ref;
3903 vn_reference_lookup_pieces (gimple_vuse (stmt), set,
3904 TREE_TYPE (rhs1),
3905 operands, &ref, VN_WALK);
3906 if (!ref)
3908 operands.release ();
3909 continue;
3912 /* If the value of the reference is not invalidated in
3913 this block until it is computed, add the expression
3914 to EXP_GEN. */
3915 if (gimple_vuse (stmt))
3917 gimple *def_stmt;
3918 bool ok = true;
3919 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
3920 while (!gimple_nop_p (def_stmt)
3921 && gimple_code (def_stmt) != GIMPLE_PHI
3922 && gimple_bb (def_stmt) == block)
3924 if (stmt_may_clobber_ref_p
3925 (def_stmt, gimple_assign_rhs1 (stmt)))
3927 ok = false;
3928 break;
3930 def_stmt
3931 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
3933 if (!ok)
3935 operands.release ();
3936 continue;
3940 /* If the load was value-numbered to another
3941 load make sure we do not use its expression
3942 for insertion if it wouldn't be a valid
3943 replacement. */
3944 /* At the momemt we have a testcase
3945 for hoist insertion of aligned vs. misaligned
3946 variants in gcc.dg/torture/pr65270-1.c thus
3947 with just alignment to be considered we can
3948 simply replace the expression in the hashtable
3949 with the most conservative one. */
3950 vn_reference_op_t ref1 = &ref->operands.last ();
3951 while (ref1->opcode != TARGET_MEM_REF
3952 && ref1->opcode != MEM_REF
3953 && ref1 != &ref->operands[0])
3954 --ref1;
3955 vn_reference_op_t ref2 = &operands.last ();
3956 while (ref2->opcode != TARGET_MEM_REF
3957 && ref2->opcode != MEM_REF
3958 && ref2 != &operands[0])
3959 --ref2;
3960 if ((ref1->opcode == TARGET_MEM_REF
3961 || ref1->opcode == MEM_REF)
3962 && (TYPE_ALIGN (ref1->type)
3963 > TYPE_ALIGN (ref2->type)))
3964 ref1->type
3965 = build_aligned_type (ref1->type,
3966 TYPE_ALIGN (ref2->type));
3967 /* TBAA behavior is an obvious part so make sure
3968 that the hashtable one covers this as well
3969 by adjusting the ref alias set and its base. */
3970 if (ref->set == set
3971 || alias_set_subset_of (set, ref->set))
3973 else if (alias_set_subset_of (ref->set, set))
3975 ref->set = set;
3976 if (ref1->opcode == MEM_REF)
3977 ref1->op0 = fold_convert (TREE_TYPE (ref2->op0),
3978 ref1->op0);
3979 else
3980 ref1->op2 = fold_convert (TREE_TYPE (ref2->op2),
3981 ref1->op2);
3983 else
3985 ref->set = 0;
3986 if (ref1->opcode == MEM_REF)
3987 ref1->op0 = fold_convert (ptr_type_node,
3988 ref1->op0);
3989 else
3990 ref1->op2 = fold_convert (ptr_type_node,
3991 ref1->op2);
3993 operands.release ();
3995 result = pre_expr_pool.allocate ();
3996 result->kind = REFERENCE;
3997 result->id = 0;
3998 PRE_EXPR_REFERENCE (result) = ref;
3999 break;
4002 default:
4003 continue;
4006 get_or_alloc_expression_id (result);
4007 add_to_value (get_expr_value_id (result), result);
4008 bitmap_value_insert_into_set (EXP_GEN (block), result);
4009 continue;
4011 default:
4012 break;
4016 if (dump_file && (dump_flags & TDF_DETAILS))
4018 print_bitmap_set (dump_file, EXP_GEN (block),
4019 "exp_gen", block->index);
4020 print_bitmap_set (dump_file, PHI_GEN (block),
4021 "phi_gen", block->index);
4022 print_bitmap_set (dump_file, TMP_GEN (block),
4023 "tmp_gen", block->index);
4024 print_bitmap_set (dump_file, AVAIL_OUT (block),
4025 "avail_out", block->index);
4028 /* Put the dominator children of BLOCK on the worklist of blocks
4029 to compute available sets for. */
4030 for (son = first_dom_son (CDI_DOMINATORS, block);
4031 son;
4032 son = next_dom_son (CDI_DOMINATORS, son))
4033 worklist[sp++] = son;
4036 free (worklist);
4040 /* Local state for the eliminate domwalk. */
4041 static vec<gimple *> el_to_remove;
4042 static vec<gimple *> el_to_fixup;
4043 static unsigned int el_todo;
4044 static vec<tree> el_avail;
4045 static vec<tree> el_avail_stack;
4047 /* Return a leader for OP that is available at the current point of the
4048 eliminate domwalk. */
4050 static tree
4051 eliminate_avail (tree op)
4053 tree valnum = VN_INFO (op)->valnum;
4054 if (TREE_CODE (valnum) == SSA_NAME)
4056 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
4057 return valnum;
4058 if (el_avail.length () > SSA_NAME_VERSION (valnum))
4059 return el_avail[SSA_NAME_VERSION (valnum)];
4061 else if (is_gimple_min_invariant (valnum))
4062 return valnum;
4063 return NULL_TREE;
4066 /* At the current point of the eliminate domwalk make OP available. */
4068 static void
4069 eliminate_push_avail (tree op)
4071 tree valnum = VN_INFO (op)->valnum;
4072 if (TREE_CODE (valnum) == SSA_NAME)
4074 if (el_avail.length () <= SSA_NAME_VERSION (valnum))
4075 el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1);
4076 tree pushop = op;
4077 if (el_avail[SSA_NAME_VERSION (valnum)])
