Merge from mainline (163495:164578).
[official-gcc/graphite-test-results.git] / gcc / tree-ssa-pre.c
blob41d0542e013fd9d0376efeb58feee70658dbc5fd
1 /* SSA-PRE for trees.
2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
5 <stevenb@suse.de>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
12 any later version.
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 #include "config.h"
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "tree.h"
28 #include "basic-block.h"
29 #include "tree-pretty-print.h"
30 #include "gimple-pretty-print.h"
31 #include "tree-inline.h"
32 #include "tree-flow.h"
33 #include "gimple.h"
34 #include "tree-dump.h"
35 #include "timevar.h"
36 #include "fibheap.h"
37 #include "hashtab.h"
38 #include "tree-iterator.h"
39 #include "alloc-pool.h"
40 #include "obstack.h"
41 #include "tree-pass.h"
42 #include "flags.h"
43 #include "bitmap.h"
44 #include "langhooks.h"
45 #include "cfgloop.h"
46 #include "tree-ssa-sccvn.h"
47 #include "tree-scalar-evolution.h"
48 #include "params.h"
49 #include "dbgcnt.h"
51 /* TODO:
53 1. Avail sets can be shared by making an avail_find_leader that
54 walks up the dominator tree and looks in those avail sets.
55 This might affect code optimality, it's unclear right now.
56 2. Strength reduction can be performed by anticipating expressions
57 we can repair later on.
58 3. We can do back-substitution or smarter value numbering to catch
59 commutative expressions split up over multiple statements.
62 /* For ease of terminology, "expression node" in the below refers to
63 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
64 represent the actual statement containing the expressions we care about,
65 and we cache the value number by putting it in the expression. */
67 /* Basic algorithm
69 First we walk the statements to generate the AVAIL sets, the
70 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
71 generation of values/expressions by a given block. We use them
72 when computing the ANTIC sets. The AVAIL sets consist of
73 SSA_NAME's that represent values, so we know what values are
74 available in what blocks. AVAIL is a forward dataflow problem. In
75 SSA, values are never killed, so we don't need a kill set, or a
76 fixpoint iteration, in order to calculate the AVAIL sets. In
77 traditional parlance, AVAIL sets tell us the downsafety of the
78 expressions/values.
80 Next, we generate the ANTIC sets. These sets represent the
81 anticipatable expressions. ANTIC is a backwards dataflow
82 problem. An expression is anticipatable in a given block if it could
83 be generated in that block. This means that if we had to perform
84 an insertion in that block, of the value of that expression, we
85 could. Calculating the ANTIC sets requires phi translation of
86 expressions, because the flow goes backwards through phis. We must
87 iterate to a fixpoint of the ANTIC sets, because we have a kill
88 set. Even in SSA form, values are not live over the entire
89 function, only from their definition point onwards. So we have to
90 remove values from the ANTIC set once we go past the definition
91 point of the leaders that make them up.
92 compute_antic/compute_antic_aux performs this computation.
94 Third, we perform insertions to make partially redundant
95 expressions fully redundant.
97 An expression is partially redundant (excluding partial
98 anticipation) if:
100 1. It is AVAIL in some, but not all, of the predecessors of a
101 given block.
102 2. It is ANTIC in all the predecessors.
104 In order to make it fully redundant, we insert the expression into
105 the predecessors where it is not available, but is ANTIC.
107 For the partial anticipation case, we only perform insertion if it
108 is partially anticipated in some block, and fully available in all
109 of the predecessors.
111 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
112 performs these steps.
114 Fourth, we eliminate fully redundant expressions.
115 This is a simple statement walk that replaces redundant
116 calculations with the now available values. */
118 /* Representations of value numbers:
120 Value numbers are represented by a representative SSA_NAME. We
121 will create fake SSA_NAME's in situations where we need a
122 representative but do not have one (because it is a complex
123 expression). In order to facilitate storing the value numbers in
124 bitmaps, and keep the number of wasted SSA_NAME's down, we also
125 associate a value_id with each value number, and create full blown
126 ssa_name's only where we actually need them (IE in operands of
127 existing expressions).
129 Theoretically you could replace all the value_id's with
130 SSA_NAME_VERSION, but this would allocate a large number of
131 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
132 It would also require an additional indirection at each point we
133 use the value id. */
135 /* Representation of expressions on value numbers:
137 Expressions consisting of value numbers are represented the same
138 way as our VN internally represents them, with an additional
139 "pre_expr" wrapping around them in order to facilitate storing all
140 of the expressions in the same sets. */
142 /* Representation of sets:
144 The dataflow sets do not need to be sorted in any particular order
145 for the majority of their lifetime, are simply represented as two
146 bitmaps, one that keeps track of values present in the set, and one
147 that keeps track of expressions present in the set.
