* g++.dg/debug/pr71432.C: Fail on AIX.
[official-gcc.git] / gcc / tree-ssa-pre.c
blob0c97f4fbcce68ffd2efb3b66f605bb58087d9888
1 /* SSA-PRE for trees.
2 Copyright (C) 2001-2016 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
4 <stevenb@suse.de>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
11 any later version.
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "backend.h"
26 #include "rtl.h"
27 #include "tree.h"
28 #include "gimple.h"
29 #include "predict.h"
30 #include "alloc-pool.h"
31 #include "tree-pass.h"
32 #include "ssa.h"
33 #include "cgraph.h"
34 #include "gimple-pretty-print.h"
35 #include "fold-const.h"
36 #include "cfganal.h"
37 #include "gimple-fold.h"
38 #include "tree-eh.h"
39 #include "gimplify.h"
40 #include "gimple-iterator.h"
41 #include "tree-cfg.h"
42 #include "tree-ssa-loop.h"
43 #include "tree-into-ssa.h"
44 #include "tree-dfa.h"
45 #include "tree-ssa.h"
46 #include "cfgloop.h"
47 #include "tree-ssa-sccvn.h"
48 #include "tree-scalar-evolution.h"
49 #include "params.h"
50 #include "dbgcnt.h"
51 #include "domwalk.h"
52 #include "tree-ssa-propagate.h"
53 #include "ipa-utils.h"
54 #include "tree-cfgcleanup.h"
55 #include "langhooks.h"
56 #include "alias.h"
58 /* TODO:
60 1. Avail sets can be shared by making an avail_find_leader that
61 walks up the dominator tree and looks in those avail sets.
62 This might affect code optimality, it's unclear right now.
63 2. Strength reduction can be performed by anticipating expressions
64 we can repair later on.
65 3. We can do back-substitution or smarter value numbering to catch
66 commutative expressions split up over multiple statements.
69 /* For ease of terminology, "expression node" in the below refers to
70 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
71 represent the actual statement containing the expressions we care about,
72 and we cache the value number by putting it in the expression. */
74 /* Basic algorithm
76 First we walk the statements to generate the AVAIL sets, the
77 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
78 generation of values/expressions by a given block. We use them
79 when computing the ANTIC sets. The AVAIL sets consist of
80 SSA_NAME's that represent values, so we know what values are
81 available in what blocks. AVAIL is a forward dataflow problem. In
82 SSA, values are never killed, so we don't need a kill set, or a
83 fixpoint iteration, in order to calculate the AVAIL sets. In
84 traditional parlance, AVAIL sets tell us the downsafety of the
85 expressions/values.
87 Next, we generate the ANTIC sets. These sets represent the
88 anticipatable expressions. ANTIC is a backwards dataflow
89 problem. An expression is anticipatable in a given block if it could
90 be generated in that block. This means that if we had to perform
91 an insertion in that block, of the value of that expression, we
92 could. Calculating the ANTIC sets requires phi translation of
93 expressions, because the flow goes backwards through phis. We must
94 iterate to a fixpoint of the ANTIC sets, because we have a kill
95 set. Even in SSA form, values are not live over the entire
96 function, only from their definition point onwards. So we have to
97 remove values from the ANTIC set once we go past the definition
98 point of the leaders that make them up.
99 compute_antic/compute_antic_aux performs this computation.
101 Third, we perform insertions to make partially redundant
102 expressions fully redundant.
104 An expression is partially redundant (excluding partial
105 anticipation) if:
107 1. It is AVAIL in some, but not all, of the predecessors of a
108 given block.
109 2. It is ANTIC in all the predecessors.
111 In order to make it fully redundant, we insert the expression into
112 the predecessors where it is not available, but is ANTIC.
114 For the partial anticipation case, we only perform insertion if it
115 is partially anticipated in some block, and fully available in all
116 of the predecessors.
118 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
119 performs these steps.
121 Fourth, we eliminate fully redundant expressions.
122 This is a simple statement walk that replaces redundant
123 calculations with the now available values. */
125 /* Representations of value numbers:
127 Value numbers are represented by a representative SSA_NAME. We
128 will create fake SSA_NAME's in situations where we need a
129 representative but do not have one (because it is a complex
130 expression). In order to facilitate storing the value numbers in
131 bitmaps, and keep the number of wasted SSA_NAME's down, we also
132 associate a value_id with each value number, and create full blown
133 ssa_name's only where we actually need them (IE in operands of
134 existing expressions).
136 Theoretically you could replace all the value_id's with
137 SSA_NAME_VERSION, but this would allocate a large number of
138 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
139 It would also require an additional indirection at each point we
140 use the value id. */
142 /* Representation of expressions on value numbers:
144 Expressions consisting of value numbers are represented the same
145 way as our VN internally represents them, with an additional
146 "pre_expr" wrapping around them in order to facilitate storing all
147 of the expressions in the same sets. */
149 /* Representation of sets:
151 The dataflow sets do not need to be sorted in any particular order
152 for the majority of their lifetime, are simply represented as two
153 bitmaps, one that keeps track of values present in the set, and one
154 that keeps track of expressions present in the set.
156 When we need them in topological order, we produce it on demand by
157 transforming the bitmap into an array and sorting it into topo
158 order. */
160 /* Type of expression, used to know which member of the PRE_EXPR union
161 is valid. */
163 enum pre_expr_kind
165 NAME,
166 NARY,
167 REFERENCE,
168 CONSTANT
171 union pre_expr_union
173 tree name;
174 tree constant;
175 vn_nary_op_t nary;
176 vn_reference_t reference;
179 typedef struct pre_expr_d : nofree_ptr_hash <pre_expr_d>
181 enum pre_expr_kind kind;
182 unsigned int id;
183 pre_expr_union u;
185 /* hash_table support. */
186 static inline hashval_t hash (const pre_expr_d *);
187 static inline int equal (const pre_expr_d *, const pre_expr_d *);
188 } *pre_expr;
190 #define PRE_EXPR_NAME(e) (e)->u.name
191 #define PRE_EXPR_NARY(e) (e)->u.nary
192 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
193 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
195 /* Compare E1 and E1 for equality. */
197 inline int
198 pre_expr_d::equal (const pre_expr_d *e1, const pre_expr_d *e2)
200 if (e1->kind != e2->kind)
201 return false;
203 switch (e1->kind)
205 case CONSTANT:
206 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
207 PRE_EXPR_CONSTANT (e2));
208 case NAME:
209 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
210 case NARY:
211 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
212 case REFERENCE:
213 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
214 PRE_EXPR_REFERENCE (e2));
215 default:
216 gcc_unreachable ();
220 /* Hash E. */
222 inline hashval_t
223 pre_expr_d::hash (const pre_expr_d *e)
225 switch (e->kind)
227 case CONSTANT:
228 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
229 case NAME:
230 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
231 case NARY:
232 return PRE_EXPR_NARY (e)->hashcode;
233 case REFERENCE:
234 return PRE_EXPR_REFERENCE (e)->hashcode;
235 default:
236 gcc_unreachable ();
240 /* Next global expression id number. */
241 static unsigned int next_expression_id;
243 /* Mapping from expression to id number we can use in bitmap sets. */
244 static vec<pre_expr> expressions;
245 static hash_table<pre_expr_d> *expression_to_id;
246 static vec<unsigned> name_to_id;
248 /* Allocate an expression id for EXPR. */
250 static inline unsigned int
251 alloc_expression_id (pre_expr expr)
253 struct pre_expr_d **slot;
254 /* Make sure we won't overflow. */
255 gcc_assert (next_expression_id + 1 > next_expression_id);
256 expr->id = next_expression_id++;
257 expressions.safe_push (expr);
258 if (expr->kind == NAME)
260 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
261 /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
262 re-allocations by using vec::reserve upfront. */
263 unsigned old_len = name_to_id.length ();
264 name_to_id.reserve (num_ssa_names - old_len);
265 name_to_id.quick_grow_cleared (num_ssa_names);
266 gcc_assert (name_to_id[version] == 0);
267 name_to_id[version] = expr->id;
269 else
271 slot = expression_to_id->find_slot (expr, INSERT);
272 gcc_assert (!*slot);
273 *slot = expr;
275 return next_expression_id - 1;
278 /* Return the expression id for tree EXPR. */
280 static inline unsigned int
281 get_expression_id (const pre_expr expr)
283 return expr->id;
286 static inline unsigned int
287 lookup_expression_id (const pre_expr expr)
289 struct pre_expr_d **slot;
291 if (expr->kind == NAME)
293 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
294 if (name_to_id.length () <= version)
295 return 0;
296 return name_to_id[version];
298 else
300 slot = expression_to_id->find_slot (expr, NO_INSERT);
301 if (!slot)
302 return 0;
303 return ((pre_expr)*slot)->id;
307 /* Return the existing expression id for EXPR, or create one if one
308 does not exist yet. */
310 static inline unsigned int
311 get_or_alloc_expression_id (pre_expr expr)
313 unsigned int id = lookup_expression_id (expr);
314 if (id == 0)
315 return alloc_expression_id (expr);
316 return expr->id = id;
319 /* Return the expression that has expression id ID */
321 static inline pre_expr
322 expression_for_id (unsigned int id)
324 return expressions[id];
327 /* Free the expression id field in all of our expressions,
328 and then destroy the expressions array. */
330 static void
331 clear_expression_ids (void)
333 expressions.release ();
336 static object_allocator<pre_expr_d> pre_expr_pool ("pre_expr nodes");
338 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
340 static pre_expr
341 get_or_alloc_expr_for_name (tree name)
343 struct pre_expr_d expr;
344 pre_expr result;
345 unsigned int result_id;
347 expr.kind = NAME;
348 expr.id = 0;
349 PRE_EXPR_NAME (&expr) = name;
350 result_id = lookup_expression_id (&expr);
351 if (result_id != 0)
352 return expression_for_id (result_id);
354 result = pre_expr_pool.allocate ();
355 result->kind = NAME;
356 PRE_EXPR_NAME (result) = name;
357 alloc_expression_id (result);
358 return result;
361 /* An unordered bitmap set. One bitmap tracks values, the other,
362 expressions. */
363 typedef struct bitmap_set
365 bitmap_head expressions;
366 bitmap_head values;
367 } *bitmap_set_t;
369 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
370 EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
372 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
373 EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
375 /* Mapping from value id to expressions with that value_id. */
376 static vec<bitmap> value_expressions;
378 /* Sets that we need to keep track of. */
379 typedef struct bb_bitmap_sets
381 /* The EXP_GEN set, which represents expressions/values generated in
382 a basic block. */
383 bitmap_set_t exp_gen;
385 /* The PHI_GEN set, which represents PHI results generated in a
386 basic block. */
387 bitmap_set_t phi_gen;
389 /* The TMP_GEN set, which represents results/temporaries generated
390 in a basic block. IE the LHS of an expression. */
391 bitmap_set_t tmp_gen;
393 /* The AVAIL_OUT set, which represents which values are available in
394 a given basic block. */
395 bitmap_set_t avail_out;
397 /* The ANTIC_IN set, which represents which values are anticipatable
398 in a given basic block. */
399 bitmap_set_t antic_in;
401 /* The PA_IN set, which represents which values are
402 partially anticipatable in a given basic block. */
403 bitmap_set_t pa_in;
405 /* The NEW_SETS set, which is used during insertion to augment the
406 AVAIL_OUT set of blocks with the new insertions performed during
407 the current iteration. */
408 bitmap_set_t new_sets;
410 /* A cache for value_dies_in_block_x. */
411 bitmap expr_dies;
413 /* The live virtual operand on successor edges. */
414 tree vop_on_exit;
416 /* True if we have visited this block during ANTIC calculation. */
417 unsigned int visited : 1;
419 /* True when the block contains a call that might not return. */
420 unsigned int contains_may_not_return_call : 1;
421 } *bb_value_sets_t;
423 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
424 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
425 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
426 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
427 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
428 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
429 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
430 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
431 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
432 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
433 #define BB_LIVE_VOP_ON_EXIT(BB) ((bb_value_sets_t) ((BB)->aux))->vop_on_exit
436 /* This structure is used to keep track of statistics on what
437 optimization PRE was able to perform. */
438 static struct
440 /* The number of RHS computations eliminated by PRE. */
441 int eliminations;
443 /* The number of new expressions/temporaries generated by PRE. */
444 int insertions;
446 /* The number of inserts found due to partial anticipation */
447 int pa_insert;
449 /* The number of new PHI nodes added by PRE. */
450 int phis;
451 } pre_stats;
453 static bool do_partial_partial;
454 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int);
455 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
456 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
457 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
458 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
459 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
460 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
461 unsigned int, bool);
462 static bitmap_set_t bitmap_set_new (void);
463 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
464 tree);
465 static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
466 static unsigned int get_expr_value_id (pre_expr);
468 /* We can add and remove elements and entries to and from sets
469 and hash tables, so we use alloc pools for them. */
471 static object_allocator<bitmap_set> bitmap_set_pool ("Bitmap sets");
472 static bitmap_obstack grand_bitmap_obstack;
474 /* Set of blocks with statements that have had their EH properties changed. */
475 static bitmap need_eh_cleanup;
477 /* Set of blocks with statements that have had their AB properties changed. */
478 static bitmap need_ab_cleanup;
480 /* A three tuple {e, pred, v} used to cache phi translations in the
481 phi_translate_table. */
483 typedef struct expr_pred_trans_d : free_ptr_hash<expr_pred_trans_d>
485 /* The expression. */
486 pre_expr e;
488 /* The predecessor block along which we translated the expression. */
489 basic_block pred;
491 /* The value that resulted from the translation. */
492 pre_expr v;
494 /* The hashcode for the expression, pred pair. This is cached for
495 speed reasons. */
496 hashval_t hashcode;
498 /* hash_table support. */
499 static inline hashval_t hash (const expr_pred_trans_d *);
500 static inline int equal (const expr_pred_trans_d *, const expr_pred_trans_d *);
501 } *expr_pred_trans_t;
502 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
504 inline hashval_t
505 expr_pred_trans_d::hash (const expr_pred_trans_d *e)
507 return e->hashcode;
510 inline int
511 expr_pred_trans_d::equal (const expr_pred_trans_d *ve1,
512 const expr_pred_trans_d *ve2)
514 basic_block b1 = ve1->pred;
515 basic_block b2 = ve2->pred;
517 /* If they are not translations for the same basic block, they can't
518 be equal. */
519 if (b1 != b2)
520 return false;
521 return pre_expr_d::equal (ve1->e, ve2->e);
524 /* The phi_translate_table caches phi translations for a given
525 expression and predecessor. */
526 static hash_table<expr_pred_trans_d> *phi_translate_table;
528 /* Add the tuple mapping from {expression E, basic block PRED} to
529 the phi translation table and return whether it pre-existed. */
531 static inline bool
532 phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred)
534 expr_pred_trans_t *slot;
535 expr_pred_trans_d tem;
536 hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e),
537 pred->index);
538 tem.e = e;
539 tem.pred = pred;
540 tem.hashcode = hash;
541 slot = phi_translate_table->find_slot_with_hash (&tem, hash, INSERT);
542 if (*slot)
544 *entry = *slot;
545 return true;
548 *entry = *slot = XNEW (struct expr_pred_trans_d);
549 (*entry)->e = e;
550 (*entry)->pred = pred;
551 (*entry)->hashcode = hash;
552 return false;
556 /* Add expression E to the expression set of value id V. */
558 static void
559 add_to_value (unsigned int v, pre_expr e)
561 bitmap set;
563 gcc_checking_assert (get_expr_value_id (e) == v);
565 if (v >= value_expressions.length ())
567 value_expressions.safe_grow_cleared (v + 1);
570 set = value_expressions[v];
571 if (!set)
573 set = BITMAP_ALLOC (&grand_bitmap_obstack);
574 value_expressions[v] = set;
577 bitmap_set_bit (set, get_or_alloc_expression_id (e));
580 /* Create a new bitmap set and return it. */
582 static bitmap_set_t
583 bitmap_set_new (void)
585 bitmap_set_t ret = bitmap_set_pool.allocate ();
586 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
587 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
588 return ret;
591 /* Return the value id for a PRE expression EXPR. */
593 static unsigned int
594 get_expr_value_id (pre_expr expr)
596 unsigned int id;
597 switch (expr->kind)
599 case CONSTANT:
600 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
601 break;
602 case NAME:
603 id = VN_INFO (PRE_EXPR_NAME (expr))->value_id;
604 break;
605 case NARY:
606 id = PRE_EXPR_NARY (expr)->value_id;
607 break;
608 case REFERENCE:
609 id = PRE_EXPR_REFERENCE (expr)->value_id;
610 break;
611 default:
612 gcc_unreachable ();
614 /* ??? We cannot assert that expr has a value-id (it can be 0), because
615 we assign value-ids only to expressions that have a result
616 in set_hashtable_value_ids. */
617 return id;
620 /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */
622 static tree
623 sccvn_valnum_from_value_id (unsigned int val)
625 bitmap_iterator bi;
626 unsigned int i;
627 bitmap exprset = value_expressions[val];
628 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
630 pre_expr vexpr = expression_for_id (i);
631 if (vexpr->kind == NAME)
632 return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
633 else if (vexpr->kind == CONSTANT)
634 return PRE_EXPR_CONSTANT (vexpr);
636 return NULL_TREE;
639 /* Remove an expression EXPR from a bitmapped set. */
641 static void
642 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
644 unsigned int val = get_expr_value_id (expr);
645 if (!value_id_constant_p (val))
647 bitmap_clear_bit (&set->values, val);
648 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
652 static void
653 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
654 unsigned int val, bool allow_constants)
656 if (allow_constants || !value_id_constant_p (val))
658 /* We specifically expect this and only this function to be able to
659 insert constants into a set. */
660 bitmap_set_bit (&set->values, val);
661 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
665 /* Insert an expression EXPR into a bitmapped set. */
667 static void
668 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
670 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
673 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
675 static void
676 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
678 bitmap_copy (&dest->expressions, &orig->expressions);
679 bitmap_copy (&dest->values, &orig->values);
683 /* Free memory used up by SET. */
684 static void
685 bitmap_set_free (bitmap_set_t set)
687 bitmap_clear (&set->expressions);
688 bitmap_clear (&set->values);
692 /* Generate an topological-ordered array of bitmap set SET. */
694 static vec<pre_expr>
695 sorted_array_from_bitmap_set (bitmap_set_t set)
697 unsigned int i, j;
698 bitmap_iterator bi, bj;
699 vec<pre_expr> result;
701 /* Pre-allocate enough space for the array. */
702 result.create (bitmap_count_bits (&set->expressions));
704 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
706 /* The number of expressions having a given value is usually
707 relatively small. Thus, rather than making a vector of all
708 the expressions and sorting it by value-id, we walk the values
709 and check in the reverse mapping that tells us what expressions
710 have a given value, to filter those in our set. As a result,
711 the expressions are inserted in value-id order, which means
712 topological order.