4078 pushop = el_avail[SSA_NAME_VERSION (valnum)];
4079 el_avail_stack.safe_push (pushop);
4080 el_avail[SSA_NAME_VERSION (valnum)] = op;
4084 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
4085 the leader for the expression if insertion was successful. */
4087 static tree
4088 eliminate_insert (gimple_stmt_iterator *gsi, tree val)
4090 gimple *stmt = gimple_seq_first_stmt (VN_INFO (val)->expr);
4091 if (!is_gimple_assign (stmt)
4092 || (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
4093 && gimple_assign_rhs_code (stmt) != VIEW_CONVERT_EXPR
4094 && gimple_assign_rhs_code (stmt) != BIT_FIELD_REF))
4095 return NULL_TREE;
4097 tree op = gimple_assign_rhs1 (stmt);
4098 if (gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR
4099 || gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
4100 op = TREE_OPERAND (op, 0);
4101 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op;
4102 if (!leader)
4103 return NULL_TREE;
4105 gimple_seq stmts = NULL;
4106 tree res;
4107 if (gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
4108 res = gimple_build (&stmts, BIT_FIELD_REF,
4109 TREE_TYPE (val), leader,
4110 TREE_OPERAND (gimple_assign_rhs1 (stmt), 1),
4111 TREE_OPERAND (gimple_assign_rhs1 (stmt), 2));
4112 else
4113 res = gimple_build (&stmts, gimple_assign_rhs_code (stmt),
4114 TREE_TYPE (val), leader);
4115 gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
4116 VN_INFO_GET (res)->valnum = val;
4118 if (TREE_CODE (leader) == SSA_NAME)
4119 gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true);
4121 pre_stats.insertions++;
4122 if (dump_file && (dump_flags & TDF_DETAILS))
4124 fprintf (dump_file, "Inserted ");
4125 print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (res), 0, 0);
4128 return res;
4131 class eliminate_dom_walker : public dom_walker
4133 public:
4134 eliminate_dom_walker (cdi_direction direction, bool do_pre_)
4135 : dom_walker (direction), do_pre (do_pre_) {}
4137 virtual edge before_dom_children (basic_block);
4138 virtual void after_dom_children (basic_block);
4140 bool do_pre;
4143 /* Perform elimination for the basic-block B during the domwalk. */
4145 edge
4146 eliminate_dom_walker::before_dom_children (basic_block b)
4148 /* Mark new bb. */
4149 el_avail_stack.safe_push (NULL_TREE);
4151 /* ??? If we do nothing for unreachable blocks then this will confuse
4152 tailmerging. Eventually we can reduce its reliance on SCCVN now
4153 that we fully copy/constant-propagate (most) things. */
4155 for (gphi_iterator gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4157 gphi *phi = gsi.phi ();
4158 tree res = PHI_RESULT (phi);
4160 if (virtual_operand_p (res))
4162 gsi_next (&gsi);
4163 continue;
4166 tree sprime = eliminate_avail (res);
4167 if (sprime
4168 && sprime != res)
4170 if (dump_file && (dump_flags & TDF_DETAILS))
4172 fprintf (dump_file, "Replaced redundant PHI node defining ");
4173 print_generic_expr (dump_file, res, 0);
4174 fprintf (dump_file, " with ");
4175 print_generic_expr (dump_file, sprime, 0);
4176 fprintf (dump_file, "\n");
4179 /* If we inserted this PHI node ourself, it's not an elimination. */
4180 if (inserted_exprs
4181 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4182 pre_stats.phis--;
4183 else
4184 pre_stats.eliminations++;
4186 /* If we will propagate into all uses don't bother to do
4187 anything. */
4188 if (may_propagate_copy (res, sprime))
4190 /* Mark the PHI for removal. */
4191 el_to_remove.safe_push (phi);
4192 gsi_next (&gsi);
4193 continue;
4196 remove_phi_node (&gsi, false);
4198 if (inserted_exprs
4199 && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4200 && TREE_CODE (sprime) == SSA_NAME)
4201 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4203 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4204 sprime = fold_convert (TREE_TYPE (res), sprime);
4205 gimple *stmt = gimple_build_assign (res, sprime);
4206 /* ??? It cannot yet be necessary (DOM walk). */
4207 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4209 gimple_stmt_iterator gsi2 = gsi_after_labels (b);
4210 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4211 continue;
4214 eliminate_push_avail (res);
4215 gsi_next (&gsi);
4218 for (gimple_stmt_iterator gsi = gsi_start_bb (b);
4219 !gsi_end_p (gsi);
4220 gsi_next (&gsi))
4222 tree sprime = NULL_TREE;
4223 gimple *stmt = gsi_stmt (gsi);
4224 tree lhs = gimple_get_lhs (stmt);
4225 if (lhs && TREE_CODE (lhs) == SSA_NAME
4226 && !gimple_has_volatile_ops (stmt)
4227 /* See PR43491. Do not replace a global register variable when
4228 it is a the RHS of an assignment. Do replace local register
4229 variables since gcc does not guarantee a local variable will
4230 be allocated in register.
