PR target/82498
[official-gcc.git] / gcc / tree-ssa-pre.c
blob5eb47a9d6d54bc8033f351f0dee25bf16aa849da
1 /* Full and partial redundancy elimination and code hoisting on SSA GIMPLE.
2 Copyright (C) 2001-2017 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
4 <stevenb@suse.de>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
11 any later version.
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "backend.h"
26 #include "rtl.h"
27 #include "tree.h"
28 #include "gimple.h"
29 #include "predict.h"
30 #include "alloc-pool.h"
31 #include "tree-pass.h"
32 #include "ssa.h"
33 #include "cgraph.h"
34 #include "gimple-pretty-print.h"
35 #include "fold-const.h"
36 #include "cfganal.h"
37 #include "gimple-fold.h"
38 #include "tree-eh.h"
39 #include "gimplify.h"
40 #include "gimple-iterator.h"
41 #include "tree-cfg.h"
42 #include "tree-ssa-loop.h"
43 #include "tree-into-ssa.h"
44 #include "tree-dfa.h"
45 #include "tree-ssa.h"
46 #include "cfgloop.h"
47 #include "tree-ssa-sccvn.h"
48 #include "tree-scalar-evolution.h"
49 #include "params.h"
50 #include "dbgcnt.h"
51 #include "domwalk.h"
52 #include "tree-ssa-propagate.h"
53 #include "ipa-utils.h"
54 #include "tree-cfgcleanup.h"
55 #include "langhooks.h"
56 #include "alias.h"
58 /* Even though this file is called tree-ssa-pre.c, we actually
59 implement a bit more than just PRE here. All of them piggy-back
60 on GVN which is implemented in tree-ssa-sccvn.c.
62 1. Full Redundancy Elimination (FRE)
63 This is the elimination phase of GVN.
65 2. Partial Redundancy Elimination (PRE)
66 This is adds computation of AVAIL_OUT and ANTIC_IN and
67 doing expression insertion to form GVN-PRE.
69 3. Code hoisting
70 This optimization uses the ANTIC_IN sets computed for PRE
71 to move expressions further up than PRE would do, to make
72 multiple computations of the same value fully redundant.
73 This pass is explained below (after the explanation of the
74 basic algorithm for PRE).
77 /* TODO:
79 1. Avail sets can be shared by making an avail_find_leader that
80 walks up the dominator tree and looks in those avail sets.
81 This might affect code optimality, it's unclear right now.
82 Currently the AVAIL_OUT sets are the remaining quadraticness in
83 memory of GVN-PRE.
84 2. Strength reduction can be performed by anticipating expressions
85 we can repair later on.
86 3. We can do back-substitution or smarter value numbering to catch
87 commutative expressions split up over multiple statements.
90 /* For ease of terminology, "expression node" in the below refers to
91 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
92 represent the actual statement containing the expressions we care about,
93 and we cache the value number by putting it in the expression. */
95 /* Basic algorithm for Partial Redundancy Elimination:
97 First we walk the statements to generate the AVAIL sets, the
98 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
99 generation of values/expressions by a given block. We use them
100 when computing the ANTIC sets. The AVAIL sets consist of
101 SSA_NAME's that represent values, so we know what values are
102 available in what blocks. AVAIL is a forward dataflow problem. In
103 SSA, values are never killed, so we don't need a kill set, or a
104 fixpoint iteration, in order to calculate the AVAIL sets. In
105 traditional parlance, AVAIL sets tell us the downsafety of the
106 expressions/values.
108 Next, we generate the ANTIC sets. These sets represent the
109 anticipatable expressions. ANTIC is a backwards dataflow
110 problem. An expression is anticipatable in a given block if it could
111 be generated in that block. This means that if we had to perform
112 an insertion in that block, of the value of that expression, we
113 could. Calculating the ANTIC sets requires phi translation of
114 expressions, because the flow goes backwards through phis. We must
115 iterate to a fixpoint of the ANTIC sets, because we have a kill
116 set. Even in SSA form, values are not live over the entire
117 function, only from their definition point onwards. So we have to
118 remove values from the ANTIC set once we go past the definition
119 point of the leaders that make them up.
120 compute_antic/compute_antic_aux performs this computation.
122 Third, we perform insertions to make partially redundant
123 expressions fully redundant.
125 An expression is partially redundant (excluding partial
126 anticipation) if:
128 1. It is AVAIL in some, but not all, of the predecessors of a
129 given block.
130 2. It is ANTIC in all the predecessors.
132 In order to make it fully redundant, we insert the expression into
133 the predecessors where it is not available, but is ANTIC.
135 When optimizing for size, we only eliminate the partial redundancy
136 if we need to insert in only one predecessor. This avoids almost
137 completely the code size increase that PRE usually causes.
139 For the partial anticipation case, we only perform insertion if it
140 is partially anticipated in some block, and fully available in all
141 of the predecessors.
143 do_pre_regular_insertion/do_pre_partial_partial_insertion
144 performs these steps, driven by insert/insert_aux.
146 Fourth, we eliminate fully redundant expressions.
147 This is a simple statement walk that replaces redundant
148 calculations with the now available values. */
150 /* Basic algorithm for Code Hoisting:
152 Code hoisting is: Moving value computations up in the control flow
153 graph to make multiple copies redundant. Typically this is a size
154 optimization, but there are cases where it also is helpful for speed.
156 A simple code hoisting algorithm is implemented that piggy-backs on
157 the PRE infrastructure. For code hoisting, we have to know ANTIC_OUT
158 which is effectively ANTIC_IN - AVAIL_OUT. The latter two have to be
159 computed for PRE, and we can use them to perform a limited version of
160 code hoisting, too.
162 For the purpose of this implementation, a value is hoistable to a basic
163 block B if the following properties are met:
165 1. The value is in ANTIC_IN(B) -- the value will be computed on all
166 paths from B to function exit and it can be computed in B);
168 2. The value is not in AVAIL_OUT(B) -- there would be no need to
169 compute the value again and make it available twice;
171 3. All successors of B are dominated by B -- makes sure that inserting
172 a computation of the value in B will make the remaining
173 computations fully redundant;
175 4. At least one successor has the value in AVAIL_OUT -- to avoid
176 hoisting values up too far;
178 5. There are at least two successors of B -- hoisting in straight
179 line code is pointless.
181 The third condition is not strictly necessary, but it would complicate
182 the hoisting pass a lot. In fact, I don't know of any code hoisting
183 algorithm that does not have this requirement. Fortunately, experiments
184 have show that most candidate hoistable values are in regions that meet
185 this condition (e.g. diamond-shape regions).
187 The forth condition is necessary to avoid hoisting things up too far
188 away from the uses of the value. Nothing else limits the algorithm
189 from hoisting everything up as far as ANTIC_IN allows. Experiments
190 with SPEC and CSiBE have shown that hoisting up too far results in more
191 spilling, less benefits for code size, and worse benchmark scores.
192 Fortunately, in practice most of the interesting hoisting opportunities
193 are caught despite this limitation.
195 For hoistable values that meet all conditions, expressions are inserted
196 to make the calculation of the hoistable value fully redundant. We
197 perform code hoisting insertions after each round of PRE insertions,
198 because code hoisting never exposes new PRE opportunities, but PRE can
199 create new code hoisting opportunities.
201 The code hoisting algorithm is implemented in do_hoist_insert, driven
202 by insert/insert_aux. */
204 /* Representations of value numbers:
206 Value numbers are represented by a representative SSA_NAME. We
207 will create fake SSA_NAME's in situations where we need a
208 representative but do not have one (because it is a complex
209 expression). In order to facilitate storing the value numbers in
210 bitmaps, and keep the number of wasted SSA_NAME's down, we also
211 associate a value_id with each value number, and create full blown
212 ssa_name's only where we actually need them (IE in operands of
213 existing expressions).
215 Theoretically you could replace all the value_id's with
216 SSA_NAME_VERSION, but this would allocate a large number of
217 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
218 It would also require an additional indirection at each point we
219 use the value id. */
221 /* Representation of expressions on value numbers:
223 Expressions consisting of value numbers are represented the same
224 way as our VN internally represents them, with an additional
225 "pre_expr" wrapping around them in order to facilitate storing all
226 of the expressions in the same sets. */
228 /* Representation of sets:
230 The dataflow sets do not need to be sorted in any particular order
231 for the majority of their lifetime, are simply represented as two
232 bitmaps, one that keeps track of values present in the set, and one
233 that keeps track of expressions present in the set.
235 When we need them in topological order, we produce it on demand by
236 transforming the bitmap into an array and sorting it into topo
237 order. */
239 /* Type of expression, used to know which member of the PRE_EXPR union
240 is valid. */
242 enum pre_expr_kind
244 NAME,
245 NARY,
246 REFERENCE,
247 CONSTANT
250 union pre_expr_union
252 tree name;
253 tree constant;
254 vn_nary_op_t nary;
255 vn_reference_t reference;
258 typedef struct pre_expr_d : nofree_ptr_hash <pre_expr_d>
260 enum pre_expr_kind kind;
261 unsigned int id;
262 pre_expr_union u;
264 /* hash_table support. */
265 static inline hashval_t hash (const pre_expr_d *);
266 static inline int equal (const pre_expr_d *, const pre_expr_d *);
267 } *pre_expr;
269 #define PRE_EXPR_NAME(e) (e)->u.name
270 #define PRE_EXPR_NARY(e) (e)->u.nary
271 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
272 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
274 /* Compare E1 and E1 for equality. */
276 inline int
277 pre_expr_d::equal (const pre_expr_d *e1, const pre_expr_d *e2)
279 if (e1->kind != e2->kind)
280 return false;
282 switch (e1->kind)
284 case CONSTANT:
285 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
286 PRE_EXPR_CONSTANT (e2));
287 case NAME:
288 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
289 case NARY:
290 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
291 case REFERENCE:
292 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
293 PRE_EXPR_REFERENCE (e2));
294 default:
295 gcc_unreachable ();
299 /* Hash E. */
301 inline hashval_t
302 pre_expr_d::hash (const pre_expr_d *e)
304 switch (e->kind)
306 case CONSTANT:
307 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
308 case NAME:
309 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
310 case NARY:
311 return PRE_EXPR_NARY (e)->hashcode;
312 case REFERENCE:
313 return PRE_EXPR_REFERENCE (e)->hashcode;
314 default:
315 gcc_unreachable ();
319 /* Next global expression id number. */
320 static unsigned int next_expression_id;
322 /* Mapping from expression to id number we can use in bitmap sets. */
323 static vec<pre_expr> expressions;
324 static hash_table<pre_expr_d> *expression_to_id;
325 static vec<unsigned> name_to_id;
327 /* Allocate an expression id for EXPR. */
329 static inline unsigned int
330 alloc_expression_id (pre_expr expr)
332 struct pre_expr_d **slot;
333 /* Make sure we won't overflow. */
334 gcc_assert (next_expression_id + 1 > next_expression_id);
335 expr->id = next_expression_id++;
336 expressions.safe_push (expr);
337 if (expr->kind == NAME)
339 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
340 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
341 re-allocations by using vec::reserve upfront. */
342 unsigned old_len = name_to_id.length ();
343 name_to_id.reserve (num_ssa_names - old_len);
344 name_to_id.quick_grow_cleared (num_ssa_names);
345 gcc_assert (name_to_id[version] == 0);
346 name_to_id[version] = expr->id;
348 else
350 slot = expression_to_id->find_slot (expr, INSERT);
351 gcc_assert (!*slot);
352 *slot = expr;
354 return next_expression_id - 1;
357 /* Return the expression id for tree EXPR. */
359 static inline unsigned int
360 get_expression_id (const pre_expr expr)
362 return expr->id;
365 static inline unsigned int
366 lookup_expression_id (const pre_expr expr)
368 struct pre_expr_d **slot;
370 if (expr->kind == NAME)
372 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
373 if (name_to_id.length () <= version)
374 return 0;
375 return name_to_id[version];
377 else
379 slot = expression_to_id->find_slot (expr, NO_INSERT);
380 if (!slot)
381 return 0;
382 return ((pre_expr)*slot)->id;
386 /* Return the existing expression id for EXPR, or create one if one
387 does not exist yet. */
389 static inline unsigned int
390 get_or_alloc_expression_id (pre_expr expr)
392 unsigned int id = lookup_expression_id (expr);
393 if (id == 0)
394 return alloc_expression_id (expr);
395 return expr->id = id;
398 /* Return the expression that has expression id ID */
400 static inline pre_expr
401 expression_for_id (unsigned int id)
403 return expressions[id];
406 /* Free the expression id field in all of our expressions,
407 and then destroy the expressions array. */
409 static void
410 clear_expression_ids (void)
412 expressions.release ();
415 static object_allocator<pre_expr_d> pre_expr_pool ("pre_expr nodes");
417 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
419 static pre_expr
420 get_or_alloc_expr_for_name (tree name)
422 struct pre_expr_d expr;
423 pre_expr result;
424 unsigned int result_id;
426 expr.kind = NAME;
427 expr.id = 0;
428 PRE_EXPR_NAME (&expr) = name;
429 result_id = lookup_expression_id (&expr);
430 if (result_id != 0)
431 return expression_for_id (result_id);
433 result = pre_expr_pool.allocate ();
434 result->kind = NAME;
435 PRE_EXPR_NAME (result) = name;
436 alloc_expression_id (result);
437 return result;
440 /* An unordered bitmap set. One bitmap tracks values, the other,
441 expressions. */
442 typedef struct bitmap_set
444 bitmap_head expressions;
445 bitmap_head values;
446 } *bitmap_set_t;
448 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
449 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
451 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
452 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
454 /* Mapping from value id to expressions with that value_id. */
455 static vec<bitmap> value_expressions;
457 /* Sets that we need to keep track of. */
458 typedef struct bb_bitmap_sets
460 /* The EXP_GEN set, which represents expressions/values generated in
461 a basic block. */
462 bitmap_set_t exp_gen;
464 /* The PHI_GEN set, which represents PHI results generated in a
465 basic block. */
466 bitmap_set_t phi_gen;
468 /* The TMP_GEN set, which represents results/temporaries generated
469 in a basic block. IE the LHS of an expression. */
470 bitmap_set_t tmp_gen;
472 /* The AVAIL_OUT set, which represents which values are available in
473 a given basic block. */
474 bitmap_set_t avail_out;
476 /* The ANTIC_IN set, which represents which values are anticipatable
477 in a given basic block. */
478 bitmap_set_t antic_in;
480 /* The PA_IN set, which represents which values are
481 partially anticipatable in a given basic block. */
482 bitmap_set_t pa_in;
484 /* The NEW_SETS set, which is used during insertion to augment the
485 AVAIL_OUT set of blocks with the new insertions performed during
486 the current iteration. */
487 bitmap_set_t new_sets;
489 /* A cache for value_dies_in_block_x. */
490 bitmap expr_dies;
492 /* The live virtual operand on successor edges. */
493 tree vop_on_exit;
495 /* True if we have visited this block during ANTIC calculation. */
496 unsigned int visited : 1;
498 /* True when the block contains a call that might not return. */
499 unsigned int contains_may_not_return_call : 1;
500 } *bb_value_sets_t;
502 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
503 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
504 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
505 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
506 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
507 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
508 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
509 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
510 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
511 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
512 #define BB_LIVE_VOP_ON_EXIT(BB) ((bb_value_sets_t) ((BB)->aux))->vop_on_exit
515 /* This structure is used to keep track of statistics on what
516 optimization PRE was able to perform. */
517 static struct
519 /* The number of RHS computations eliminated by PRE. */
520 int eliminations;
522 /* The number of new expressions/temporaries generated by PRE. */
523 int insertions;
525 /* The number of inserts found due to partial anticipation */
526 int pa_insert;
528 /* The number of inserts made for code hoisting. */
529 int hoist_insert;
531 /* The number of new PHI nodes added by PRE. */
532 int phis;
533 } pre_stats;
535 static bool do_partial_partial;
536 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int);
537 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
538 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
539 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
540 static void bitmap_set_and (bitmap_set_t, bitmap_set_t);
541 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
542 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
543 static bitmap_set_t bitmap_set_new (void);
544 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
545 tree);
546 static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
547 static unsigned int get_expr_value_id (pre_expr);
549 /* We can add and remove elements and entries to and from sets
550 and hash tables, so we use alloc pools for them. */
552 static object_allocator<bitmap_set> bitmap_set_pool ("Bitmap sets");
553 static bitmap_obstack grand_bitmap_obstack;
555 /* Set of blocks with statements that have had their EH properties changed. */
556 static bitmap need_eh_cleanup;
558 /* Set of blocks with statements that have had their AB properties changed. */
559 static bitmap need_ab_cleanup;
561 /* A three tuple {e, pred, v} used to cache phi translations in the
562 phi_translate_table. */
564 typedef struct expr_pred_trans_d : free_ptr_hash<expr_pred_trans_d>
566 /* The expression. */
567 pre_expr e;
569 /* The predecessor block along which we translated the expression. */
570 basic_block pred;
572 /* The value that resulted from the translation. */
573 pre_expr v;
575 /* The hashcode for the expression, pred pair. This is cached for
576 speed reasons. */
577 hashval_t hashcode;
579 /* hash_table support. */
580 static inline hashval_t hash (const expr_pred_trans_d *);
581 static inline int equal (const expr_pred_trans_d *, const expr_pred_trans_d *);
582 } *expr_pred_trans_t;
583 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
585 inline hashval_t
586 expr_pred_trans_d::hash (const expr_pred_trans_d *e)
588 return e->hashcode;
591 inline int
592 expr_pred_trans_d::equal (const expr_pred_trans_d *ve1,
593 const expr_pred_trans_d *ve2)
595 basic_block b1 = ve1->pred;
596 basic_block b2 = ve2->pred;
598 /* If they are not translations for the same basic block, they can't
599 be equal. */
600 if (b1 != b2)
601 return false;
602 return pre_expr_d::equal (ve1->e, ve2->e);
605 /* The phi_translate_table caches phi translations for a given
606 expression and predecessor. */
607 static hash_table<expr_pred_trans_d> *phi_translate_table;
609 /* Add the tuple mapping from {expression E, basic block PRED} to
610 the phi translation table and return whether it pre-existed. */
612 static inline bool
613 phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred)
615 expr_pred_trans_t *slot;
616 expr_pred_trans_d tem;
617 hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e),
618 pred->index);
619 tem.e = e;
620 tem.pred = pred;
621 tem.hashcode = hash;
622 slot = phi_translate_table->find_slot_with_hash (&tem, hash, INSERT);
623 if (*slot)
625 *entry = *slot;
626 return true;
629 *entry = *slot = XNEW (struct expr_pred_trans_d);
630 (*entry)->e = e;
631 (*entry)->pred = pred;
632 (*entry)->hashcode = hash;
633 return false;
637 /* Add expression E to the expression set of value id V. */
639 static void
640 add_to_value (unsigned int v, pre_expr e)
642 bitmap set;
644 gcc_checking_assert (get_expr_value_id (e) == v);
646 if (v >= value_expressions.length ())
648 value_expressions.safe_grow_cleared (v + 1);
651 set = value_expressions[v];
652 if (!set)
654 set = BITMAP_ALLOC (&grand_bitmap_obstack);
655 value_expressions[v] = set;
658 bitmap_set_bit (set, get_or_alloc_expression_id (e));
661 /* Create a new bitmap set and return it. */
663 static bitmap_set_t
664 bitmap_set_new (void)
666 bitmap_set_t ret = bitmap_set_pool.allocate ();
667 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
668 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
669 return ret;
672 /* Return the value id for a PRE expression EXPR. */
674 static unsigned int
675 get_expr_value_id (pre_expr expr)
677 unsigned int id;
678 switch (expr->kind)
680 case CONSTANT:
681 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
682 break;
683 case NAME:
684 id = VN_INFO (PRE_EXPR_NAME (expr))->value_id;
685 break;
686 case NARY:
687 id = PRE_EXPR_NARY (expr)->value_id;
688 break;
689 case REFERENCE:
690 id = PRE_EXPR_REFERENCE (expr)->value_id;
691 break;
692 default:
693 gcc_unreachable ();
695 /* ??? We cannot assert that expr has a value-id (it can be 0), because
696 we assign value-ids only to expressions that have a result
697 in set_hashtable_value_ids. */
698 return id;
701 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
703 static tree
704 sccvn_valnum_from_value_id (unsigned int val)
706 bitmap_iterator bi;
707 unsigned int i;
708 bitmap exprset = value_expressions[val];
709 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
711 pre_expr vexpr = expression_for_id (i);
712 if (vexpr->kind == NAME)
713 return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
714 else if (vexpr->kind == CONSTANT)
715 return PRE_EXPR_CONSTANT (vexpr);
717 return NULL_TREE;
720 /* Remove an expression EXPR from a bitmapped set. */
722 static void
723 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
725 unsigned int val = get_expr_value_id (expr);
726 if (!value_id_constant_p (val))
728 bitmap_clear_bit (&set->values, val);
729 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
733 /* Insert an expression EXPR into a bitmapped set. */
735 static void
736 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
738 unsigned int val = get_expr_value_id (expr);
739 if (! value_id_constant_p (val))
741 /* Note this is the only function causing multiple expressions
742 for the same value to appear in a set. This is needed for
743 TMP_GEN, PHI_GEN and NEW_SETs. */
744 bitmap_set_bit (&set->values, val);
745 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
749 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
751 static void
752 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
754 bitmap_copy (&dest->expressions, &orig->expressions);
755 bitmap_copy (&dest->values, &orig->values);
759 /* Free memory used up by SET. */
760 static void
761 bitmap_set_free (bitmap_set_t set)
763 bitmap_clear (&set->expressions);
764 bitmap_clear (&set->values);
768 /* Generate an topological-ordered array of bitmap set SET. */
770 static vec<pre_expr>
771 sorted_array_from_bitmap_set (bitmap_set_t set)
773 unsigned int i, j;
774 bitmap_iterator bi, bj;
775 vec<pre_expr> result;
777 /* Pre-allocate enough space for the array. */
778 result.create (bitmap_count_bits (&set->expressions));
780 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
782 /* The number of expressions having a given value is usually
783 relatively small. Thus, rather than making a vector of all
784 the expressions and sorting it by value-id, we walk the values
785 and check in the reverse mapping that tells us what expressions
786 have a given value, to filter those in our set. As a result,
787 the expressions are inserted in value-id order, which means
788 topological order.