714 If this is somehow a significant lose for some cases, we can
715 choose which set to walk based on the set size. */
716 bitmap exprset = value_expressions[i];
717 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj)
719 if (bitmap_bit_p (&set->expressions, j))
720 result.quick_push (expression_for_id (j));
724 return result;
727 /* Perform bitmapped set operation DEST &= ORIG. */
729 static void
730 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
732 bitmap_iterator bi;
733 unsigned int i;
735 if (dest != orig)
737 bitmap_head temp;
738 bitmap_initialize (&temp, &grand_bitmap_obstack);
740 bitmap_and_into (&dest->values, &orig->values);
741 bitmap_copy (&temp, &dest->expressions);
742 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
744 pre_expr expr = expression_for_id (i);
745 unsigned int value_id = get_expr_value_id (expr);
746 if (!bitmap_bit_p (&dest->values, value_id))
747 bitmap_clear_bit (&dest->expressions, i);
749 bitmap_clear (&temp);
753 /* Subtract all values and expressions contained in ORIG from DEST. */
755 static bitmap_set_t
756 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
758 bitmap_set_t result = bitmap_set_new ();
759 bitmap_iterator bi;
760 unsigned int i;
762 bitmap_and_compl (&result->expressions, &dest->expressions,
763 &orig->expressions);
765 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
767 pre_expr expr = expression_for_id (i);
768 unsigned int value_id = get_expr_value_id (expr);
769 bitmap_set_bit (&result->values, value_id);
772 return result;
775 /* Subtract all the values in bitmap set B from bitmap set A. */
777 static void
778 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
780 unsigned int i;
781 bitmap_iterator bi;
782 bitmap_head temp;
784 bitmap_initialize (&temp, &grand_bitmap_obstack);
786 bitmap_copy (&temp, &a->expressions);
787 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
789 pre_expr expr = expression_for_id (i);
790 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
791 bitmap_remove_from_set (a, expr);
793 bitmap_clear (&temp);
797 /* Return true if bitmapped set SET contains the value VALUE_ID. */
799 static bool
800 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
802 if (value_id_constant_p (value_id))
803 return true;
805 if (!set || bitmap_empty_p (&set->expressions))
806 return false;
808 return bitmap_bit_p (&set->values, value_id);
811 static inline bool
812 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
814 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
817 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
819 static void
820 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
821 const pre_expr expr)
823 bitmap exprset;
824 unsigned int i;
825 bitmap_iterator bi;
827 if (value_id_constant_p (lookfor))
828 return;
830 if (!bitmap_set_contains_value (set, lookfor))
831 return;
833 /* The number of expressions having a given value is usually
834 significantly less than the total number of expressions in SET.
835 Thus, rather than check, for each expression in SET, whether it
836 has the value LOOKFOR, we walk the reverse mapping that tells us
837 what expressions have a given value, and see if any of those
838 expressions are in our set. For large testcases, this is about
839 5-10x faster than walking the bitmap. If this is somehow a
840 significant lose for some cases, we can choose which set to walk
841 based on the set size. */
842 exprset = value_expressions[lookfor];
843 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
845 if (bitmap_clear_bit (&set->expressions, i))
847 bitmap_set_bit (&set->expressions, get_expression_id (expr));
848 return;
852 gcc_unreachable ();
855 /* Return true if two bitmap sets are equal. */
857 static bool
858 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
860 return bitmap_equal_p (&a->values, &b->values);
863 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
864 and add it otherwise. */
866 static void
867 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
869 unsigned int val = get_expr_value_id (expr);
871 if (bitmap_set_contains_value (set, val))
872 bitmap_set_replace_value (set, val, expr);
873 else
874 bitmap_insert_into_set (set, expr);
877 /* Insert EXPR into SET if EXPR's value is not already present in
878 SET. */
880 static void
881 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
883 unsigned int val = get_expr_value_id (expr);
885 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
887 /* Constant values are always considered to be part of the set. */
888 if (value_id_constant_p (val))
889 return;
891 /* If the value membership changed, add the expression. */
892 if (bitmap_set_bit (&set->values, val))
893 bitmap_set_bit (&set->expressions, expr->id);
896 /* Print out EXPR to outfile. */
898 static void
899 print_pre_expr (FILE *outfile, const pre_expr expr)
901 switch (expr->kind)
903 case CONSTANT:
904 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
905 break;
906 case NAME:
907 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
908 break;
909 case NARY:
911 unsigned int i;
912 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
913 fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode));
914 for (i = 0; i < nary->length; i++)
916 print_generic_expr (outfile, nary->op[i], 0);
917 if (i != (unsigned) nary->length - 1)
918 fprintf (outfile, ",");
920 fprintf (outfile, "}");
922 break;
924 case REFERENCE:
926 vn_reference_op_t vro;
927 unsigned int i;
928 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
929 fprintf (outfile, "{");
930 for (i = 0;
931 ref->operands.iterate (i, &vro);
932 i++)
934 bool closebrace = false;
935 if (vro->opcode != SSA_NAME
936 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
938 fprintf (outfile, "%s", get_tree_code_name (vro->opcode));
939 if (vro->op0)
941 fprintf (outfile, "<");
942 closebrace = true;
945 if (vro->op0)
947 print_generic_expr (outfile, vro->op0, 0);
948 if (vro->op1)
950 fprintf (outfile, ",");
951 print_generic_expr (outfile, vro->op1, 0);
953 if (vro->op2)
955 fprintf (outfile, ",");
956 print_generic_expr (outfile, vro->op2, 0);
959 if (closebrace)
960 fprintf (outfile, ">");
961 if (i != ref->operands.length () - 1)
962 fprintf (outfile, ",");
964 fprintf (outfile, "}");
965 if (ref->vuse)
967 fprintf (outfile, "@");
968 print_generic_expr (outfile, ref->vuse, 0);
971 break;
974 void debug_pre_expr (pre_expr);
976 /* Like print_pre_expr but always prints to stderr. */
977 DEBUG_FUNCTION void
978 debug_pre_expr (pre_expr e)
980 print_pre_expr (stderr, e);
981 fprintf (stderr, "\n");
984 /* Print out SET to OUTFILE. */
986 static void
987 print_bitmap_set (FILE *outfile, bitmap_set_t set,
988 const char *setname, int blockindex)
990 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
991 if (set)
993 bool first = true;
994 unsigned i;
995 bitmap_iterator bi;
997 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
999 const pre_expr expr = expression_for_id (i);
1001 if (!first)
1002 fprintf (outfile, ", ");
1003 first = false;
1004 print_pre_expr (outfile, expr);
1006 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1009 fprintf (outfile, " }\n");
1012 void debug_bitmap_set (bitmap_set_t);
1014 DEBUG_FUNCTION void
1015 debug_bitmap_set (bitmap_set_t set)
1017 print_bitmap_set (stderr, set, "debug", 0);
1020 void debug_bitmap_sets_for (basic_block);
1022 DEBUG_FUNCTION void
1023 debug_bitmap_sets_for (basic_block bb)
1025 print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1026 print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1027 print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1028 print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1029 print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1030 if (do_partial_partial)
1031 print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1032 print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1035 /* Print out the expressions that have VAL to OUTFILE. */
1037 static void
1038 print_value_expressions (FILE *outfile, unsigned int val)
1040 bitmap set = value_expressions[val];
1041 if (set)
1043 bitmap_set x;
1044 char s[10];
1045 sprintf (s, "%04d", val);
1046 x.expressions = *set;
1047 print_bitmap_set (outfile, &x, s, 0);
1052 DEBUG_FUNCTION void
1053 debug_value_expressions (unsigned int val)
1055 print_value_expressions (stderr, val);
1058 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1059 represent it. */
1061 static pre_expr
1062 get_or_alloc_expr_for_constant (tree constant)
1064 unsigned int result_id;
1065 unsigned int value_id;
1066 struct pre_expr_d expr;
1067 pre_expr newexpr;
1069 expr.kind = CONSTANT;
1070 PRE_EXPR_CONSTANT (&expr) = constant;
1071 result_id = lookup_expression_id (&expr);
1072 if (result_id != 0)
1073 return expression_for_id (result_id);
1075 newexpr = pre_expr_pool.allocate ();
1076 newexpr->kind = CONSTANT;
1077 PRE_EXPR_CONSTANT (newexpr) = constant;
1078 alloc_expression_id (newexpr);
1079 value_id = get_or_alloc_constant_value_id (constant);
1080 add_to_value (value_id, newexpr);
1081 return newexpr;
1084 /* Given a value id V, find the actual tree representing the constant
1085 value if there is one, and return it. Return NULL if we can't find
1086 a constant. */
1088 static tree
1089 get_constant_for_value_id (unsigned int v)
1091 if (value_id_constant_p (v))
1093 unsigned int i;
1094 bitmap_iterator bi;
1095 bitmap exprset = value_expressions[v];
1097 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1099 pre_expr expr = expression_for_id (i);
1100 if (expr->kind == CONSTANT)
1101 return PRE_EXPR_CONSTANT (expr);
1104 return NULL;
1107 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1108 Currently only supports constants and SSA_NAMES. */
1109 static pre_expr
1110 get_or_alloc_expr_for (tree t)
1112 if (TREE_CODE (t) == SSA_NAME)
1113 return get_or_alloc_expr_for_name (t);
1114 else if (is_gimple_min_invariant (t))
1115 return get_or_alloc_expr_for_constant (t);
1116 else
1118 /* More complex expressions can result from SCCVN expression
1119 simplification that inserts values for them. As they all
1120 do not have VOPs the get handled by the nary ops struct. */
1121 vn_nary_op_t result;
1122 unsigned int result_id;
1123 vn_nary_op_lookup (t, &result);
1124 if (result != NULL)
1126 pre_expr e = pre_expr_pool.allocate ();
1127 e->kind = NARY;
1128 PRE_EXPR_NARY (e) = result;
1129 result_id = lookup_expression_id (e);
1130 if (result_id != 0)
1132 pre_expr_pool.remove (e);
1133 e = expression_for_id (result_id);
1134 return e;
1136 alloc_expression_id (e);
1137 return e;
1140 return NULL;
1143 /* Return the folded version of T if T, when folded, is a gimple
1144 min_invariant. Otherwise, return T. */
1146 static pre_expr
1147 fully_constant_expression (pre_expr e)
1149 switch (e->kind)
1151 case CONSTANT:
1152 return e;
1153 case NARY:
1155 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1156 switch (TREE_CODE_CLASS (nary->opcode))
1158 case tcc_binary:
1159 case tcc_comparison:
1161 /* We have to go from trees to pre exprs to value ids to
1162 constants. */
1163 tree naryop0 = nary->op[0];
1164 tree naryop1 = nary->op[1];
1165 tree result;
1166 if (!is_gimple_min_invariant (naryop0))
1168 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1169 unsigned int vrep0 = get_expr_value_id (rep0);
1170 tree const0 = get_constant_for_value_id (vrep0);
1171 if (const0)
1172 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1174 if (!is_gimple_min_invariant (naryop1))
1176 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1177 unsigned int vrep1 = get_expr_value_id (rep1);
1178 tree const1 = get_constant_for_value_id (vrep1);
1179 if (const1)
1180 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1182 result = fold_binary (nary->opcode, nary->type,
1183 naryop0, naryop1);
1184 if (result && is_gimple_min_invariant (result))
1185 return get_or_alloc_expr_for_constant (result);
1186 /* We might have simplified the expression to a
1187 SSA_NAME for example from x_1 * 1. But we cannot
1188 insert a PHI for x_1 unconditionally as x_1 might
1189 not be available readily. */
1190 return e;
1192 case tcc_reference:
1193 if (nary->opcode != REALPART_EXPR
1194 && nary->opcode != IMAGPART_EXPR
1195 && nary->opcode != VIEW_CONVERT_EXPR)
1196 return e;
1197 /* Fallthrough. */
1198 case tcc_unary:
1200 /* We have to go from trees to pre exprs to value ids to
1201 constants. */
1202 tree naryop0 = nary->op[0];
1203 tree const0, result;
1204 if (is_gimple_min_invariant (naryop0))
1205 const0 = naryop0;
1206 else
1208 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1209 unsigned int vrep0 = get_expr_value_id (rep0);
1210 const0 = get_constant_for_value_id (vrep0);
1212 result = NULL;
1213 if (const0)
1215 tree type1 = TREE_TYPE (nary->op[0]);
1216 const0 = fold_convert (type1, const0);
1217 result = fold_unary (nary->opcode, nary->type, const0);
1219 if (result && is_gimple_min_invariant (result))
1220 return get_or_alloc_expr_for_constant (result);
1221 return e;
1223 default:
1224 return e;
1227 case REFERENCE:
1229 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1230 tree folded;
1231 if ((folded = fully_constant_vn_reference_p (ref)))
1232 return get_or_alloc_expr_for_constant (folded);
1233 return e;
1235 default:
1236 return e;
1238 return e;
1241 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1242 it has the value it would have in BLOCK. Set *SAME_VALID to true
1243 in case the new vuse doesn't change the value id of the OPERANDS. */
1245 static tree
1246 translate_vuse_through_block (vec<vn_reference_op_s> operands,
1247 alias_set_type set, tree type, tree vuse,
1248 basic_block phiblock,
1249 basic_block block, bool *same_valid)
1251 gimple *phi = SSA_NAME_DEF_STMT (vuse);
1252 ao_ref ref;
1253 edge e = NULL;
1254 bool use_oracle;
1256 *same_valid = true;
1258 if (gimple_bb (phi) != phiblock)
1259 return vuse;
1261 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1263 /* Use the alias-oracle to find either the PHI node in this block,
1264 the first VUSE used in this block that is equivalent to vuse or
1265 the first VUSE which definition in this block kills the value. */
1266 if (gimple_code (phi) == GIMPLE_PHI)
1267 e = find_edge (block, phiblock);
1268 else if (use_oracle)
1269 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1271 vuse = gimple_vuse (phi);
1272 phi = SSA_NAME_DEF_STMT (vuse);
1273 if (gimple_bb (phi) != phiblock)
1274 return vuse;
1275 if (gimple_code (phi) == GIMPLE_PHI)
1277 e = find_edge (block, phiblock);
1278 break;
1281 else
1282 return NULL_TREE;
1284 if (e)
1286 if (use_oracle)
1288 bitmap visited = NULL;
1289 unsigned int cnt;
1290 /* Try to find a vuse that dominates this phi node by skipping
1291 non-clobbering statements. */
1292 vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false,
1293 NULL, NULL);
1294 if (visited)
1295 BITMAP_FREE (visited);
1297 else
1298 vuse = NULL_TREE;
1299 if (!vuse)
1301 /* If we didn't find any, the value ID can't stay the same,
1302 but return the translated vuse. */
1303 *same_valid = false;
1304 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1306 /* ??? We would like to return vuse here as this is the canonical
1307 upmost vdef that this reference is associated with. But during
1308 insertion of the references into the hash tables we only ever
1309 directly insert with their direct gimple_vuse, hence returning
1310 something else would make us not find the other expression. */
1311 return PHI_ARG_DEF (phi, e->dest_idx);
1314 return NULL_TREE;
1317 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1318 SET2. This is used to avoid making a set consisting of the union
1319 of PA_IN and ANTIC_IN during insert. */
1321 static inline pre_expr
1322 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1324 pre_expr result;
1326 result = bitmap_find_leader (set1, val);
1327 if (!result && set2)
1328 result = bitmap_find_leader (set2, val);
1329 return result;
1332 /* Get the tree type for our PRE expression e. */
1334 static tree
1335 get_expr_type (const pre_expr e)
1337 switch (e->kind)
1339 case NAME:
1340 return TREE_TYPE (PRE_EXPR_NAME (e));
1341 case CONSTANT:
1342 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1343 case REFERENCE:
1344 return PRE_EXPR_REFERENCE (e)->type;
1345 case NARY:
1346 return PRE_EXPR_NARY (e)->type;
1348 gcc_unreachable ();
1351 /* Get a representative SSA_NAME for a given expression.
1352 Since all of our sub-expressions are treated as values, we require
1353 them to be SSA_NAME's for simplicity.
1354 Prior versions of GVNPRE used to use "value handles" here, so that
1355 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1356 either case, the operands are really values (IE we do not expect
1357 them to be usable without finding leaders). */
1359 static tree
1360 get_representative_for (const pre_expr e)
1362 tree name;
1363 unsigned int value_id = get_expr_value_id (e);
1365 switch (e->kind)
1367 case NAME:
1368 return PRE_EXPR_NAME (e);
1369 case CONSTANT:
1370 return PRE_EXPR_CONSTANT (e);
1371 case NARY:
1372 case REFERENCE:
1374 /* Go through all of the expressions representing this value
1375 and pick out an SSA_NAME. */
1376 unsigned int i;
1377 bitmap_iterator bi;
1378 bitmap exprs = value_expressions[value_id];
1379 EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1381 pre_expr rep = expression_for_id (i);
1382 if (rep->kind == NAME)
1383 return PRE_EXPR_NAME (rep);
1384 else if (rep->kind == CONSTANT)
1385 return PRE_EXPR_CONSTANT (rep);
1388 break;
1391 /* If we reached here we couldn't find an SSA_NAME. This can
1392 happen when we've discovered a value that has never appeared in
1393 the program as set to an SSA_NAME, as the result of phi translation.