4231 ??? The fix isn't effective here. This should instead
4232 be ensured by not value-numbering them the same but treating
4233 them like volatiles? */
4234 && !(gimple_assign_single_p (stmt)
4235 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL
4236 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt))
4237 && is_global_var (gimple_assign_rhs1 (stmt)))))
4239 sprime = eliminate_avail (lhs);
4240 if (!sprime)
4242 /* If there is no existing usable leader but SCCVN thinks
4243 it has an expression it wants to use as replacement,
4244 insert that. */
4245 tree val = VN_INFO (lhs)->valnum;
4246 if (val != VN_TOP
4247 && TREE_CODE (val) == SSA_NAME
4248 && VN_INFO (val)->needs_insertion
4249 && VN_INFO (val)->expr != NULL
4250 && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE)
4251 eliminate_push_avail (sprime);
4254 /* If this now constitutes a copy duplicate points-to
4255 and range info appropriately. This is especially
4256 important for inserted code. See tree-ssa-copy.c
4257 for similar code. */
4258 if (sprime
4259 && TREE_CODE (sprime) == SSA_NAME)
4261 basic_block sprime_b = gimple_bb (SSA_NAME_DEF_STMT (sprime));
4262 if (POINTER_TYPE_P (TREE_TYPE (lhs))
4263 && VN_INFO_PTR_INFO (lhs)
4264 && ! VN_INFO_PTR_INFO (sprime))
4266 duplicate_ssa_name_ptr_info (sprime,
4267 VN_INFO_PTR_INFO (lhs));
4268 if (b != sprime_b)
4269 mark_ptr_info_alignment_unknown
4270 (SSA_NAME_PTR_INFO (sprime));
4272 else if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))
4273 && VN_INFO_RANGE_INFO (lhs)
4274 && ! VN_INFO_RANGE_INFO (sprime)
4275 && b == sprime_b)
4276 duplicate_ssa_name_range_info (sprime,
4277 VN_INFO_RANGE_TYPE (lhs),
4278 VN_INFO_RANGE_INFO (lhs));
4281 /* Inhibit the use of an inserted PHI on a loop header when
4282 the address of the memory reference is a simple induction
4283 variable. In other cases the vectorizer won't do anything
4284 anyway (either it's loop invariant or a complicated
4285 expression). */
4286 if (sprime
4287 && TREE_CODE (sprime) == SSA_NAME
4288 && do_pre
4289 && (flag_tree_loop_vectorize || flag_tree_parallelize_loops > 1)
4290 && loop_outer (b->loop_father)
4291 && has_zero_uses (sprime)
4292 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))
4293 && gimple_assign_load_p (stmt))
4295 gimple *def_stmt = SSA_NAME_DEF_STMT (sprime);
4296 basic_block def_bb = gimple_bb (def_stmt);
4297 if (gimple_code (def_stmt) == GIMPLE_PHI
4298 && def_bb->loop_father->header == def_bb)
4300 loop_p loop = def_bb->loop_father;
4301 ssa_op_iter iter;
4302 tree op;
4303 bool found = false;
4304 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4306 affine_iv iv;
4307 def_bb = gimple_bb (SSA_NAME_DEF_STMT (op));
4308 if (def_bb
4309 && flow_bb_inside_loop_p (loop, def_bb)
4310 && simple_iv (loop, loop, op, &iv, true))
4312 found = true;
4313 break;
4316 if (found)
4318 if (dump_file && (dump_flags & TDF_DETAILS))
4320 fprintf (dump_file, "Not replacing ");
4321 print_gimple_expr (dump_file, stmt, 0, 0);
4322 fprintf (dump_file, " with ");
4323 print_generic_expr (dump_file, sprime, 0);
4324 fprintf (dump_file, " which would add a loop"
4325 " carried dependence to loop %d\n",
4326 loop->num);
4328 /* Don't keep sprime available. */
4329 sprime = NULL_TREE;
4334 if (sprime)
4336 /* If we can propagate the value computed for LHS into
4337 all uses don't bother doing anything with this stmt. */
4338 if (may_propagate_copy (lhs, sprime))
4340 /* Mark it for removal. */
4341 el_to_remove.safe_push (stmt);
4343 /* ??? Don't count copy/constant propagations. */
4344 if (gimple_assign_single_p (stmt)
4345 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4346 || gimple_assign_rhs1 (stmt) == sprime))
4347 continue;
4349 if (dump_file && (dump_flags & TDF_DETAILS))
4351 fprintf (dump_file, "Replaced ");
4352 print_gimple_expr (dump_file, stmt, 0, 0);
4353 fprintf (dump_file, " with ");
4354 print_generic_expr (dump_file, sprime, 0);
4355 fprintf (dump_file, " in all uses of ");
4356 print_gimple_stmt (dump_file, stmt, 0, 0);
4359 pre_stats.eliminations++;
4360 continue;
4363 /* If this is an assignment from our leader (which
4364 happens in the case the value-number is a constant)
4365 then there is nothing to do. */
4366 if (gimple_assign_single_p (stmt)
4367 && sprime == gimple_assign_rhs1 (stmt))