790 If this is somehow a significant lose for some cases, we can
791 choose which set to walk based on the set size. */
792 bitmap exprset = value_expressions[i];
793 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj)
795 if (bitmap_bit_p (&set->expressions, j))
796 result.quick_push (expression_for_id (j));
800 return result;
803 /* Perform bitmapped set operation DEST &= ORIG. */
805 static void
806 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
808 bitmap_iterator bi;
809 unsigned int i;
811 if (dest != orig)
813 bitmap_and_into (&dest->values, &orig->values);
815 unsigned int to_clear = -1U;
816 FOR_EACH_EXPR_ID_IN_SET (dest, i, bi)
818 if (to_clear != -1U)
820 bitmap_clear_bit (&dest->expressions, to_clear);
821 to_clear = -1U;
823 pre_expr expr = expression_for_id (i);
824 unsigned int value_id = get_expr_value_id (expr);
825 if (!bitmap_bit_p (&dest->values, value_id))
826 to_clear = i;
828 if (to_clear != -1U)
829 bitmap_clear_bit (&dest->expressions, to_clear);
833 /* Subtract all expressions contained in ORIG from DEST. */
835 static bitmap_set_t
836 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
838 bitmap_set_t result = bitmap_set_new ();
839 bitmap_iterator bi;
840 unsigned int i;
842 bitmap_and_compl (&result->expressions, &dest->expressions,
843 &orig->expressions);
845 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
847 pre_expr expr = expression_for_id (i);
848 unsigned int value_id = get_expr_value_id (expr);
849 bitmap_set_bit (&result->values, value_id);
852 return result;
855 /* Subtract all values in bitmap set B from bitmap set A. */
857 static void
858 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
860 unsigned int i;
861 bitmap_iterator bi;
862 pre_expr to_remove = NULL;
863 FOR_EACH_EXPR_ID_IN_SET (a, i, bi)
865 if (to_remove)
867 bitmap_remove_from_set (a, to_remove);
868 to_remove = NULL;
870 pre_expr expr = expression_for_id (i);
871 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
872 to_remove = expr;
874 if (to_remove)
875 bitmap_remove_from_set (a, to_remove);
879 /* Return true if bitmapped set SET contains the value VALUE_ID. */
881 static bool
882 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
884 if (value_id_constant_p (value_id))
885 return true;
887 if (!set || bitmap_empty_p (&set->expressions))
888 return false;
890 return bitmap_bit_p (&set->values, value_id);
893 static inline bool
894 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
896 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
899 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
901 static void
902 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
903 const pre_expr expr)
905 bitmap exprset;
906 unsigned int i;
907 bitmap_iterator bi;
909 if (value_id_constant_p (lookfor))
910 return;
912 if (!bitmap_set_contains_value (set, lookfor))
913 return;
915 /* The number of expressions having a given value is usually
916 significantly less than the total number of expressions in SET.
917 Thus, rather than check, for each expression in SET, whether it
918 has the value LOOKFOR, we walk the reverse mapping that tells us
919 what expressions have a given value, and see if any of those
920 expressions are in our set. For large testcases, this is about
921 5-10x faster than walking the bitmap. If this is somehow a
922 significant lose for some cases, we can choose which set to walk
923 based on the set size. */
924 exprset = value_expressions[lookfor];
925 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
927 if (bitmap_clear_bit (&set->expressions, i))
929 bitmap_set_bit (&set->expressions, get_expression_id (expr));
930 return;
934 gcc_unreachable ();
937 /* Return true if two bitmap sets are equal. */
939 static bool
940 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
942 return bitmap_equal_p (&a->values, &b->values);
945 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
946 and add it otherwise. */
948 static void
949 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
951 unsigned int val = get_expr_value_id (expr);
953 if (bitmap_set_contains_value (set, val))
954 bitmap_set_replace_value (set, val, expr);
955 else
956 bitmap_insert_into_set (set, expr);
959 /* Insert EXPR into SET if EXPR's value is not already present in
960 SET. */
962 static void
963 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
965 unsigned int val = get_expr_value_id (expr);
967 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
969 /* Constant values are always considered to be part of the set. */
970 if (value_id_constant_p (val))
971 return;
973 /* If the value membership changed, add the expression. */
974 if (bitmap_set_bit (&set->values, val))
975 bitmap_set_bit (&set->expressions, expr->id);
978 /* Print out EXPR to outfile. */
980 static void
981 print_pre_expr (FILE *outfile, const pre_expr expr)
983 if (! expr)
985 fprintf (outfile, "NULL");
986 return;
988 switch (expr->kind)
990 case CONSTANT:
991 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr));
992 break;
993 case NAME:
994 print_generic_expr (outfile, PRE_EXPR_NAME (expr));
995 break;
996 case NARY:
998 unsigned int i;
999 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1000 fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode));
1001 for (i = 0; i < nary->length; i++)
1003 print_generic_expr (outfile, nary->op[i]);
1004 if (i != (unsigned) nary->length - 1)
1005 fprintf (outfile, ",");
1007 fprintf (outfile, "}");
1009 break;
1011 case REFERENCE:
1013 vn_reference_op_t vro;
1014 unsigned int i;
1015 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1016 fprintf (outfile, "{");
1017 for (i = 0;
1018 ref->operands.iterate (i, &vro);
1019 i++)
1021 bool closebrace = false;
1022 if (vro->opcode != SSA_NAME
1023 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
1025 fprintf (outfile, "%s", get_tree_code_name (vro->opcode));
1026 if (vro->op0)
1028 fprintf (outfile, "<");
1029 closebrace = true;
1032 if (vro->op0)
1034 print_generic_expr (outfile, vro->op0);
1035 if (vro->op1)
1037 fprintf (outfile, ",");
1038 print_generic_expr (outfile, vro->op1);
1040 if (vro->op2)
1042 fprintf (outfile, ",");
1043 print_generic_expr (outfile, vro->op2);
1046 if (closebrace)
1047 fprintf (outfile, ">");
1048 if (i != ref->operands.length () - 1)
1049 fprintf (outfile, ",");
1051 fprintf (outfile, "}");
1052 if (ref->vuse)
1054 fprintf (outfile, "@");
1055 print_generic_expr (outfile, ref->vuse);
1058 break;
1061 void debug_pre_expr (pre_expr);
1063 /* Like print_pre_expr but always prints to stderr. */
1064 DEBUG_FUNCTION void
1065 debug_pre_expr (pre_expr e)
1067 print_pre_expr (stderr, e);
1068 fprintf (stderr, "\n");
1071 /* Print out SET to OUTFILE. */
1073 static void
1074 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1075 const char *setname, int blockindex)
1077 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1078 if (set)
1080 bool first = true;
1081 unsigned i;
1082 bitmap_iterator bi;
1084 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1086 const pre_expr expr = expression_for_id (i);
1088 if (!first)
1089 fprintf (outfile, ", ");
1090 first = false;
1091 print_pre_expr (outfile, expr);
1093 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1096 fprintf (outfile, " }\n");
1099 void debug_bitmap_set (bitmap_set_t);
1101 DEBUG_FUNCTION void
1102 debug_bitmap_set (bitmap_set_t set)
1104 print_bitmap_set (stderr, set, "debug", 0);
1107 void debug_bitmap_sets_for (basic_block);
1109 DEBUG_FUNCTION void
1110 debug_bitmap_sets_for (basic_block bb)
1112 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1113 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1114 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1115 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1116 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1117 if (do_partial_partial)
1118 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1119 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1122 /* Print out the expressions that have VAL to OUTFILE. */
1124 static void
1125 print_value_expressions (FILE *outfile, unsigned int val)
1127 bitmap set = value_expressions[val];
1128 if (set)
1130 bitmap_set x;
1131 char s[10];
1132 sprintf (s, "%04d", val);
1133 x.expressions = *set;
1134 print_bitmap_set (outfile, &x, s, 0);
1139 DEBUG_FUNCTION void
1140 debug_value_expressions (unsigned int val)
1142 print_value_expressions (stderr, val);
1145 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1146 represent it. */
1148 static pre_expr
1149 get_or_alloc_expr_for_constant (tree constant)
1151 unsigned int result_id;
1152 unsigned int value_id;
1153 struct pre_expr_d expr;
1154 pre_expr newexpr;
1156 expr.kind = CONSTANT;
1157 PRE_EXPR_CONSTANT (&expr) = constant;
1158 result_id = lookup_expression_id (&expr);
1159 if (result_id != 0)
1160 return expression_for_id (result_id);
1162 newexpr = pre_expr_pool.allocate ();
1163 newexpr->kind = CONSTANT;
1164 PRE_EXPR_CONSTANT (newexpr) = constant;
1165 alloc_expression_id (newexpr);
1166 value_id = get_or_alloc_constant_value_id (constant);
1167 add_to_value (value_id, newexpr);
1168 return newexpr;
1171 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1172 Currently only supports constants and SSA_NAMES. */
1173 static pre_expr
1174 get_or_alloc_expr_for (tree t)
1176 if (TREE_CODE (t) == SSA_NAME)
1177 return get_or_alloc_expr_for_name (t);
1178 else if (is_gimple_min_invariant (t))
1179 return get_or_alloc_expr_for_constant (t);
1180 gcc_unreachable ();
1183 /* Return the folded version of T if T, when folded, is a gimple
1184 min_invariant or an SSA name. Otherwise, return T. */
1186 static pre_expr
1187 fully_constant_expression (pre_expr e)
1189 switch (e->kind)
1191 case CONSTANT:
1192 return e;
1193 case NARY:
1195 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1196 tree res = vn_nary_simplify (nary);
1197 if (!res)
1198 return e;
1199 if (is_gimple_min_invariant (res))
1200 return get_or_alloc_expr_for_constant (res);
1201 if (TREE_CODE (res) == SSA_NAME)
1202 return get_or_alloc_expr_for_name (res);
1203 return e;
1205 case REFERENCE:
1207 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1208 tree folded;
1209 if ((folded = fully_constant_vn_reference_p (ref)))
1210 return get_or_alloc_expr_for_constant (folded);
1211 return e;
1213 default:
1214 return e;
1216 return e;
1219 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1220 it has the value it would have in BLOCK. Set *SAME_VALID to true
1221 in case the new vuse doesn't change the value id of the OPERANDS. */
1223 static tree
1224 translate_vuse_through_block (vec<vn_reference_op_s> operands,
1225 alias_set_type set, tree type, tree vuse,
1226 basic_block phiblock,
1227 basic_block block, bool *same_valid)
1229 gimple *phi = SSA_NAME_DEF_STMT (vuse);
1230 ao_ref ref;
1231 edge e = NULL;
1232 bool use_oracle;
1234 *same_valid = true;
1236 if (gimple_bb (phi) != phiblock)
1237 return vuse;
1239 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1241 /* Use the alias-oracle to find either the PHI node in this block,
1242 the first VUSE used in this block that is equivalent to vuse or
1243 the first VUSE which definition in this block kills the value. */
1244 if (gimple_code (phi) == GIMPLE_PHI)
1245 e = find_edge (block, phiblock);
1246 else if (use_oracle)
1247 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1249 vuse = gimple_vuse (phi);
1250 phi = SSA_NAME_DEF_STMT (vuse);
1251 if (gimple_bb (phi) != phiblock)
1252 return vuse;
1253 if (gimple_code (phi) == GIMPLE_PHI)
1255 e = find_edge (block, phiblock);
1256 break;
1259 else
1260 return NULL_TREE;
1262 if (e)
1264 if (use_oracle)
1266 bitmap visited = NULL;
1267 unsigned int cnt;
1268 /* Try to find a vuse that dominates this phi node by skipping
1269 non-clobbering statements. */
1270 vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false,
1271 NULL, NULL);
1272 if (visited)
1273 BITMAP_FREE (visited);
1275 else
1276 vuse = NULL_TREE;
1277 if (!vuse)
1279 /* If we didn't find any, the value ID can't stay the same,
1280 but return the translated vuse. */
1281 *same_valid = false;
1282 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1284 /* ??? We would like to return vuse here as this is the canonical
1285 upmost vdef that this reference is associated with. But during
1286 insertion of the references into the hash tables we only ever
1287 directly insert with their direct gimple_vuse, hence returning
1288 something else would make us not find the other expression. */
1289 return PHI_ARG_DEF (phi, e->dest_idx);
1292 return NULL_TREE;
1295 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1296 SET2 *or* SET3. This is used to avoid making a set consisting of the union
1297 of PA_IN and ANTIC_IN during insert and phi-translation. */
1299 static inline pre_expr
1300 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2,
1301 bitmap_set_t set3 = NULL)
1303 pre_expr result;
1305 result = bitmap_find_leader (set1, val);
1306 if (!result && set2)
1307 result = bitmap_find_leader (set2, val);
1308 if (!result && set3)
1309 result = bitmap_find_leader (set3, val);
1310 return result;
1313 /* Get the tree type for our PRE expression e. */
1315 static tree
1316 get_expr_type (const pre_expr e)
1318 switch (e->kind)
1320 case NAME:
1321 return TREE_TYPE (PRE_EXPR_NAME (e));
1322 case CONSTANT:
1323 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1324 case REFERENCE:
1325 return PRE_EXPR_REFERENCE (e)->type;
1326 case NARY:
1327 return PRE_EXPR_NARY (e)->type;
1329 gcc_unreachable ();
1332 /* Get a representative SSA_NAME for a given expression.
1333 Since all of our sub-expressions are treated as values, we require
1334 them to be SSA_NAME's for simplicity.
1335 Prior versions of GVNPRE used to use "value handles" here, so that
1336 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1337 either case, the operands are really values (IE we do not expect
1338 them to be usable without finding leaders). */
1340 static tree
1341 get_representative_for (const pre_expr e)
1343 tree name;
1344 unsigned int value_id = get_expr_value_id (e);
1346 switch (e->kind)
1348 case NAME:
1349 return VN_INFO (PRE_EXPR_NAME (e))->valnum;
1350 case CONSTANT:
1351 return PRE_EXPR_CONSTANT (e);
1352 case NARY:
1353 case REFERENCE:
1355 /* Go through all of the expressions representing this value
1356 and pick out an SSA_NAME. */
1357 unsigned int i;
1358 bitmap_iterator bi;
1359 bitmap exprs = value_expressions[value_id];
1360 EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1362 pre_expr rep = expression_for_id (i);
1363 if (rep->kind == NAME)
1364 return VN_INFO (PRE_EXPR_NAME (rep))->valnum;
1365 else if (rep->kind == CONSTANT)
1366 return PRE_EXPR_CONSTANT (rep);
1369 break;
1372 /* If we reached here we couldn't find an SSA_NAME. This can
1373 happen when we've discovered a value that has never appeared in
1374 the program as set to an SSA_NAME, as the result of phi translation.
1375 Create one here.
1376 ??? We should be able to re-use this when we insert the statement
1377 to compute it. */
1378 name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1379 VN_INFO_GET (name)->value_id = value_id;
1380 VN_INFO (name)->valnum = name;
1381 /* ??? For now mark this SSA name for release by SCCVN. */
1382 VN_INFO (name)->needs_insertion = true;
1383 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1384 if (dump_file && (dump_flags & TDF_DETAILS))
1386 fprintf (dump_file, "Created SSA_NAME representative ");
1387 print_generic_expr (dump_file, name);
1388 fprintf (dump_file, " for expression:");
1389 print_pre_expr (dump_file, e);
1390 fprintf (dump_file, " (%04d)\n", value_id);
1393 return name;
1398 static pre_expr
1399 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1400 basic_block pred, basic_block phiblock);
1402 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1403 the phis in PRED. Return NULL if we can't find a leader for each part
1404 of the translated expression. */
1406 static pre_expr
1407 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1408 basic_block pred, basic_block phiblock)
1410 switch (expr->kind)
1412 case NARY:
1414 unsigned int i;
1415 bool changed = false;
1416 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1417 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1418 sizeof_vn_nary_op (nary->length));
1419 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1421 for (i = 0; i < newnary->length; i++)
1423 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1424 continue;
1425 else
1427 pre_expr leader, result;
1428 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1429 leader = find_leader_in_sets (op_val_id, set1, set2);
1430 result = phi_translate (leader, set1, set2, pred, phiblock);
1431 if (result && result != leader)
1432 newnary->op[i] = get_representative_for (result);
1433 else if (!result)
1434 return NULL;
1436 changed |= newnary->op[i] != nary->op[i];
1439 if (changed)
1441 pre_expr constant;
1442 unsigned int new_val_id;
1444 PRE_EXPR_NARY (expr) = newnary;
1445 constant = fully_constant_expression (expr);
1446 PRE_EXPR_NARY (expr) = nary;
1447 if (constant != expr)
1449 /* For non-CONSTANTs we have to make sure we can eventually
1450 insert the expression. Which means we need to have a
1451 leader for it. */
1452 if (constant->kind != CONSTANT)
1454 /* Do not allow simplifications to non-constants over
1455 backedges as this will likely result in a loop PHI node
1456 to be inserted and increased register pressure.