1394 Create one here.
1395 ??? We should be able to re-use this when we insert the statement
1396 to compute it. */
1397 name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1398 VN_INFO_GET (name)->value_id = value_id;
1399 VN_INFO (name)->valnum = name;
1400 /* ??? For now mark this SSA name for release by SCCVN. */
1401 VN_INFO (name)->needs_insertion = true;
1402 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1403 if (dump_file && (dump_flags & TDF_DETAILS))
1405 fprintf (dump_file, "Created SSA_NAME representative ");
1406 print_generic_expr (dump_file, name, 0);
1407 fprintf (dump_file, " for expression:");
1408 print_pre_expr (dump_file, e);
1409 fprintf (dump_file, " (%04d)\n", value_id);
1412 return name;
1417 static pre_expr
1418 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1419 basic_block pred, basic_block phiblock);
1421 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1422 the phis in PRED. Return NULL if we can't find a leader for each part
1423 of the translated expression. */
1425 static pre_expr
1426 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1427 basic_block pred, basic_block phiblock)
1429 switch (expr->kind)
1431 case NARY:
1433 unsigned int i;
1434 bool changed = false;
1435 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1436 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1437 sizeof_vn_nary_op (nary->length));
1438 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1440 for (i = 0; i < newnary->length; i++)
1442 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1443 continue;
1444 else
1446 pre_expr leader, result;
1447 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1448 leader = find_leader_in_sets (op_val_id, set1, set2);
1449 result = phi_translate (leader, set1, set2, pred, phiblock);
1450 if (result && result != leader)
1452 tree name = get_representative_for (result);
1453 if (!name)
1454 return NULL;
1455 newnary->op[i] = name;
1457 else if (!result)
1458 return NULL;
1460 changed |= newnary->op[i] != nary->op[i];
1463 if (changed)
1465 pre_expr constant;
1466 unsigned int new_val_id;
1468 PRE_EXPR_NARY (expr) = newnary;
1469 constant = fully_constant_expression (expr);
1470 PRE_EXPR_NARY (expr) = nary;
1471 if (constant != expr)
1472 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);
1491 else
1493 new_val_id = get_next_value_id ();
1494 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
1495 nary = vn_nary_op_insert_pieces (newnary->length,
1496 newnary->opcode,
1497 newnary->type,
1498 &newnary->op[0],
1499 result, new_val_id);
1500 PRE_EXPR_NARY (expr) = nary;
1501 get_or_alloc_expression_id (expr);
1503 add_to_value (new_val_id, expr);
1505 return expr;
1507 break;
1509 case REFERENCE:
1511 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1512 vec<vn_reference_op_s> operands = ref->operands;
1513 tree vuse = ref->vuse;
1514 tree newvuse = vuse;
1515 vec<vn_reference_op_s> newoperands = vNULL;
1516 bool changed = false, same_valid = true;
1517 unsigned int i, n;
1518 vn_reference_op_t operand;
1519 vn_reference_t newref;
1521 for (i = 0; operands.iterate (i, &operand); i++)
1523 pre_expr opresult;
1524 pre_expr leader;
1525 tree op[3];
1526 tree type = operand->type;
1527 vn_reference_op_s newop = *operand;
1528 op[0] = operand->op0;
1529 op[1] = operand->op1;
1530 op[2] = operand->op2;
1531 for (n = 0; n < 3; ++n)
1533 unsigned int op_val_id;
1534 if (!op[n])
1535 continue;
1536 if (TREE_CODE (op[n]) != SSA_NAME)
1538 /* We can't possibly insert these. */
1539 if (n != 0
1540 && !is_gimple_min_invariant (op[n]))
1541 break;
1542 continue;
1544 op_val_id = VN_INFO (op[n])->value_id;
1545 leader = find_leader_in_sets (op_val_id, set1, set2);
1546 if (!leader)
1547 break;
1548 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1549 if (!opresult)
1550 break;
1551 if (opresult != leader)
1553 tree name = get_representative_for (opresult);
1554 if (!name)
1555 break;
1556 changed |= name != op[n];
1557 op[n] = name;
1560 if (n != 3)
1562 newoperands.release ();
1563 return NULL;
1565 if (!changed)
1566 continue;
1567 if (!newoperands.exists ())
1568 newoperands = operands.copy ();
1569 /* We may have changed from an SSA_NAME to a constant */
1570 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1571 newop.opcode = TREE_CODE (op[0]);
1572 newop.type = type;
1573 newop.op0 = op[0];
1574 newop.op1 = op[1];
1575 newop.op2 = op[2];
1576 newoperands[i] = newop;
1578 gcc_checking_assert (i == operands.length ());
1580 if (vuse)
1582 newvuse = translate_vuse_through_block (newoperands.exists ()
1583 ? newoperands : operands,
1584 ref->set, ref->type,
1585 vuse, phiblock, pred,
1586 &same_valid);
1587 if (newvuse == NULL_TREE)
1589 newoperands.release ();
1590 return NULL;
1594 if (changed || newvuse != vuse)
1596 unsigned int new_val_id;
1597 pre_expr constant;
1599 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1600 ref->type,
1601 newoperands.exists ()
1602 ? newoperands : operands,
1603 &newref, VN_WALK);
1604 if (result)
1605 newoperands.release ();
1607 /* We can always insert constants, so if we have a partial
1608 redundant constant load of another type try to translate it
1609 to a constant of appropriate type. */
1610 if (result && is_gimple_min_invariant (result))
1612 tree tem = result;
1613 if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1615 tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1616 if (tem && !is_gimple_min_invariant (tem))
1617 tem = NULL_TREE;
1619 if (tem)
1620 return get_or_alloc_expr_for_constant (tem);
1623 /* If we'd have to convert things we would need to validate
1624 if we can insert the translated expression. So fail
1625 here for now - we cannot insert an alias with a different
1626 type in the VN tables either, as that would assert. */
1627 if (result
1628 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1629 return NULL;
1630 else if (!result && newref
1631 && !useless_type_conversion_p (ref->type, newref->type))
1633 newoperands.release ();
1634 return NULL;
1637 expr = pre_expr_pool.allocate ();
1638 expr->kind = REFERENCE;
1639 expr->id = 0;
1641 if (newref)
1643 PRE_EXPR_REFERENCE (expr) = newref;
1644 constant = fully_constant_expression (expr);
1645 if (constant != expr)
1646 return constant;
1648 new_val_id = newref->value_id;
1649 get_or_alloc_expression_id (expr);
1651 else
1653 if (changed || !same_valid)
1655 new_val_id = get_next_value_id ();
1656 value_expressions.safe_grow_cleared
1657 (get_max_value_id () + 1);
1659 else
1660 new_val_id = ref->value_id;
1661 if (!newoperands.exists ())
1662 newoperands = operands.copy ();
1663 newref = vn_reference_insert_pieces (newvuse, ref->set,
1664 ref->type,
1665 newoperands,
1666 result, new_val_id);
1667 newoperands = vNULL;
1668 PRE_EXPR_REFERENCE (expr) = newref;
1669 constant = fully_constant_expression (expr);
1670 if (constant != expr)
1671 return constant;
1672 get_or_alloc_expression_id (expr);
1674 add_to_value (new_val_id, expr);
1676 newoperands.release ();
1677 return expr;
1679 break;
1681 case NAME:
1683 tree name = PRE_EXPR_NAME (expr);
1684 gimple *def_stmt = SSA_NAME_DEF_STMT (name);
1685 /* If the SSA name is defined by a PHI node in this block,
1686 translate it. */
1687 if (gimple_code (def_stmt) == GIMPLE_PHI
1688 && gimple_bb (def_stmt) == phiblock)
1690 edge e = find_edge (pred, gimple_bb (def_stmt));
1691 tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1693 /* Handle constant. */
1694 if (is_gimple_min_invariant (def))
1695 return get_or_alloc_expr_for_constant (def);
1697 return get_or_alloc_expr_for_name (def);
1699 /* Otherwise return it unchanged - it will get removed if its
1700 value is not available in PREDs AVAIL_OUT set of expressions
1701 by the subtraction of TMP_GEN. */
1702 return expr;
1705 default:
1706 gcc_unreachable ();
1710 /* Wrapper around phi_translate_1 providing caching functionality. */
1712 static pre_expr
1713 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1714 basic_block pred, basic_block phiblock)
1716 expr_pred_trans_t slot = NULL;
1717 pre_expr phitrans;
1719 if (!expr)
1720 return NULL;
1722 /* Constants contain no values that need translation. */
1723 if (expr->kind == CONSTANT)
1724 return expr;
1726 if (value_id_constant_p (get_expr_value_id (expr)))
1727 return expr;
1729 /* Don't add translations of NAMEs as those are cheap to translate. */
1730 if (expr->kind != NAME)
1732 if (phi_trans_add (&slot, expr, pred))
1733 return slot->v;
1734 /* Store NULL for the value we want to return in the case of
1735 recursing. */
1736 slot->v = NULL;
1739 /* Translate. */
1740 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1742 if (slot)
1744 if (phitrans)
1745 slot->v = phitrans;
1746 else
1747 /* Remove failed translations again, they cause insert
1748 iteration to not pick up new opportunities reliably. */
1749 phi_translate_table->remove_elt_with_hash (slot, slot->hashcode);
1752 return phitrans;
1756 /* For each expression in SET, translate the values through phi nodes
1757 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1758 expressions in DEST. */
1760 static void
1761 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1762 basic_block phiblock)
1764 vec<pre_expr> exprs;
1765 pre_expr expr;
1766 int i;
1768 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1770 bitmap_set_copy (dest, set);
1771 return;
1774 exprs = sorted_array_from_bitmap_set (set);
1775 FOR_EACH_VEC_ELT (exprs, i, expr)
1777 pre_expr translated;
1778 translated = phi_translate (expr, set, NULL, pred, phiblock);
1779 if (!translated)
1780 continue;
1782 /* We might end up with multiple expressions from SET being
1783 translated to the same value. In this case we do not want
1784 to retain the NARY or REFERENCE expression but prefer a NAME
1785 which would be the leader. */
1786 if (translated->kind == NAME)
1787 bitmap_value_replace_in_set (dest, translated);
1788 else
1789 bitmap_value_insert_into_set (dest, translated);
1791 exprs.release ();
1794 /* Find the leader for a value (i.e., the name representing that
1795 value) in a given set, and return it. Return NULL if no leader
1796 is found. */
1798 static pre_expr
1799 bitmap_find_leader (bitmap_set_t set, unsigned int val)
1801 if (value_id_constant_p (val))
1803 unsigned int i;
1804 bitmap_iterator bi;
1805 bitmap exprset = value_expressions[val];
1807 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1809 pre_expr expr = expression_for_id (i);
1810 if (expr->kind == CONSTANT)
1811 return expr;
1814 if (bitmap_set_contains_value (set, val))
1816 /* Rather than walk the entire bitmap of expressions, and see
1817 whether any of them has the value we are looking for, we look
1818 at the reverse mapping, which tells us the set of expressions
1819 that have a given value (IE value->expressions with that
1820 value) and see if any of those expressions are in our set.
1821 The number of expressions per value is usually significantly
1822 less than the number of expressions in the set. In fact, for
1823 large testcases, doing it this way is roughly 5-10x faster
1824 than walking the bitmap.
1825 If this is somehow a significant lose for some cases, we can
1826 choose which set to walk based on which set is smaller. */
1827 unsigned int i;
1828 bitmap_iterator bi;
1829 bitmap exprset = value_expressions[val];
1831 EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1832 return expression_for_id (i);
1834 return NULL;
1837 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1838 BLOCK by seeing if it is not killed in the block. Note that we are
1839 only determining whether there is a store that kills it. Because
1840 of the order in which clean iterates over values, we are guaranteed
1841 that altered operands will have caused us to be eliminated from the
1842 ANTIC_IN set already. */
1844 static bool
1845 value_dies_in_block_x (pre_expr expr, basic_block block)
1847 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1848 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1849 gimple *def;
1850 gimple_stmt_iterator gsi;
1851 unsigned id = get_expression_id (expr);
1852 bool res = false;
1853 ao_ref ref;
1855 if (!vuse)
1856 return false;
1858 /* Lookup a previously calculated result. */
1859 if (EXPR_DIES (block)
1860 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1861 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1863 /* A memory expression {e, VUSE} dies in the block if there is a
1864 statement that may clobber e. If, starting statement walk from the
1865 top of the basic block, a statement uses VUSE there can be no kill
1866 inbetween that use and the original statement that loaded {e, VUSE},
1867 so we can stop walking. */
1868 ref.base = NULL_TREE;
1869 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1871 tree def_vuse, def_vdef;
1872 def = gsi_stmt (gsi);
1873 def_vuse = gimple_vuse (def);
1874 def_vdef = gimple_vdef (def);
1876 /* Not a memory statement. */
1877 if (!def_vuse)
1878 continue;
1880 /* Not a may-def. */
1881 if (!def_vdef)
1883 /* A load with the same VUSE, we're done. */
1884 if (def_vuse == vuse)
1885 break;
1887 continue;
1890 /* Init ref only if we really need it. */
1891 if (ref.base == NULL_TREE
1892 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1893 refx->operands))
1895 res = true;
1896 break;
1898 /* If the statement may clobber expr, it dies. */
1899 if (stmt_may_clobber_ref_p_1 (def, &ref))
1901 res = true;
1902 break;
1906 /* Remember the result. */
1907 if (!EXPR_DIES (block))
1908 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1909 bitmap_set_bit (EXPR_DIES (block), id * 2);
1910 if (res)
1911 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1913 return res;
1917 /* Determine if OP is valid in SET1 U SET2, which it is when the union
1918 contains its value-id. */
1920 static bool
1921 op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1923 if (op && TREE_CODE (op) == SSA_NAME)
1925 unsigned int value_id = VN_INFO (op)->value_id;
1926 if (!(bitmap_set_contains_value (set1, value_id)
1927 || (set2 && bitmap_set_contains_value (set2, value_id))))
1928 return false;
1930 return true;
1933 /* Determine if the expression EXPR is valid in SET1 U SET2.
1934 ONLY SET2 CAN BE NULL.
1935 This means that we have a leader for each part of the expression
1936 (if it consists of values), or the expression is an SSA_NAME.
1937 For loads/calls, we also see if the vuse is killed in this block. */
1939 static bool
1940 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr)
1942 switch (expr->kind)
1944 case NAME:
1945 /* By construction all NAMEs are available. Non-available
1946 NAMEs are removed by subtracting TMP_GEN from the sets. */
1947 return true;
1948 case NARY:
1950 unsigned int i;
1951 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1952 for (i = 0; i < nary->length; i++)
1953 if (!op_valid_in_sets (set1, set2, nary->op[i]))
1954 return false;
1955 return true;
1957 break;
1958 case REFERENCE:
1960 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1961 vn_reference_op_t vro;
1962 unsigned int i;
1964 FOR_EACH_VEC_ELT (ref->operands, i, vro)
1966 if (!op_valid_in_sets (set1, set2, vro->op0)
1967 || !op_valid_in_sets (set1, set2, vro->op1)
1968 || !op_valid_in_sets (set1, set2, vro->op2))
1969 return false;
1971 return true;
1973 default:
1974 gcc_unreachable ();
1978 /* Clean the set of expressions that are no longer valid in SET1 or
1979 SET2. This means expressions that are made up of values we have no
1980 leaders for in SET1 or SET2. This version is used for partial
1981 anticipation, which means it is not valid in either ANTIC_IN or
1982 PA_IN. */
1984 static void
1985 dependent_clean (bitmap_set_t set1, bitmap_set_t set2)
1987 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
1988 pre_expr expr;
1989 int i;
1991 FOR_EACH_VEC_ELT (exprs, i, expr)
1993 if (!valid_in_sets (set1, set2, expr))
1994 bitmap_remove_from_set (set1, expr);
1996 exprs.release ();
1999 /* Clean the set of expressions that are no longer valid in SET. This
2000 means expressions that are made up of values we have no leaders for
2001 in SET. */
2003 static void
2004 clean (bitmap_set_t set)
2006 vec<pre_expr> exprs = sorted_array_from_bitmap_set (set);
2007 pre_expr expr;
2008 int i;
2010 FOR_EACH_VEC_ELT (exprs, i, expr)
2012 if (!valid_in_sets (set, NULL, expr))
2013 bitmap_remove_from_set (set, expr);
2015 exprs.release ();
2018 /* Clean the set of expressions that are no longer valid in SET because
2019 they are clobbered in BLOCK or because they trap and may not be executed. */
2021 static void
2022 prune_clobbered_mems (bitmap_set_t set, basic_block block)
2024 bitmap_iterator bi;
2025 unsigned i;
2027 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2029 pre_expr expr = expression_for_id (i);
2030 if (expr->kind == REFERENCE)
2032 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2033 if (ref->vuse)
2035 gimple *def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2036 if (!gimple_nop_p (def_stmt)
2037 && ((gimple_bb (def_stmt) != block
2038 && !dominated_by_p (CDI_DOMINATORS,
2039 block, gimple_bb (def_stmt)))
2040 || (gimple_bb (def_stmt) == block
2041 && value_dies_in_block_x (expr, block))))
2042 bitmap_remove_from_set (set, expr);
2045 else if (expr->kind == NARY)
2047 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2048 /* If the NARY may trap make sure the block does not contain
2049 a possible exit point.
2050 ??? This is overly conservative if we translate AVAIL_OUT
2051 as the available expression might be after the exit point. */
2052 if (BB_MAY_NOTRETURN (block)
2053 && vn_nary_may_trap (nary))
2054 bitmap_remove_from_set (set, expr);
2059 static sbitmap has_abnormal_preds;
2061 /* Compute the ANTIC set for BLOCK.