4368 continue;
4370 /* Else replace its RHS. */
4371 bool can_make_abnormal_goto
4372 = is_gimple_call (stmt)
4373 && stmt_can_make_abnormal_goto (stmt);
4375 if (dump_file && (dump_flags & TDF_DETAILS))
4377 fprintf (dump_file, "Replaced ");
4378 print_gimple_expr (dump_file, stmt, 0, 0);
4379 fprintf (dump_file, " with ");
4380 print_generic_expr (dump_file, sprime, 0);
4381 fprintf (dump_file, " in ");
4382 print_gimple_stmt (dump_file, stmt, 0, 0);
4385 if (TREE_CODE (sprime) == SSA_NAME)
4386 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4387 NECESSARY, true);
4389 pre_stats.eliminations++;
4390 gimple *orig_stmt = stmt;
4391 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4392 TREE_TYPE (sprime)))
4393 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4394 tree vdef = gimple_vdef (stmt);
4395 tree vuse = gimple_vuse (stmt);
4396 propagate_tree_value_into_stmt (&gsi, sprime);
4397 stmt = gsi_stmt (gsi);
4398 update_stmt (stmt);
4399 if (vdef != gimple_vdef (stmt))
4400 VN_INFO (vdef)->valnum = vuse;
4402 /* If we removed EH side-effects from the statement, clean
4403 its EH information. */
4404 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4406 bitmap_set_bit (need_eh_cleanup,
4407 gimple_bb (stmt)->index);
4408 if (dump_file && (dump_flags & TDF_DETAILS))
4409 fprintf (dump_file, " Removed EH side-effects.\n");
4412 /* Likewise for AB side-effects. */
4413 if (can_make_abnormal_goto
4414 && !stmt_can_make_abnormal_goto (stmt))
4416 bitmap_set_bit (need_ab_cleanup,
4417 gimple_bb (stmt)->index);
4418 if (dump_file && (dump_flags & TDF_DETAILS))
4419 fprintf (dump_file, " Removed AB side-effects.\n");
4422 continue;
4426 /* If the statement is a scalar store, see if the expression
4427 has the same value number as its rhs. If so, the store is
4428 dead. */
4429 if (gimple_assign_single_p (stmt)
4430 && !gimple_has_volatile_ops (stmt)
4431 && !is_gimple_reg (gimple_assign_lhs (stmt))
4432 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4433 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4435 tree val;
4436 tree rhs = gimple_assign_rhs1 (stmt);
4437 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4438 gimple_vuse (stmt), VN_WALK, NULL, false);
4439 if (TREE_CODE (rhs) == SSA_NAME)
4440 rhs = VN_INFO (rhs)->valnum;
4441 if (val
4442 && operand_equal_p (val, rhs, 0))
4444 if (dump_file && (dump_flags & TDF_DETAILS))
4446 fprintf (dump_file, "Deleted redundant store ");
4447 print_gimple_stmt (dump_file, stmt, 0, 0);
4450 /* Queue stmt for removal. */
4451 el_to_remove.safe_push (stmt);
4452 continue;
4456 /* If this is a control statement value numbering left edges
4457 unexecuted on force the condition in a way consistent with
4458 that. */
4459 if (gcond *cond = dyn_cast <gcond *> (stmt))
4461 if ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE)
4462 ^ (EDGE_SUCC (b, 1)->flags & EDGE_EXECUTABLE))
4464 if (dump_file && (dump_flags & TDF_DETAILS))
4466 fprintf (dump_file, "Removing unexecutable edge from ");
4467 print_gimple_stmt (dump_file, stmt, 0, 0);
4469 if (((EDGE_SUCC (b, 0)->flags & EDGE_TRUE_VALUE) != 0)
4470 == ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE) != 0))
4471 gimple_cond_make_true (cond);
4472 else
4473 gimple_cond_make_false (cond);
4474 update_stmt (cond);
4475 el_todo |= TODO_cleanup_cfg;
4476 continue;
4480 bool can_make_abnormal_goto = stmt_can_make_abnormal_goto (stmt);
4481 bool was_noreturn = (is_gimple_call (stmt)
4482 && gimple_call_noreturn_p (stmt));
4483 tree vdef = gimple_vdef (stmt);
4484 tree vuse = gimple_vuse (stmt);
4486 /* If we didn't replace the whole stmt (or propagate the result
4487 into all uses), replace all uses on this stmt with their
4488 leaders. */
4489 use_operand_p use_p;
4490 ssa_op_iter iter;
4491 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
4493 tree use = USE_FROM_PTR (use_p);
4494 /* ??? The call code above leaves stmt operands un-updated. */
4495 if (TREE_CODE (use) != SSA_NAME)
4496 continue;
4497 tree sprime = eliminate_avail (use);
4498 if (sprime && sprime != use
4499 && may_propagate_copy (use, sprime)
4500 /* We substitute into debug stmts to avoid excessive
4501 debug temporaries created by removed stmts, but we need
4502 to avoid doing so for inserted sprimes as we never want
4503 to create debug temporaries for them. */