1457 See PR77498 - this avoids doing predcoms work in
1458 a less efficient way. */
1459 if (find_edge (pred, phiblock)->flags & EDGE_DFS_BACK)
1461 else
1463 unsigned value_id = get_expr_value_id (constant);
1464 constant = find_leader_in_sets (value_id, set1, set2,
1465 AVAIL_OUT (pred));
1466 if (constant)
1467 return constant;
1470 else
1471 return constant;
1474 tree result = vn_nary_op_lookup_pieces (newnary->length,
1475 newnary->opcode,
1476 newnary->type,
1477 &newnary->op[0],
1478 &nary);
1479 if (result && is_gimple_min_invariant (result))
1480 return get_or_alloc_expr_for_constant (result);
1482 expr = pre_expr_pool.allocate ();
1483 expr->kind = NARY;
1484 expr->id = 0;
1485 if (nary)
1487 PRE_EXPR_NARY (expr) = nary;
1488 new_val_id = nary->value_id;
1489 get_or_alloc_expression_id (expr);
1490 /* When we end up re-using a value number make sure that
1491 doesn't have unrelated (which we can't check here)
1492 range or points-to info on it. */
1493 if (result
1494 && INTEGRAL_TYPE_P (TREE_TYPE (result))
1495 && SSA_NAME_RANGE_INFO (result)
1496 && ! SSA_NAME_IS_DEFAULT_DEF (result))
1498 if (! VN_INFO (result)->info.range_info)
1500 VN_INFO (result)->info.range_info
1501 = SSA_NAME_RANGE_INFO (result);
1502 VN_INFO (result)->range_info_anti_range_p
1503 = SSA_NAME_ANTI_RANGE_P (result);
1505 if (dump_file && (dump_flags & TDF_DETAILS))
1507 fprintf (dump_file, "clearing range info of ");
1508 print_generic_expr (dump_file, result);
1509 fprintf (dump_file, "\n");
1511 SSA_NAME_RANGE_INFO (result) = NULL;
1513 else if (result
1514 && POINTER_TYPE_P (TREE_TYPE (result))
1515 && SSA_NAME_PTR_INFO (result)
1516 && ! SSA_NAME_IS_DEFAULT_DEF (result))
1518 if (! VN_INFO (result)->info.ptr_info)
1519 VN_INFO (result)->info.ptr_info
1520 = SSA_NAME_PTR_INFO (result);
1521 if (dump_file && (dump_flags & TDF_DETAILS))
1523 fprintf (dump_file, "clearing points-to info of ");
1524 print_generic_expr (dump_file, result);
1525 fprintf (dump_file, "\n");
1527 SSA_NAME_PTR_INFO (result) = NULL;
1530 else
1532 new_val_id = get_next_value_id ();
1533 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
1534 nary = vn_nary_op_insert_pieces (newnary->length,
1535 newnary->opcode,
1536 newnary->type,
1537 &newnary->op[0],
1538 result, new_val_id);
1539 PRE_EXPR_NARY (expr) = nary;
1540 get_or_alloc_expression_id (expr);
1542 add_to_value (new_val_id, expr);
1544 return expr;
1546 break;
1548 case REFERENCE:
1550 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1551 vec<vn_reference_op_s> operands = ref->operands;
1552 tree vuse = ref->vuse;
1553 tree newvuse = vuse;
1554 vec<vn_reference_op_s> newoperands = vNULL;
1555 bool changed = false, same_valid = true;
1556 unsigned int i, n;
1557 vn_reference_op_t operand;
1558 vn_reference_t newref;
1560 for (i = 0; operands.iterate (i, &operand); i++)
1562 pre_expr opresult;
1563 pre_expr leader;
1564 tree op[3];
1565 tree type = operand->type;
1566 vn_reference_op_s newop = *operand;
1567 op[0] = operand->op0;
1568 op[1] = operand->op1;
1569 op[2] = operand->op2;
1570 for (n = 0; n < 3; ++n)
1572 unsigned int op_val_id;
1573 if (!op[n])
1574 continue;
1575 if (TREE_CODE (op[n]) != SSA_NAME)
1577 /* We can't possibly insert these. */
1578 if (n != 0
1579 && !is_gimple_min_invariant (op[n]))
1580 break;
1581 continue;
1583 op_val_id = VN_INFO (op[n])->value_id;
1584 leader = find_leader_in_sets (op_val_id, set1, set2);
1585 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1586 if (opresult && opresult != leader)
1588 tree name = get_representative_for (opresult);
1589 changed |= name != op[n];
1590 op[n] = name;
1592 else if (!opresult)
1593 break;
1595 if (n != 3)
1597 newoperands.release ();
1598 return NULL;
1600 if (!changed)
1601 continue;
1602 if (!newoperands.exists ())
1603 newoperands = operands.copy ();
1604 /* We may have changed from an SSA_NAME to a constant */
1605 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1606 newop.opcode = TREE_CODE (op[0]);
1607 newop.type = type;
1608 newop.op0 = op[0];
1609 newop.op1 = op[1];
1610 newop.op2 = op[2];
1611 newoperands[i] = newop;
1613 gcc_checking_assert (i == operands.length ());
1615 if (vuse)
1617 newvuse = translate_vuse_through_block (newoperands.exists ()
1618 ? newoperands : operands,
1619 ref->set, ref->type,
1620 vuse, phiblock, pred,
1621 &same_valid);
1622 if (newvuse == NULL_TREE)
1624 newoperands.release ();
1625 return NULL;
1629 if (changed || newvuse != vuse)
1631 unsigned int new_val_id;
1632 pre_expr constant;
1634 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1635 ref->type,
1636 newoperands.exists ()
1637 ? newoperands : operands,
1638 &newref, VN_WALK);
1639 if (result)
1640 newoperands.release ();
1642 /* We can always insert constants, so if we have a partial
1643 redundant constant load of another type try to translate it
1644 to a constant of appropriate type. */
1645 if (result && is_gimple_min_invariant (result))
1647 tree tem = result;
1648 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1650 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1651 if (tem && !is_gimple_min_invariant (tem))
1652 tem = NULL_TREE;
1654 if (tem)
1655 return get_or_alloc_expr_for_constant (tem);
1658 /* If we'd have to convert things we would need to validate
1659 if we can insert the translated expression. So fail
1660 here for now - we cannot insert an alias with a different
1661 type in the VN tables either, as that would assert. */
1662 if (result
1663 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1664 return NULL;
1665 else if (!result && newref
1666 && !useless_type_conversion_p (ref->type, newref->type))
1668 newoperands.release ();
1669 return NULL;
1672 expr = pre_expr_pool.allocate ();
1673 expr->kind = REFERENCE;
1674 expr->id = 0;
1676 if (newref)
1678 PRE_EXPR_REFERENCE (expr) = newref;
1679 constant = fully_constant_expression (expr);
1680 if (constant != expr)
1681 return constant;
1683 new_val_id = newref->value_id;
1684 get_or_alloc_expression_id (expr);
1686 else
1688 if (changed || !same_valid)
1690 new_val_id = get_next_value_id ();
1691 value_expressions.safe_grow_cleared
1692 (get_max_value_id () + 1);
1694 else
1695 new_val_id = ref->value_id;
1696 if (!newoperands.exists ())
1697 newoperands = operands.copy ();
1698 newref = vn_reference_insert_pieces (newvuse, ref->set,
1699 ref->type,
1700 newoperands,
1701 result, new_val_id);
1702 newoperands = vNULL;
1703 PRE_EXPR_REFERENCE (expr) = newref;
1704 constant = fully_constant_expression (expr);
1705 if (constant != expr)
1706 return constant;
1707 get_or_alloc_expression_id (expr);
1709 add_to_value (new_val_id, expr);
1711 newoperands.release ();
1712 return expr;
1714 break;
1716 case NAME:
1718 tree name = PRE_EXPR_NAME (expr);
1719 gimple *def_stmt = SSA_NAME_DEF_STMT (name);
1720 /* If the SSA name is defined by a PHI node in this block,
1721 translate it. */
1722 if (gimple_code (def_stmt) == GIMPLE_PHI
1723 && gimple_bb (def_stmt) == phiblock)
1725 edge e = find_edge (pred, gimple_bb (def_stmt));
1726 tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1728 /* Handle constant. */
1729 if (is_gimple_min_invariant (def))
1730 return get_or_alloc_expr_for_constant (def);
1732 return get_or_alloc_expr_for_name (def);
1734 /* Otherwise return it unchanged - it will get removed if its
1735 value is not available in PREDs AVAIL_OUT set of expressions
1736 by the subtraction of TMP_GEN. */
1737 return expr;
1740 default:
1741 gcc_unreachable ();
1745 /* Wrapper around phi_translate_1 providing caching functionality. */
1747 static pre_expr
1748 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1749 basic_block pred, basic_block phiblock)
1751 expr_pred_trans_t slot = NULL;
1752 pre_expr phitrans;
1754 if (!expr)
1755 return NULL;
1757 /* Constants contain no values that need translation. */
1758 if (expr->kind == CONSTANT)
1759 return expr;
1761 if (value_id_constant_p (get_expr_value_id (expr)))
1762 return expr;
1764 /* Don't add translations of NAMEs as those are cheap to translate. */
1765 if (expr->kind != NAME)
1767 if (phi_trans_add (&slot, expr, pred))
1768 return slot->v;
1769 /* Store NULL for the value we want to return in the case of
1770 recursing. */
1771 slot->v = NULL;
1774 /* Translate. */
1775 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1777 if (slot)
1779 if (phitrans)
1780 slot->v = phitrans;
1781 else
1782 /* Remove failed translations again, they cause insert
1783 iteration to not pick up new opportunities reliably. */
1784 phi_translate_table->remove_elt_with_hash (slot, slot->hashcode);
1787 return phitrans;
1791 /* For each expression in SET, translate the values through phi nodes
1792 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1793 expressions in DEST. */
1795 static void
1796 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1797 basic_block phiblock)
1799 vec<pre_expr> exprs;
1800 pre_expr expr;
1801 int i;
1803 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1805 bitmap_set_copy (dest, set);
1806 return;
1809 exprs = sorted_array_from_bitmap_set (set);
1810 FOR_EACH_VEC_ELT (exprs, i, expr)
1812 pre_expr translated;
1813 translated = phi_translate (expr, set, NULL, pred, phiblock);
1814 if (!translated)
1815 continue;
1817 /* We might end up with multiple expressions from SET being
1818 translated to the same value. In this case we do not want
1819 to retain the NARY or REFERENCE expression but prefer a NAME
1820 which would be the leader. */
1821 if (translated->kind == NAME)
1822 bitmap_value_replace_in_set (dest, translated);
1823 else
1824 bitmap_value_insert_into_set (dest, translated);
1826 exprs.release ();
1829 /* Find the leader for a value (i.e., the name representing that
1830 value) in a given set, and return it. Return NULL if no leader
1831 is found. */
1833 static pre_expr
1834 bitmap_find_leader (bitmap_set_t set, unsigned int val)
1836 if (value_id_constant_p (val))
1838 unsigned int i;
1839 bitmap_iterator bi;
1840 bitmap exprset = value_expressions[val];
1842 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1844 pre_expr expr = expression_for_id (i);
1845 if (expr->kind == CONSTANT)
1846 return expr;
1849 if (bitmap_set_contains_value (set, val))
1851 /* Rather than walk the entire bitmap of expressions, and see
1852 whether any of them has the value we are looking for, we look
1853 at the reverse mapping, which tells us the set of expressions
1854 that have a given value (IE value->expressions with that
1855 value) and see if any of those expressions are in our set.
1856 The number of expressions per value is usually significantly
1857 less than the number of expressions in the set. In fact, for
1858 large testcases, doing it this way is roughly 5-10x faster
1859 than walking the bitmap.
1860 If this is somehow a significant lose for some cases, we can
1861 choose which set to walk based on which set is smaller. */
1862 unsigned int i;
1863 bitmap_iterator bi;
1864 bitmap exprset = value_expressions[val];
1866 EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1867 return expression_for_id (i);
1869 return NULL;
1872 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1873 BLOCK by seeing if it is not killed in the block. Note that we are
1874 only determining whether there is a store that kills it. Because
1875 of the order in which clean iterates over values, we are guaranteed
1876 that altered operands will have caused us to be eliminated from the
1877 ANTIC_IN set already. */
1879 static bool
1880 value_dies_in_block_x (pre_expr expr, basic_block block)
1882 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1883 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1884 gimple *def;
1885 gimple_stmt_iterator gsi;
1886 unsigned id = get_expression_id (expr);
1887 bool res = false;
1888 ao_ref ref;
1890 if (!vuse)
1891 return false;
1893 /* Lookup a previously calculated result. */
1894 if (EXPR_DIES (block)
1895 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1896 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1898 /* A memory expression {e, VUSE} dies in the block if there is a
1899 statement that may clobber e. If, starting statement walk from the
1900 top of the basic block, a statement uses VUSE there can be no kill
1901 inbetween that use and the original statement that loaded {e, VUSE},
1902 so we can stop walking. */
1903 ref.base = NULL_TREE;
1904 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1906 tree def_vuse, def_vdef;
1907 def = gsi_stmt (gsi);
1908 def_vuse = gimple_vuse (def);
1909 def_vdef = gimple_vdef (def);
1911 /* Not a memory statement. */
1912 if (!def_vuse)
1913 continue;
1915 /* Not a may-def. */
1916 if (!def_vdef)
1918 /* A load with the same VUSE, we're done. */
1919 if (def_vuse == vuse)
1920 break;
1922 continue;
1925 /* Init ref only if we really need it. */
1926 if (ref.base == NULL_TREE
1927 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1928 refx->operands))
1930 res = true;
1931 break;
1933 /* If the statement may clobber expr, it dies. */
1934 if (stmt_may_clobber_ref_p_1 (def, &ref))
1936 res = true;
1937 break;
1941 /* Remember the result. */
1942 if (!EXPR_DIES (block))
1943 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1944 bitmap_set_bit (EXPR_DIES (block), id * 2);
1945 if (res)
1946 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1948 return res;
1952 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1953 contains its value-id. */
1955 static bool
1956 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1958 if (op && TREE_CODE (op) == SSA_NAME)
1960 unsigned int value_id = VN_INFO (op)->value_id;
1961 if (!(bitmap_set_contains_value (set1, value_id)
1962 || (set2 && bitmap_set_contains_value (set2, value_id))))
1963 return false;
1965 return true;
1968 /* Determine if the expression EXPR is valid in SET1 U SET2.
1969 ONLY SET2 CAN BE NULL.
1970 This means that we have a leader for each part of the expression
1971 (if it consists of values), or the expression is an SSA_NAME.
1972 For loads/calls, we also see if the vuse is killed in this block. */
1974 static bool
1975 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr)
1977 switch (expr->kind)
1979 case NAME:
1980 /* By construction all NAMEs are available. Non-available
1981 NAMEs are removed by subtracting TMP_GEN from the sets. */
1982 return true;
1983 case NARY:
1985 unsigned int i;
1986 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1987 for (i = 0; i < nary->length; i++)
1988 if (!op_valid_in_sets (set1, set2, nary->op[i]))
1989 return false;
1990 return true;
1992 break;
1993 case REFERENCE:
1995 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1996 vn_reference_op_t vro;
1997 unsigned int i;
1999 FOR_EACH_VEC_ELT (ref->operands, i, vro)
2001 if (!op_valid_in_sets (set1, set2, vro->op0)
2002 || !op_valid_in_sets (set1, set2, vro->op1)
2003 || !op_valid_in_sets (set1, set2, vro->op2))
2004 return false;
2006 return true;
2008 default:
2009 gcc_unreachable ();
2013 /* Clean the set of expressions that are no longer valid in SET1 or
2014 SET2. This means expressions that are made up of values we have no
2015 leaders for in SET1 or SET2. This version is used for partial
2016 anticipation, which means it is not valid in either ANTIC_IN or
2017 PA_IN. */
2019 static void
2020 dependent_clean (bitmap_set_t set1, bitmap_set_t set2)
2022 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
2023 pre_expr expr;
2024 int i;
2026 FOR_EACH_VEC_ELT (exprs, i, expr)
2028 if (!valid_in_sets (set1, set2, expr))
2029 bitmap_remove_from_set (set1, expr);
2031 exprs.release ();
2034 /* Clean the set of expressions that are no longer valid in SET. This
2035 means expressions that are made up of values we have no leaders for
2036 in SET. */
2038 static void
2039 clean (bitmap_set_t set)
2041 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set);
2042 pre_expr expr;
2043 int i;
2045 FOR_EACH_VEC_ELT (exprs, i, expr)
2047 if (!valid_in_sets (set, NULL, expr))
2048 bitmap_remove_from_set (set, expr);
2050 exprs.release ();
2053 /* Clean the set of expressions that are no longer valid in SET because
2054 they are clobbered in BLOCK or because they trap and may not be executed. */
2056 static void
2057 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2059 bitmap_iterator bi;
2060 unsigned i;
2061 pre_expr to_remove = NULL;
2063 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2065 /* Remove queued expr. */
2066 if (to_remove)
2068 bitmap_remove_from_set (set, to_remove);
2069 to_remove = NULL;
2072 pre_expr expr = expression_for_id (i);
2073 if (expr->kind == REFERENCE)
2075 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2076 if (ref->vuse)
2078 gimple *def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2079 if (!gimple_nop_p (def_stmt)
2080 && ((gimple_bb (def_stmt) != block
2081 && !dominated_by_p (CDI_DOMINATORS,
2082 block, gimple_bb (def_stmt)))
2083 || (gimple_bb (def_stmt) == block
2084 && value_dies_in_block_x (expr, block))))
2085 to_remove = expr;
2088 else if (expr->kind == NARY)
2090 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2091 /* If the NARY may trap make sure the block does not contain
2092 a possible exit point.
2093 ??? This is overly conservative if we translate AVAIL_OUT
2094 as the available expression might be after the exit point. */
2095 if (BB_MAY_NOTRETURN (block)
2096 && vn_nary_may_trap (nary))
2097 to_remove = expr;
2101 /* Remove queued expr. */
2102 if (to_remove)
2103 bitmap_remove_from_set (set, to_remove);
2106 static sbitmap has_abnormal_preds;
2108 /* Compute the ANTIC set for BLOCK.
2110 If succs(BLOCK) > 1 then
2111 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2112 else if succs(BLOCK) == 1 then
2113 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2115 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2117 Note that clean() is deferred until after the iteration. */
2119 static bool
2120 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2122 bitmap_set_t S, old, ANTIC_OUT;
2123 bitmap_iterator bi;
2124 unsigned int bii;
2125 edge e;
2126 edge_iterator ei;
2128 bool changed = ! BB_VISITED (block);
2129 BB_VISITED (block) = 1;
2130 old = ANTIC_OUT = S = NULL;
2132 /* If any edges from predecessors are abnormal, antic_in is empty,
2133 so do nothing. */
2134 if (block_has_abnormal_pred_edge)
2135 goto maybe_dump_sets;
2137 old = ANTIC_IN (block);
2138 ANTIC_OUT = bitmap_set_new ();
2140 /* If the block has no successors, ANTIC_OUT is empty. */
2141 if (EDGE_COUNT (block->succs) == 0)
2143 /* If we have one successor, we could have some phi nodes to
2144 translate through. */
2145 else if (single_succ_p (block))
2147 basic_block succ_bb = single_succ (block);
2148 gcc_assert (BB_VISITED (succ_bb));
2149 phi_translate_set (ANTIC_OUT, ANTIC_IN (succ_bb), block, succ_bb);
2151 /* If we have multiple successors, we take the intersection of all of
2152 them. Note that in the case of loop exit phi nodes, we may have
2153 phis to translate through. */
2154 else
2156 size_t i;
2157 basic_block bprime, first = NULL;
2159 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2160 FOR_EACH_EDGE (e, ei, block->succs)
2162 if (!first
2163 && BB_VISITED (e->dest))
2164 first = e->dest;
2165 else if (BB_VISITED (e->dest))
2166 worklist.quick_push (e->dest);
2167 else
2169 /* Unvisited successors get their ANTIC_IN replaced by the
2170 maximal set to arrive at a maximum ANTIC_IN solution.