2063 If succs(BLOCK) > 1 then
2064 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2065 else if succs(BLOCK) == 1 then
2066 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2068 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2071 static bool
2072 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2074 bool changed = false;
2075 bitmap_set_t S, old, ANTIC_OUT;
2076 bitmap_iterator bi;
2077 unsigned int bii;
2078 edge e;
2079 edge_iterator ei;
2080 bool was_visited = BB_VISITED (block);
2082 old = ANTIC_OUT = S = NULL;
2083 BB_VISITED (block) = 1;
2085 /* If any edges from predecessors are abnormal, antic_in is empty,
2086 so do nothing. */
2087 if (block_has_abnormal_pred_edge)
2088 goto maybe_dump_sets;
2090 old = ANTIC_IN (block);
2091 ANTIC_OUT = bitmap_set_new ();
2093 /* If the block has no successors, ANTIC_OUT is empty. */
2094 if (EDGE_COUNT (block->succs) == 0)
2096 /* If we have one successor, we could have some phi nodes to
2097 translate through. */
2098 else if (single_succ_p (block))
2100 basic_block succ_bb = single_succ (block);
2101 gcc_assert (BB_VISITED (succ_bb));
2102 phi_translate_set (ANTIC_OUT, ANTIC_IN (succ_bb), block, succ_bb);
2104 /* If we have multiple successors, we take the intersection of all of
2105 them. Note that in the case of loop exit phi nodes, we may have
2106 phis to translate through. */
2107 else
2109 size_t i;
2110 basic_block bprime, first = NULL;
2112 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2113 FOR_EACH_EDGE (e, ei, block->succs)
2115 if (!first
2116 && BB_VISITED (e->dest))
2117 first = e->dest;
2118 else if (BB_VISITED (e->dest))
2119 worklist.quick_push (e->dest);
2120 else
2122 /* Unvisited successors get their ANTIC_IN replaced by the
2123 maximal set to arrive at a maximum ANTIC_IN solution.
2124 We can ignore them in the intersection operation and thus
2125 need not explicitely represent that maximum solution. */
2126 if (dump_file && (dump_flags & TDF_DETAILS))
2127 fprintf (dump_file, "ANTIC_IN is MAX on %d->%d\n",
2128 e->src->index, e->dest->index);
2132 /* Of multiple successors we have to have visited one already
2133 which is guaranteed by iteration order. */
2134 gcc_assert (first != NULL);
2136 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2138 FOR_EACH_VEC_ELT (worklist, i, bprime)
2140 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2142 bitmap_set_t tmp = bitmap_set_new ();
2143 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2144 bitmap_set_and (ANTIC_OUT, tmp);
2145 bitmap_set_free (tmp);
2147 else
2148 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2152 /* Prune expressions that are clobbered in block and thus become
2153 invalid if translated from ANTIC_OUT to ANTIC_IN. */
2154 prune_clobbered_mems (ANTIC_OUT, block);
2156 /* Generate ANTIC_OUT - TMP_GEN. */
2157 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2159 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2160 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2161 TMP_GEN (block));
2163 /* Then union in the ANTIC_OUT - TMP_GEN values,
2164 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2165 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2166 bitmap_value_insert_into_set (ANTIC_IN (block),
2167 expression_for_id (bii));
2169 clean (ANTIC_IN (block));
2171 if (!was_visited || !bitmap_set_equal (old, ANTIC_IN (block)))
2172 changed = true;
2174 maybe_dump_sets:
2175 if (dump_file && (dump_flags & TDF_DETAILS))
2177 if (ANTIC_OUT)
2178 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2180 if (changed)
2181 fprintf (dump_file, "[changed] ");
2182 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2183 block->index);
2185 if (S)
2186 print_bitmap_set (dump_file, S, "S", block->index);
2188 if (old)
2189 bitmap_set_free (old);
2190 if (S)
2191 bitmap_set_free (S);
2192 if (ANTIC_OUT)
2193 bitmap_set_free (ANTIC_OUT);
2194 return changed;
2197 /* Compute PARTIAL_ANTIC for BLOCK.
2199 If succs(BLOCK) > 1 then
2200 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2201 in ANTIC_OUT for all succ(BLOCK)
2202 else if succs(BLOCK) == 1 then
2203 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2205 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2206 - ANTIC_IN[BLOCK])
2209 static void
2210 compute_partial_antic_aux (basic_block block,
2211 bool block_has_abnormal_pred_edge)
2213 bitmap_set_t old_PA_IN;
2214 bitmap_set_t PA_OUT;
2215 edge e;
2216 edge_iterator ei;
2217 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2219 old_PA_IN = PA_OUT = NULL;
2221 /* If any edges from predecessors are abnormal, antic_in is empty,
2222 so do nothing. */
2223 if (block_has_abnormal_pred_edge)
2224 goto maybe_dump_sets;
2226 /* If there are too many partially anticipatable values in the
2227 block, phi_translate_set can take an exponential time: stop
2228 before the translation starts. */
2229 if (max_pa
2230 && single_succ_p (block)
2231 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2232 goto maybe_dump_sets;
2234 old_PA_IN = PA_IN (block);
2235 PA_OUT = bitmap_set_new ();
2237 /* If the block has no successors, ANTIC_OUT is empty. */
2238 if (EDGE_COUNT (block->succs) == 0)
2240 /* If we have one successor, we could have some phi nodes to
2241 translate through. Note that we can't phi translate across DFS
2242 back edges in partial antic, because it uses a union operation on
2243 the successors. For recurrences like IV's, we will end up
2244 generating a new value in the set on each go around (i + 3 (VH.1)
2245 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2246 else if (single_succ_p (block))
2248 basic_block succ = single_succ (block);
2249 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2250 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2252 /* If we have multiple successors, we take the union of all of
2253 them. */
2254 else
2256 size_t i;
2257 basic_block bprime;
2259 auto_vec<basic_block> worklist (EDGE_COUNT (block->succs));
2260 FOR_EACH_EDGE (e, ei, block->succs)
2262 if (e->flags & EDGE_DFS_BACK)
2263 continue;
2264 worklist.quick_push (e->dest);
2266 if (worklist.length () > 0)
2268 FOR_EACH_VEC_ELT (worklist, i, bprime)
2270 unsigned int i;
2271 bitmap_iterator bi;
2273 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2274 bitmap_value_insert_into_set (PA_OUT,
2275 expression_for_id (i));
2276 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2278 bitmap_set_t pa_in = bitmap_set_new ();
2279 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2280 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2281 bitmap_value_insert_into_set (PA_OUT,
2282 expression_for_id (i));
2283 bitmap_set_free (pa_in);
2285 else
2286 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2287 bitmap_value_insert_into_set (PA_OUT,
2288 expression_for_id (i));
2293 /* Prune expressions that are clobbered in block and thus become
2294 invalid if translated from PA_OUT to PA_IN. */
2295 prune_clobbered_mems (PA_OUT, block);
2297 /* PA_IN starts with PA_OUT - TMP_GEN.
2298 Then we subtract things from ANTIC_IN. */
2299 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2301 /* For partial antic, we want to put back in the phi results, since
2302 we will properly avoid making them partially antic over backedges. */
2303 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2304 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2306 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2307 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2309 dependent_clean (PA_IN (block), ANTIC_IN (block));
2311 maybe_dump_sets:
2312 if (dump_file && (dump_flags & TDF_DETAILS))
2314 if (PA_OUT)
2315 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2317 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2319 if (old_PA_IN)
2320 bitmap_set_free (old_PA_IN);
2321 if (PA_OUT)
2322 bitmap_set_free (PA_OUT);
2325 /* Compute ANTIC and partial ANTIC sets. */
2327 static void
2328 compute_antic (void)
2330 bool changed = true;
2331 int num_iterations = 0;
2332 basic_block block;
2333 int i;
2334 edge_iterator ei;
2335 edge e;
2337 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2338 We pre-build the map of blocks with incoming abnormal edges here. */
2339 has_abnormal_preds = sbitmap_alloc (last_basic_block_for_fn (cfun));
2340 bitmap_clear (has_abnormal_preds);
2342 FOR_ALL_BB_FN (block, cfun)
2344 BB_VISITED (block) = 0;
2346 FOR_EACH_EDGE (e, ei, block->preds)
2347 if (e->flags & EDGE_ABNORMAL)
2349 bitmap_set_bit (has_abnormal_preds, block->index);
2351 /* We also anticipate nothing. */
2352 BB_VISITED (block) = 1;
2353 break;
2356 /* While we are here, give empty ANTIC_IN sets to each block. */
2357 ANTIC_IN (block) = bitmap_set_new ();
2358 if (do_partial_partial)
2359 PA_IN (block) = bitmap_set_new ();
2362 /* At the exit block we anticipate nothing. */
2363 BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun)) = 1;
2365 /* For ANTIC computation we need a postorder that also guarantees that
2366 a block with a single successor is visited after its successor.
2367 RPO on the inverted CFG has this property. */
2368 int *postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
2369 int postorder_num = inverted_post_order_compute (postorder);
2371 sbitmap worklist = sbitmap_alloc (last_basic_block_for_fn (cfun) + 1);
2372 bitmap_ones (worklist);
2373 while (changed)
2375 if (dump_file && (dump_flags & TDF_DETAILS))
2376 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2377 /* ??? We need to clear our PHI translation cache here as the
2378 ANTIC sets shrink and we restrict valid translations to
2379 those having operands with leaders in ANTIC. Same below
2380 for PA ANTIC computation. */
2381 num_iterations++;
2382 changed = false;
2383 for (i = postorder_num - 1; i >= 0; i--)
2385 if (bitmap_bit_p (worklist, postorder[i]))
2387 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2388 bitmap_clear_bit (worklist, block->index);
2389 if (compute_antic_aux (block,
2390 bitmap_bit_p (has_abnormal_preds,
2391 block->index)))
2393 FOR_EACH_EDGE (e, ei, block->preds)
2394 bitmap_set_bit (worklist, e->src->index);
2395 changed = true;
2399 /* Theoretically possible, but *highly* unlikely. */
2400 gcc_checking_assert (num_iterations < 500);
2403 statistics_histogram_event (cfun, "compute_antic iterations",
2404 num_iterations);
2406 if (do_partial_partial)
2408 /* For partial antic we ignore backedges and thus we do not need
2409 to perform any iteration when we process blocks in postorder. */
2410 postorder_num = pre_and_rev_post_order_compute (NULL, postorder, false);
2411 for (i = postorder_num - 1 ; i >= 0; i--)
2413 basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]);
2414 compute_partial_antic_aux (block,
2415 bitmap_bit_p (has_abnormal_preds,
2416 block->index));
2420 sbitmap_free (has_abnormal_preds);
2421 sbitmap_free (worklist);
2422 free (postorder);
2426 /* Inserted expressions are placed onto this worklist, which is used
2427 for performing quick dead code elimination of insertions we made
2428 that didn't turn out to be necessary. */
2429 static bitmap inserted_exprs;
2431 /* The actual worker for create_component_ref_by_pieces. */
2433 static tree
2434 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2435 unsigned int *operand, gimple_seq *stmts)
2437 vn_reference_op_t currop = &ref->operands[*operand];
2438 tree genop;
2439 ++*operand;
2440 switch (currop->opcode)
2442 case CALL_EXPR:
2443 gcc_unreachable ();
2445 case MEM_REF:
2447 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2448 stmts);
2449 if (!baseop)
2450 return NULL_TREE;
2451 tree offset = currop->op0;
2452 if (TREE_CODE (baseop) == ADDR_EXPR
2453 && handled_component_p (TREE_OPERAND (baseop, 0)))
2455 HOST_WIDE_INT off;
2456 tree base;
2457 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2458 &off);
2459 gcc_assert (base);
2460 offset = int_const_binop (PLUS_EXPR, offset,
2461 build_int_cst (TREE_TYPE (offset),
2462 off));
2463 baseop = build_fold_addr_expr (base);
2465 genop = build2 (MEM_REF, currop->type, baseop, offset);
2466 MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2467 MR_DEPENDENCE_BASE (genop) = currop->base;
2468 REF_REVERSE_STORAGE_ORDER (genop) = currop->reverse;
2469 return genop;
2472 case TARGET_MEM_REF:
2474 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2475 vn_reference_op_t nextop = &ref->operands[++*operand];
2476 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2477 stmts);
2478 if (!baseop)
2479 return NULL_TREE;
2480 if (currop->op0)
2482 genop0 = find_or_generate_expression (block, currop->op0, stmts);
2483 if (!genop0)
2484 return NULL_TREE;
2486 if (nextop->op0)
2488 genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2489 if (!genop1)
2490 return NULL_TREE;
2492 genop = build5 (TARGET_MEM_REF, currop->type,
2493 baseop, currop->op2, genop0, currop->op1, genop1);
2495 MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2496 MR_DEPENDENCE_BASE (genop) = currop->base;
2497 return genop;
2500 case ADDR_EXPR:
2501 if (currop->op0)
2503 gcc_assert (is_gimple_min_invariant (currop->op0));
2504 return currop->op0;
2506 /* Fallthrough. */
2507 case REALPART_EXPR:
2508 case IMAGPART_EXPR:
2509 case VIEW_CONVERT_EXPR:
2511 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2512 stmts);
2513 if (!genop0)
2514 return NULL_TREE;
2515 return fold_build1 (currop->opcode, currop->type, genop0);
2518 case WITH_SIZE_EXPR:
2520 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2521 stmts);
2522 if (!genop0)
2523 return NULL_TREE;
2524 tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2525 if (!genop1)
2526 return NULL_TREE;
2527 return fold_build2 (currop->opcode, currop->type, genop0, genop1);
2530 case BIT_FIELD_REF:
2532 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2533 stmts);
2534 if (!genop0)
2535 return NULL_TREE;
2536 tree op1 = currop->op0;
2537 tree op2 = currop->op1;
2538 tree t = build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2539 REF_REVERSE_STORAGE_ORDER (t) = currop->reverse;
2540 return fold (t);
2543 /* For array ref vn_reference_op's, operand 1 of the array ref
2544 is op0 of the reference op and operand 3 of the array ref is
2545 op1. */
2546 case ARRAY_RANGE_REF:
2547 case ARRAY_REF:
2549 tree genop0;
2550 tree genop1 = currop->op0;
2551 tree genop2 = currop->op1;
2552 tree genop3 = currop->op2;
2553 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2554 stmts);
2555 if (!genop0)
2556 return NULL_TREE;
2557 genop1 = find_or_generate_expression (block, genop1, stmts);
2558 if (!genop1)
2559 return NULL_TREE;
2560 if (genop2)
2562 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2563 /* Drop zero minimum index if redundant. */
2564 if (integer_zerop (genop2)
2565 && (!domain_type
2566 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2567 genop2 = NULL_TREE;
2568 else
2570 genop2 = find_or_generate_expression (block, genop2, stmts);
2571 if (!genop2)
2572 return NULL_TREE;
2575 if (genop3)
2577 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2578 /* We can't always put a size in units of the element alignment
2579 here as the element alignment may be not visible. See
2580 PR43783. Simply drop the element size for constant
2581 sizes. */
2582 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2583 genop3 = NULL_TREE;
2584 else
2586 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2587 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2588 genop3 = find_or_generate_expression (block, genop3, stmts);
2589 if (!genop3)
2590 return NULL_TREE;
2593 return build4 (currop->opcode, currop->type, genop0, genop1,
2594 genop2, genop3);
2596 case COMPONENT_REF:
2598 tree op0;
2599 tree op1;
2600 tree genop2 = currop->op1;
2601 op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2602 if (!op0)
2603 return NULL_TREE;
2604 /* op1 should be a FIELD_DECL, which are represented by themselves. */
2605 op1 = currop->op0;
2606 if (genop2)
2608 genop2 = find_or_generate_expression (block, genop2, stmts);
2609 if (!genop2)
2610 return NULL_TREE;
2612 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2615 case SSA_NAME:
2617 genop = find_or_generate_expression (block, currop->op0, stmts);
2618 return genop;
2620 case STRING_CST:
2621 case INTEGER_CST:
2622 case COMPLEX_CST:
2623 case VECTOR_CST:
2624 case REAL_CST:
2625 case CONSTRUCTOR:
2626 case VAR_DECL:
2627 case PARM_DECL:
2628 case CONST_DECL:
2629 case RESULT_DECL:
2630 case FUNCTION_DECL:
2631 return currop->op0;
2633 default:
2634 gcc_unreachable ();
2638 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2639 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2640 trying to rename aggregates into ssa form directly, which is a no no.
2642 Thus, this routine doesn't create temporaries, it just builds a
2643 single access expression for the array, calling
2644 find_or_generate_expression to build the innermost pieces.
2646 This function is a subroutine of create_expression_by_pieces, and
2647 should not be called on it's own unless you really know what you
2648 are doing. */
2650 static tree
2651 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2652 gimple_seq *stmts)
2654 unsigned int op = 0;
2655 return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2658 /* Find a simple leader for an expression, or generate one using
2659 create_expression_by_pieces from a NARY expression for the value.
2660 BLOCK is the basic_block we are looking for leaders in.
2661 OP is the tree expression to find a leader for or generate.
2662 Returns the leader or NULL_TREE on failure. */
2664 static tree
2665 find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2667 pre_expr expr = get_or_alloc_expr_for (op);
2668 unsigned int lookfor = get_expr_value_id (expr);
2669 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2670 if (leader)
2672 if (leader->kind == NAME)
2673 return PRE_EXPR_NAME (leader);
2674 else if (leader->kind == CONSTANT)
2675 return PRE_EXPR_CONSTANT (leader);
2677 /* Defer. */
2678 return NULL_TREE;
2681 /* It must be a complex expression, so generate it recursively. Note
2682 that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2683 where the insert algorithm fails to insert a required expression. */
2684 bitmap exprset = value_expressions[lookfor];
2685 bitmap_iterator bi;
2686 unsigned int i;
2687 EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2689 pre_expr temp = expression_for_id (i);
2690 /* We cannot insert random REFERENCE expressions at arbitrary
2691 places. We can insert NARYs which eventually re-materializes
2692 its operand values. */
2693 if (temp->kind == NARY)
2694 return create_expression_by_pieces (block, temp, stmts,
2695 get_expr_type (expr));
2698 /* Defer. */
2699 return NULL_TREE;
2702 #define NECESSARY GF_PLF_1
2704 /* Create an expression in pieces, so that we can handle very complex
2705 expressions that may be ANTIC, but not necessary GIMPLE.
2706 BLOCK is the basic block the expression will be inserted into,
2707 EXPR is the expression to insert (in value form)
2708 STMTS is a statement list to append the necessary insertions into.
2710 This function will die if we hit some value that shouldn't be
2711 ANTIC but is (IE there is no leader for it, or its components).
2712 The function returns NULL_TREE in case a different antic expression
2713 has to be inserted first.