4504 && (!inserted_exprs
4505 || TREE_CODE (sprime) != SSA_NAME
4506 || !is_gimple_debug (stmt)
4507 || !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))))
4509 propagate_value (use_p, sprime);
4510 gimple_set_modified (stmt, true);
4511 if (TREE_CODE (sprime) == SSA_NAME
4512 && !is_gimple_debug (stmt))
4513 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4514 NECESSARY, true);
4518 /* Visit indirect calls and turn them into direct calls if
4519 possible using the devirtualization machinery. */
4520 if (gcall *call_stmt = dyn_cast <gcall *> (stmt))
4522 tree fn = gimple_call_fn (call_stmt);
4523 if (fn
4524 && flag_devirtualize
4525 && virtual_method_call_p (fn))
4527 tree otr_type = obj_type_ref_class (fn);
4528 tree instance;
4529 ipa_polymorphic_call_context context (current_function_decl, fn, stmt, &instance);
4530 bool final;
4532 context.get_dynamic_type (instance, OBJ_TYPE_REF_OBJECT (fn), otr_type, stmt);
4534 vec <cgraph_node *>targets
4535 = possible_polymorphic_call_targets (obj_type_ref_class (fn),
4536 tree_to_uhwi
4537 (OBJ_TYPE_REF_TOKEN (fn)),
4538 context,
4539 &final);
4540 if (dump_file)
4541 dump_possible_polymorphic_call_targets (dump_file,
4542 obj_type_ref_class (fn),
4543 tree_to_uhwi
4544 (OBJ_TYPE_REF_TOKEN (fn)),
4545 context);
4546 if (final && targets.length () <= 1 && dbg_cnt (devirt))
4548 tree fn;
4549 if (targets.length () == 1)
4550 fn = targets[0]->decl;
4551 else
4552 fn = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
4553 if (dump_enabled_p ())
4555 location_t loc = gimple_location_safe (stmt);
4556 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc,
4557 "converting indirect call to "
4558 "function %s\n",
4559 lang_hooks.decl_printable_name (fn, 2));
4561 gimple_call_set_fndecl (call_stmt, fn);
4562 /* If changing the call to __builtin_unreachable
4563 or similar noreturn function, adjust gimple_call_fntype
4564 too. */
4565 if (gimple_call_noreturn_p (call_stmt)
4566 && VOID_TYPE_P (TREE_TYPE (TREE_TYPE (fn)))
4567 && TYPE_ARG_TYPES (TREE_TYPE (fn))
4568 && (TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fn)))
4569 == void_type_node))
4570 gimple_call_set_fntype (call_stmt, TREE_TYPE (fn));
4571 maybe_remove_unused_call_args (cfun, call_stmt);
4572 gimple_set_modified (stmt, true);
4577 if (gimple_modified_p (stmt))
4579 /* If a formerly non-invariant ADDR_EXPR is turned into an
4580 invariant one it was on a separate stmt. */
4581 if (gimple_assign_single_p (stmt)
4582 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR)
4583 recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
4584 gimple *old_stmt = stmt;
4585 if (is_gimple_call (stmt))
4587 /* ??? Only fold calls inplace for now, this may create new
4588 SSA names which in turn will confuse free_scc_vn SSA name
4589 release code. */
4590 fold_stmt_inplace (&gsi);
4591 /* When changing a call into a noreturn call, cfg cleanup
4592 is needed to fix up the noreturn call. */
4593 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4594 el_to_fixup.safe_push (stmt);
4596 else
4598 fold_stmt (&gsi);
4599 stmt = gsi_stmt (gsi);
4600 if ((gimple_code (stmt) == GIMPLE_COND
4601 && (gimple_cond_true_p (as_a <gcond *> (stmt))
4602 || gimple_cond_false_p (as_a <gcond *> (stmt))))
4603 || (gimple_code (stmt) == GIMPLE_SWITCH
4604 && TREE_CODE (gimple_switch_index (
4605 as_a <gswitch *> (stmt)))
4606 == INTEGER_CST))
4607 el_todo |= TODO_cleanup_cfg;
4609 /* If we removed EH side-effects from the statement, clean
4610 its EH information. */
4611 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
4613 bitmap_set_bit (need_eh_cleanup,
4614 gimple_bb (stmt)->index);
4615 if (dump_file && (dump_flags & TDF_DETAILS))
4616 fprintf (dump_file, " Removed EH side-effects.\n");
4618 /* Likewise for AB side-effects. */
4619 if (can_make_abnormal_goto
4620 && !stmt_can_make_abnormal_goto (stmt))
4622 bitmap_set_bit (need_ab_cleanup,
4623 gimple_bb (stmt)->index);
4624 if (dump_file && (dump_flags & TDF_DETAILS))
4625 fprintf (dump_file, " Removed AB side-effects.\n");
4627 update_stmt (stmt);
4628 if (vdef != gimple_vdef (stmt))
4629 VN_INFO (vdef)->valnum = vuse;
4632 /* Make new values available - for fully redundant LHS we
4633 continue with the next stmt above and skip this. */
4634 def_operand_p defp;