2171 We can ignore them in the intersection operation and thus
2172 need not explicitely represent that maximum solution. */
2173 if (dump_file && (dump_flags & TDF_DETAILS))
2174 fprintf (dump_file, "ANTIC_IN is MAX on %d->%d\n",
2175 e->src->index, e->dest->index);
2179 /* Of multiple successors we have to have visited one already
2180 which is guaranteed by iteration order. */
2181 gcc_assert (first != NULL);
2183 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2185 FOR_EACH_VEC_ELT (worklist, i, bprime)
2187 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2189 bitmap_set_t tmp = bitmap_set_new ();
2190 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2191 bitmap_set_and (ANTIC_OUT, tmp);
2192 bitmap_set_free (tmp);
2194 else
2195 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2199 /* Prune expressions that are clobbered in block and thus become
2200 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2201 prune_clobbered_mems (ANTIC_OUT, block);
2203 /* Generate ANTIC_OUT - TMP_GEN. */
2204 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2206 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2207 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2208 TMP_GEN (block));
2210 /* Then union in the ANTIC_OUT - TMP_GEN values,
2211 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2212 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2213 bitmap_value_insert_into_set (ANTIC_IN (block),
2214 expression_for_id (bii));
2216 /* clean (ANTIC_IN (block)) is defered to after the iteration converged
2217 because it can cause non-convergence, see for example PR81181. */
2219 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2220 changed = true;
2222 maybe_dump_sets:
2223 if (dump_file && (dump_flags & TDF_DETAILS))
2225 if (ANTIC_OUT)
2226 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2228 if (changed)
2229 fprintf (dump_file, "[changed] ");
2230 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2231 block->index);
2233 if (S)
2234 print_bitmap_set (dump_file, S, "S", block->index);
2236 if (old)
2237 bitmap_set_free (old);
2238 if (S)
2239 bitmap_set_free (S);
2240 if (ANTIC_OUT)
2241 bitmap_set_free (ANTIC_OUT);
2242 return changed;
2245 /* Compute PARTIAL_ANTIC for BLOCK.
2247 If succs(BLOCK) > 1 then
2248 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2249 in ANTIC_OUT for all succ(BLOCK)
2250 else if succs(BLOCK) == 1 then
2251 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2253 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2254 - ANTIC_IN[BLOCK])
2257 static void
2258 compute_partial_antic_aux (basic_block block,
2259 bool block_has_abnormal_pred_edge)
2261 bitmap_set_t old_PA_IN;
2262 bitmap_set_t PA_OUT;
2263 edge e;
2264 edge_iterator ei;
2265 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2267 old_PA_IN = PA_OUT = NULL;
2269 /* If any edges from predecessors are abnormal, antic_in is empty,
2270 so do nothing. */
2271 if (block_has_abnormal_pred_edge)
2272 goto maybe_dump_sets;
2274 /* If there are too many partially anticipatable values in the
2275 block, phi_translate_set can take an exponential time: stop
2276 before the translation starts. */
2277 if (max_pa
2278 && single_succ_p (block)
2279 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2280 goto maybe_dump_sets;
2282 old_PA_IN = PA_IN (block);
2283 PA_OUT = bitmap_set_new ();
2285 /* If the block has no successors, ANTIC_OUT is empty. */
2286 if (EDGE_COUNT (block->succs) == 0)
2288 /* If we have one successor, we could have some phi nodes to
2289 translate through. Note that we can't phi translate across DFS
2290 back edges in partial antic, because it uses a union operation on
2291 the successors. For recurrences like IV's, we will end up
2292 generating a new value in the set on each go around (i + 3 (VH.1)
2293 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2294 else if (single_succ_p (block))
2296 basic_block succ = single_succ (block);
2297 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2298 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2300 /* If we have multiple successors, we take the union of all of
2301 them. */
2302 else
2304 size_t i;
2305 basic_block bprime;
2307 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2308 FOR_EACH_EDGE (e, ei, block->succs)
2310 if (e->flags & EDGE_DFS_BACK)
2311 continue;
2312 worklist.quick_push (e->dest);
2314 if (worklist.length () > 0)
2316 FOR_EACH_VEC_ELT (worklist, i, bprime)
2318 unsigned int i;
2319 bitmap_iterator bi;
2321 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2322 bitmap_value_insert_into_set (PA_OUT,
2323 expression_for_id (i));
2324 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2326 bitmap_set_t pa_in = bitmap_set_new ();
2327 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2328 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2329 bitmap_value_insert_into_set (PA_OUT,
2330 expression_for_id (i));
2331 bitmap_set_free (pa_in);
2333 else
2334 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2335 bitmap_value_insert_into_set (PA_OUT,
2336 expression_for_id (i));
2341 /* Prune expressions that are clobbered in block and thus become
2342 invalid if translated from PA_OUT to PA_IN. */
2343 prune_clobbered_mems (PA_OUT, block);
2345 /* PA_IN starts with PA_OUT - TMP_GEN.
2346 Then we subtract things from ANTIC_IN. */
2347 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2349 /* For partial antic, we want to put back in the phi results, since
2350 we will properly avoid making them partially antic over backedges. */
2351 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2352 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2354 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2355 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2357 dependent_clean (PA_IN (block), ANTIC_IN (block));
2359 maybe_dump_sets:
2360 if (dump_file && (dump_flags & TDF_DETAILS))
2362 if (PA_OUT)
2363 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2365 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2367 if (old_PA_IN)
2368 bitmap_set_free (old_PA_IN);
2369 if (PA_OUT)
2370 bitmap_set_free (PA_OUT);
2373 /* Compute ANTIC and partial ANTIC sets. */
2375 static void
2376 compute_antic (void)
2378 bool changed = true;
2379 int num_iterations = 0;
2380 basic_block block;
2381 int i;
2382 edge_iterator ei;
2383 edge e;
2385 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2386 We pre-build the map of blocks with incoming abnormal edges here. */
2387 has_abnormal_preds = sbitmap_alloc (last_basic_block_for_fn (cfun));
2388 bitmap_clear (has_abnormal_preds);
2390 FOR_ALL_BB_FN (block, cfun)
2392 BB_VISITED (block) = 0;
2394 FOR_EACH_EDGE (e, ei, block->preds)
2395 if (e->flags & EDGE_ABNORMAL)
2397 bitmap_set_bit (has_abnormal_preds, block->index);
2398 break;
2401 /* While we are here, give empty ANTIC_IN sets to each block. */
2402 ANTIC_IN (block) = bitmap_set_new ();
2403 if (do_partial_partial)
2404 PA_IN (block) = bitmap_set_new ();
2407 /* At the exit block we anticipate nothing. */
2408 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun)) = 1;
2410 /* For ANTIC computation we need a postorder that also guarantees that
2411 a block with a single successor is visited after its successor.
2412 RPO on the inverted CFG has this property. */
2413 auto_vec<int, 20> postorder;
2414 inverted_post_order_compute (&postorder);
2416 auto_sbitmap worklist (last_basic_block_for_fn (cfun) + 1);
2417 bitmap_clear (worklist);
2418 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
2419 bitmap_set_bit (worklist, e->src->index);
2420 while (changed)
2422 if (dump_file && (dump_flags & TDF_DETAILS))
2423 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2424 /* ??? We need to clear our PHI translation cache here as the
2425 ANTIC sets shrink and we restrict valid translations to
2426 those having operands with leaders in ANTIC. Same below
2427 for PA ANTIC computation. */
2428 num_iterations++;
2429 changed = false;
2430 for (i = postorder.length () - 1; i >= 0; i--)
2432 if (bitmap_bit_p (worklist, postorder[i]))
2434 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2435 bitmap_clear_bit (worklist, block->index);
2436 if (compute_antic_aux (block,
2437 bitmap_bit_p (has_abnormal_preds,
2438 block->index)))
2440 FOR_EACH_EDGE (e, ei, block->preds)
2441 bitmap_set_bit (worklist, e->src->index);
2442 changed = true;
2446 /* Theoretically possible, but *highly* unlikely. */
2447 gcc_checking_assert (num_iterations < 500);
2450 /* We have to clean after the dataflow problem converged as cleaning
2451 can cause non-convergence because it is based on expressions
2452 rather than values. */
2453 FOR_EACH_BB_FN (block, cfun)
2454 clean (ANTIC_IN (block));
2456 statistics_histogram_event (cfun, "compute_antic iterations",
2457 num_iterations);
2459 if (do_partial_partial)
2461 /* For partial antic we ignore backedges and thus we do not need
2462 to perform any iteration when we process blocks in postorder. */
2463 int postorder_num
2464 = pre_and_rev_post_order_compute (NULL, postorder.address (), false);
2465 for (i = postorder_num - 1 ; i >= 0; i--)
2467 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2468 compute_partial_antic_aux (block,
2469 bitmap_bit_p (has_abnormal_preds,
2470 block->index));
2474 sbitmap_free (has_abnormal_preds);
2478 /* Inserted expressions are placed onto this worklist, which is used
2479 for performing quick dead code elimination of insertions we made
2480 that didn't turn out to be necessary. */
2481 static bitmap inserted_exprs;
2483 /* The actual worker for create_component_ref_by_pieces. */
2485 static tree
2486 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2487 unsigned int *operand, gimple_seq *stmts)
2489 vn_reference_op_t currop = &ref->operands[*operand];
2490 tree genop;
2491 ++*operand;
2492 switch (currop->opcode)
2494 case CALL_EXPR:
2495 gcc_unreachable ();
2497 case MEM_REF:
2499 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2500 stmts);
2501 if (!baseop)
2502 return NULL_TREE;
2503 tree offset = currop->op0;
2504 if (TREE_CODE (baseop) == ADDR_EXPR
2505 && handled_component_p (TREE_OPERAND (baseop, 0)))
2507 HOST_WIDE_INT off;
2508 tree base;
2509 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2510 &off);
2511 gcc_assert (base);
2512 offset = int_const_binop (PLUS_EXPR, offset,
2513 build_int_cst (TREE_TYPE (offset),
2514 off));
2515 baseop = build_fold_addr_expr (base);
2517 genop = build2 (MEM_REF, currop->type, baseop, offset);
2518 MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2519 MR_DEPENDENCE_BASE (genop) = currop->base;
2520 REF_REVERSE_STORAGE_ORDER (genop) = currop->reverse;
2521 return genop;
2524 case TARGET_MEM_REF:
2526 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2527 vn_reference_op_t nextop = &ref->operands[++*operand];
2528 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2529 stmts);
2530 if (!baseop)
2531 return NULL_TREE;
2532 if (currop->op0)
2534 genop0 = find_or_generate_expression (block, currop->op0, stmts);
2535 if (!genop0)
2536 return NULL_TREE;
2538 if (nextop->op0)
2540 genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2541 if (!genop1)
2542 return NULL_TREE;
2544 genop = build5 (TARGET_MEM_REF, currop->type,
2545 baseop, currop->op2, genop0, currop->op1, genop1);
2547 MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2548 MR_DEPENDENCE_BASE (genop) = currop->base;
2549 return genop;
2552 case ADDR_EXPR:
2553 if (currop->op0)
2555 gcc_assert (is_gimple_min_invariant (currop->op0));
2556 return currop->op0;
2558 /* Fallthrough. */
2559 case REALPART_EXPR:
2560 case IMAGPART_EXPR:
2561 case VIEW_CONVERT_EXPR:
2563 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2564 stmts);
2565 if (!genop0)
2566 return NULL_TREE;
2567 return fold_build1 (currop->opcode, currop->type, genop0);
2570 case WITH_SIZE_EXPR:
2572 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2573 stmts);
2574 if (!genop0)
2575 return NULL_TREE;
2576 tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2577 if (!genop1)
2578 return NULL_TREE;
2579 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2582 case BIT_FIELD_REF:
2584 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2585 stmts);
2586 if (!genop0)
2587 return NULL_TREE;
2588 tree op1 = currop->op0;
2589 tree op2 = currop->op1;
2590 tree t = build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2591 REF_REVERSE_STORAGE_ORDER (t) = currop->reverse;
2592 return fold (t);
2595 /* For array ref vn_reference_op's, operand 1 of the array ref
2596 is op0 of the reference op and operand 3 of the array ref is
2597 op1. */
2598 case ARRAY_RANGE_REF:
2599 case ARRAY_REF:
2601 tree genop0;
2602 tree genop1 = currop->op0;
2603 tree genop2 = currop->op1;
2604 tree genop3 = currop->op2;
2605 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2606 stmts);
2607 if (!genop0)
2608 return NULL_TREE;
2609 genop1 = find_or_generate_expression (block, genop1, stmts);
2610 if (!genop1)
2611 return NULL_TREE;
2612 if (genop2)
2614 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2615 /* Drop zero minimum index if redundant. */
2616 if (integer_zerop (genop2)
2617 && (!domain_type
2618 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2619 genop2 = NULL_TREE;
2620 else
2622 genop2 = find_or_generate_expression (block, genop2, stmts);
2623 if (!genop2)
2624 return NULL_TREE;
2627 if (genop3)
2629 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2630 /* We can't always put a size in units of the element alignment
2631 here as the element alignment may be not visible. See
2632 PR43783. Simply drop the element size for constant
2633 sizes. */
2634 if (TREE_CODE (genop3) == INTEGER_CST
2635 && TREE_CODE (TYPE_SIZE_UNIT (elmt_type)) == INTEGER_CST
2636 && wi::eq_p (wi::to_offset (TYPE_SIZE_UNIT (elmt_type)),
2637 (wi::to_offset (genop3)
2638 * vn_ref_op_align_unit (currop))))
2639 genop3 = NULL_TREE;
2640 else
2642 genop3 = find_or_generate_expression (block, genop3, stmts);
2643 if (!genop3)
2644 return NULL_TREE;
2647 return build4 (currop->opcode, currop->type, genop0, genop1,
2648 genop2, genop3);
2650 case COMPONENT_REF:
2652 tree op0;
2653 tree op1;
2654 tree genop2 = currop->op1;
2655 op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2656 if (!op0)
2657 return NULL_TREE;
2658 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2659 op1 = currop->op0;
2660 if (genop2)
2662 genop2 = find_or_generate_expression (block, genop2, stmts);
2663 if (!genop2)
2664 return NULL_TREE;
2666 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2669 case SSA_NAME:
2671 genop = find_or_generate_expression (block, currop->op0, stmts);
2672 return genop;
2674 case STRING_CST:
2675 case INTEGER_CST:
2676 case COMPLEX_CST:
2677 case VECTOR_CST:
2678 case REAL_CST:
2679 case CONSTRUCTOR:
2680 case VAR_DECL:
2681 case PARM_DECL:
2682 case CONST_DECL:
2683 case RESULT_DECL:
2684 case FUNCTION_DECL:
2685 return currop->op0;
2687 default:
2688 gcc_unreachable ();
2692 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2693 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2694 trying to rename aggregates into ssa form directly, which is a no no.
2696 Thus, this routine doesn't create temporaries, it just builds a
2697 single access expression for the array, calling
2698 find_or_generate_expression to build the innermost pieces.
2700 This function is a subroutine of create_expression_by_pieces, and
2701 should not be called on it's own unless you really know what you
2702 are doing. */
2704 static tree
2705 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2706 gimple_seq *stmts)
2708 unsigned int op = 0;
2709 return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2712 /* Find a simple leader for an expression, or generate one using
2713 create_expression_by_pieces from a NARY expression for the value.
2714 BLOCK is the basic_block we are looking for leaders in.
2715 OP is the tree expression to find a leader for or generate.
2716 Returns the leader or NULL_TREE on failure. */
2718 static tree
2719 find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2721 pre_expr expr = get_or_alloc_expr_for (op);
2722 unsigned int lookfor = get_expr_value_id (expr);
2723 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2724 if (leader)
2726 if (leader->kind == NAME)
2727 return PRE_EXPR_NAME (leader);
2728 else if (leader->kind == CONSTANT)
2729 return PRE_EXPR_CONSTANT (leader);
2731 /* Defer. */
2732 return NULL_TREE;
2735 /* It must be a complex expression, so generate it recursively. Note
2736 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2737 where the insert algorithm fails to insert a required expression. */
2738 bitmap exprset = value_expressions[lookfor];
2739 bitmap_iterator bi;
2740 unsigned int i;
2741 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2743 pre_expr temp = expression_for_id (i);
2744 /* We cannot insert random REFERENCE expressions at arbitrary
2745 places. We can insert NARYs which eventually re-materializes
2746 its operand values. */
2747 if (temp->kind == NARY)
2748 return create_expression_by_pieces (block, temp, stmts,
2749 get_expr_type (expr));
2752 /* Defer. */
2753 return NULL_TREE;
2756 /* Create an expression in pieces, so that we can handle very complex
2757 expressions that may be ANTIC, but not necessary GIMPLE.
2758 BLOCK is the basic block the expression will be inserted into,
2759 EXPR is the expression to insert (in value form)
2760 STMTS is a statement list to append the necessary insertions into.
2762 This function will die if we hit some value that shouldn't be
2763 ANTIC but is (IE there is no leader for it, or its components).
2764 The function returns NULL_TREE in case a different antic expression
2765 has to be inserted first.