2714 This function may also generate expressions that are themselves
2715 partially or fully redundant. Those that are will be either made
2716 fully redundant during the next iteration of insert (for partially
2717 redundant ones), or eliminated by eliminate (for fully redundant
2718 ones). */
2720 static tree
2721 create_expression_by_pieces (basic_block block, pre_expr expr,
2722 gimple_seq *stmts, tree type)
2724 tree name;
2725 tree folded;
2726 gimple_seq forced_stmts = NULL;
2727 unsigned int value_id;
2728 gimple_stmt_iterator gsi;
2729 tree exprtype = type ? type : get_expr_type (expr);
2730 pre_expr nameexpr;
2731 gassign *newstmt;
2733 switch (expr->kind)
2735 /* We may hit the NAME/CONSTANT case if we have to convert types
2736 that value numbering saw through. */
2737 case NAME:
2738 folded = PRE_EXPR_NAME (expr);
2739 if (useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2740 return folded;
2741 break;
2742 case CONSTANT:
2744 folded = PRE_EXPR_CONSTANT (expr);
2745 tree tem = fold_convert (exprtype, folded);
2746 if (is_gimple_min_invariant (tem))
2747 return tem;
2748 break;
2750 case REFERENCE:
2751 if (PRE_EXPR_REFERENCE (expr)->operands[0].opcode == CALL_EXPR)
2753 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2754 unsigned int operand = 1;
2755 vn_reference_op_t currop = &ref->operands[0];
2756 tree sc = NULL_TREE;
2757 tree fn;
2758 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2759 fn = currop->op0;
2760 else
2761 fn = find_or_generate_expression (block, currop->op0, stmts);
2762 if (!fn)
2763 return NULL_TREE;
2764 if (currop->op1)
2766 sc = find_or_generate_expression (block, currop->op1, stmts);
2767 if (!sc)
2768 return NULL_TREE;
2770 auto_vec<tree> args (ref->operands.length () - 1);
2771 while (operand < ref->operands.length ())
2773 tree arg = create_component_ref_by_pieces_1 (block, ref,
2774 &operand, stmts);
2775 if (!arg)
2776 return NULL_TREE;
2777 args.quick_push (arg);
2779 gcall *call
2780 = gimple_build_call_vec ((TREE_CODE (fn) == FUNCTION_DECL
2781 ? build_fold_addr_expr (fn) : fn), args);
2782 gimple_call_set_with_bounds (call, currop->with_bounds);
2783 if (sc)
2784 gimple_call_set_chain (call, sc);
2785 tree forcedname = make_ssa_name (currop->type);
2786 gimple_call_set_lhs (call, forcedname);
2787 gimple_set_vuse (call, BB_LIVE_VOP_ON_EXIT (block));
2788 gimple_seq_add_stmt_without_update (&forced_stmts, call);
2789 folded = forcedname;
2791 else
2793 folded = create_component_ref_by_pieces (block,
2794 PRE_EXPR_REFERENCE (expr),
2795 stmts);
2796 if (!folded)
2797 return NULL_TREE;
2798 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2799 newstmt = gimple_build_assign (name, folded);
2800 gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2801 gimple_set_vuse (newstmt, BB_LIVE_VOP_ON_EXIT (block));
2802 folded = name;
2804 break;
2805 case NARY:
2807 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2808 tree *genop = XALLOCAVEC (tree, nary->length);
2809 unsigned i;
2810 for (i = 0; i < nary->length; ++i)
2812 genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2813 if (!genop[i])
2814 return NULL_TREE;
2815 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2816 may have conversions stripped. */
2817 if (nary->opcode == POINTER_PLUS_EXPR)
2819 if (i == 0)
2820 genop[i] = gimple_convert (&forced_stmts,
2821 nary->type, genop[i]);
2822 else if (i == 1)
2823 genop[i] = gimple_convert (&forced_stmts,
2824 sizetype, genop[i]);
2826 else
2827 genop[i] = gimple_convert (&forced_stmts,
2828 TREE_TYPE (nary->op[i]), genop[i]);
2830 if (nary->opcode == CONSTRUCTOR)
2832 vec<constructor_elt, va_gc> *elts = NULL;
2833 for (i = 0; i < nary->length; ++i)
2834 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2835 folded = build_constructor (nary->type, elts);
2836 name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2837 newstmt = gimple_build_assign (name, folded);
2838 gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2839 folded = name;
2841 else
2843 switch (nary->length)
2845 case 1:
2846 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2847 genop[0]);
2848 break;
2849 case 2:
2850 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2851 genop[0], genop[1]);
2852 break;
2853 case 3:
2854 folded = gimple_build (&forced_stmts, nary->opcode, nary->type,
2855 genop[0], genop[1], genop[2]);
2856 break;
2857 default:
2858 gcc_unreachable ();
2862 break;
2863 default:
2864 gcc_unreachable ();
2867 folded = gimple_convert (&forced_stmts, exprtype, folded);
2869 /* If there is nothing to insert, return the simplified result. */
2870 if (gimple_seq_empty_p (forced_stmts))
2871 return folded;
2872 /* If we simplified to a constant return it and discard eventually
2873 built stmts. */
2874 if (is_gimple_min_invariant (folded))
2876 gimple_seq_discard (forced_stmts);
2877 return folded;
2880 gcc_assert (TREE_CODE (folded) == SSA_NAME);
2882 /* If we have any intermediate expressions to the value sets, add them
2883 to the value sets and chain them in the instruction stream. */
2884 if (forced_stmts)
2886 gsi = gsi_start (forced_stmts);
2887 for (; !gsi_end_p (gsi); gsi_next (&gsi))
2889 gimple *stmt = gsi_stmt (gsi);
2890 tree forcedname = gimple_get_lhs (stmt);
2891 pre_expr nameexpr;
2893 if (forcedname != folded)
2895 VN_INFO_GET (forcedname)->valnum = forcedname;
2896 VN_INFO (forcedname)->value_id = get_next_value_id ();
2897 nameexpr = get_or_alloc_expr_for_name (forcedname);
2898 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
2899 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2900 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2903 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
2904 gimple_set_plf (stmt, NECESSARY, false);
2906 gimple_seq_add_seq (stmts, forced_stmts);
2909 name = folded;
2911 /* Fold the last statement. */
2912 gsi = gsi_last (*stmts);
2913 if (fold_stmt_inplace (&gsi))
2914 update_stmt (gsi_stmt (gsi));
2916 /* Add a value number to the temporary.
2917 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
2918 we are creating the expression by pieces, and this particular piece of
2919 the expression may have been represented. There is no harm in replacing
2920 here. */
2921 value_id = get_expr_value_id (expr);
2922 VN_INFO_GET (name)->value_id = value_id;
2923 VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id);
2924 if (VN_INFO (name)->valnum == NULL_TREE)
2925 VN_INFO (name)->valnum = name;
2926 gcc_assert (VN_INFO (name)->valnum != NULL_TREE);
2927 nameexpr = get_or_alloc_expr_for_name (name);
2928 add_to_value (value_id, nameexpr);
2929 if (NEW_SETS (block))
2930 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
2931 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
2933 pre_stats.insertions++;
2934 if (dump_file && (dump_flags & TDF_DETAILS))
2936 fprintf (dump_file, "Inserted ");
2937 print_gimple_stmt (dump_file, gsi_stmt (gsi_last (*stmts)), 0, 0);
2938 fprintf (dump_file, " in predecessor %d (%04d)\n",
2939 block->index, value_id);
2942 return name;
2946 /* Insert the to-be-made-available values of expression EXPRNUM for each
2947 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
2948 merge the result with a phi node, given the same value number as
2949 NODE. Return true if we have inserted new stuff. */
2951 static bool
2952 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
2953 vec<pre_expr> avail)
2955 pre_expr expr = expression_for_id (exprnum);
2956 pre_expr newphi;
2957 unsigned int val = get_expr_value_id (expr);
2958 edge pred;
2959 bool insertions = false;
2960 bool nophi = false;
2961 basic_block bprime;
2962 pre_expr eprime;
2963 edge_iterator ei;
2964 tree type = get_expr_type (expr);
2965 tree temp;
2966 gphi *phi;
2968 /* Make sure we aren't creating an induction variable. */
2969 if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
2971 bool firstinsideloop = false;
2972 bool secondinsideloop = false;
2973 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
2974 EDGE_PRED (block, 0)->src);
2975 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
2976 EDGE_PRED (block, 1)->src);
2977 /* Induction variables only have one edge inside the loop. */
2978 if ((firstinsideloop ^ secondinsideloop)
2979 && expr->kind != REFERENCE)
2981 if (dump_file && (dump_flags & TDF_DETAILS))
2982 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
2983 nophi = true;
2987 /* Make the necessary insertions. */
2988 FOR_EACH_EDGE (pred, ei, block->preds)
2990 gimple_seq stmts = NULL;
2991 tree builtexpr;
2992 bprime = pred->src;
2993 eprime = avail[pred->dest_idx];
2994 builtexpr = create_expression_by_pieces (bprime, eprime,
2995 &stmts, type);
2996 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
2997 if (!gimple_seq_empty_p (stmts))
2999 gsi_insert_seq_on_edge (pred, stmts);
3000 insertions = true;
3002 if (!builtexpr)
3004 /* We cannot insert a PHI node if we failed to insert
3005 on one edge. */
3006 nophi = true;
3007 continue;
3009 if (is_gimple_min_invariant (builtexpr))
3010 avail[pred->dest_idx] = get_or_alloc_expr_for_constant (builtexpr);
3011 else
3012 avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3014 /* If we didn't want a phi node, and we made insertions, we still have
3015 inserted new stuff, and thus return true. If we didn't want a phi node,
3016 and didn't make insertions, we haven't added anything new, so return
3017 false. */
3018 if (nophi && insertions)
3019 return true;
3020 else if (nophi && !insertions)
3021 return false;
3023 /* Now build a phi for the new variable. */
3024 temp = make_temp_ssa_name (type, NULL, "prephitmp");
3025 phi = create_phi_node (temp, block);
3027 gimple_set_plf (phi, NECESSARY, false);
3028 VN_INFO_GET (temp)->value_id = val;
3029 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3030 if (VN_INFO (temp)->valnum == NULL_TREE)
3031 VN_INFO (temp)->valnum = temp;
3032 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3033 FOR_EACH_EDGE (pred, ei, block->preds)
3035 pre_expr ae = avail[pred->dest_idx];
3036 gcc_assert (get_expr_type (ae) == type
3037 || useless_type_conversion_p (type, get_expr_type (ae)));
3038 if (ae->kind == CONSTANT)
3039 add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3040 pred, UNKNOWN_LOCATION);
3041 else
3042 add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3045 newphi = get_or_alloc_expr_for_name (temp);
3046 add_to_value (val, newphi);
3048 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3049 this insertion, since we test for the existence of this value in PHI_GEN
3050 before proceeding with the partial redundancy checks in insert_aux.
3052 The value may exist in AVAIL_OUT, in particular, it could be represented
3053 by the expression we are trying to eliminate, in which case we want the
3054 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3055 inserted there.
3057 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3058 this block, because if it did, it would have existed in our dominator's
3059 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3062 bitmap_insert_into_set (PHI_GEN (block), newphi);
3063 bitmap_value_replace_in_set (AVAIL_OUT (block),
3064 newphi);
3065 bitmap_insert_into_set (NEW_SETS (block),
3066 newphi);
3068 /* If we insert a PHI node for a conversion of another PHI node
3069 in the same basic-block try to preserve range information.
3070 This is important so that followup loop passes receive optimal
3071 number of iteration analysis results. See PR61743. */
3072 if (expr->kind == NARY
3073 && CONVERT_EXPR_CODE_P (expr->u.nary->opcode)
3074 && TREE_CODE (expr->u.nary->op[0]) == SSA_NAME
3075 && gimple_bb (SSA_NAME_DEF_STMT (expr->u.nary->op[0])) == block
3076 && INTEGRAL_TYPE_P (type)
3077 && INTEGRAL_TYPE_P (TREE_TYPE (expr->u.nary->op[0]))
3078 && (TYPE_PRECISION (type)
3079 >= TYPE_PRECISION (TREE_TYPE (expr->u.nary->op[0])))
3080 && SSA_NAME_RANGE_INFO (expr->u.nary->op[0]))
3082 wide_int min, max;
3083 if (get_range_info (expr->u.nary->op[0], &min, &max) == VR_RANGE
3084 && !wi::neg_p (min, SIGNED)
3085 && !wi::neg_p (max, SIGNED))
3086 /* Just handle extension and sign-changes of all-positive ranges. */
3087 set_range_info (temp,
3088 SSA_NAME_RANGE_TYPE (expr->u.nary->op[0]),
3089 wide_int_storage::from (min, TYPE_PRECISION (type),
3090 TYPE_SIGN (type)),
3091 wide_int_storage::from (max, TYPE_PRECISION (type),
3092 TYPE_SIGN (type)));
3095 if (dump_file && (dump_flags & TDF_DETAILS))
3097 fprintf (dump_file, "Created phi ");
3098 print_gimple_stmt (dump_file, phi, 0, 0);
3099 fprintf (dump_file, " in block %d (%04d)\n", block->index, val);
3101 pre_stats.phis++;
3102 return true;
3107 /* Perform insertion of partially redundant values.
3108 For BLOCK, do the following:
3109 1. Propagate the NEW_SETS of the dominator into the current block.
3110 If the block has multiple predecessors,
3111 2a. Iterate over the ANTIC expressions for the block to see if
3112 any of them are partially redundant.
3113 2b. If so, insert them into the necessary predecessors to make
3114 the expression fully redundant.
3115 2c. Insert a new PHI merging the values of the predecessors.
3116 2d. Insert the new PHI, and the new expressions, into the
3117 NEW_SETS set.
3118 3. Recursively call ourselves on the dominator children of BLOCK.
3120 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3121 do_regular_insertion and do_partial_insertion.
3125 static bool
3126 do_regular_insertion (basic_block block, basic_block dom)
3128 bool new_stuff = false;
3129 vec<pre_expr> exprs;
3130 pre_expr expr;
3131 auto_vec<pre_expr> avail;
3132 int i;
3134 exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3135 avail.safe_grow (EDGE_COUNT (block->preds));
3137 FOR_EACH_VEC_ELT (exprs, i, expr)
3139 if (expr->kind == NARY
3140 || expr->kind == REFERENCE)
3142 unsigned int val;
3143 bool by_some = false;
3144 bool cant_insert = false;
3145 bool all_same = true;
3146 pre_expr first_s = NULL;
3147 edge pred;
3148 basic_block bprime;
3149 pre_expr eprime = NULL;
3150 edge_iterator ei;
3151 pre_expr edoubleprime = NULL;
3152 bool do_insertion = false;
3154 val = get_expr_value_id (expr);
3155 if (bitmap_set_contains_value (PHI_GEN (block), val))
3156 continue;
3157 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3159 if (dump_file && (dump_flags & TDF_DETAILS))
3161 fprintf (dump_file, "Found fully redundant value: ");
3162 print_pre_expr (dump_file, expr);
3163 fprintf (dump_file, "\n");
3165 continue;
3168 FOR_EACH_EDGE (pred, ei, block->preds)
3170 unsigned int vprime;
3172 /* We should never run insertion for the exit block
3173 and so not come across fake pred edges. */
3174 gcc_assert (!(pred->flags & EDGE_FAKE));
3175 bprime = pred->src;
3176 /* We are looking at ANTIC_OUT of bprime. */
3177 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3178 bprime, block);
3180 /* eprime will generally only be NULL if the
3181 value of the expression, translated
3182 through the PHI for this predecessor, is
3183 undefined. If that is the case, we can't
3184 make the expression fully redundant,
3185 because its value is undefined along a
3186 predecessor path. We can thus break out
3187 early because it doesn't matter what the
3188 rest of the results are. */
3189 if (eprime == NULL)
3191 avail[pred->dest_idx] = NULL;
3192 cant_insert = true;
3193 break;
3196 eprime = fully_constant_expression (eprime);
3197 vprime = get_expr_value_id (eprime);
3198 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3199 vprime);
3200 if (edoubleprime == NULL)
3202 avail[pred->dest_idx] = eprime;
3203 all_same = false;
3205 else
3207 avail[pred->dest_idx] = edoubleprime;
3208 by_some = true;
3209 /* We want to perform insertions to remove a redundancy on
3210 a path in the CFG we want to optimize for speed. */
3211 if (optimize_edge_for_speed_p (pred))
3212 do_insertion = true;
3213 if (first_s == NULL)
3214 first_s = edoubleprime;
3215 else if (!pre_expr_d::equal (first_s, edoubleprime))
3216 all_same = false;
3219 /* If we can insert it, it's not the same value
3220 already existing along every predecessor, and
3221 it's defined by some predecessor, it is
3222 partially redundant. */
3223 if (!cant_insert && !all_same && by_some)
3225 if (!do_insertion)
3227 if (dump_file && (dump_flags & TDF_DETAILS))
3229 fprintf (dump_file, "Skipping partial redundancy for "
3230 "expression ");
3231 print_pre_expr (dump_file, expr);
3232 fprintf (dump_file, " (%04d), no redundancy on to be "
3233 "optimized for speed edge\n", val);
3236 else if (dbg_cnt (treepre_insert))
3238 if (dump_file && (dump_flags & TDF_DETAILS))
3240 fprintf (dump_file, "Found partial redundancy for "
3241 "expression ");
3242 print_pre_expr (dump_file, expr);
3243 fprintf (dump_file, " (%04d)\n",
3244 get_expr_value_id (expr));
3246 if (insert_into_preds_of_block (block,
3247 get_expression_id (expr),
3248 avail))
3249 new_stuff = true;
3252 /* If all edges produce the same value and that value is
3253 an invariant, then the PHI has the same value on all
3254 edges. Note this. */
3255 else if (!cant_insert && all_same)
3257 gcc_assert (edoubleprime->kind == CONSTANT
3258 || edoubleprime->kind == NAME);
3260 tree temp = make_temp_ssa_name (get_expr_type (expr),
3261 NULL, "pretmp");
3262 gassign *assign
3263 = gimple_build_assign (temp,
3264 edoubleprime->kind == CONSTANT ?