4635 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_DEF)
4636 eliminate_push_avail (DEF_FROM_PTR (defp));
4639 /* Replace destination PHI arguments. */
4640 edge_iterator ei;
4641 edge e;
4642 FOR_EACH_EDGE (e, ei, b->succs)
4644 for (gphi_iterator gsi = gsi_start_phis (e->dest);
4645 !gsi_end_p (gsi);
4646 gsi_next (&gsi))
4648 gphi *phi = gsi.phi ();
4649 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
4650 tree arg = USE_FROM_PTR (use_p);
4651 if (TREE_CODE (arg) != SSA_NAME
4652 || virtual_operand_p (arg))
4653 continue;
4654 tree sprime = eliminate_avail (arg);
4655 if (sprime && may_propagate_copy (arg, sprime))
4657 propagate_value (use_p, sprime);
4658 if (TREE_CODE (sprime) == SSA_NAME)
4659 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4663 return NULL;
4666 /* Make no longer available leaders no longer available. */
4668 void
4669 eliminate_dom_walker::after_dom_children (basic_block)
4671 tree entry;
4672 while ((entry = el_avail_stack.pop ()) != NULL_TREE)
4674 tree valnum = VN_INFO (entry)->valnum;
4675 tree old = el_avail[SSA_NAME_VERSION (valnum)];
4676 if (old == entry)
4677 el_avail[SSA_NAME_VERSION (valnum)] = NULL_TREE;
4678 else
4679 el_avail[SSA_NAME_VERSION (valnum)] = entry;
4683 /* Eliminate fully redundant computations. */
4685 static unsigned int
4686 eliminate (bool do_pre)
4688 gimple_stmt_iterator gsi;
4689 gimple *stmt;
4691 need_eh_cleanup = BITMAP_ALLOC (NULL);
4692 need_ab_cleanup = BITMAP_ALLOC (NULL);
4694 el_to_remove.create (0);
4695 el_to_fixup.create (0);
4696 el_todo = 0;
4697 el_avail.create (num_ssa_names);
4698 el_avail_stack.create (0);
4700 eliminate_dom_walker (CDI_DOMINATORS,
4701 do_pre).walk (cfun->cfg->x_entry_block_ptr);
4703 el_avail.release ();
4704 el_avail_stack.release ();
4706 /* We cannot remove stmts during BB walk, especially not release SSA
4707 names there as this confuses the VN machinery. The stmts ending
4708 up in el_to_remove are either stores or simple copies.
4709 Remove stmts in reverse order to make debug stmt creation possible. */
4710 while (!el_to_remove.is_empty ())
4712 stmt = el_to_remove.pop ();
4714 if (dump_file && (dump_flags & TDF_DETAILS))
4716 fprintf (dump_file, "Removing dead stmt ");
4717 print_gimple_stmt (dump_file, stmt, 0, 0);
4720 tree lhs;
4721 if (gimple_code (stmt) == GIMPLE_PHI)
4722 lhs = gimple_phi_result (stmt);
4723 else
4724 lhs = gimple_get_lhs (stmt);
4726 if (inserted_exprs
4727 && TREE_CODE (lhs) == SSA_NAME)
4728 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4730 gsi = gsi_for_stmt (stmt);
4731 if (gimple_code (stmt) == GIMPLE_PHI)
4732 remove_phi_node (&gsi, true);
4733 else
4735 basic_block bb = gimple_bb (stmt);
4736 unlink_stmt_vdef (stmt);
4737 if (gsi_remove (&gsi, true))
4738 bitmap_set_bit (need_eh_cleanup, bb->index);
4739 if (is_gimple_call (stmt) && stmt_can_make_abnormal_goto (stmt))
4740 bitmap_set_bit (need_ab_cleanup, bb->index);
4741 release_defs (stmt);
4744 /* Removing a stmt may expose a forwarder block. */
4745 el_todo |= TODO_cleanup_cfg;
4747 el_to_remove.release ();
4749 /* Fixup stmts that became noreturn calls. This may require splitting
4750 blocks and thus isn't possible during the dominator walk. Do this
4751 in reverse order so we don't inadvertedly remove a stmt we want to
4752 fixup by visiting a dominating now noreturn call first. */
4753 while (!el_to_fixup.is_empty ())
4755 stmt = el_to_fixup.pop ();
4757 if (dump_file && (dump_flags & TDF_DETAILS))
4759 fprintf (dump_file, "Fixing up noreturn call ");
4760 print_gimple_stmt (dump_file, stmt, 0, 0);
4763 if (fixup_noreturn_call (stmt))
4764 el_todo |= TODO_cleanup_cfg;
4766 el_to_fixup.release ();
4768 return el_todo;
4771 /* Perform CFG cleanups made necessary by elimination. */
4773 static unsigned
4774 fini_eliminate (void)
4776 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4777 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4779 if (do_eh_cleanup)
4780 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4782 if (do_ab_cleanup)
4783 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4785 BITMAP_FREE (need_eh_cleanup);
4786 BITMAP_FREE (need_ab_cleanup);
4788 if (do_eh_cleanup || do_ab_cleanup)
4789 return TODO_cleanup_cfg;
4790 return 0;
4793 /* Borrow a bit of tree-ssa-dce.c for the moment.