2766 This function may also generate expressions that are themselves
2767 partially or fully redundant. Those that are will be either made
2768 fully redundant during the next iteration of insert (for partially
2769 redundant ones), or eliminated by eliminate (for fully redundant
2770 ones). */
2772 static tree
2773 create_expression_by_pieces (basic_block block, pre_expr expr,
2774 gimple_seq *stmts, tree type)
2776 tree name;
2777 tree folded;
2778 gimple_seq forced_stmts = NULL;
2779 unsigned int value_id;
2780 gimple_stmt_iterator gsi;
2781 tree exprtype = type ? type : get_expr_type (expr);
2782 pre_expr nameexpr;
2783 gassign *newstmt;
2785 switch (expr->kind)
2787 /* We may hit the NAME/CONSTANT case if we have to convert types
2788 that value numbering saw through. */
2789 case NAME:
2790 folded = PRE_EXPR_NAME (expr);
2791 if (useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2792 return folded;
2793 break;
2794 case CONSTANT:
2796 folded = PRE_EXPR_CONSTANT (expr);
2797 tree tem = fold_convert (exprtype, folded);
2798 if (is_gimple_min_invariant (tem))
2799 return tem;
2800 break;
2802 case REFERENCE:
2803 if (PRE_EXPR_REFERENCE (expr)->operands[0].opcode == CALL_EXPR)
2805 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2806 unsigned int operand = 1;
2807 vn_reference_op_t currop = &ref->operands[0];
2808 tree sc = NULL_TREE;
2809 tree fn;
2810 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2811 fn = currop->op0;
2812 else
2813 fn = find_or_generate_expression (block, currop->op0, stmts);
2814 if (!fn)
2815 return NULL_TREE;
2816 if (currop->op1)
2818 sc = find_or_generate_expression (block, currop->op1, stmts);
2819 if (!sc)
2820 return NULL_TREE;
2822 auto_vec<tree> args (ref->operands.length () - 1);
2823 while (operand < ref->operands.length ())
2825 tree arg = create_component_ref_by_pieces_1 (block, ref,
2826 &operand, stmts);
2827 if (!arg)
2828 return NULL_TREE;
2829 args.quick_push (arg);
2831 gcall *call
2832 = gimple_build_call_vec ((TREE_CODE (fn) == FUNCTION_DECL
2833 ? build_fold_addr_expr (fn) : fn), args);
2834 gimple_call_set_with_bounds (call, currop->with_bounds);
2835 if (sc)
2836 gimple_call_set_chain (call, sc);
2837 tree forcedname = make_ssa_name (currop->type);
2838 gimple_call_set_lhs (call, forcedname);
2839 gimple_set_vuse (call, BB_LIVE_VOP_ON_EXIT (block));
2840 gimple_seq_add_stmt_without_update (&forced_stmts, call);
2841 folded = forcedname;
2843 else
2845 folded = create_component_ref_by_pieces (block,
2846 PRE_EXPR_REFERENCE (expr),
2847 stmts);
2848 if (!folded)
2849 return NULL_TREE;
2850 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2851 newstmt = gimple_build_assign (name, folded);
2852 gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2853 gimple_set_vuse (newstmt, BB_LIVE_VOP_ON_EXIT (block));
2854 folded = name;
2856 break;
2857 case NARY:
2859 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2860 tree *genop = XALLOCAVEC (tree, nary->length);
2861 unsigned i;
2862 for (i = 0; i < nary->length; ++i)
2864 genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2865 if (!genop[i])
2866 return NULL_TREE;
2867 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2868 may have conversions stripped. */
2869 if (nary->opcode == POINTER_PLUS_EXPR)
2871 if (i == 0)
2872 genop[i] = gimple_convert (&forced_stmts,
2873 nary->type, genop[i]);
2874 else if (i == 1)
2875 genop[i] = gimple_convert (&forced_stmts,
2876 sizetype, genop[i]);
2878 else
2879 genop[i] = gimple_convert (&forced_stmts,
2880 TREE_TYPE (nary->op[i]), genop[i]);
2882 if (nary->opcode == CONSTRUCTOR)
2884 vec<constructor_elt, va_gc> *elts = NULL;
2885 for (i = 0; i < nary->length; ++i)
2886 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2887 folded = build_constructor (nary->type, elts);
2888 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2889 newstmt = gimple_build_assign (name, folded);
2890 gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2891 folded = name;
2893 else
2895 switch (nary->length)
2897 case 1:
2898 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2899 genop[0]);
2900 break;
2901 case 2:
2902 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2903 genop[0], genop[1]);
2904 break;
2905 case 3:
2906 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2907 genop[0], genop[1], genop[2]);
2908 break;
2909 default:
2910 gcc_unreachable ();
2914 break;
2915 default:
2916 gcc_unreachable ();
2919 folded = gimple_convert (&forced_stmts, exprtype, folded);
2921 /* If there is nothing to insert, return the simplified result. */
2922 if (gimple_seq_empty_p (forced_stmts))
2923 return folded;
2924 /* If we simplified to a constant return it and discard eventually
2925 built stmts. */
2926 if (is_gimple_min_invariant (folded))
2928 gimple_seq_discard (forced_stmts);
2929 return folded;
2931 /* Likewise if we simplified to sth not queued for insertion. */
2932 bool found = false;
2933 gsi = gsi_last (forced_stmts);
2934 for (; !gsi_end_p (gsi); gsi_prev (&gsi))
2936 gimple *stmt = gsi_stmt (gsi);
2937 tree forcedname = gimple_get_lhs (stmt);
2938 if (forcedname == folded)
2940 found = true;
2941 break;
2944 if (! found)
2946 gimple_seq_discard (forced_stmts);
2947 return folded;
2949 gcc_assert (TREE_CODE (folded) == SSA_NAME);
2951 /* If we have any intermediate expressions to the value sets, add them
2952 to the value sets and chain them in the instruction stream. */
2953 if (forced_stmts)
2955 gsi = gsi_start (forced_stmts);
2956 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2958 gimple *stmt = gsi_stmt (gsi);
2959 tree forcedname = gimple_get_lhs (stmt);
2960 pre_expr nameexpr;
2962 if (forcedname != folded)
2964 VN_INFO_GET (forcedname)->valnum = forcedname;
2965 VN_INFO (forcedname)->value_id = get_next_value_id ();
2966 nameexpr = get_or_alloc_expr_for_name (forcedname);
2967 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2968 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2969 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2972 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
2974 gimple_seq_add_seq (stmts, forced_stmts);
2977 name = folded;
2979 /* Fold the last statement. */
2980 gsi = gsi_last (*stmts);
2981 if (fold_stmt_inplace (&gsi))
2982 update_stmt (gsi_stmt (gsi));
2984 /* Add a value number to the temporary.
2985 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2986 we are creating the expression by pieces, and this particular piece of
2987 the expression may have been represented. There is no harm in replacing
2988 here. */
2989 value_id = get_expr_value_id (expr);
2990 VN_INFO_GET (name)->value_id = value_id;
2991 VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id);
2992 if (VN_INFO (name)->valnum == NULL_TREE)
2993 VN_INFO (name)->valnum = name;
2994 gcc_assert (VN_INFO (name)->valnum != NULL_TREE);
2995 nameexpr = get_or_alloc_expr_for_name (name);
2996 add_to_value (value_id, nameexpr);
2997 if (NEW_SETS (block))
2998 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2999 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3001 pre_stats.insertions++;
3002 if (dump_file && (dump_flags & TDF_DETAILS))
3004 fprintf (dump_file, "Inserted ");
3005 print_gimple_stmt (dump_file, gsi_stmt (gsi_last (*stmts)), 0);
3006 fprintf (dump_file, " in predecessor %d (%04d)\n",
3007 block->index, value_id);
3010 return name;
3014 /* Insert the to-be-made-available values of expression EXPRNUM for each
3015 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3016 merge the result with a phi node, given the same value number as
3017 NODE. Return true if we have inserted new stuff. */
3019 static bool
3020 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3021 vec<pre_expr> avail)
3023 pre_expr expr = expression_for_id (exprnum);
3024 pre_expr newphi;
3025 unsigned int val = get_expr_value_id (expr);
3026 edge pred;
3027 bool insertions = false;
3028 bool nophi = false;
3029 basic_block bprime;
3030 pre_expr eprime;
3031 edge_iterator ei;
3032 tree type = get_expr_type (expr);
3033 tree temp;
3034 gphi *phi;
3036 /* Make sure we aren't creating an induction variable. */
3037 if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
3039 bool firstinsideloop = false;
3040 bool secondinsideloop = false;
3041 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3042 EDGE_PRED (block, 0)->src);
3043 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3044 EDGE_PRED (block, 1)->src);
3045 /* Induction variables only have one edge inside the loop. */
3046 if ((firstinsideloop ^ secondinsideloop)
3047 && expr->kind != REFERENCE)
3049 if (dump_file && (dump_flags & TDF_DETAILS))
3050 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3051 nophi = true;
3055 /* Make the necessary insertions. */
3056 FOR_EACH_EDGE (pred, ei, block->preds)
3058 gimple_seq stmts = NULL;
3059 tree builtexpr;
3060 bprime = pred->src;
3061 eprime = avail[pred->dest_idx];
3062 builtexpr = create_expression_by_pieces (bprime, eprime,
3063 &stmts, type);
3064 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3065 if (!gimple_seq_empty_p (stmts))
3067 gsi_insert_seq_on_edge (pred, stmts);
3068 insertions = true;
3070 if (!builtexpr)
3072 /* We cannot insert a PHI node if we failed to insert
3073 on one edge. */
3074 nophi = true;
3075 continue;
3077 if (is_gimple_min_invariant (builtexpr))
3078 avail[pred->dest_idx] = get_or_alloc_expr_for_constant (builtexpr);
3079 else
3080 avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3082 /* If we didn't want a phi node, and we made insertions, we still have
3083 inserted new stuff, and thus return true. If we didn't want a phi node,
3084 and didn't make insertions, we haven't added anything new, so return
3085 false. */
3086 if (nophi && insertions)
3087 return true;
3088 else if (nophi && !insertions)
3089 return false;
3091 /* Now build a phi for the new variable. */
3092 temp = make_temp_ssa_name (type, NULL, "prephitmp");
3093 phi = create_phi_node (temp, block);
3095 VN_INFO_GET (temp)->value_id = val;
3096 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3097 if (VN_INFO (temp)->valnum == NULL_TREE)
3098 VN_INFO (temp)->valnum = temp;
3099 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3100 FOR_EACH_EDGE (pred, ei, block->preds)
3102 pre_expr ae = avail[pred->dest_idx];
3103 gcc_assert (get_expr_type (ae) == type
3104 || useless_type_conversion_p (type, get_expr_type (ae)));
3105 if (ae->kind == CONSTANT)
3106 add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3107 pred, UNKNOWN_LOCATION);
3108 else
3109 add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3112 newphi = get_or_alloc_expr_for_name (temp);
3113 add_to_value (val, newphi);
3115 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3116 this insertion, since we test for the existence of this value in PHI_GEN
3117 before proceeding with the partial redundancy checks in insert_aux.
3119 The value may exist in AVAIL_OUT, in particular, it could be represented
3120 by the expression we are trying to eliminate, in which case we want the
3121 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3122 inserted there.
3124 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3125 this block, because if it did, it would have existed in our dominator's
3126 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3129 bitmap_insert_into_set (PHI_GEN (block), newphi);
3130 bitmap_value_replace_in_set (AVAIL_OUT (block),
3131 newphi);
3132 bitmap_insert_into_set (NEW_SETS (block),
3133 newphi);
3135 /* If we insert a PHI node for a conversion of another PHI node
3136 in the same basic-block try to preserve range information.
3137 This is important so that followup loop passes receive optimal
3138 number of iteration analysis results. See PR61743. */
3139 if (expr->kind == NARY
3140 && CONVERT_EXPR_CODE_P (expr->u.nary->opcode)
3141 && TREE_CODE (expr->u.nary->op[0]) == SSA_NAME
3142 && gimple_bb (SSA_NAME_DEF_STMT (expr->u.nary->op[0])) == block
3143 && INTEGRAL_TYPE_P (type)
3144 && INTEGRAL_TYPE_P (TREE_TYPE (expr->u.nary->op[0]))
3145 && (TYPE_PRECISION (type)
3146 >= TYPE_PRECISION (TREE_TYPE (expr->u.nary->op[0])))
3147 && SSA_NAME_RANGE_INFO (expr->u.nary->op[0]))
3149 wide_int min, max;
3150 if (get_range_info (expr->u.nary->op[0], &min, &max) == VR_RANGE
3151 && !wi::neg_p (min, SIGNED)
3152 && !wi::neg_p (max, SIGNED))
3153 /* Just handle extension and sign-changes of all-positive ranges. */
3154 set_range_info (temp,
3155 SSA_NAME_RANGE_TYPE (expr->u.nary->op[0]),
3156 wide_int_storage::from (min, TYPE_PRECISION (type),
3157 TYPE_SIGN (type)),
3158 wide_int_storage::from (max, TYPE_PRECISION (type),
3159 TYPE_SIGN (type)));
3162 if (dump_file && (dump_flags & TDF_DETAILS))
3164 fprintf (dump_file, "Created phi ");
3165 print_gimple_stmt (dump_file, phi, 0);
3166 fprintf (dump_file, " in block %d (%04d)\n", block->index, val);
3168 pre_stats.phis++;
3169 return true;
3174 /* Perform insertion of partially redundant or hoistable values.
3175 For BLOCK, do the following:
3176 1. Propagate the NEW_SETS of the dominator into the current block.
3177 If the block has multiple predecessors,
3178 2a. Iterate over the ANTIC expressions for the block to see if
3179 any of them are partially redundant.
3180 2b. If so, insert them into the necessary predecessors to make
3181 the expression fully redundant.
3182 2c. Insert a new PHI merging the values of the predecessors.
3183 2d. Insert the new PHI, and the new expressions, into the
3184 NEW_SETS set.
3185 If the block has multiple successors,
3186 3a. Iterate over the ANTIC values for the block to see if
3187 any of them are good candidates for hoisting.
3188 3b. If so, insert expressions computing the values in BLOCK,
3189 and add the new expressions into the NEW_SETS set.
3190 4. Recursively call ourselves on the dominator children of BLOCK.
3192 Steps 1, 2a, and 4 are done by insert_aux. 2b, 2c and 2d are done by
3193 do_pre_regular_insertion and do_partial_insertion. 3a and 3b are
3194 done in do_hoist_insertion.
3197 static bool
3198 do_pre_regular_insertion (basic_block block, basic_block dom)
3200 bool new_stuff = false;
3201 vec<pre_expr> exprs;
3202 pre_expr expr;
3203 auto_vec<pre_expr> avail;
3204 int i;
3206 exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3207 avail.safe_grow (EDGE_COUNT (block->preds));
3209 FOR_EACH_VEC_ELT (exprs, i, expr)
3211 if (expr->kind == NARY
3212 || expr->kind == REFERENCE)
3214 unsigned int val;
3215 bool by_some = false;
3216 bool cant_insert = false;
3217 bool all_same = true;
3218 pre_expr first_s = NULL;
3219 edge pred;
3220 basic_block bprime;
3221 pre_expr eprime = NULL;
3222 edge_iterator ei;
3223 pre_expr edoubleprime = NULL;
3224 bool do_insertion = false;
3226 val = get_expr_value_id (expr);
3227 if (bitmap_set_contains_value (PHI_GEN (block), val))
3228 continue;
3229 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3231 if (dump_file && (dump_flags & TDF_DETAILS))
3233 fprintf (dump_file, "Found fully redundant value: ");
3234 print_pre_expr (dump_file, expr);
3235 fprintf (dump_file, "\n");
3237 continue;
3240 FOR_EACH_EDGE (pred, ei, block->preds)
3242 unsigned int vprime;
3244 /* We should never run insertion for the exit block
3245 and so not come across fake pred edges. */
3246 gcc_assert (!(pred->flags & EDGE_FAKE));
3247 bprime = pred->src;
3248 /* We are looking at ANTIC_OUT of bprime. */
3249 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3250 bprime, block);
3252 /* eprime will generally only be NULL if the
3253 value of the expression, translated
3254 through the PHI for this predecessor, is
3255 undefined. If that is the case, we can't
3256 make the expression fully redundant,
3257 because its value is undefined along a
3258 predecessor path. We can thus break out
3259 early because it doesn't matter what the
3260 rest of the results are. */
3261 if (eprime == NULL)
3263 avail[pred->dest_idx] = NULL;
3264 cant_insert = true;
3265 break;
3268 vprime = get_expr_value_id (eprime);
3269 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3270 vprime);
3271 if (edoubleprime == NULL)
3273 avail[pred->dest_idx] = eprime;
3274 all_same = false;
3276 else
3278 avail[pred->dest_idx] = edoubleprime;
3279 by_some = true;
3280 /* We want to perform insertions to remove a redundancy on
3281 a path in the CFG we want to optimize for speed. */
3282 if (optimize_edge_for_speed_p (pred))
3283 do_insertion = true;
3284 if (first_s == NULL)
3285 first_s = edoubleprime;
3286 else if (!pre_expr_d::equal (first_s, edoubleprime))
3287 all_same = false;
3290 /* If we can insert it, it's not the same value
3291 already existing along every predecessor, and
3292 it's defined by some predecessor, it is
3293 partially redundant. */
3294 if (!cant_insert && !all_same && by_some)
3296 if (!do_insertion)
3298 if (dump_file && (dump_flags & TDF_DETAILS))
3300 fprintf (dump_file, "Skipping partial redundancy for "
3301 "expression ");
3302 print_pre_expr (dump_file, expr);
3303 fprintf (dump_file, " (%04d), no redundancy on to be "
3304 "optimized for speed edge\n", val);
3307 else if (dbg_cnt (treepre_insert))
3309 if (dump_file && (dump_flags & TDF_DETAILS))
3311 fprintf (dump_file, "Found partial redundancy for "
3312 "expression ");
3313 print_pre_expr (dump_file, expr);
3314 fprintf (dump_file, " (%04d)\n",
3315 get_expr_value_id (expr));
3317 if (insert_into_preds_of_block (block,
3318 get_expression_id (expr),
3319 avail))
3320 new_stuff = true;
3323 /* If all edges produce the same value and that value is
3324 an invariant, then the PHI has the same value on all
3325 edges. Note this. */
3326 else if (!cant_insert && all_same)
3328 gcc_assert (edoubleprime->kind == CONSTANT
3329 || edoubleprime->kind == NAME);
3331 tree temp = make_temp_ssa_name (get_expr_type (expr),
3332 NULL, "pretmp");
3333 gassign *assign
3334 = gimple_build_assign (temp,
3335 edoubleprime->kind == CONSTANT ?
3336 PRE_EXPR_CONSTANT (edoubleprime) :
3337 PRE_EXPR_NAME (edoubleprime));
3338 gimple_stmt_iterator gsi = gsi_after_labels (block);
3339 gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3341 VN_INFO_GET (temp)->value_id = val;
3342 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3343 if (VN_INFO (temp)->valnum == NULL_TREE)
3344 VN_INFO (temp)->valnum = temp;
3345 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3346 pre_expr newe = get_or_alloc_expr_for_name (temp);
3347 add_to_value (val, newe);
3348 bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3349 bitmap_insert_into_set (NEW_SETS (block), newe);
3354 exprs.release ();
3355 return new_stuff;
3359 /* Perform insertion for partially anticipatable expressions. There
3360 is only one case we will perform insertion for these. This case is
3361 if the expression is partially anticipatable, and fully available.
3362 In this case, we know that putting it earlier will enable us to
3363 remove the later computation. */
3365 static bool
3366 do_pre_partial_partial_insertion (basic_block block, basic_block dom)
3368 bool new_stuff = false;
3369 vec<pre_expr> exprs;
3370 pre_expr expr;
3371 auto_vec<pre_expr> avail;
3372 int i;
3374 exprs = sorted_array_from_bitmap_set (PA_IN (block));
3375 avail.safe_grow (EDGE_COUNT (block->preds));
3377 FOR_EACH_VEC_ELT (exprs, i, expr)
3379 if (expr->kind == NARY
3380 || expr->kind == REFERENCE)
3382 unsigned int val;
3383 bool by_all = true;
3384 bool cant_insert = false;
3385 edge pred;
3386 basic_block bprime;
3387 pre_expr eprime = NULL;
3388 edge_iterator ei;
3390 val = get_expr_value_id (expr);
3391 if (bitmap_set_contains_value (PHI_GEN (block), val))
3392 continue;
3393 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3394 continue;
3396 FOR_EACH_EDGE (pred, ei, block->preds)
3398 unsigned int vprime;
3399 pre_expr edoubleprime;
3401 /* We should never run insertion for the exit block
3402 and so not come across fake pred edges. */
3403 gcc_assert (!(pred->flags & EDGE_FAKE));
3404 bprime = pred->src;
3405 eprime = phi_translate (expr, ANTIC_IN (block),
3406 PA_IN (block),
3407 bprime, block);
3409 /* eprime will generally only be NULL if the
3410 value of the expression, translated
3411 through the PHI for this predecessor, is
3412 undefined. If that is the case, we can't
3413 make the expression fully redundant,
3414 because its value is undefined along a
3415 predecessor path. We can thus break out
3416 early because it doesn't matter what the
3417 rest of the results are. */
3418 if (eprime == NULL)
3420 avail[pred->dest_idx] = NULL;
3421 cant_insert = true;
3422 break;
3425 vprime = get_expr_value_id (eprime);
3426 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3427 avail[pred->dest_idx] = edoubleprime;
3428 if (edoubleprime == NULL)
3430 by_all = false;
3431 break;
3435 /* If we can insert it, it's not the same value
3436 already existing along every predecessor, and
3437 it's defined by some predecessor, it is
3438 partially redundant. */
3439 if (!cant_insert && by_all)
3441 edge succ;
3442 bool do_insertion = false;
3444 /* Insert only if we can remove a later expression on a path
3445 that we want to optimize for speed.