3265 PRE_EXPR_CONSTANT (edoubleprime) :
3266 PRE_EXPR_NAME (edoubleprime));
3267 gimple_stmt_iterator gsi = gsi_after_labels (block);
3268 gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3270 gimple_set_plf (assign, NECESSARY, false);
3271 VN_INFO_GET (temp)->value_id = val;
3272 VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val);
3273 if (VN_INFO (temp)->valnum == NULL_TREE)
3274 VN_INFO (temp)->valnum = temp;
3275 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3276 pre_expr newe = get_or_alloc_expr_for_name (temp);
3277 add_to_value (val, newe);
3278 bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3279 bitmap_insert_into_set (NEW_SETS (block), newe);
3284 exprs.release ();
3285 return new_stuff;
3289 /* Perform insertion for partially anticipatable expressions. There
3290 is only one case we will perform insertion for these. This case is
3291 if the expression is partially anticipatable, and fully available.
3292 In this case, we know that putting it earlier will enable us to
3293 remove the later computation. */
3296 static bool
3297 do_partial_partial_insertion (basic_block block, basic_block dom)
3299 bool new_stuff = false;
3300 vec<pre_expr> exprs;
3301 pre_expr expr;
3302 auto_vec<pre_expr> avail;
3303 int i;
3305 exprs = sorted_array_from_bitmap_set (PA_IN (block));
3306 avail.safe_grow (EDGE_COUNT (block->preds));
3308 FOR_EACH_VEC_ELT (exprs, i, expr)
3310 if (expr->kind == NARY
3311 || expr->kind == REFERENCE)
3313 unsigned int val;
3314 bool by_all = true;
3315 bool cant_insert = false;
3316 edge pred;
3317 basic_block bprime;
3318 pre_expr eprime = NULL;
3319 edge_iterator ei;
3321 val = get_expr_value_id (expr);
3322 if (bitmap_set_contains_value (PHI_GEN (block), val))
3323 continue;
3324 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3325 continue;
3327 FOR_EACH_EDGE (pred, ei, block->preds)
3329 unsigned int vprime;
3330 pre_expr edoubleprime;
3332 /* We should never run insertion for the exit block
3333 and so not come across fake pred edges. */
3334 gcc_assert (!(pred->flags & EDGE_FAKE));
3335 bprime = pred->src;
3336 eprime = phi_translate (expr, ANTIC_IN (block),
3337 PA_IN (block),
3338 bprime, block);
3340 /* eprime will generally only be NULL if the
3341 value of the expression, translated
3342 through the PHI for this predecessor, is
3343 undefined. If that is the case, we can't
3344 make the expression fully redundant,
3345 because its value is undefined along a
3346 predecessor path. We can thus break out
3347 early because it doesn't matter what the
3348 rest of the results are. */
3349 if (eprime == NULL)
3351 avail[pred->dest_idx] = NULL;
3352 cant_insert = true;
3353 break;
3356 eprime = fully_constant_expression (eprime);
3357 vprime = get_expr_value_id (eprime);
3358 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3359 avail[pred->dest_idx] = edoubleprime;
3360 if (edoubleprime == NULL)
3362 by_all = false;
3363 break;
3367 /* If we can insert it, it's not the same value
3368 already existing along every predecessor, and
3369 it's defined by some predecessor, it is
3370 partially redundant. */
3371 if (!cant_insert && by_all)
3373 edge succ;
3374 bool do_insertion = false;
3376 /* Insert only if we can remove a later expression on a path
3377 that we want to optimize for speed.
3378 The phi node that we will be inserting in BLOCK is not free,
3379 and inserting it for the sake of !optimize_for_speed successor
3380 may cause regressions on the speed path. */
3381 FOR_EACH_EDGE (succ, ei, block->succs)
3383 if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3384 || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3386 if (optimize_edge_for_speed_p (succ))
3387 do_insertion = true;
3391 if (!do_insertion)
3393 if (dump_file && (dump_flags & TDF_DETAILS))
3395 fprintf (dump_file, "Skipping partial partial redundancy "
3396 "for expression ");
3397 print_pre_expr (dump_file, expr);
3398 fprintf (dump_file, " (%04d), not (partially) anticipated "
3399 "on any to be optimized for speed edges\n", val);
3402 else if (dbg_cnt (treepre_insert))
3404 pre_stats.pa_insert++;
3405 if (dump_file && (dump_flags & TDF_DETAILS))
3407 fprintf (dump_file, "Found partial partial redundancy "
3408 "for expression ");
3409 print_pre_expr (dump_file, expr);
3410 fprintf (dump_file, " (%04d)\n",
3411 get_expr_value_id (expr));
3413 if (insert_into_preds_of_block (block,
3414 get_expression_id (expr),
3415 avail))
3416 new_stuff = true;
3422 exprs.release ();
3423 return new_stuff;
3426 static bool
3427 insert_aux (basic_block block)
3429 basic_block son;
3430 bool new_stuff = false;
3432 if (block)
3434 basic_block dom;
3435 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3436 if (dom)
3438 unsigned i;
3439 bitmap_iterator bi;
3440 bitmap_set_t newset = NEW_SETS (dom);
3441 if (newset)
3443 /* Note that we need to value_replace both NEW_SETS, and
3444 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3445 represented by some non-simple expression here that we want
3446 to replace it with. */
3447 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3449 pre_expr expr = expression_for_id (i);
3450 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3451 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3454 if (!single_pred_p (block))
3456 new_stuff |= do_regular_insertion (block, dom);
3457 if (do_partial_partial)
3458 new_stuff |= do_partial_partial_insertion (block, dom);
3462 for (son = first_dom_son (CDI_DOMINATORS, block);
3463 son;
3464 son = next_dom_son (CDI_DOMINATORS, son))
3466 new_stuff |= insert_aux (son);
3469 return new_stuff;
3472 /* Perform insertion of partially redundant values. */
3474 static void
3475 insert (void)
3477 bool new_stuff = true;
3478 basic_block bb;
3479 int num_iterations = 0;
3481 FOR_ALL_BB_FN (bb, cfun)
3482 NEW_SETS (bb) = bitmap_set_new ();
3484 while (new_stuff)
3486 num_iterations++;
3487 if (dump_file && dump_flags & TDF_DETAILS)
3488 fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3489 new_stuff = insert_aux (ENTRY_BLOCK_PTR_FOR_FN (cfun));
3491 /* Clear the NEW sets before the next iteration. We have already
3492 fully propagated its contents. */
3493 if (new_stuff)
3494 FOR_ALL_BB_FN (bb, cfun)
3495 bitmap_set_free (NEW_SETS (bb));
3497 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3501 /* Compute the AVAIL set for all basic blocks.
3503 This function performs value numbering of the statements in each basic
3504 block. The AVAIL sets are built from information we glean while doing
3505 this value numbering, since the AVAIL sets contain only one entry per
3506 value.
3508 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3509 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3511 static void
3512 compute_avail (void)
3515 basic_block block, son;
3516 basic_block *worklist;
3517 size_t sp = 0;
3518 unsigned i;
3520 /* We pretend that default definitions are defined in the entry block.
3521 This includes function arguments and the static chain decl. */
3522 for (i = 1; i < num_ssa_names; ++i)
3524 tree name = ssa_name (i);
3525 pre_expr e;
3526 if (!name
3527 || !SSA_NAME_IS_DEFAULT_DEF (name)
3528 || has_zero_uses (name)
3529 || virtual_operand_p (name))
3530 continue;
3532 e = get_or_alloc_expr_for_name (name);
3533 add_to_value (get_expr_value_id (e), e);
3534 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)), e);
3535 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3539 if (dump_file && (dump_flags & TDF_DETAILS))
3541 print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3542 "tmp_gen", ENTRY_BLOCK);
3543 print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)),
3544 "avail_out", ENTRY_BLOCK);
3547 /* Allocate the worklist. */
3548 worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
3550 /* Seed the algorithm by putting the dominator children of the entry
3551 block on the worklist. */
3552 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR_FOR_FN (cfun));
3553 son;
3554 son = next_dom_son (CDI_DOMINATORS, son))
3555 worklist[sp++] = son;
3557 BB_LIVE_VOP_ON_EXIT (ENTRY_BLOCK_PTR_FOR_FN (cfun))
3558 = ssa_default_def (cfun, gimple_vop (cfun));
3560 /* Loop until the worklist is empty. */
3561 while (sp)
3563 gimple *stmt;
3564 basic_block dom;
3566 /* Pick a block from the worklist. */
3567 block = worklist[--sp];
3569 /* Initially, the set of available values in BLOCK is that of
3570 its immediate dominator. */
3571 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3572 if (dom)
3574 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3575 BB_LIVE_VOP_ON_EXIT (block) = BB_LIVE_VOP_ON_EXIT (dom);
3578 /* Generate values for PHI nodes. */
3579 for (gphi_iterator gsi = gsi_start_phis (block); !gsi_end_p (gsi);
3580 gsi_next (&gsi))
3582 tree result = gimple_phi_result (gsi.phi ());
3584 /* We have no need for virtual phis, as they don't represent
3585 actual computations. */
3586 if (virtual_operand_p (result))
3588 BB_LIVE_VOP_ON_EXIT (block) = result;
3589 continue;
3592 pre_expr e = get_or_alloc_expr_for_name (result);
3593 add_to_value (get_expr_value_id (e), e);
3594 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3595 bitmap_insert_into_set (PHI_GEN (block), e);
3598 BB_MAY_NOTRETURN (block) = 0;
3600 /* Now compute value numbers and populate value sets with all
3601 the expressions computed in BLOCK. */
3602 for (gimple_stmt_iterator gsi = gsi_start_bb (block); !gsi_end_p (gsi);
3603 gsi_next (&gsi))
3605 ssa_op_iter iter;
3606 tree op;
3608 stmt = gsi_stmt (gsi);
3610 /* Cache whether the basic-block has any non-visible side-effect
3611 or control flow.
3612 If this isn't a call or it is the last stmt in the
3613 basic-block then the CFG represents things correctly. */
3614 if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
3616 /* Non-looping const functions always return normally.
3617 Otherwise the call might not return or have side-effects
3618 that forbids hoisting possibly trapping expressions
3619 before it. */
3620 int flags = gimple_call_flags (stmt);
3621 if (!(flags & ECF_CONST)
3622 || (flags & ECF_LOOPING_CONST_OR_PURE))
3623 BB_MAY_NOTRETURN (block) = 1;
3626 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3628 pre_expr e = get_or_alloc_expr_for_name (op);
3630 add_to_value (get_expr_value_id (e), e);
3631 bitmap_insert_into_set (TMP_GEN (block), e);
3632 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3635 if (gimple_vdef (stmt))
3636 BB_LIVE_VOP_ON_EXIT (block) = gimple_vdef (stmt);
3638 if (gimple_has_side_effects (stmt)
3639 || stmt_could_throw_p (stmt)
3640 || is_gimple_debug (stmt))
3641 continue;
3643 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3645 if (ssa_undefined_value_p (op))
3646 continue;
3647 pre_expr e = get_or_alloc_expr_for_name (op);
3648 bitmap_value_insert_into_set (EXP_GEN (block), e);
3651 switch (gimple_code (stmt))
3653 case GIMPLE_RETURN:
3654 continue;
3656 case GIMPLE_CALL:
3658 vn_reference_t ref;
3659 vn_reference_s ref1;
3660 pre_expr result = NULL;
3662 /* We can value number only calls to real functions. */
3663 if (gimple_call_internal_p (stmt))
3664 continue;
3666 vn_reference_lookup_call (as_a <gcall *> (stmt), &ref, &ref1);
3667 if (!ref)
3668 continue;
3670 /* If the value of the call is not invalidated in
3671 this block until it is computed, add the expression
3672 to EXP_GEN. */
3673 if (!gimple_vuse (stmt)
3674 || gimple_code
3675 (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
3676 || gimple_bb (SSA_NAME_DEF_STMT
3677 (gimple_vuse (stmt))) != block)
3679 result = pre_expr_pool.allocate ();
3680 result->kind = REFERENCE;
3681 result->id = 0;
3682 PRE_EXPR_REFERENCE (result) = ref;
3684 get_or_alloc_expression_id (result);
3685 add_to_value (get_expr_value_id (result), result);
3686 bitmap_value_insert_into_set (EXP_GEN (block), result);
3688 continue;
3691 case GIMPLE_ASSIGN:
3693 pre_expr result = NULL;
3694 switch (vn_get_stmt_kind (stmt))
3696 case VN_NARY:
3698 enum tree_code code = gimple_assign_rhs_code (stmt);
3699 vn_nary_op_t nary;
3701 /* COND_EXPR and VEC_COND_EXPR are awkward in
3702 that they contain an embedded complex expression.
3703 Don't even try to shove those through PRE. */
3704 if (code == COND_EXPR
3705 || code == VEC_COND_EXPR)
3706 continue;
3708 vn_nary_op_lookup_stmt (stmt, &nary);
3709 if (!nary)
3710 continue;
3712 /* If the NARY traps and there was a preceding
3713 point in the block that might not return avoid
3714 adding the nary to EXP_GEN. */
3715 if (BB_MAY_NOTRETURN (block)
3716 && vn_nary_may_trap (nary))
3717 continue;
3719 result = pre_expr_pool.allocate ();
3720 result->kind = NARY;
3721 result->id = 0;
3722 PRE_EXPR_NARY (result) = nary;
3723 break;
3726 case VN_REFERENCE:
3728 tree rhs1 = gimple_assign_rhs1 (stmt);
3729 alias_set_type set = get_alias_set (rhs1);
3730 vec<vn_reference_op_s> operands
3731 = vn_reference_operands_for_lookup (rhs1);
3732 vn_reference_t ref;
3733 vn_reference_lookup_pieces (gimple_vuse (stmt), set,
3734 TREE_TYPE (rhs1),
3735 operands, &ref, VN_WALK);
3736 if (!ref)
3738 operands.release ();
3739 continue;
3742 /* If the value of the reference is not invalidated in
3743 this block until it is computed, add the expression
3744 to EXP_GEN. */
3745 if (gimple_vuse (stmt))
3747 gimple *def_stmt;
3748 bool ok = true;
3749 def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
3750 while (!gimple_nop_p (def_stmt)
3751 && gimple_code (def_stmt) != GIMPLE_PHI
3752 && gimple_bb (def_stmt) == block)
3754 if (stmt_may_clobber_ref_p
3755 (def_stmt, gimple_assign_rhs1 (stmt)))
3757 ok = false;
3758 break;
3760 def_stmt
3761 = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
3763 if (!ok)
3765 operands.release ();
3766 continue;
3770 /* If the load was value-numbered to another
3771 load make sure we do not use its expression
3772 for insertion if it wouldn't be a valid
3773 replacement. */
3774 /* At the momemt we have a testcase
3775 for hoist insertion of aligned vs. misaligned
3776 variants in gcc.dg/torture/pr65270-1.c thus
3777 with just alignment to be considered we can
3778 simply replace the expression in the hashtable
3779 with the most conservative one. */
3780 vn_reference_op_t ref1 = &ref->operands.last ();
3781 while (ref1->opcode != TARGET_MEM_REF
3782 && ref1->opcode != MEM_REF
3783 && ref1 != &ref->operands[0])
3784 --ref1;
3785 vn_reference_op_t ref2 = &operands.last ();
3786 while (ref2->opcode != TARGET_MEM_REF
3787 && ref2->opcode != MEM_REF
3788 && ref2 != &operands[0])
3789 --ref2;
3790 if ((ref1->opcode == TARGET_MEM_REF
3791 || ref1->opcode == MEM_REF)
3792 && (TYPE_ALIGN (ref1->type)
3793 > TYPE_ALIGN (ref2->type)))
3795 ref->operands.release ();
3796 ref->operands = operands;
3797 ref1 = ref2;
3799 else
3800 operands.release ();
3801 /* TBAA behavior is an obvious part so make sure
3802 that the hashtable one covers this as well
3803 by adjusting the ref alias set and its base. */
3804 if (ref->set == set
3805 || alias_set_subset_of (set, ref->set))
3807 else if (alias_set_subset_of (ref->set, set))
3809 ref->set = set;
3810 if (ref1->opcode == MEM_REF)
3811 ref1->op0 = fold_convert (TREE_TYPE (ref2->op0),
3812 ref1->op0);
3813 else
3814 ref1->op2 = fold_convert (TREE_TYPE (ref2->op2),
3815 ref1->op2);
3817 else
3819 ref->set = 0;
3820 if (ref1->opcode == MEM_REF)
3821 ref1->op0 = fold_convert (ptr_type_node,
3822 ref1->op0);
3823 else
3824 ref1->op2 = fold_convert (ptr_type_node,
3825 ref1->op2);
3828 result = pre_expr_pool.allocate ();
3829 result->kind = REFERENCE;
3830 result->id = 0;
3831 PRE_EXPR_REFERENCE (result) = ref;
3832 break;
3835 default:
3836 continue;
3839 get_or_alloc_expression_id (result);
3840 add_to_value (get_expr_value_id (result), result);
3841 bitmap_value_insert_into_set (EXP_GEN (block), result);
3842 continue;
3844 default:
3845 break;
3849 if (dump_file && (dump_flags & TDF_DETAILS))
3851 print_bitmap_set (dump_file, EXP_GEN (block),
3852 "exp_gen", block->index);
3853 print_bitmap_set (dump_file, PHI_GEN (block),
3854 "phi_gen", block->index);
3855 print_bitmap_set (dump_file, TMP_GEN (block),
3856 "tmp_gen", block->index);
3857 print_bitmap_set (dump_file, AVAIL_OUT (block),
3858 "avail_out", block->index);
3861 /* Put the dominator children of BLOCK on the worklist of blocks
3862 to compute available sets for. */
3863 for (son = first_dom_son (CDI_DOMINATORS, block);
3864 son;
3865 son = next_dom_son (CDI_DOMINATORS, son))
3866 worklist[sp++] = son;
3869 free (worklist);
3873 /* Local state for the eliminate domwalk. */
3874 static vec<gimple *> el_to_remove;
3875 static vec<gimple *> el_to_fixup;
3876 static unsigned int el_todo;
3877 static vec<tree> el_avail;
3878 static vec<tree> el_avail_stack;
3880 /* Return a leader for OP that is available at the current point of the
3881 eliminate domwalk. */
3883 static tree
3884 eliminate_avail (tree op)
3886 tree valnum = VN_INFO (op)->valnum;
3887 if (TREE_CODE (valnum) == SSA_NAME)
3889 if (SSA_NAME_IS_DEFAULT_DEF (valnum))
3890 return valnum;
3891 if (el_avail.length () > SSA_NAME_VERSION (valnum))
3892 return el_avail[SSA_NAME_VERSION (valnum)];
3894 else if (is_gimple_min_invariant (valnum))
3895 return valnum;
3896 return NULL_TREE;
3899 /* At the current point of the eliminate domwalk make OP available. */
3901 static void
3902 eliminate_push_avail (tree op)
3904 tree valnum = VN_INFO (op)->valnum;
3905 if (TREE_CODE (valnum) == SSA_NAME)
3907 if (el_avail.length () <= SSA_NAME_VERSION (valnum))
3908 el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1);
3909 tree pushop = op;
3910 if (el_avail[SSA_NAME_VERSION (valnum)])
3911 pushop = el_avail[SSA_NAME_VERSION (valnum)];
3912 el_avail_stack.safe_push (pushop);
3913 el_avail[SSA_NAME_VERSION (valnum)] = op;
3917 /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns
3918 the leader for the expression if insertion was successful. */
3920 static tree
3921 eliminate_insert (gimple_stmt_iterator *gsi, tree val)
3923 gimple *stmt = gimple_seq_first_stmt (VN_INFO (val)->expr);
3924 if (!is_gimple_assign (stmt)
3925 || (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
3926 && gimple_assign_rhs_code (stmt) != VIEW_CONVERT_EXPR
3927 && gimple_assign_rhs_code (stmt) != BIT_FIELD_REF))
3928 return NULL_TREE;
3930 tree op = gimple_assign_rhs1 (stmt);
3931 if (gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR
3932 || gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