4794 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4795 this may be a bit faster, and we may want critical edges kept split. */
4797 /* If OP's defining statement has not already been determined to be necessary,
4798 mark that statement necessary. Return the stmt, if it is newly
4799 necessary. */
4801 static inline gimple *
4802 mark_operand_necessary (tree op)
4804 gimple *stmt;
4806 gcc_assert (op);
4808 if (TREE_CODE (op) != SSA_NAME)
4809 return NULL;
4811 stmt = SSA_NAME_DEF_STMT (op);
4812 gcc_assert (stmt);
4814 if (gimple_plf (stmt, NECESSARY)
4815 || gimple_nop_p (stmt))
4816 return NULL;
4818 gimple_set_plf (stmt, NECESSARY, true);
4819 return stmt;
4822 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4823 to insert PHI nodes sometimes, and because value numbering of casts isn't
4824 perfect, we sometimes end up inserting dead code. This simple DCE-like
4825 pass removes any insertions we made that weren't actually used. */
4827 static void
4828 remove_dead_inserted_code (void)
4830 bitmap worklist;
4831 unsigned i;
4832 bitmap_iterator bi;
4833 gimple *t;
4835 worklist = BITMAP_ALLOC (NULL);
4836 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4838 t = SSA_NAME_DEF_STMT (ssa_name (i));
4839 if (gimple_plf (t, NECESSARY))
4840 bitmap_set_bit (worklist, i);
4842 while (!bitmap_empty_p (worklist))
4844 i = bitmap_first_set_bit (worklist);
4845 bitmap_clear_bit (worklist, i);
4846 t = SSA_NAME_DEF_STMT (ssa_name (i));
4848 /* PHI nodes are somewhat special in that each PHI alternative has
4849 data and control dependencies. All the statements feeding the
4850 PHI node's arguments are always necessary. */
4851 if (gimple_code (t) == GIMPLE_PHI)
4853 unsigned k;
4855 for (k = 0; k < gimple_phi_num_args (t); k++)
4857 tree arg = PHI_ARG_DEF (t, k);
4858 if (TREE_CODE (arg) == SSA_NAME)
4860 gimple *n = mark_operand_necessary (arg);
4861 if (n)
4862 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4866 else
4868 /* Propagate through the operands. Examine all the USE, VUSE and
4869 VDEF operands in this statement. Mark all the statements
4870 which feed this statement's uses as necessary. */
4871 ssa_op_iter iter;
4872 tree use;
4874 /* The operands of VDEF expressions are also needed as they
4875 represent potential definitions that may reach this
4876 statement (VDEF operands allow us to follow def-def
4877 links). */
4879 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4881 gimple *n = mark_operand_necessary (use);
4882 if (n)
4883 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4888 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4890 t = SSA_NAME_DEF_STMT (ssa_name (i));
4891 if (!gimple_plf (t, NECESSARY))
4893 gimple_stmt_iterator gsi;
4895 if (dump_file && (dump_flags & TDF_DETAILS))
4897 fprintf (dump_file, "Removing unnecessary insertion:");
4898 print_gimple_stmt (dump_file, t, 0, 0);
4901 gsi = gsi_for_stmt (t);
4902 if (gimple_code (t) == GIMPLE_PHI)
4903 remove_phi_node (&gsi, true);
4904 else
4906 gsi_remove (&gsi, true);
4907 release_defs (t);
4911 BITMAP_FREE (worklist);
4915 /* Initialize data structures used by PRE. */
4917 static void
4918 init_pre (void)
4920 basic_block bb;
4922 next_expression_id = 1;
4923 expressions.create (0);
4924 expressions.safe_push (NULL);
4925 value_expressions.create (get_max_value_id () + 1);
4926 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
4927 name_to_id.create (0);
4929 inserted_exprs = BITMAP_ALLOC (NULL);
4931 connect_infinite_loops_to_exit ();
4932 memset (&pre_stats, 0, sizeof (pre_stats));
4934 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4936 calculate_dominance_info (CDI_DOMINATORS);
4938 bitmap_obstack_initialize (&grand_bitmap_obstack);
4939 phi_translate_table = new hash_table<expr_pred_trans_d> (5110);
4940 expression_to_id = new hash_table<pre_expr_d> (num_ssa_names * 3);
4941 FOR_ALL_BB_FN (bb, cfun)
4943 EXP_GEN (bb) = bitmap_set_new ();
4944 PHI_GEN (bb) = bitmap_set_new ();
4945 TMP_GEN (bb) = bitmap_set_new ();
4946 AVAIL_OUT (bb) = bitmap_set_new ();
4951 /* Deallocate data structures used by PRE. */
4953 static void
4954 fini_pre ()
4956 value_expressions.release ();
4957 BITMAP_FREE (inserted_exprs);
4958 bitmap_obstack_release (&grand_bitmap_obstack);
4959 bitmap_set_pool.release ();
4960 pre_expr_pool.release ();
4961 delete phi_translate_table;
4962 phi_translate_table = NULL;
4963 delete expression_to_id;
4964 expression_to_id = NULL;
4965 name_to_id.release ();
4967 free_aux_for_blocks ();
4970 namespace {
4972 const pass_data pass_data_pre =
4974 GIMPLE_PASS, /* type */
4975 "pre", /* name */
4976 OPTGROUP_NONE, /* optinfo_flags */
4977 TV_TREE_PRE, /* tv_id */
4978 /* PROP_no_crit_edges is ensured by placing pass_split_crit_edges before
4979 pass_pre. */
4980 ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */
4981 0, /* properties_provided */
4982 PROP_no_crit_edges, /* properties_destroyed */
4983 TODO_rebuild_alias, /* todo_flags_start */
4984 0, /* todo_flags_finish */
4987 class pass_pre : public gimple_opt_pass
4989 public:
4990 pass_pre (gcc::context *ctxt)
4991 : gimple_opt_pass (pass_data_pre, ctxt)
4994 /* opt_pass methods: */
4995 virtual bool gate (function *)
4996 { return flag_tree_pre != 0 || flag_code_hoisting != 0; }
4997 virtual unsigned int execute (function *);
4999 }; // class pass_pre
5001 unsigned int
5002 pass_pre::execute (function *fun)
5004 unsigned int todo = 0;
5006 do_partial_partial =
5007 flag_tree_partial_pre && optimize_function_for_speed_p (fun);
5009 /* This has to happen before SCCVN runs because
5010 loop_optimizer_init may create new phis, etc. */
5011 loop_optimizer_init (LOOPS_NORMAL);
5013 if (!run_scc_vn (VN_WALK))
5015 loop_optimizer_finalize ();
5016 return 0;
5019 init_pre ();
5020 scev_initialize ();
5022 /* Collect and value number expressions computed in each basic block. */
5023 compute_avail ();
5025 /* Insert can get quite slow on an incredibly large number of basic
5026 blocks due to some quadratic behavior. Until this behavior is
5027 fixed, don't run it when he have an incredibly large number of
5028 bb's. If we aren't going to run insert, there is no point in
5029 computing ANTIC, either, even though it's plenty fast. */
5030 if (n_basic_blocks_for_fn (fun) < 4000)
5032 compute_antic ();
5033 insert ();
5036 /* Make sure to remove fake edges before committing our inserts.
5037 This makes sure we don't end up with extra critical edges that
5038 we would need to split. */
5039 remove_fake_exit_edges ();
5040 gsi_commit_edge_inserts ();
5042 /* Eliminate folds statements which might (should not...) end up
5043 not keeping virtual operands up-to-date. */
5044 gcc_assert (!need_ssa_update_p (fun));
5046 /* Remove all the redundant expressions. */
5047 todo |= eliminate (true);
5049 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
5050 statistics_counter_event (fun, "PA inserted", pre_stats.pa_insert);
5051 statistics_counter_event (fun, "HOIST inserted", pre_stats.hoist_insert);
5052 statistics_counter_event (fun, "New PHIs", pre_stats.phis);
5053 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
5055 clear_expression_ids ();
5056 remove_dead_inserted_code ();
5058 scev_finalize ();
5059 fini_pre ();
5060 todo |= fini_eliminate ();
5061 loop_optimizer_finalize ();
5063 /* Restore SSA info before tail-merging as that resets it as well. */
5064 scc_vn_restore_ssa_info ();
5066 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
5067 case we can merge the block with the remaining predecessor of the block.
5068 It should either:
5069 - call merge_blocks after each tail merge iteration
5070 - call merge_blocks after all tail merge iterations
5071 - mark TODO_cleanup_cfg when necessary
5072 - share the cfg cleanup with fini_pre. */
5073 todo |= tail_merge_optimize (todo);
5075 free_scc_vn ();
5077 /* Tail merging invalidates the virtual SSA web, together with
5078 cfg-cleanup opportunities exposed by PRE this will wreck the
5079 SSA updating machinery. So make sure to run update-ssa
5080 manually, before eventually scheduling cfg-cleanup as part of
5081 the todo. */
5082 update_ssa (TODO_update_ssa_only_virtuals);
5084 return todo;
5087 } // anon namespace
5089 gimple_opt_pass *
5090 make_pass_pre (gcc::context *ctxt)
5092 return new pass_pre (ctxt);
5095 namespace {
5097 const pass_data pass_data_fre =
5099 GIMPLE_PASS, /* type */
5100 "fre", /* name */
5101 OPTGROUP_NONE, /* optinfo_flags */
5102 TV_TREE_FRE, /* tv_id */
5103 ( PROP_cfg | PROP_ssa ), /* properties_required */
5104 0, /* properties_provided */
5105 0, /* properties_destroyed */
5106 0, /* todo_flags_start */
5107 0, /* todo_flags_finish */
5110 class pass_fre : public gimple_opt_pass
5112 public:
5113 pass_fre (gcc::context *ctxt)
5114 : gimple_opt_pass (pass_data_fre, ctxt)
5117 /* opt_pass methods: */
5118 opt_pass * clone () { return new pass_fre (m_ctxt); }
5119 virtual bool gate (function *) { return flag_tree_fre != 0; }
5120 virtual unsigned int execute (function *);
5122 }; // class pass_fre
5124 unsigned int
5125 pass_fre::execute (function *fun)
5127 unsigned int todo = 0;
5129 if (!run_scc_vn (VN_WALKREWRITE))
5130 return 0;
5132 memset (&pre_stats, 0, sizeof (pre_stats));
5134 /* Remove all the redundant expressions. */
5135 todo |= eliminate (false);
5137 todo |= fini_eliminate ();
5139 scc_vn_restore_ssa_info ();
5140 free_scc_vn ();
5142 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
5143 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
5145 return todo;
5148 } // anon namespace
5150 gimple_opt_pass *
5151 make_pass_fre (gcc::context *ctxt)
5153 return new pass_fre (ctxt);