149 When we need them in topological order, we produce it on demand by
150 transforming the bitmap into an array and sorting it into topo
151 order. */
153 /* Type of expression, used to know which member of the PRE_EXPR union
154 is valid. */
156 enum pre_expr_kind
158 NAME,
159 NARY,
160 REFERENCE,
161 CONSTANT
164 typedef union pre_expr_union_d
166 tree name;
167 tree constant;
168 vn_nary_op_t nary;
169 vn_reference_t reference;
170 } pre_expr_union;
172 typedef struct pre_expr_d
174 enum pre_expr_kind kind;
175 unsigned int id;
176 pre_expr_union u;
177 } *pre_expr;
179 #define PRE_EXPR_NAME(e) (e)->u.name
180 #define PRE_EXPR_NARY(e) (e)->u.nary
181 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
182 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
184 static int
185 pre_expr_eq (const void *p1, const void *p2)
187 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
188 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
190 if (e1->kind != e2->kind)
191 return false;
193 switch (e1->kind)
195 case CONSTANT:
196 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
197 PRE_EXPR_CONSTANT (e2));
198 case NAME:
199 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
200 case NARY:
201 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
202 case REFERENCE:
203 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
204 PRE_EXPR_REFERENCE (e2));
205 default:
206 gcc_unreachable ();
210 static hashval_t
211 pre_expr_hash (const void *p1)
213 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
214 switch (e->kind)
216 case CONSTANT:
217 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
218 case NAME:
219 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
220 case NARY:
221 return PRE_EXPR_NARY (e)->hashcode;
222 case REFERENCE:
223 return PRE_EXPR_REFERENCE (e)->hashcode;
224 default:
225 gcc_unreachable ();
230 /* Next global expression id number. */
231 static unsigned int next_expression_id;
233 /* Mapping from expression to id number we can use in bitmap sets. */
234 DEF_VEC_P (pre_expr);
235 DEF_VEC_ALLOC_P (pre_expr, heap);
236 static VEC(pre_expr, heap) *expressions;
237 static htab_t expression_to_id;
238 static VEC(unsigned, heap) *name_to_id;
240 /* Allocate an expression id for EXPR. */
242 static inline unsigned int
243 alloc_expression_id (pre_expr expr)
245 void **slot;
246 /* Make sure we won't overflow. */
247 gcc_assert (next_expression_id + 1 > next_expression_id);
248 expr->id = next_expression_id++;
249 VEC_safe_push (pre_expr, heap, expressions, expr);
250 if (expr->kind == NAME)
252 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
253 /* VEC_safe_grow_cleared allocates no headroom. Avoid frequent
254 re-allocations by using VEC_reserve upfront. There is no
255 VEC_quick_grow_cleared unfortunately. */
256 VEC_reserve (unsigned, heap, name_to_id, num_ssa_names);
257 VEC_safe_grow_cleared (unsigned, heap, name_to_id, num_ssa_names);
258 gcc_assert (VEC_index (unsigned, name_to_id, version) == 0);
259 VEC_replace (unsigned, name_to_id, version, expr->id);
261 else
263 slot = htab_find_slot (expression_to_id, expr, INSERT);
264 gcc_assert (!*slot);
265 *slot = expr;
267 return next_expression_id - 1;
270 /* Return the expression id for tree EXPR. */
272 static inline unsigned int
273 get_expression_id (const pre_expr expr)
275 return expr->id;
278 static inline unsigned int
279 lookup_expression_id (const pre_expr expr)
281 void **slot;
283 if (expr->kind == NAME)
285 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
286 if (VEC_length (unsigned, name_to_id) <= version)
287 return 0;
288 return VEC_index (unsigned, name_to_id, version);
290 else
292 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
293 if (!slot)
294 return 0;
295 return ((pre_expr)*slot)->id;
299 /* Return the existing expression id for EXPR, or create one if one
300 does not exist yet. */
302 static inline unsigned int
303 get_or_alloc_expression_id (pre_expr expr)
305 unsigned int id = lookup_expression_id (expr);
306 if (id == 0)
307 return alloc_expression_id (expr);
308 return expr->id = id;
311 /* Return the expression that has expression id ID */
313 static inline pre_expr
314 expression_for_id (unsigned int id)
316 return VEC_index (pre_expr, expressions, id);
319 /* Free the expression id field in all of our expressions,
320 and then destroy the expressions array. */
322 static void
323 clear_expression_ids (void)
325 VEC_free (pre_expr, heap, expressions);
328 static alloc_pool pre_expr_pool;
330 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
332 static pre_expr
333 get_or_alloc_expr_for_name (tree name)
335 struct pre_expr_d expr;
336 pre_expr result;
337 unsigned int result_id;
339 expr.kind = NAME;
340 expr.id = 0;
341 PRE_EXPR_NAME (&expr) = name;
342 result_id = lookup_expression_id (&expr);
343 if (result_id != 0)
344 return expression_for_id (result_id);
346 result = (pre_expr) pool_alloc (pre_expr_pool);
347 result->kind = NAME;
348 PRE_EXPR_NAME (result) = name;
349 alloc_expression_id (result);
350 return result;
353 static bool in_fre = false;
355 /* An unordered bitmap set. One bitmap tracks values, the other,
356 expressions. */
357 typedef struct bitmap_set
359 bitmap_head expressions;
360 bitmap_head values;