3933 op = TREE_OPERAND (op, 0);
3934 tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op;
3935 if (!leader)
3936 return NULL_TREE;
3938 gimple_seq stmts = NULL;
3939 tree res;
3940 if (gimple_assign_rhs_code (stmt) == BIT_FIELD_REF)
3941 res = gimple_build (&stmts, BIT_FIELD_REF,
3942 TREE_TYPE (val), leader,
3943 TREE_OPERAND (gimple_assign_rhs1 (stmt), 1),
3944 TREE_OPERAND (gimple_assign_rhs1 (stmt), 2));
3945 else
3946 res = gimple_build (&stmts, gimple_assign_rhs_code (stmt),
3947 TREE_TYPE (val), leader);
3948 gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
3949 VN_INFO_GET (res)->valnum = val;
3951 if (TREE_CODE (leader) == SSA_NAME)
3952 gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true);
3954 pre_stats.insertions++;
3955 if (dump_file && (dump_flags & TDF_DETAILS))
3957 fprintf (dump_file, "Inserted ");
3958 print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (res), 0, 0);
3961 return res;
3964 class eliminate_dom_walker : public dom_walker
3966 public:
3967 eliminate_dom_walker (cdi_direction direction, bool do_pre_)
3968 : dom_walker (direction), do_pre (do_pre_) {}
3970 virtual edge before_dom_children (basic_block);
3971 virtual void after_dom_children (basic_block);
3973 bool do_pre;
3976 /* Perform elimination for the basic-block B during the domwalk. */
3978 edge
3979 eliminate_dom_walker::before_dom_children (basic_block b)
3981 /* Mark new bb. */
3982 el_avail_stack.safe_push (NULL_TREE);
3984 /* ??? If we do nothing for unreachable blocks then this will confuse
3985 tailmerging. Eventually we can reduce its reliance on SCCVN now
3986 that we fully copy/constant-propagate (most) things. */
3988 for (gphi_iterator gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
3990 gphi *phi = gsi.phi ();
3991 tree res = PHI_RESULT (phi);
3993 if (virtual_operand_p (res))
3995 gsi_next (&gsi);
3996 continue;
3999 tree sprime = eliminate_avail (res);
4000 if (sprime
4001 && sprime != res)
4003 if (dump_file && (dump_flags & TDF_DETAILS))
4005 fprintf (dump_file, "Replaced redundant PHI node defining ");
4006 print_generic_expr (dump_file, res, 0);
4007 fprintf (dump_file, " with ");
4008 print_generic_expr (dump_file, sprime, 0);
4009 fprintf (dump_file, "\n");
4012 /* If we inserted this PHI node ourself, it's not an elimination. */
4013 if (inserted_exprs
4014 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4015 pre_stats.phis--;
4016 else
4017 pre_stats.eliminations++;
4019 /* If we will propagate into all uses don't bother to do
4020 anything. */
4021 if (may_propagate_copy (res, sprime))
4023 /* Mark the PHI for removal. */
4024 el_to_remove.safe_push (phi);
4025 gsi_next (&gsi);
4026 continue;
4029 remove_phi_node (&gsi, false);
4031 if (inserted_exprs
4032 && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4033 && TREE_CODE (sprime) == SSA_NAME)
4034 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4036 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4037 sprime = fold_convert (TREE_TYPE (res), sprime);
4038 gimple *stmt = gimple_build_assign (res, sprime);
4039 /* ??? It cannot yet be necessary (DOM walk). */
4040 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4042 gimple_stmt_iterator gsi2 = gsi_after_labels (b);
4043 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4044 continue;
4047 eliminate_push_avail (res);
4048 gsi_next (&gsi);
4051 for (gimple_stmt_iterator gsi = gsi_start_bb (b);
4052 !gsi_end_p (gsi);
4053 gsi_next (&gsi))
4055 tree sprime = NULL_TREE;
4056 gimple *stmt = gsi_stmt (gsi);
4057 tree lhs = gimple_get_lhs (stmt);
4058 if (lhs && TREE_CODE (lhs) == SSA_NAME
4059 && !gimple_has_volatile_ops (stmt)
4060 /* See PR43491. Do not replace a global register variable when
4061 it is a the RHS of an assignment. Do replace local register
4062 variables since gcc does not guarantee a local variable will
4063 be allocated in register.
4064 ??? The fix isn't effective here. This should instead
4065 be ensured by not value-numbering them the same but treating
4066 them like volatiles? */
4067 && !(gimple_assign_single_p (stmt)
4068 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL
4069 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt))
4070 && is_global_var (gimple_assign_rhs1 (stmt)))))
4072 sprime = eliminate_avail (lhs);
4073 if (!sprime)
4075 /* If there is no existing usable leader but SCCVN thinks
4076 it has an expression it wants to use as replacement,
4077 insert that. */
4078 tree val = VN_INFO (lhs)->valnum;
4079 if (val != VN_TOP
4080 && TREE_CODE (val) == SSA_NAME
4081 && VN_INFO (val)->needs_insertion
4082 && VN_INFO (val)->expr != NULL
4083 && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE)
4084 eliminate_push_avail (sprime);
4087 /* If this now constitutes a copy duplicate points-to
4088 and range info appropriately. This is especially
4089 important for inserted code. See tree-ssa-copy.c
4090 for similar code. */
4091 if (sprime
4092 && TREE_CODE (sprime) == SSA_NAME)
4094 basic_block sprime_b = gimple_bb (SSA_NAME_DEF_STMT (sprime));
4095 if (POINTER_TYPE_P (TREE_TYPE (lhs))
4096 && VN_INFO_PTR_INFO (lhs)
4097 && ! VN_INFO_PTR_INFO (sprime))
4099 duplicate_ssa_name_ptr_info (sprime,
4100 VN_INFO_PTR_INFO (lhs));
4101 if (b != sprime_b)
4102 mark_ptr_info_alignment_unknown
4103 (SSA_NAME_PTR_INFO (sprime));
4105 else if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))
4106 && VN_INFO_RANGE_INFO (lhs)
4107 && ! VN_INFO_RANGE_INFO (sprime)
4108 && b == sprime_b)
4109 duplicate_ssa_name_range_info (sprime,
4110 VN_INFO_RANGE_TYPE (lhs),
4111 VN_INFO_RANGE_INFO (lhs));
4114 /* Inhibit the use of an inserted PHI on a loop header when
4115 the address of the memory reference is a simple induction
4116 variable. In other cases the vectorizer won't do anything
4117 anyway (either it's loop invariant or a complicated
4118 expression). */
4119 if (sprime
4120 && TREE_CODE (sprime) == SSA_NAME
4121 && do_pre
4122 && flag_tree_loop_vectorize
4123 && loop_outer (b->loop_father)
4124 && has_zero_uses (sprime)
4125 && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))
4126 && gimple_assign_load_p (stmt))
4128 gimple *def_stmt = SSA_NAME_DEF_STMT (sprime);
4129 basic_block def_bb = gimple_bb (def_stmt);
4130 if (gimple_code (def_stmt) == GIMPLE_PHI
4131 && def_bb->loop_father->header == def_bb)
4133 loop_p loop = def_bb->loop_father;
4134 ssa_op_iter iter;
4135 tree op;
4136 bool found = false;
4137 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4139 affine_iv iv;
4140 def_bb = gimple_bb (SSA_NAME_DEF_STMT (op));
4141 if (def_bb
4142 && flow_bb_inside_loop_p (loop, def_bb)
4143 && simple_iv (loop, loop, op, &iv, true))
4145 found = true;
4146 break;
4149 if (found)
4151 if (dump_file && (dump_flags & TDF_DETAILS))
4153 fprintf (dump_file, "Not replacing ");
4154 print_gimple_expr (dump_file, stmt, 0, 0);
4155 fprintf (dump_file, " with ");
4156 print_generic_expr (dump_file, sprime, 0);
4157 fprintf (dump_file, " which would add a loop"
4158 " carried dependence to loop %d\n",
4159 loop->num);
4161 /* Don't keep sprime available. */
4162 sprime = NULL_TREE;
4167 if (sprime)
4169 /* If we can propagate the value computed for LHS into
4170 all uses don't bother doing anything with this stmt. */
4171 if (may_propagate_copy (lhs, sprime))
4173 /* Mark it for removal. */
4174 el_to_remove.safe_push (stmt);
4176 /* ??? Don't count copy/constant propagations. */
4177 if (gimple_assign_single_p (stmt)
4178 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4179 || gimple_assign_rhs1 (stmt) == sprime))
4180 continue;
4182 if (dump_file && (dump_flags & TDF_DETAILS))
4184 fprintf (dump_file, "Replaced ");
4185 print_gimple_expr (dump_file, stmt, 0, 0);
4186 fprintf (dump_file, " with ");
4187 print_generic_expr (dump_file, sprime, 0);
4188 fprintf (dump_file, " in all uses of ");
4189 print_gimple_stmt (dump_file, stmt, 0, 0);
4192 pre_stats.eliminations++;
4193 continue;
4196 /* If this is an assignment from our leader (which
4197 happens in the case the value-number is a constant)
4198 then there is nothing to do. */
4199 if (gimple_assign_single_p (stmt)
4200 && sprime == gimple_assign_rhs1 (stmt))
4201 continue;
4203 /* Else replace its RHS. */
4204 bool can_make_abnormal_goto
4205 = is_gimple_call (stmt)
4206 && stmt_can_make_abnormal_goto (stmt);
4208 if (dump_file && (dump_flags & TDF_DETAILS))
4210 fprintf (dump_file, "Replaced ");
4211 print_gimple_expr (dump_file, stmt, 0, 0);
4212 fprintf (dump_file, " with ");
4213 print_generic_expr (dump_file, sprime, 0);
4214 fprintf (dump_file, " in ");
4215 print_gimple_stmt (dump_file, stmt, 0, 0);
4218 if (TREE_CODE (sprime) == SSA_NAME)
4219 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4220 NECESSARY, true);
4222 pre_stats.eliminations++;
4223 gimple *orig_stmt = stmt;
4224 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4225 TREE_TYPE (sprime)))
4226 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4227 tree vdef = gimple_vdef (stmt);
4228 tree vuse = gimple_vuse (stmt);
4229 propagate_tree_value_into_stmt (&gsi, sprime);
4230 stmt = gsi_stmt (gsi);
4231 update_stmt (stmt);
4232 if (vdef != gimple_vdef (stmt))
4233 VN_INFO (vdef)->valnum = vuse;
4235 /* If we removed EH side-effects from the statement, clean
4236 its EH information. */
4237 if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
4239 bitmap_set_bit (need_eh_cleanup,
4240 gimple_bb (stmt)->index);
4241 if (dump_file && (dump_flags & TDF_DETAILS))
4242 fprintf (dump_file, " Removed EH side-effects.\n");
4245 /* Likewise for AB side-effects. */
4246 if (can_make_abnormal_goto
4247 && !stmt_can_make_abnormal_goto (stmt))
4249 bitmap_set_bit (need_ab_cleanup,
4250 gimple_bb (stmt)->index);
4251 if (dump_file && (dump_flags & TDF_DETAILS))
4252 fprintf (dump_file, " Removed AB side-effects.\n");
4255 continue;
4259 /* If the statement is a scalar store, see if the expression
4260 has the same value number as its rhs. If so, the store is
4261 dead. */
4262 if (gimple_assign_single_p (stmt)
4263 && !gimple_has_volatile_ops (stmt)
4264 && !is_gimple_reg (gimple_assign_lhs (stmt))
4265 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4266 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4268 tree val;
4269 tree rhs = gimple_assign_rhs1 (stmt);
4270 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4271 gimple_vuse (stmt), VN_WALK, NULL, false);
4272 if (TREE_CODE (rhs) == SSA_NAME)
4273 rhs = VN_INFO (rhs)->valnum;
4274 if (val
4275 && operand_equal_p (val, rhs, 0))
4277 if (dump_file && (dump_flags & TDF_DETAILS))
4279 fprintf (dump_file, "Deleted redundant store ");
4280 print_gimple_stmt (dump_file, stmt, 0, 0);
4283 /* Queue stmt for removal. */
4284 el_to_remove.safe_push (stmt);
4285 continue;
4289 /* If this is a control statement value numbering left edges
4290 unexecuted on force the condition in a way consistent with
4291 that. */
4292 if (gcond *cond = dyn_cast <gcond *> (stmt))
4294 if ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE)
4295 ^ (EDGE_SUCC (b, 1)->flags & EDGE_EXECUTABLE))
4297 if (dump_file && (dump_flags & TDF_DETAILS))
4299 fprintf (dump_file, "Removing unexecutable edge from ");
4300 print_gimple_stmt (dump_file, stmt, 0, 0);
4302 if (((EDGE_SUCC (b, 0)->flags & EDGE_TRUE_VALUE) != 0)
4303 == ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE) != 0))
4304 gimple_cond_make_true (cond);
4305 else
4306 gimple_cond_make_false (cond);
4307 update_stmt (cond);
4308 el_todo |= TODO_cleanup_cfg;
4309 continue;
4313 bool can_make_abnormal_goto = stmt_can_make_abnormal_goto (stmt);
4314 bool was_noreturn = (is_gimple_call (stmt)
4315 && gimple_call_noreturn_p (stmt));
4316 tree vdef = gimple_vdef (stmt);
4317 tree vuse = gimple_vuse (stmt);
4319 /* If we didn't replace the whole stmt (or propagate the result
4320 into all uses), replace all uses on this stmt with their
4321 leaders. */
4322 use_operand_p use_p;
4323 ssa_op_iter iter;
4324 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
4326 tree use = USE_FROM_PTR (use_p);
4327 /* ??? The call code above leaves stmt operands un-updated. */
4328 if (TREE_CODE (use) != SSA_NAME)
4329 continue;
4330 tree sprime = eliminate_avail (use);
4331 if (sprime && sprime != use
4332 && may_propagate_copy (use, sprime)
4333 /* We substitute into debug stmts to avoid excessive
4334 debug temporaries created by removed stmts, but we need
4335 to avoid doing so for inserted sprimes as we never want
4336 to create debug temporaries for them. */
4337 && (!inserted_exprs
4338 || TREE_CODE (sprime) != SSA_NAME
4339 || !is_gimple_debug (stmt)
4340 || !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))))
4342 propagate_value (use_p, sprime);
4343 gimple_set_modified (stmt, true);
4344 if (TREE_CODE (sprime) == SSA_NAME
4345 && !is_gimple_debug (stmt))
4346 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4347 NECESSARY, true);
4351 /* Visit indirect calls and turn them into direct calls if
4352 possible using the devirtualization machinery. */
4353 if (gcall *call_stmt = dyn_cast <gcall *> (stmt))
4355 tree fn = gimple_call_fn (call_stmt);
4356 if (fn
4357 && flag_devirtualize
4358 && virtual_method_call_p (fn))
4360 tree otr_type = obj_type_ref_class (fn);
4361 tree instance;
4362 ipa_polymorphic_call_context context (current_function_decl, fn, stmt, &instance);
4363 bool final;
4365 context.get_dynamic_type (instance, OBJ_TYPE_REF_OBJECT (fn), otr_type, stmt);
4367 vec <cgraph_node *>targets
4368 = possible_polymorphic_call_targets (obj_type_ref_class (fn),
4369 tree_to_uhwi
4370 (OBJ_TYPE_REF_TOKEN (fn)),
4371 context,
4372 &final);
4373 if (dump_file)
4374 dump_possible_polymorphic_call_targets (dump_file,
4375 obj_type_ref_class (fn),
4376 tree_to_uhwi
4377 (OBJ_TYPE_REF_TOKEN (fn)),
4378 context);
4379 if (final && targets.length () <= 1 && dbg_cnt (devirt))
4381 tree fn;
4382 if (targets.length () == 1)
4383 fn = targets[0]->decl;
4384 else
4385 fn = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
4386 if (dump_enabled_p ())
4388 location_t loc = gimple_location_safe (stmt);
4389 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc,
4390 "converting indirect call to "
4391 "function %s\n",
4392 lang_hooks.decl_printable_name (fn, 2));
4394 gimple_call_set_fndecl (call_stmt, fn);
4395 maybe_remove_unused_call_args (cfun, call_stmt);
4396 gimple_set_modified (stmt, true);
4401 if (gimple_modified_p (stmt))
4403 /* If a formerly non-invariant ADDR_EXPR is turned into an
4404 invariant one it was on a separate stmt. */
4405 if (gimple_assign_single_p (stmt)
4406 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR)
4407 recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
4408 gimple *old_stmt = stmt;
4409 if (is_gimple_call (stmt))
4411 /* ??? Only fold calls inplace for now, this may create new
4412 SSA names which in turn will confuse free_scc_vn SSA name
4413 release code. */
4414 fold_stmt_inplace (&gsi);
4415 /* When changing a call into a noreturn call, cfg cleanup
4416 is needed to fix up the noreturn call. */
4417 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4418 el_to_fixup.safe_push (stmt);
4420 else
4422 fold_stmt (&gsi);
4423 stmt = gsi_stmt (gsi);
4424 if ((gimple_code (stmt) == GIMPLE_COND
4425 && (gimple_cond_true_p (as_a <gcond *> (stmt))
4426 || gimple_cond_false_p (as_a <gcond *> (stmt))))
4427 || (gimple_code (stmt) == GIMPLE_SWITCH
4428 && TREE_CODE (gimple_switch_index (
4429 as_a <gswitch *> (stmt)))
4430 == INTEGER_CST))
4431 el_todo |= TODO_cleanup_cfg;
4433 /* If we removed EH side-effects from the statement, clean
4434 its EH information. */
4435 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
4437 bitmap_set_bit (need_eh_cleanup,
4438 gimple_bb (stmt)->index);
4439 if (dump_file && (dump_flags & TDF_DETAILS))
4440 fprintf (dump_file, " Removed EH side-effects.\n");
4442 /* Likewise for AB side-effects. */
4443 if (can_make_abnormal_goto
4444 && !stmt_can_make_abnormal_goto (stmt))
4446 bitmap_set_bit (need_ab_cleanup,
4447 gimple_bb (stmt)->index);
4448 if (dump_file && (dump_flags & TDF_DETAILS))
4449 fprintf (dump_file, " Removed AB side-effects.\n");
4451 update_stmt (stmt);
4452 if (vdef != gimple_vdef (stmt))
4453 VN_INFO (vdef)->valnum = vuse;
4456 /* Make new values available - for fully redundant LHS we
4457 continue with the next stmt above and skip this. */
4458 def_operand_p defp;
4459 FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_DEF)
4460 eliminate_push_avail (DEF_FROM_PTR (defp));
4463 /* Replace destination PHI arguments. */
4464 edge_iterator ei;
4465 edge e;
4466 FOR_EACH_EDGE (e, ei, b->succs)
4468 for (gphi_iterator gsi = gsi_start_phis (e->dest);
4469 !gsi_end_p (gsi);
4470 gsi_next (&gsi))
4472 gphi *phi = gsi.phi ();
4473 use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
4474 tree arg = USE_FROM_PTR (use_p);
4475 if (TREE_CODE (arg) != SSA_NAME
4476 || virtual_operand_p (arg))
4477 continue;
4478 tree sprime = eliminate_avail (arg);
4479 if (sprime && may_propagate_copy (arg, sprime))
4481 propagate_value (use_p, sprime);
4482 if (TREE_CODE (sprime) == SSA_NAME)
4483 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4487 return NULL;
4490 /* Make no longer available leaders no longer available. */
4492 void
4493 eliminate_dom_walker::after_dom_children (basic_block)
4495 tree entry;
4496 while ((entry = el_avail_stack.pop ()) != NULL_TREE)
4498 tree valnum = VN_INFO (entry)->valnum;
4499 tree old = el_avail[SSA_NAME_VERSION (valnum)];
4500 if (old == entry)
4501 el_avail[SSA_NAME_VERSION (valnum)] = NULL_TREE;
4502 else
4503 el_avail[SSA_NAME_VERSION (valnum)] = entry;
4507 /* Eliminate fully redundant computations. */
4509 static unsigned int
4510 eliminate (bool do_pre)
4512 gimple_stmt_iterator gsi;
4513 gimple *stmt;
4515 need_eh_cleanup = BITMAP_ALLOC (NULL);
4516 need_ab_cleanup = BITMAP_ALLOC (NULL);
4518 el_to_remove.create (0);
4519 el_to_fixup.create (0);
4520 el_todo = 0;
4521 el_avail.create (num_ssa_names);
4522 el_avail_stack.create (0);
4524 eliminate_dom_walker (CDI_DOMINATORS,
4525 do_pre).walk (cfun->cfg->x_entry_block_ptr);
4527 el_avail.release ();
4528 el_avail_stack.release ();
4530 /* We cannot remove stmts during BB walk, especially not release SSA
4531 names there as this confuses the VN machinery. The stmts ending
4532 up in el_to_remove are either stores or simple copies.