3446 The phi node that we will be inserting in BLOCK is not free,
3447 and inserting it for the sake of !optimize_for_speed successor
3448 may cause regressions on the speed path. */
3449 FOR_EACH_EDGE (succ, ei, block->succs)
3451 if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3452 || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3454 if (optimize_edge_for_speed_p (succ))
3455 do_insertion = true;
3459 if (!do_insertion)
3461 if (dump_file && (dump_flags & TDF_DETAILS))
3463 fprintf (dump_file, "Skipping partial partial redundancy "
3464 "for expression ");
3465 print_pre_expr (dump_file, expr);
3466 fprintf (dump_file, " (%04d), not (partially) anticipated "
3467 "on any to be optimized for speed edges\n", val);
3470 else if (dbg_cnt (treepre_insert))
3472 pre_stats.pa_insert++;
3473 if (dump_file && (dump_flags & TDF_DETAILS))
3475 fprintf (dump_file, "Found partial partial redundancy "
3476 "for expression ");
3477 print_pre_expr (dump_file, expr);
3478 fprintf (dump_file, " (%04d)\n",
3479 get_expr_value_id (expr));
3481 if (insert_into_preds_of_block (block,
3482 get_expression_id (expr),
3483 avail))
3484 new_stuff = true;
3490 exprs.release ();
3491 return new_stuff;
3494 /* Insert expressions in BLOCK to compute hoistable values up.
3495 Return TRUE if something was inserted, otherwise return FALSE.
3496 The caller has to make sure that BLOCK has at least two successors. */
3498 static bool
3499 do_hoist_insertion (basic_block block)
3501 edge e;
3502 edge_iterator ei;
3503 bool new_stuff = false;
3504 unsigned i;
3505 gimple_stmt_iterator last;
3507 /* At least two successors, or else... */
3508 gcc_assert (EDGE_COUNT (block->succs) >= 2);
3510 /* Check that all successors of BLOCK are dominated by block.
3511 We could use dominated_by_p() for this, but actually there is a much
3512 quicker check: any successor that is dominated by BLOCK can't have
3513 more than one predecessor edge. */
3514 FOR_EACH_EDGE (e, ei, block->succs)
3515 if (! single_pred_p (e->dest))
3516 return false;
3518 /* Determine the insertion point. If we cannot safely insert before
3519 the last stmt if we'd have to, bail out. */
3520 last = gsi_last_bb (block);
3521 if (!gsi_end_p (last)
3522 && !is_ctrl_stmt (gsi_stmt (last))
3523 && stmt_ends_bb_p (gsi_stmt (last)))
3524 return false;
3526 /* Compute the set of hoistable expressions from ANTIC_IN. First compute
3527 hoistable values. */
3528 bitmap_set hoistable_set;
3530 /* A hoistable value must be in ANTIC_IN(block)
3531 but not in AVAIL_OUT(BLOCK). */
3532 bitmap_initialize (&hoistable_set.values, &grand_bitmap_obstack);
3533 bitmap_and_compl (&hoistable_set.values,
3534 &ANTIC_IN (block)->values, &AVAIL_OUT (block)->values);
3536 /* Short-cut for a common case: hoistable_set is empty. */
3537 if (bitmap_empty_p (&hoistable_set.values))
3538 return false;
3540 /* Compute which of the hoistable values is in AVAIL_OUT of
3541 at least one of the successors of BLOCK. */
3542 bitmap_head availout_in_some;
3543 bitmap_initialize (&availout_in_some, &grand_bitmap_obstack);
3544 FOR_EACH_EDGE (e, ei, block->succs)
3545 /* Do not consider expressions solely because their availability
3546 on loop exits. They'd be ANTIC-IN throughout the whole loop
3547 and thus effectively hoisted across loops by combination of
3548 PRE and hoisting. */
3549 if (! loop_exit_edge_p (block->loop_father, e))
3550 bitmap_ior_and_into (&availout_in_some, &hoistable_set.values,
3551 &AVAIL_OUT (e->dest)->values);
3552 bitmap_clear (&hoistable_set.values);
3554 /* Short-cut for a common case: availout_in_some is empty. */
3555 if (bitmap_empty_p (&availout_in_some))
3556 return false;
3558 /* Hack hoitable_set in-place so we can use sorted_array_from_bitmap_set. */
3559 hoistable_set.values = availout_in_some;
3560 hoistable_set.expressions = ANTIC_IN (block)->expressions;
3562 /* Now finally construct the topological-ordered expression set. */
3563 vec<pre_expr> exprs = sorted_array_from_bitmap_set (&hoistable_set);
3565 bitmap_clear (&hoistable_set.values);
3567 /* If there are candidate values for hoisting, insert expressions
3568 strategically to make the hoistable expressions fully redundant. */
3569 pre_expr expr;
3570 FOR_EACH_VEC_ELT (exprs, i, expr)
3572 /* While we try to sort expressions topologically above the
3573 sorting doesn't work out perfectly. Catch expressions we
3574 already inserted. */
3575 unsigned int value_id = get_expr_value_id (expr);
3576 if (bitmap_set_contains_value (AVAIL_OUT (block), value_id))
3578 if (dump_file && (dump_flags & TDF_DETAILS))
3580 fprintf (dump_file,
3581 "Already inserted expression for ");
3582 print_pre_expr (dump_file, expr);
3583 fprintf (dump_file, " (%04d)\n", value_id);
3585 continue;
3588 /* OK, we should hoist this value. Perform the transformation. */
3589 pre_stats.hoist_insert++;
3590 if (dump_file && (dump_flags & TDF_DETAILS))
3592 fprintf (dump_file,
3593 "Inserting expression in block %d for code hoisting: ",
3594 block->index);
3595 print_pre_expr (dump_file, expr);
3596 fprintf (dump_file, " (%04d)\n", value_id);
3599 gimple_seq stmts = NULL;
3600 tree res = create_expression_by_pieces (block, expr, &stmts,
3601 get_expr_type (expr));
3603 /* Do not return true if expression creation ultimately
3604 did not insert any statements. */
3605 if (gimple_seq_empty_p (stmts))
3606 res = NULL_TREE;
3607 else
3609 if (gsi_end_p (last) || is_ctrl_stmt (gsi_stmt (last)))
3610 gsi_insert_seq_before (&last, stmts, GSI_SAME_STMT);
3611 else
3612 gsi_insert_seq_after (&last, stmts, GSI_NEW_STMT);
3615 /* Make sure to not return true if expression creation ultimately
3616 failed but also make sure to insert any stmts produced as they
3617 are tracked in inserted_exprs. */
3618 if (! res)
3619 continue;
3621 new_stuff = true;
3624 exprs.release ();
3626 return new_stuff;
3629 /* Do a dominator walk on the control flow graph, and insert computations
3630 of values as necessary for PRE and hoisting. */
3632 static bool
3633 insert_aux (basic_block block, bool do_pre, bool do_hoist)
3635 basic_block son;
3636 bool new_stuff = false;
3638 if (block)
3640 basic_block dom;
3641 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3642 if (dom)
3644 unsigned i;
3645 bitmap_iterator bi;
3646 bitmap_set_t newset;
3648 /* First, update the AVAIL_OUT set with anything we may have
3649 inserted higher up in the dominator tree. */
3650 newset = NEW_SETS (dom);
3651 if (newset)
3653 /* Note that we need to value_replace both NEW_SETS, and
3654 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3655 represented by some non-simple expression here that we want
3656 to replace it with. */
3657 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3659 pre_expr expr = expression_for_id (i);
3660 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3661 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3665 /* Insert expressions for partial redundancies. */
3666 if (do_pre && !single_pred_p (block))
3668 new_stuff |= do_pre_regular_insertion (block, dom);
3669 if (do_partial_partial)
3670 new_stuff |= do_pre_partial_partial_insertion (block, dom);
3673 /* Insert expressions for hoisting. */
3674 if (do_hoist && EDGE_COUNT (block->succs) >= 2)
3675 new_stuff |= do_hoist_insertion (block);
3678 for (son = first_dom_son (CDI_DOMINATORS, block);
3679 son;
3680 son = next_dom_son (CDI_DOMINATORS, son))
3682 new_stuff |= insert_aux (son, do_pre, do_hoist);
3685 return new_stuff;
3688 /* Perform insertion of partially redundant and hoistable values. */
3690 static void
3691 insert (void)
3693 bool new_stuff = true;
3694 basic_block bb;
3695 int num_iterations = 0;
3697 FOR_ALL_BB_FN (bb, cfun)
3698 NEW_SETS (bb) = bitmap_set_new ();
3700 while (new_stuff)
3702 num_iterations++;
3703 if (dump_file && dump_flags & TDF_DETAILS)
3704 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3705 new_stuff = insert_aux (ENTRY_BLOCK_PTR_FOR_FN (cfun), flag_tree_pre,
3706 flag_code_hoisting);
3708 /* Clear the NEW sets before the next iteration. We have already
3709 fully propagated its contents. */
3710 if (new_stuff)
3711 FOR_ALL_BB_FN (bb, cfun)
3712 bitmap_set_free (NEW_SETS (bb));
3714 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3718 /* Compute the AVAIL set for all basic blocks.
3720 This function performs value numbering of the statements in each basic
3721 block. The AVAIL sets are built from information we glean while doing
3722 this value numbering, since the AVAIL sets contain only one entry per
3723 value.
3725 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3726 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3728 static void
3729 compute_avail (void)
3732 basic_block block, son;
3733 basic_block *worklist;
3734 size_t sp = 0;
3735 unsigned i;
3736 tree name;
3738 /* We pretend that default definitions are defined in the entry block.
3739 This includes function arguments and the static chain decl. */
3740 FOR_EACH_SSA_NAME (i, name, cfun)
3742 pre_expr e;
3743 if (!SSA_NAME_IS_DEFAULT_DEF (name)
3744 || has_zero_uses (name)
3745 || virtual_operand_p (name))
3746 continue;
3748 e = get_or_alloc_expr_for_name (name);
3749 add_to_value (get_expr_value_id (e), e);
3750 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)), e);
3751 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3755 if (dump_file && (dump_flags & TDF_DETAILS))
3757 print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3758 "tmp_gen", ENTRY_BLOCK);
3759 print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3760 "avail_out", ENTRY_BLOCK);
3763 /* Allocate the worklist. */
3764 worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
3766 /* Seed the algorithm by putting the dominator children of the entry
3767 block on the worklist. */
3768 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR_FOR_FN (cfun));
3769 son;
3770 son = next_dom_son (CDI_DOMINATORS, son))
3771 worklist[sp++] = son;
3773 BB_LIVE_VOP_ON_EXIT (ENTRY_BLOCK_PTR_FOR_FN (cfun))
3774 = ssa_default_def (cfun, gimple_vop (cfun));
3776 /* Loop until the worklist is empty. */
3777 while (sp)
3779 gimple *stmt;
3780 basic_block dom;
3782 /* Pick a block from the worklist. */
3783 block = worklist[--sp];
3785 /* Initially, the set of available values in BLOCK is that of
3786 its immediate dominator. */
3787 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3788 if (dom)
3790 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3791 BB_LIVE_VOP_ON_EXIT (block) = BB_LIVE_VOP_ON_EXIT (dom);
3794 /* Generate values for PHI nodes. */
3795 for (gphi_iterator gsi = gsi_start_phis (block); !gsi_end_p (gsi);
3796 gsi_next (&gsi))
3798 tree result = gimple_phi_result (gsi.phi ());
3800 /* We have no need for virtual phis, as they don't represent
3801 actual computations. */
3802 if (virtual_operand_p (result))
3804 BB_LIVE_VOP_ON_EXIT (block) = result;
3805 continue;
3808 pre_expr e = get_or_alloc_expr_for_name (result);
3809 add_to_value (get_expr_value_id (e), e);
3810 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3811 bitmap_insert_into_set (PHI_GEN (block), e);
3814 BB_MAY_NOTRETURN (block) = 0;
3816 /* Now compute value numbers and populate value sets with all
3817 the expressions computed in BLOCK. */
3818 for (gimple_stmt_iterator gsi = gsi_start_bb (block); !gsi_end_p (gsi);
3819 gsi_next (&gsi))
3821 ssa_op_iter iter;
3822 tree op;
3824 stmt = gsi_stmt (gsi);
3826 /* Cache whether the basic-block has any non-visible side-effect
3827 or control flow.
3828 If this isn't a call or it is the last stmt in the
3829 basic-block then the CFG represents things correctly. */
3830 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3832 /* Non-looping const functions always return normally.
3833 Otherwise the call might not return or have side-effects
3834 that forbids hoisting possibly trapping expressions
3835 before it. */
3836 int flags = gimple_call_flags (stmt);
3837 if (!(flags & ECF_CONST)
3838 || (flags & ECF_LOOPING_CONST_OR_PURE))
3839 BB_MAY_NOTRETURN (block) = 1;
3842 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3844 pre_expr e = get_or_alloc_expr_for_name (op);
3846 add_to_value (get_expr_value_id (e), e);
3847 bitmap_insert_into_set (TMP_GEN (block), e);
3848 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3851 if (gimple_vdef (stmt))
3852 BB_LIVE_VOP_ON_EXIT (block) = gimple_vdef (stmt);
3854 if (gimple_has_side_effects (stmt)
3855 || stmt_could_throw_p (stmt)
3856 || is_gimple_debug (stmt))
3857 continue;
3859 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3861 if (ssa_undefined_value_p (op))
3862 continue;
3863 pre_expr e = get_or_alloc_expr_for_name (op);
3864 bitmap_value_insert_into_set (EXP_GEN (block), e);
3867 switch (gimple_code (stmt))
3869 case GIMPLE_RETURN:
3870 continue;
3872 case GIMPLE_CALL:
3874 vn_reference_t ref;
3875 vn_reference_s ref1;
3876 pre_expr result = NULL;
3878 /* We can value number only calls to real functions. */
3879 if (gimple_call_internal_p (stmt))
3880 continue;
3882 vn_reference_lookup_call (as_a <gcall *> (stmt), &ref, &ref1);
3883 if (!ref)
3884 continue;
3886 /* If the value of the call is not invalidated in
3887 this block until it is computed, add the expression
3888 to EXP_GEN. */
3889 if (!gimple_vuse (stmt)
3890 || gimple_code
3891 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
3892 || gimple_bb (SSA_NAME_DEF_STMT
3893 (gimple_vuse (stmt))) != block)
3895 result = pre_expr_pool.allocate ();
3896 result->kind = REFERENCE;
3897 result->id = 0;
3898 PRE_EXPR_REFERENCE (result) = ref;
3900 get_or_alloc_expression_id (result);
3901 add_to_value (get_expr_value_id (result), result);
3902 bitmap_value_insert_into_set (EXP_GEN (block), result);
3904 continue;
3907 case GIMPLE_ASSIGN:
3909 pre_expr result = NULL;
3910 switch (vn_get_stmt_kind (stmt))
3912 case VN_NARY:
3914 enum tree_code code = gimple_assign_rhs_code (stmt);
3915 vn_nary_op_t nary;
3917 /* COND_EXPR and VEC_COND_EXPR are awkward in
3918 that they contain an embedded complex expression.