4533 Remove stmts in reverse order to make debug stmt creation possible. */
4534 while (!el_to_remove.is_empty ())
4536 stmt = el_to_remove.pop ();
4538 if (dump_file && (dump_flags & TDF_DETAILS))
4540 fprintf (dump_file, "Removing dead stmt ");
4541 print_gimple_stmt (dump_file, stmt, 0, 0);
4544 tree lhs;
4545 if (gimple_code (stmt) == GIMPLE_PHI)
4546 lhs = gimple_phi_result (stmt);
4547 else
4548 lhs = gimple_get_lhs (stmt);
4550 if (inserted_exprs
4551 && TREE_CODE (lhs) == SSA_NAME)
4552 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4554 gsi = gsi_for_stmt (stmt);
4555 if (gimple_code (stmt) == GIMPLE_PHI)
4556 remove_phi_node (&gsi, true);
4557 else
4559 basic_block bb = gimple_bb (stmt);
4560 unlink_stmt_vdef (stmt);
4561 if (gsi_remove (&gsi, true))
4562 bitmap_set_bit (need_eh_cleanup, bb->index);
4563 if (is_gimple_call (stmt) && stmt_can_make_abnormal_goto (stmt))
4564 bitmap_set_bit (need_ab_cleanup, bb->index);
4565 release_defs (stmt);
4568 /* Removing a stmt may expose a forwarder block. */
4569 el_todo |= TODO_cleanup_cfg;
4571 el_to_remove.release ();
4573 /* Fixup stmts that became noreturn calls. This may require splitting
4574 blocks and thus isn't possible during the dominator walk. Do this
4575 in reverse order so we don't inadvertedly remove a stmt we want to
4576 fixup by visiting a dominating now noreturn call first. */
4577 while (!el_to_fixup.is_empty ())
4579 stmt = el_to_fixup.pop ();
4581 if (dump_file && (dump_flags & TDF_DETAILS))
4583 fprintf (dump_file, "Fixing up noreturn call ");
4584 print_gimple_stmt (dump_file, stmt, 0, 0);
4587 if (fixup_noreturn_call (stmt))
4588 el_todo |= TODO_cleanup_cfg;
4590 el_to_fixup.release ();
4592 return el_todo;
4595 /* Perform CFG cleanups made necessary by elimination. */
4597 static unsigned
4598 fini_eliminate (void)
4600 bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup);
4601 bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup);
4603 if (do_eh_cleanup)
4604 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4606 if (do_ab_cleanup)
4607 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4609 BITMAP_FREE (need_eh_cleanup);
4610 BITMAP_FREE (need_ab_cleanup);
4612 if (do_eh_cleanup || do_ab_cleanup)
4613 return TODO_cleanup_cfg;
4614 return 0;
4617 /* Borrow a bit of tree-ssa-dce.c for the moment.
4618 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4619 this may be a bit faster, and we may want critical edges kept split. */
4621 /* If OP's defining statement has not already been determined to be necessary,
4622 mark that statement necessary. Return the stmt, if it is newly
4623 necessary. */
4625 static inline gimple *
4626 mark_operand_necessary (tree op)
4628 gimple *stmt;
4630 gcc_assert (op);
4632 if (TREE_CODE (op) != SSA_NAME)
4633 return NULL;
4635 stmt = SSA_NAME_DEF_STMT (op);
4636 gcc_assert (stmt);
4638 if (gimple_plf (stmt, NECESSARY)
4639 || gimple_nop_p (stmt))
4640 return NULL;
4642 gimple_set_plf (stmt, NECESSARY, true);
4643 return stmt;
4646 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4647 to insert PHI nodes sometimes, and because value numbering of casts isn't
4648 perfect, we sometimes end up inserting dead code. This simple DCE-like
4649 pass removes any insertions we made that weren't actually used. */
4651 static void
4652 remove_dead_inserted_code (void)
4654 bitmap worklist;
4655 unsigned i;
4656 bitmap_iterator bi;
4657 gimple *t;
4659 worklist = BITMAP_ALLOC (NULL);
4660 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4662 t = SSA_NAME_DEF_STMT (ssa_name (i));
4663 if (gimple_plf (t, NECESSARY))
4664 bitmap_set_bit (worklist, i);
4666 while (!bitmap_empty_p (worklist))
4668 i = bitmap_first_set_bit (worklist);
4669 bitmap_clear_bit (worklist, i);
4670 t = SSA_NAME_DEF_STMT (ssa_name (i));
4672 /* PHI nodes are somewhat special in that each PHI alternative has
4673 data and control dependencies. All the statements feeding the
4674 PHI node's arguments are always necessary. */
4675 if (gimple_code (t) == GIMPLE_PHI)
4677 unsigned k;
4679 for (k = 0; k < gimple_phi_num_args (t); k++)
4681 tree arg = PHI_ARG_DEF (t, k);
4682 if (TREE_CODE (arg) == SSA_NAME)
4684 gimple *n = mark_operand_necessary (arg);
4685 if (n)
4686 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4690 else
4692 /* Propagate through the operands. Examine all the USE, VUSE and
4693 VDEF operands in this statement. Mark all the statements
4694 which feed this statement's uses as necessary. */
4695 ssa_op_iter iter;
4696 tree use;
4698 /* The operands of VDEF expressions are also needed as they
4699 represent potential definitions that may reach this
4700 statement (VDEF operands allow us to follow def-def
4701 links). */
4703 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4705 gimple *n = mark_operand_necessary (use);
4706 if (n)
4707 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4712 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4714 t = SSA_NAME_DEF_STMT (ssa_name (i));
4715 if (!gimple_plf (t, NECESSARY))
4717 gimple_stmt_iterator gsi;
4719 if (dump_file && (dump_flags & TDF_DETAILS))
4721 fprintf (dump_file, "Removing unnecessary insertion:");
4722 print_gimple_stmt (dump_file, t, 0, 0);
4725 gsi = gsi_for_stmt (t);
4726 if (gimple_code (t) == GIMPLE_PHI)
4727 remove_phi_node (&gsi, true);
4728 else
4730 gsi_remove (&gsi, true);
4731 release_defs (t);
4735 BITMAP_FREE (worklist);
4739 /* Initialize data structures used by PRE. */
4741 static void
4742 init_pre (void)
4744 basic_block bb;
4746 next_expression_id = 1;
4747 expressions.create (0);
4748 expressions.safe_push (NULL);
4749 value_expressions.create (get_max_value_id () + 1);
4750 value_expressions.safe_grow_cleared (get_max_value_id () + 1);
4751 name_to_id.create (0);
4753 inserted_exprs = BITMAP_ALLOC (NULL);
4755 connect_infinite_loops_to_exit ();
4756 memset (&pre_stats, 0, sizeof (pre_stats));
4758 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4760 calculate_dominance_info (CDI_DOMINATORS);
4762 bitmap_obstack_initialize (&grand_bitmap_obstack);
4763 phi_translate_table = new hash_table<expr_pred_trans_d> (5110);
4764 expression_to_id = new hash_table<pre_expr_d> (num_ssa_names * 3);
4765 FOR_ALL_BB_FN (bb, cfun)
4767 EXP_GEN (bb) = bitmap_set_new ();
4768 PHI_GEN (bb) = bitmap_set_new ();
4769 TMP_GEN (bb) = bitmap_set_new ();
4770 AVAIL_OUT (bb) = bitmap_set_new ();
4775 /* Deallocate data structures used by PRE. */
4777 static void
4778 fini_pre ()
4780 value_expressions.release ();
4781 BITMAP_FREE (inserted_exprs);
4782 bitmap_obstack_release (&grand_bitmap_obstack);
4783 bitmap_set_pool.release ();
4784 pre_expr_pool.release ();
4785 delete phi_translate_table;
4786 phi_translate_table = NULL;
4787 delete expression_to_id;
4788 expression_to_id = NULL;
4789 name_to_id.release ();
4791 free_aux_for_blocks ();
4794 namespace {
4796 const pass_data pass_data_pre =
4798 GIMPLE_PASS, /* type */
4799 "pre", /* name */
4800 OPTGROUP_NONE, /* optinfo_flags */
4801 TV_TREE_PRE, /* tv_id */
4802 /* PROP_no_crit_edges is ensured by placing pass_split_crit_edges before
4803 pass_pre. */
4804 ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */
4805 0, /* properties_provided */
4806 PROP_no_crit_edges, /* properties_destroyed */
4807 TODO_rebuild_alias, /* todo_flags_start */
4808 0, /* todo_flags_finish */
4811 class pass_pre : public gimple_opt_pass
4813 public:
4814 pass_pre (gcc::context *ctxt)
4815 : gimple_opt_pass (pass_data_pre, ctxt)
4818 /* opt_pass methods: */
4819 virtual bool gate (function *) { return flag_tree_pre != 0; }
4820 virtual unsigned int execute (function *);
4822 }; // class pass_pre
4824 unsigned int
4825 pass_pre::execute (function *fun)
4827 unsigned int todo = 0;
4829 do_partial_partial =
4830 flag_tree_partial_pre && optimize_function_for_speed_p (fun);
4832 /* This has to happen before SCCVN runs because
4833 loop_optimizer_init may create new phis, etc. */
4834 loop_optimizer_init (LOOPS_NORMAL);
4836 if (!run_scc_vn (VN_WALK))
4838 loop_optimizer_finalize ();
4839 return 0;
4842 init_pre ();
4843 scev_initialize ();
4845 /* Collect and value number expressions computed in each basic block. */
4846 compute_avail ();
4848 /* Insert can get quite slow on an incredibly large number of basic
4849 blocks due to some quadratic behavior. Until this behavior is
4850 fixed, don't run it when he have an incredibly large number of
4851 bb's. If we aren't going to run insert, there is no point in
4852 computing ANTIC, either, even though it's plenty fast. */
4853 if (n_basic_blocks_for_fn (fun) < 4000)
4855 compute_antic ();
4856 insert ();
4859 /* Make sure to remove fake edges before committing our inserts.
4860 This makes sure we don't end up with extra critical edges that
4861 we would need to split. */
4862 remove_fake_exit_edges ();
4863 gsi_commit_edge_inserts ();
4865 /* Eliminate folds statements which might (should not...) end up
4866 not keeping virtual operands up-to-date. */
4867 gcc_assert (!need_ssa_update_p (fun));
4869 /* Remove all the redundant expressions. */
4870 todo |= eliminate (true);
4872 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
4873 statistics_counter_event (fun, "PA inserted", pre_stats.pa_insert);
4874 statistics_counter_event (fun, "New PHIs", pre_stats.phis);
4875 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
4877 clear_expression_ids ();
4878 remove_dead_inserted_code ();
4880 scev_finalize ();
4881 fini_pre ();
4882 todo |= fini_eliminate ();
4883 loop_optimizer_finalize ();
4885 /* Restore SSA info before tail-merging as that resets it as well. */
4886 scc_vn_restore_ssa_info ();
4888 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4889 case we can merge the block with the remaining predecessor of the block.
4890 It should either:
4891 - call merge_blocks after each tail merge iteration
4892 - call merge_blocks after all tail merge iterations
4893 - mark TODO_cleanup_cfg when necessary
4894 - share the cfg cleanup with fini_pre. */
4895 todo |= tail_merge_optimize (todo);
4897 free_scc_vn ();
4899 /* Tail merging invalidates the virtual SSA web, together with
4900 cfg-cleanup opportunities exposed by PRE this will wreck the
4901 SSA updating machinery. So make sure to run update-ssa
4902 manually, before eventually scheduling cfg-cleanup as part of
4903 the todo. */
4904 update_ssa (TODO_update_ssa_only_virtuals);
4906 return todo;
4909 } // anon namespace
4911 gimple_opt_pass *
4912 make_pass_pre (gcc::context *ctxt)
4914 return new pass_pre (ctxt);
4917 namespace {
4919 const pass_data pass_data_fre =
4921 GIMPLE_PASS, /* type */
4922 "fre", /* name */
4923 OPTGROUP_NONE, /* optinfo_flags */
4924 TV_TREE_FRE, /* tv_id */
4925 ( PROP_cfg | PROP_ssa ), /* properties_required */
4926 0, /* properties_provided */
4927 0, /* properties_destroyed */
4928 0, /* todo_flags_start */
4929 0, /* todo_flags_finish */
4932 class pass_fre : public gimple_opt_pass
4934 public:
4935 pass_fre (gcc::context *ctxt)
4936 : gimple_opt_pass (pass_data_fre, ctxt)
4939 /* opt_pass methods: */
4940 opt_pass * clone () { return new pass_fre (m_ctxt); }
4941 virtual bool gate (function *) { return flag_tree_fre != 0; }
4942 virtual unsigned int execute (function *);
4944 }; // class pass_fre
4946 unsigned int
4947 pass_fre::execute (function *fun)
4949 unsigned int todo = 0;
4951 if (!run_scc_vn (VN_WALKREWRITE))
4952 return 0;
4954 memset (&pre_stats, 0, sizeof (pre_stats));
4956 /* Remove all the redundant expressions. */
4957 todo |= eliminate (false);
4959 todo |= fini_eliminate ();
4961 scc_vn_restore_ssa_info ();
4962 free_scc_vn ();
4964 statistics_counter_event (fun, "Insertions", pre_stats.insertions);
4965 statistics_counter_event (fun, "Eliminated", pre_stats.eliminations);
4967 return todo;
4970 } // anon namespace
4972 gimple_opt_pass *
4973 make_pass_fre (gcc::context *ctxt)
4975 return new pass_fre (ctxt);