3919 Don't even try to shove those through PRE. */
3920 if (code == COND_EXPR
3921 || code == VEC_COND_EXPR)
3922 continue;
3924 vn_nary_op_lookup_stmt (stmt, &nary);
3925 if (!nary)
3926 continue;
3928 /* If the NARY traps and there was a preceding
3929 point in the block that might not return avoid
3930 adding the nary to EXP_GEN. */
3931 if (BB_MAY_NOTRETURN (block)
3932 && vn_nary_may_trap (nary))
3933 continue;
3935 result = pre_expr_pool.allocate ();
3936 result->kind = NARY;
3937 result->id = 0;
3938 PRE_EXPR_NARY (result) = nary;
3939 break;
3942 case VN_REFERENCE:
3944 tree rhs1 = gimple_assign_rhs1 (stmt);
3945 alias_set_type set = get_alias_set (rhs1);
3946 vec<vn_reference_op_s> operands
3947 = vn_reference_operands_for_lookup (rhs1);
3948 vn_reference_t ref;
3949 vn_reference_lookup_pieces (gimple_vuse (stmt), set,
3950 TREE_TYPE (rhs1),
3951 operands, &ref, VN_WALK);
3952 if (!ref)
3954 operands.release ();
3955 continue;
3958 /* If the value of the reference is not invalidated in
3959 this block until it is computed, add the expression
3960 to EXP_GEN. */
3961 if (gimple_vuse (stmt))
3963 gimple *def_stmt;
3964 bool ok = true;
3965 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
3966 while (!gimple_nop_p (def_stmt)
3967 && gimple_code (def_stmt) != GIMPLE_PHI
3968 && gimple_bb (def_stmt) == block)
3970 if (stmt_may_clobber_ref_p
3971 (def_stmt, gimple_assign_rhs1 (stmt)))
3973 ok = false;
3974 break;
3976 def_stmt
3977 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
3979 if (!ok)
3981 operands.release ();
3982 continue;
3986 /* If the load was value-numbered to another
3987 load make sure we do not use its expression
3988 for insertion if it wouldn't be a valid
3989 replacement. */
3990 /* At the momemt we have a testcase
3991 for hoist insertion of aligned vs. misaligned
3992 variants in gcc.dg/torture/pr65270-1.c thus
3993 with just alignment to be considered we can
3994 simply replace the expression in the hashtable
3995 with the most conservative one. */
3996 vn_reference_op_t ref1 = &ref->operands.last ();
3997 while (ref1->opcode != TARGET_MEM_REF
3998 && ref1->opcode != MEM_REF
3999 && ref1 != &ref->operands[0])
4000 --ref1;
4001 vn_reference_op_t ref2 = &operands.last ();
4002 while (ref2->opcode != TARGET_MEM_REF
4003 && ref2->opcode != MEM_REF
4004 && ref2 != &operands[0])
4005 --ref2;
4006 if ((ref1->opcode == TARGET_MEM_REF
4007 || ref1->opcode == MEM_REF)
4008 && (TYPE_ALIGN (ref1->type)
4009 > TYPE_ALIGN (ref2->type)))
4010 ref1->type
4011 = build_aligned_type (ref1->type,
4012 TYPE_ALIGN (ref2->type));
4013 /* TBAA behavior is an obvious part so make sure
4014 that the hashtable one covers this as well
4015 by adjusting the ref alias set and its base. */
4016 if (ref->set == set
4017 || alias_set_subset_of (set, ref->set))
4019 else if (alias_set_subset_of (ref->set, set))
4021 ref->set = set;
4022 if (ref1->opcode == MEM_REF)
4023 ref1->op0
4024 = wide_int_to_tree (TREE_TYPE (ref2->op0),
4025 wi::to_wide (ref1->op0));
4026 else
4027 ref1->op2
4028 = wide_int_to_tree (TREE_TYPE (ref2->op2),
4029 wi::to_wide (ref1->op2));
4031 else
4033 ref->set = 0;
4034 if (ref1->opcode == MEM_REF)
4035 ref1->op0
4036 = wide_int_to_tree (ptr_type_node,
4037 wi::to_wide (ref1->op0));
4038 else
4039 ref1->op2
4040 = wide_int_to_tree (ptr_type_node,
4041 wi::to_wide (ref1->op2));
4043 operands.release ();
4045 result = pre_expr_pool.allocate ();
4046 result->kind = REFERENCE;
4047 result->id = 0;
4048 PRE_EXPR_REFERENCE (result) = ref;
4049 break;
4052 default:
4053 continue;
4056 get_or_alloc_expression_id (result);
4057 add_to_value (get_expr_value_id (result), result);
4058 bitmap_value_insert_into_set (EXP_GEN (block), result);
4059 continue;
4061 default:
4062 break;
4066 if (dump_file && (dump_flags & TDF_DETAILS))
4068 print_bitmap_set (dump_file, EXP_GEN (block),
4069 "exp_gen", block->index);
4070 print_bitmap_set (dump_file, PHI_GEN (block),
4071 "phi_gen", block->index);
4072 print_bitmap_set (dump_file, TMP_GEN (block),
4073 "tmp_gen", block->index);
4074 print_bitmap_set (dump_file, AVAIL_OUT (block),
4075 "avail_out", block->index);
4078 /* Put the dominator children of BLOCK on the worklist of blocks
4079 to compute available sets for. */
4080 for (son = first_dom_son (CDI_DOMINATORS, block);
4081 son;
4082 son = next_dom_son (CDI_DOMINATORS, son))
4083 worklist[sp++] = son;
4086 free (worklist);
4090 /* Local state for the eliminate domwalk. */
4091 static vec<gimple *> el_to_remove;
4092 static vec<gimple *> el_to_fixup;
4093 static unsigned int el_todo;
4094 static vec<tree> el_avail;
4095 static vec<tree> el_avail_stack;
4097 /* Return a leader for OP that is available at the current point of the
4098 eliminate domwalk. */
4100 static tree
4101 eliminate_avail (tree op)
4103 tree valnum = VN_INFO (op)->valnum;
4104 if (TREE_CODE (valnum) == SSA_NAME)
4106 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
4107 return valnum;
4108 if (el_avail.length () > SSA_NAME_VERSION (valnum))
4109 return el_avail[SSA_NAME_VERSION (valnum)];
4111 else if (is_gimple_min_invariant (valnum))
4112 return valnum;
4113 return NULL_TREE;
4116 /* At the current point of the eliminate domwalk make OP available. */
4118 static void
4119 eliminate_push_avail (tree op)
4121 tree valnum = VN_INFO (op)->valnum;
4122 if (TREE_CODE (valnum) == SSA_NAME)
4124 if (el_avail.length () <= SSA_NAME_VERSION (valnum))
4125 el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1);
4126 tree pushop = op;
4127 if (el_avail[SSA_NAME_VERSION (valnum)])
4128 pushop = el_avail[SSA_NAME_VERSION (valnum)];
4129 el_avail_stack.safe_push (pushop);
4130 el_avail[SSA_NAME_VERSION (valnum)] = op;
4134 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
4135 the leader for the expression if insertion was successful. */
4137 static tree
4138 eliminate_insert (gimple_stmt_iterator *gsi, tree val)
4140 /* We can insert a sequence with a single assignment only. */
4141 gimple_seq stmts = VN_INFO (val)->expr;
4142 if (!gimple_seq_singleton_p (stmts))
4143 return NULL_TREE;
4144 gassign *stmt = dyn_cast <gassign *> (gimple_seq_first_stmt (stmts));
4145 if (!stmt
4146 || (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
4147 && gimple_assign_rhs_code (stmt) != VIEW_CONVERT_EXPR
4148 && gimple_assign_rhs_code (stmt) != BIT_FIELD_REF
4149 && (gimple_assign_rhs_code (stmt) != BIT_AND_EXPR
4150 || TREE_CODE (gimple_assign_rhs2 (stmt)) != INTEGER_CST)))
4151 return NULL_TREE;
4153 tree op = gimple_assign_rhs1 (stmt);
4154 if (gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR
4155 || gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
4156 op = TREE_OPERAND (op, 0);
4157 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op;
4158 if (!leader)
4159 return NULL_TREE;
4161 tree res;
4162 stmts = NULL;
4163 if (gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
4164 res = gimple_build (&stmts, BIT_FIELD_REF,
4165 TREE_TYPE (val), leader,
4166 TREE_OPERAND (gimple_assign_rhs1 (stmt), 1),
4167 TREE_OPERAND (gimple_assign_rhs1 (stmt), 2));
4168 else if (gimple_assign_rhs_code (stmt) == BIT_AND_EXPR)
4169 res = gimple_build (&stmts, BIT_AND_EXPR,
4170 TREE_TYPE (val), leader, gimple_assign_rhs2 (stmt));
4171 else
4172 res = gimple_build (&stmts, gimple_assign_rhs_code (stmt),
4173 TREE_TYPE (val), leader);
4174 if (TREE_CODE (res) != SSA_NAME
4175 || SSA_NAME_IS_DEFAULT_DEF (res)
4176 || gimple_bb (SSA_NAME_DEF_STMT (res)))
4178 gimple_seq_discard (stmts);
4180 /* During propagation we have to treat SSA info conservatively
4181 and thus we can end up simplifying the inserted expression
4182 at elimination time to sth not defined in stmts. */
4183 /* But then this is a redundancy we failed to detect. Which means
4184 res now has two values. That doesn't play well with how
4185 we track availability here, so give up. */
4186 if (dump_file && (dump_flags & TDF_DETAILS))
4188 if (TREE_CODE (res) == SSA_NAME)
4189 res = eliminate_avail (res);
4190 if (res)
4192 fprintf (dump_file, "Failed to insert expression for value ");
4193 print_generic_expr (dump_file, val);
4194 fprintf (dump_file, " which is really fully redundant to ");
4195 print_generic_expr (dump_file, res);
4196 fprintf (dump_file, "\n");
4200 return NULL_TREE;
4202 else
4204 gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
4205 VN_INFO_GET (res)->valnum = val;
4208 pre_stats.insertions++;
4209 if (dump_file && (dump_flags & TDF_DETAILS))
4211 fprintf (dump_file, "Inserted ");
4212 print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (res), 0);
4215 return res;
4218 class eliminate_dom_walker : public dom_walker
4220 public:
4221 eliminate_dom_walker (cdi_direction direction, bool do_pre_)
4222 : dom_walker (direction), do_pre (do_pre_) {}
4224 virtual edge before_dom_children (basic_block);
4225 virtual void after_dom_children (basic_block);
4227 bool do_pre;
4230 /* Perform elimination for the basic-block B during the domwalk. */
4232 edge
4233 eliminate_dom_walker::before_dom_children (basic_block b)
4235 /* Mark new bb. */
4236 el_avail_stack.safe_push (NULL_TREE);
4238 /* Skip unreachable blocks marked unreachable during the SCCVN domwalk. */
4239 edge_iterator ei;
4240 edge e;
4241 FOR_EACH_EDGE (e, ei, b->preds)
4242 if (e->flags & EDGE_EXECUTABLE)
4243 break;
4244 if (! e)
4245 return NULL;
4247 for (gphi_iterator gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4249 gphi *phi = gsi.phi ();
4250 tree res = PHI_RESULT (phi);
4252 if (virtual_operand_p (res))
4254 gsi_next (&gsi);
4255 continue;
4258 tree sprime = eliminate_avail (res);
4259 if (sprime
4260 && sprime != res)
4262 if (dump_file && (dump_flags & TDF_DETAILS))
4264 fprintf (dump_file, "Replaced redundant PHI node defining ");
4265 print_generic_expr (dump_file, res);
4266 fprintf (dump_file, " with ");
4267 print_generic_expr (dump_file, sprime);
4268 fprintf (dump_file, "\n");
4271 /* If we inserted this PHI node ourself, it's not an elimination. */
4272 if (inserted_exprs
4273 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4274 pre_stats.phis--;
4275 else
4276 pre_stats.eliminations++;
4278 /* If we will propagate into all uses don't bother to do
4279 anything. */
4280 if (may_propagate_copy (res, sprime))
4282 /* Mark the PHI for removal. */
4283 el_to_remove.safe_push (phi);
4284 gsi_next (&gsi);
4285 continue;
4288 remove_phi_node (&gsi, false);
4290 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4291 sprime = fold_convert (TREE_TYPE (res), sprime);
4292 gimple *stmt = gimple_build_assign (res, sprime);
4293 gimple_stmt_iterator gsi2 = gsi_after_labels (b);
4294 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4295 continue;
4298 eliminate_push_avail (res);
4299 gsi_next (&gsi);
4302 for (gimple_stmt_iterator gsi = gsi_start_bb (b);
4303 !gsi_end_p (gsi);
4304 gsi_next (&gsi))
4306 tree sprime = NULL_TREE;
4307 gimple *stmt = gsi_stmt (gsi);
4308 tree lhs = gimple_get_lhs (stmt);
4309 if (lhs && TREE_CODE (lhs) == SSA_NAME
4310 && !gimple_has_volatile_ops (stmt)
4311 /* See PR43491. Do not replace a global register variable when
4312 it is a the RHS of an assignment. Do replace local register
4313 variables since gcc does not guarantee a local variable will
4314 be allocated in register.
4315 ??? The fix isn't effective here. This should instead
4316 be ensured by not value-numbering them the same but treating
4317 them like volatiles? */
4318 && !(gimple_assign_single_p (stmt)
4319 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL
4320 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt))
4321 && is_global_var (gimple_assign_rhs1 (stmt)))))
4323 sprime = eliminate_avail (lhs);
4324 if (!sprime)
4326 /* If there is no existing usable leader but SCCVN thinks
4327 it has an expression it wants to use as replacement,
4328 insert that. */
4329 tree val = VN_INFO (lhs)->valnum;
4330 if (val != VN_TOP
4331 && TREE_CODE (val) == SSA_NAME
4332 && VN_INFO (val)->needs_insertion
4333 && VN_INFO (val)->expr != NULL
4334 && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE)
4335 eliminate_push_avail (sprime);
4338 /* If this now constitutes a copy duplicate points-to
4339 and range info appropriately. This is especially
4340 important for inserted code. See tree-ssa-copy.c
4341 for similar code. */
4342 if (sprime
4343 && TREE_CODE (sprime) == SSA_NAME)
4345 basic_block sprime_b = gimple_bb (SSA_NAME_DEF_STMT (sprime));
4346 if (POINTER_TYPE_P (TREE_TYPE (lhs))
4347 && VN_INFO_PTR_INFO (lhs)
4348 && ! VN_INFO_PTR_INFO (sprime))
4350 duplicate_ssa_name_ptr_info (sprime,
4351 VN_INFO_PTR_INFO (lhs));
4352 if (b != sprime_b)
4353 mark_ptr_info_alignment_unknown
4354 (SSA_NAME_PTR_INFO (sprime));
4356 else if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))
4357 && VN_INFO_RANGE_INFO (lhs)
4358 && ! VN_INFO_RANGE_INFO (sprime)
4359 && b == sprime_b)
4360 duplicate_ssa_name_range_info (sprime,
4361 VN_INFO_RANGE_TYPE (lhs),
4362 VN_INFO_RANGE_INFO (lhs));
4365 /* Inhibit the use of an inserted PHI on a loop header when
4366 the address of the memory reference is a simple induction
4367 variable. In other cases the vectorizer won't do anything
4368 anyway (either it's loop invariant or a complicated
4369 expression). */
4370 if (sprime
4371 && TREE_CODE (sprime) == SSA_NAME
4372 && do_pre
4373 && (flag_tree_loop_vectorize || flag_tree_parallelize_loops > 1)
4374 && loop_outer (b->loop_father)
4375 && has_zero_uses (sprime)
4376 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))
4377 && gimple_assign_load_p (stmt))
4379 gimple *def_stmt = SSA_NAME_DEF_STMT (sprime);
4380 basic_block def_bb = gimple_bb (def_stmt);
4381 if (gimple_code (def_stmt) == GIMPLE_PHI
4382 && def_bb->loop_father->header == def_bb)
4384 loop_p loop = def_bb->loop_father;
4385 ssa_op_iter iter;
4386 tree op;
4387 bool found = false;
4388 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4390 affine_iv iv;
4391 def_bb = gimple_bb (SSA_NAME_DEF_STMT (op));
4392 if (def_bb
4393 && flow_bb_inside_loop_p (loop, def_bb)
4394 && simple_iv (loop, loop, op, &iv, true))
4396 found = true;
4397 break;
4400 if (found)
4402 if (dump_file && (dump_flags & TDF_DETAILS))
4404 fprintf (dump_file, "Not replacing ");
4405 print_gimple_expr (dump_file, stmt, 0);
4406 fprintf (dump_file, " with ");
4407 print_generic_expr (dump_file, sprime);
4408 fprintf (dump_file, " which would add a loop"
4409 " carried dependence to loop %d\n",
4410 loop->num);
4412 /* Don't keep sprime available. */
4413 sprime = NULL_TREE;
4418 if (sprime)
4420 /* If we can propagate the value computed for LHS into
4421 all uses don't bother doing anything with this stmt. */
4422 if (may_propagate_copy (lhs, sprime))
4424 /* Mark it for removal. */
4425 el_to_remove.safe_push (stmt);
4427 /* ??? Don't count copy/constant propagations. */
4428 if (gimple_assign_single_p (stmt)
4429 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4430 || gimple_assign_rhs1 (stmt) == sprime))
4431 continue;
4433 if (dump_file && (dump_flags & TDF_DETAILS))
4435 fprintf (dump_file, "Replaced ");
4436 print_gimple_expr (dump_file, stmt, 0);
4437 fprintf (dump_file, " with ");
4438 print_generic_expr (dump_file, sprime);
4439 fprintf (dump_file, " in all uses of ");
4440 print_gimple_stmt (dump_file, stmt, 0);
4443 pre_stats.eliminations++;
4444 continue;
4447 /* If this is an assignment from our leader (which
4448 happens in the case the value-number is a constant)
4449 then there is nothing to do. */
4450 if (gimple_assign_single_p (stmt)
4451 && sprime == gimple_assign_rhs1 (stmt))
4452 continue;
4454 /* Else replace its RHS. */
4455 bool can_make_abnormal_goto
4456 = is_gimple_call (stmt)
4457 && stmt_can_make_abnormal_goto (stmt);
4459 if (dump_file && (dump_flags & TDF_DETAILS))
4461 fprintf (dump_file, "Replaced ");
4462 print_gimple_expr (dump_file, stmt, 0);
4463 fprintf (dump_file, " with ");
4464 print_generic_expr (dump_file, sprime);
4465 fprintf (dump_file, " in ");
4466 print_gimple_stmt (dump_file, stmt, 0);
4469 pre_stats.eliminations++;
4470 gimple *orig_stmt = stmt;
4471 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4472 TREE_TYPE (sprime)))
4473 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4474 tree vdef = gimple_vdef (stmt);
4475 tree vuse = gimple_vuse (stmt);
4476 propagate_tree_value_into_stmt (&gsi, sprime);
4477 stmt = gsi_stmt (gsi);
4478 update_stmt (stmt);
4479 if (vdef != gimple_vdef (stmt))
4480 VN_INFO (vdef)->valnum = vuse;
4482 /* If we removed EH side-effects from the statement, clean
4483 its EH information. */
4484 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4486 bitmap_set_bit (need_eh_cleanup,
4487 gimple_bb (stmt)->index);
4488 if (dump_file && (dump_flags & TDF_DETAILS))
4489 fprintf (dump_file, " Removed EH side-effects.\n");
4492 /* Likewise for AB side-effects. */
4493 if (can_make_abnormal_goto
4494 && !stmt_can_make_abnormal_goto (stmt))
4496 bitmap_set_bit (need_ab_cleanup,
4497 gimple_bb (stmt)->index);
4498 if (dump_file && (dump_flags & TDF_DETAILS))
4499 fprintf (dump_file, " Removed AB side-effects.\n");
4502 continue;
4506 /* If the statement is a scalar store, see if the expression
4507 has the same value number as its rhs. If so, the store is
4508 dead. */
4509 if (gimple_assign_single_p (stmt)
4510 && !gimple_has_volatile_ops (stmt)
4511 && !is_gimple_reg (gimple_assign_lhs (stmt))
4512 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4513 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4515 tree val;
4516 tree rhs = gimple_assign_rhs1 (stmt);
4517 vn_reference_t vnresult;
4518 val = vn_reference_lookup (lhs, gimple_vuse (stmt), VN_WALKREWRITE,
4519 &vnresult, false);
4520 if (TREE_CODE (rhs) == SSA_NAME)
4521 rhs = VN_INFO (rhs)->valnum;
4522 if (val
4523 && operand_equal_p (val, rhs, 0))
4525 /* We can only remove the later store if the former aliases
4526 at least all accesses the later one does or if the store
4527 was to readonly memory storing the same value. */
4528 alias_set_type set = get_alias_set (lhs);
4529 if (! vnresult
4530 || vnresult->set == set
4531 || alias_set_subset_of (set, vnresult->set))
4533 if (dump_file && (dump_flags & TDF_DETAILS))
4535 fprintf (dump_file, "Deleted redundant store ");
4536 print_gimple_stmt (dump_file, stmt, 0);
4539 /* Queue stmt for removal. */
4540 el_to_remove.safe_push (stmt);
4541 continue;
4546 /* If this is a control statement value numbering left edges
4547 unexecuted on force the condition in a way consistent with
4548 that. */
4549 if (gcond *cond = dyn_cast <gcond *> (stmt))
4551 if ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE)
4552 ^ (EDGE_SUCC (b, 1)->flags & EDGE_EXECUTABLE))
4554 if (dump_file && (dump_flags & TDF_DETAILS))
4556 fprintf (dump_file, "Removing unexecutable edge from ");
4557 print_gimple_stmt (dump_file, stmt, 0);
4559 if (((EDGE_SUCC (b, 0)->flags & EDGE_TRUE_VALUE) != 0)
4560 == ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE) != 0))
4561 gimple_cond_make_true (cond);
4562 else
4563 gimple_cond_make_false (cond);
4564 update_stmt (cond);
4565 el_todo |= TODO_cleanup_cfg;
4566 continue;
4570 bool can_make_abnormal_goto = stmt_can_make_abnormal_goto (stmt);
4571 bool was_noreturn = (is_gimple_call (stmt)
4572 && gimple_call_noreturn_p (stmt));
4573 tree vdef = gimple_vdef (stmt);
4574 tree vuse = gimple_vuse (stmt);
4576 /* If we didn't replace the whole stmt (or propagate the result
4577 into all uses), replace all uses on this stmt with their
4578 leaders. */
4579 bool modified = false;
4580 use_operand_p use_p;
4581 ssa_op_iter iter;
4582 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
4584 tree use = USE_FROM_PTR (use_p);
4585 /* ??? The call code above leaves stmt operands un-updated. */
4586 if (TREE_CODE (use) != SSA_NAME)
4587 continue;
4588 tree sprime = eliminate_avail (use);
4589 if (sprime && sprime != use
4590 && may_propagate_copy (use, sprime)
4591 /* We substitute into debug stmts to avoid excessive
4592 debug temporaries created by removed stmts, but we need
4593 to avoid doing so for inserted sprimes as we never want
4594 to create debug temporaries for them. */
4595 && (!inserted_exprs
4596 || TREE_CODE (sprime) != SSA_NAME
4597 || !is_gimple_debug (stmt)
4598 || !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))))
4600 propagate_value (use_p, sprime);
4601 modified = true;
4605 /* Fold the stmt if modified, this canonicalizes MEM_REFs we propagated
4606 into which is a requirement for the IPA devirt machinery. */
4607 gimple *old_stmt = stmt;
4608 if (modified)
4610 /* If a formerly non-invariant ADDR_EXPR is turned into an
4611 invariant one it was on a separate stmt. */
4612 if (gimple_assign_single_p (stmt)
4613 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR)
4614 recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
4615 gimple_stmt_iterator prev = gsi;
4616 gsi_prev (&prev);
4617 if (fold_stmt (&gsi))
4619 /* fold_stmt may have created new stmts inbetween
4620 the previous stmt and the folded stmt. Mark
4621 all defs created there as varying to not confuse
4622 the SCCVN machinery as we're using that even during
4623 elimination. */
4624 if (gsi_end_p (prev))
4625 prev = gsi_start_bb (b);
4626 else
4627 gsi_next (&prev);
4628 if (gsi_stmt (prev) != gsi_stmt (gsi))
4631 tree def;
4632 ssa_op_iter dit;
4633 FOR_EACH_SSA_TREE_OPERAND (def, gsi_stmt (prev),
4634 dit, SSA_OP_ALL_DEFS)
4635 /* As existing DEFs may move between stmts
4636 we have to guard VN_INFO_GET. */
4637 if (! has_VN_INFO (def))
4638 VN_INFO_GET (def)->valnum = def;
4639 if (gsi_stmt (prev) == gsi_stmt (gsi))
4640 break;
4641 gsi_next (&prev);
4643 while (1);
4645 stmt = gsi_stmt (gsi);
4646 /* In case we folded the stmt away schedule the NOP for removal. */
4647 if (gimple_nop_p (stmt))
4648 el_to_remove.safe_push (stmt);
4651 /* Visit indirect calls and turn them into direct calls if
4652 possible using the devirtualization machinery. Do this before
4653 checking for required EH/abnormal/noreturn cleanup as devird
4654 may expose more of those. */
4655 if (gcall *call_stmt = dyn_cast <gcall *> (stmt))
4657 tree fn = gimple_call_fn (call_stmt);
4658 if (fn
4659 && flag_devirtualize
4660 && virtual_method_call_p (fn))
4662 tree otr_type = obj_type_ref_class (fn);
4663 unsigned HOST_WIDE_INT otr_tok
4664 = tree_to_uhwi (OBJ_TYPE_REF_TOKEN (fn));
4665 tree instance;
4666 ipa_polymorphic_call_context context (current_function_decl,
4667 fn, stmt, &instance);
4668 context.get_dynamic_type (instance, OBJ_TYPE_REF_OBJECT (fn),
4669 otr_type, stmt);
4670 bool final;
4671 vec <cgraph_node *> targets
4672 = possible_polymorphic_call_targets (obj_type_ref_class (fn),
4673 otr_tok, context, &final);
4674 if (dump_file)
4675 dump_possible_polymorphic_call_targets (dump_file,
4676 obj_type_ref_class (fn),
4677 otr_tok, context);
4678 if (final && targets.length () <= 1 && dbg_cnt (devirt))
4680 tree fn;
4681 if (targets.length () == 1)
4682 fn = targets[0]->decl;
4683 else
4684 fn = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
4685 if (dump_enabled_p ())
4687 location_t loc = gimple_location (stmt);
4688 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc,
4689 "converting indirect call to "
4690 "function %s\n",
4691 lang_hooks.decl_printable_name (fn, 2));
4693 gimple_call_set_fndecl (call_stmt, fn);
4694 /* If changing the call to __builtin_unreachable
4695 or similar noreturn function, adjust gimple_call_fntype
4696 too. */
4697 if (gimple_call_noreturn_p (call_stmt)
4698 && VOID_TYPE_P (TREE_TYPE (TREE_TYPE (fn)))
4699 && TYPE_ARG_TYPES (TREE_TYPE (fn))
4700 && (TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fn)))
4701 == void_type_node))
4702 gimple_call_set_fntype (call_stmt, TREE_TYPE (fn));
4703 maybe_remove_unused_call_args (cfun, call_stmt);
4704 modified = true;
4709 if (modified)
4711 /* When changing a call into a noreturn call, cfg cleanup
4712 is needed to fix up the noreturn call. */
4713 if (!was_noreturn
4714 && is_gimple_call (stmt) && gimple_call_noreturn_p (stmt))
4715 el_to_fixup.safe_push (stmt);
4716 /* When changing a condition or switch into one we know what
4717 edge will be executed, schedule a cfg cleanup. */
4718 if ((gimple_code (stmt) == GIMPLE_COND
4719 && (gimple_cond_true_p (as_a <gcond *> (stmt))
4720 || gimple_cond_false_p (as_a <gcond *> (stmt))))
4721 || (gimple_code (stmt) == GIMPLE_SWITCH
4722 && TREE_CODE (gimple_switch_index
4723 (as_a <gswitch *> (stmt))) == INTEGER_CST))
4724 el_todo |= TODO_cleanup_cfg;
4725 /* If we removed EH side-effects from the statement, clean
4726 its EH information. */
4727 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
4729 bitmap_set_bit (need_eh_cleanup,
4730 gimple_bb (stmt)->index);
4731 if (dump_file && (dump_flags & TDF_DETAILS))
4732 fprintf (dump_file, " Removed EH side-effects.\n");
4734 /* Likewise for AB side-effects. */
4735 if (can_make_abnormal_goto
4736 && !stmt_can_make_abnormal_goto (stmt))
4738 bitmap_set_bit (need_ab_cleanup,
4739 gimple_bb (stmt)->index);
4740 if (dump_file && (dump_flags & TDF_DETAILS))
4741 fprintf (dump_file, " Removed AB side-effects.\n");
4743 update_stmt (stmt);
4744 if (vdef != gimple_vdef (stmt))
4745 VN_INFO (vdef)->valnum = vuse;
4748 /* Make new values available - for fully redundant LHS we
4749 continue with the next stmt above and skip this. */
4750 def_operand_p defp;
4751 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_DEF)
4752 eliminate_push_avail (DEF_FROM_PTR (defp));
4755 /* Replace destination PHI arguments. */
4756 FOR_EACH_EDGE (e, ei, b->succs)
4757 if (e->flags & EDGE_EXECUTABLE)
4758 for (gphi_iterator gsi = gsi_start_phis (e->dest);
4759 !gsi_end_p (gsi);
4760 gsi_next (&gsi))
4762 gphi *phi = gsi.phi ();
4763 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
4764 tree arg = USE_FROM_PTR (use_p);
4765 if (TREE_CODE (arg) != SSA_NAME
4766 || virtual_operand_p (arg))
4767 continue;
4768 tree sprime = eliminate_avail (arg);
4769 if (sprime && may_propagate_copy (arg, sprime))
4770 propagate_value (use_p, sprime);
4772 return NULL;
4775 /* Make no longer available leaders no longer available. */
4777 void
4778 eliminate_dom_walker::after_dom_children (basic_block)
4780 tree entry;
4781 while ((entry = el_avail_stack.pop ()) != NULL_TREE)
4783 tree valnum = VN_INFO (entry)->valnum;
4784 tree old = el_avail[SSA_NAME_VERSION (valnum)];
4785 if (old == entry)
4786 el_avail[SSA_NAME_VERSION (valnum)] = NULL_TREE;
4787 else
4788 el_avail[SSA_NAME_VERSION (valnum)] = entry;
4792 /* Eliminate fully redundant computations. */
4794 static unsigned int
4795 eliminate (bool do_pre)
4797 need_eh_cleanup = BITMAP_ALLOC (NULL);
4798 need_ab_cleanup = BITMAP_ALLOC (NULL);
4800 el_to_remove.create (0);
4801 el_to_fixup.create (0);
4802 el_todo = 0;
4803 el_avail.create (num_ssa_names);
4804 el_avail_stack.create (0);
4806 eliminate_dom_walker (CDI_DOMINATORS,
4807 do_pre).walk (cfun->cfg->x_entry_block_ptr);
4809 el_avail.release ();
4810 el_avail_stack.release ();
4812 return el_todo;
4815 /* Perform CFG cleanups made necessary by elimination. */
4817 static unsigned
4818 fini_eliminate (void)
4820 gimple_stmt_iterator gsi;
4821 gimple *stmt;
4822 unsigned todo = 0;
4824 /* We cannot remove stmts during BB walk, especially not release SSA
4825 names there as this confuses the VN machinery. The stmts ending
4826 up in el_to_remove are either stores or simple copies.
4827 Remove stmts in reverse order to make debug stmt creation possible. */
4828 while (!el_to_remove.is_empty ())
4830 stmt = el_to_remove.pop ();
4832 if (dump_file && (dump_flags & TDF_DETAILS))
4834 fprintf (dump_file, "Removing dead stmt ");
4835 print_gimple_stmt (dump_file, stmt, 0, 0);
4838 gsi = gsi_for_stmt (stmt);
4839 if (gimple_code (stmt) == GIMPLE_PHI)
4840 remove_phi_node (&gsi, true);
4841 else
4843 basic_block bb = gimple_bb (stmt);
4844 unlink_stmt_vdef (stmt);
4845 if (gsi_remove (&gsi, true))
4846 bitmap_set_bit (need_eh_cleanup, bb->index);
4847 if (is_gimple_call (stmt) && stmt_can_make_abnormal_goto (stmt))
4848 bitmap_set_bit (need_ab_cleanup, bb->index);
4849 release_defs (stmt);
4852 /* Removing a stmt may expose a forwarder block. */
4853 todo |= TODO_cleanup_cfg;
4855 el_to_remove.release ();
4857 /* Fixup stmts that became noreturn calls. This may require splitting
4858 blocks and thus isn't possible during the dominator walk. Do this
4859 in reverse order so we don't inadvertedly remove a stmt we want to
4860 fixup by visiting a dominating now noreturn call first. */
4861 while (!el_to_fixup.is_empty ())
4863 stmt = el_to_fixup.pop ();
4865 if (dump_file && (dump_flags & TDF_DETAILS))
4867 fprintf (dump_file, "Fixing up noreturn call ");
4868 print_gimple_stmt (dump_file, stmt, 0);
4871 if (fixup_noreturn_call (stmt))
4872 todo |= TODO_cleanup_cfg;
4874 el_to_fixup.release ();
4876 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4877 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4879 if (do_eh_cleanup)
4880 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4882 if (do_ab_cleanup)
4883 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4885 BITMAP_FREE (need_eh_cleanup);
4886 BITMAP_FREE (need_ab_cleanup);
4888 if (do_eh_cleanup || do_ab_cleanup)
4889 todo |= TODO_cleanup_cfg;
4890 return todo;
4893 /* Cheap DCE of a known set of possibly dead stmts.
4895 Because we don't follow exactly the standard PRE algorithm, and decide not
4896 to insert PHI nodes sometimes, and because value numbering of casts isn't
4897 perfect, we sometimes end up inserting dead code. This simple DCE-like
4898 pass removes any insertions we made that weren't actually used. */
4900 static void
4901 remove_dead_inserted_code (void)
4903 /* ??? Re-use inserted_exprs as worklist not only as initial set.
4904 This may end up removing non-inserted code as well. If we
4905 keep inserted_exprs unchanged we could restrict new worklist
4906 elements to members of inserted_exprs. */
4907 bitmap worklist = inserted_exprs;
4908 while (! bitmap_empty_p (worklist))
4910 /* Pop item. */
4911 unsigned i = bitmap_first_set_bit (worklist);
4912 bitmap_clear_bit (worklist, i);
4914 tree def = ssa_name (i);
4915 /* Removed by somebody else or still in use. */
4916 if (! def || ! has_zero_uses (def))
4917 continue;
4919 gimple *t = SSA_NAME_DEF_STMT (def);
4920 if (gimple_has_side_effects (t))
4921 continue;
4923 /* Add uses to the worklist. */
4924 ssa_op_iter iter;
4925 use_operand_p use_p;
4926 FOR_EACH_PHI_OR_STMT_USE (use_p, t, iter, SSA_OP_USE)
4928 tree use = USE_FROM_PTR (use_p);
4929 if (TREE_CODE (use) == SSA_NAME
4930 && ! SSA_NAME_IS_DEFAULT_DEF (use))
4931 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4934 /* Remove stmt. */
4935 if (dump_file && (dump_flags & TDF_DETAILS))
4937 fprintf (dump_file, "Removing unnecessary insertion:");
4938 print_gimple_stmt (dump_file, t, 0);
4940 gimple_stmt_iterator gsi = gsi_for_stmt (t);
4941 if (gimple_code (t) == GIMPLE_PHI)
4942 remove_phi_node (&gsi, true);
4943 else
4945 gsi_remove (&gsi, true);
4946 release_defs (t);
4952 /* Initialize data structures used by PRE. */
4954 static void
4955 init_pre (void)
4957 basic_block bb;
4959 next_expression_id = 1;
4960 expressions.create (0);
4961 expressions.safe_push (NULL);
4962 value_expressions.create (get_max_value_id () + 1);
4963 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
4964 name_to_id.create (0);
4966 inserted_exprs = BITMAP_ALLOC (NULL);
4968 connect_infinite_loops_to_exit ();
4969 memset (&pre_stats, 0, sizeof (pre_stats));
4971 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4973 calculate_dominance_info (CDI_DOMINATORS);
4975 bitmap_obstack_initialize (&grand_bitmap_obstack);
4976 phi_translate_table = new hash_table<expr_pred_trans_d> (5110);
4977 expression_to_id = new hash_table<pre_expr_d> (num_ssa_names * 3);
4978 FOR_ALL_BB_FN (bb, cfun)
4980 EXP_GEN (bb) = bitmap_set_new ();
4981 PHI_GEN (bb) = bitmap_set_new ();
4982 TMP_GEN (bb) = bitmap_set_new ();
4983 AVAIL_OUT (bb) = bitmap_set_new ();
4988 /* Deallocate data structures used by PRE. */
4990 static void
4991 fini_pre ()
4993 value_expressions.release ();
4994 BITMAP_FREE (inserted_exprs);
4995 bitmap_obstack_release (&grand_bitmap_obstack);
4996 bitmap_set_pool.release ();
4997 pre_expr_pool.release ();
4998 delete phi_translate_table;
4999 phi_translate_table = NULL;
5000 delete expression_to_id;
5001 expression_to_id = NULL;
5002 name_to_id.release ();
5004 free_aux_for_blocks ();
5007 namespace {
5009 const pass_data pass_data_pre =
5011 GIMPLE_PASS, /* type */
5012 "pre", /* name */
5013 OPTGROUP_NONE, /* optinfo_flags */
5014 TV_TREE_PRE, /* tv_id */
5015 ( PROP_cfg | PROP_ssa ), /* properties_required */
5016 0, /* properties_provided */
5017 0, /* properties_destroyed */
5018 TODO_rebuild_alias, /* todo_flags_start */
5019 0, /* todo_flags_finish */
5022 class pass_pre : public gimple_opt_pass
5024 public:
5025 pass_pre (gcc::context *ctxt)
5026 : gimple_opt_pass (pass_data_pre, ctxt)
5029 /* opt_pass methods: */
5030 virtual bool gate (function *)
5031 { return flag_tree_pre != 0 || flag_code_hoisting != 0; }
5032 virtual unsigned int execute (function *);
5034 }; // class pass_pre
5036 unsigned int
5037 pass_pre::execute (function *fun)
5039 unsigned int todo = 0;
5041 do_partial_partial =
5042 flag_tree_partial_pre && optimize_function_for_speed_p (fun);
5044 /* This has to happen before SCCVN runs because
5045 loop_optimizer_init may create new phis, etc. */
5046 loop_optimizer_init (LOOPS_NORMAL);
5047 split_critical_edges ();
5049 run_scc_vn (VN_WALK);
5051 init_pre ();
5052 scev_initialize ();
5054 /* Collect and value number expressions computed in each basic block. */
5055 compute_avail ();
5057 /* Insert can get quite slow on an incredibly large number of basic
5058 blocks due to some quadratic behavior. Until this behavior is
5059 fixed, don't run it when he have an incredibly large number of
5060 bb's. If we aren't going to run insert, there is no point in
5061 computing ANTIC, either, even though it's plenty fast. */
5062 if (n_basic_blocks_for_fn (fun) < 4000)
5064 compute_antic ();
5065 insert ();
5068 /* Make sure to remove fake edges before committing our inserts.
5069 This makes sure we don't end up with extra critical edges that
5070 we would need to split. */
5071 remove_fake_exit_edges ();
5072 gsi_commit_edge_inserts ();
5074 /* Eliminate folds statements which might (should not...) end up
5075 not keeping virtual operands up-to-date. */
5076 gcc_assert (!need_ssa_update_p (fun));
5078 /* Remove all the redundant expressions. */
5079 todo |= eliminate (true);
5081 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
5082 statistics_counter_event (fun, "PA inserted", pre_stats.pa_insert);
5083 statistics_counter_event (fun, "HOIST inserted", pre_stats.hoist_insert);
5084 statistics_counter_event (fun, "New PHIs", pre_stats.phis);
5085 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
5087 clear_expression_ids ();
5089 scev_finalize ();
5090 todo |= fini_eliminate ();
5091 remove_dead_inserted_code ();
5092 fini_pre ();
5093 loop_optimizer_finalize ();
5095 /* Restore SSA info before tail-merging as that resets it as well. */
5096 scc_vn_restore_ssa_info ();
5098 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
5099 case we can merge the block with the remaining predecessor of the block.
5100 It should either:
5101 - call merge_blocks after each tail merge iteration
5102 - call merge_blocks after all tail merge iterations
5103 - mark TODO_cleanup_cfg when necessary
5104 - share the cfg cleanup with fini_pre. */
5105 todo |= tail_merge_optimize (todo);
5107 free_scc_vn ();
5109 /* Tail merging invalidates the virtual SSA web, together with
5110 cfg-cleanup opportunities exposed by PRE this will wreck the
5111 SSA updating machinery. So make sure to run update-ssa
5112 manually, before eventually scheduling cfg-cleanup as part of
5113 the todo. */
5114 update_ssa (TODO_update_ssa_only_virtuals);
5116 return todo;
5119 } // anon namespace
5121 gimple_opt_pass *
5122 make_pass_pre (gcc::context *ctxt)
5124 return new pass_pre (ctxt);
5127 namespace {
5129 const pass_data pass_data_fre =
5131 GIMPLE_PASS, /* type */
5132 "fre", /* name */
5133 OPTGROUP_NONE, /* optinfo_flags */
5134 TV_TREE_FRE, /* tv_id */
5135 ( PROP_cfg | PROP_ssa ), /* properties_required */
5136 0, /* properties_provided */
5137 0, /* properties_destroyed */
5138 0, /* todo_flags_start */
5139 0, /* todo_flags_finish */
5142 class pass_fre : public gimple_opt_pass
5144 public:
5145 pass_fre (gcc::context *ctxt)
5146 : gimple_opt_pass (pass_data_fre, ctxt)
5149 /* opt_pass methods: */
5150 opt_pass * clone () { return new pass_fre (m_ctxt); }
5151 virtual bool gate (function *) { return flag_tree_fre != 0; }
5152 virtual unsigned int execute (function *);
5154 }; // class pass_fre
5156 unsigned int
5157 pass_fre::execute (function *fun)
5159 unsigned int todo = 0;
5161 run_scc_vn (VN_WALKREWRITE);
5163 memset (&pre_stats, 0, sizeof (pre_stats));
5165 /* Remove all the redundant expressions. */
5166 todo |= eliminate (false);
5168 todo |= fini_eliminate ();
5170 scc_vn_restore_ssa_info ();
5171 free_scc_vn ();
5173 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
5174 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
5176 return todo;
5179 } // anon namespace
5181 gimple_opt_pass *
5182 make_pass_fre (gcc::context *ctxt)
5184 return new pass_fre (ctxt);