361 } *bitmap_set_t;
363 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
364 EXECUTE_IF_SET_IN_BITMAP(&(set)->expressions, 0, (id), (bi))
366 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
367 EXECUTE_IF_SET_IN_BITMAP(&(set)->values, 0, (id), (bi))
369 /* Mapping from value id to expressions with that value_id. */
370 DEF_VEC_P (bitmap_set_t);
371 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
372 static VEC(bitmap_set_t, heap) *value_expressions;
374 /* Sets that we need to keep track of. */
375 typedef struct bb_bitmap_sets
377 /* The EXP_GEN set, which represents expressions/values generated in
378 a basic block. */
379 bitmap_set_t exp_gen;
381 /* The PHI_GEN set, which represents PHI results generated in a
382 basic block. */
383 bitmap_set_t phi_gen;
385 /* The TMP_GEN set, which represents results/temporaries generated
386 in a basic block. IE the LHS of an expression. */
387 bitmap_set_t tmp_gen;
389 /* The AVAIL_OUT set, which represents which values are available in
390 a given basic block. */
391 bitmap_set_t avail_out;
393 /* The ANTIC_IN set, which represents which values are anticipatable
394 in a given basic block. */
395 bitmap_set_t antic_in;
397 /* The PA_IN set, which represents which values are
398 partially anticipatable in a given basic block. */
399 bitmap_set_t pa_in;
401 /* The NEW_SETS set, which is used during insertion to augment the
402 AVAIL_OUT set of blocks with the new insertions performed during
403 the current iteration. */
404 bitmap_set_t new_sets;
406 /* A cache for value_dies_in_block_x. */
407 bitmap expr_dies;
409 /* True if we have visited this block during ANTIC calculation. */
410 unsigned int visited : 1;
412 /* True we have deferred processing this block during ANTIC
413 calculation until its successor is processed. */
414 unsigned int deferred : 1;
416 /* True when the block contains a call that might not return. */
417 unsigned int contains_may_not_return_call : 1;
418 } *bb_value_sets_t;
420 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
421 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
422 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
423 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
424 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
425 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
426 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
427 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
428 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
429 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
430 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
433 /* Basic block list in postorder. */
434 static int *postorder;
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;
452 /* The number of values found constant. */
453 int constified;
455 } pre_stats;
457 static bool do_partial_partial;
458 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
459 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
460 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
461 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
462 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
463 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
464 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
465 unsigned int, bool);
466 static bitmap_set_t bitmap_set_new (void);
467 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
468 gimple, tree);
469 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
470 gimple);
471 static unsigned int get_expr_value_id (pre_expr);
473 /* We can add and remove elements and entries to and from sets
474 and hash tables, so we use alloc pools for them. */
476 static alloc_pool bitmap_set_pool;
477 static bitmap_obstack grand_bitmap_obstack;
479 /* To avoid adding 300 temporary variables when we only need one, we
480 only create one temporary variable, on demand, and build ssa names
481 off that. We do have to change the variable if the types don't
482 match the current variable's type. */
483 static tree pretemp;
484 static tree storetemp;
485 static tree prephitemp;
487 /* Set of blocks with statements that have had its EH information
488 cleaned up. */
489 static bitmap need_eh_cleanup;
491 /* The phi_translate_table caches phi translations for a given
492 expression and predecessor. */
494 static htab_t phi_translate_table;
496 /* A three tuple {e, pred, v} used to cache phi translations in the
497 phi_translate_table. */
499 typedef struct expr_pred_trans_d
501 /* The expression. */
502 pre_expr e;
504 /* The predecessor block along which we translated the expression. */
505 basic_block pred;
507 /* The value that resulted from the translation. */
508 pre_expr v;
510 /* The hashcode for the expression, pred pair. This is cached for
511 speed reasons. */
512 hashval_t hashcode;
513 } *expr_pred_trans_t;
514 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
516 /* Return the hash value for a phi translation table entry. */
518 static hashval_t
519 expr_pred_trans_hash (const void *p)
521 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
522 return ve->hashcode;
525 /* Return true if two phi translation table entries are the same.
526 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
528 static int
529 expr_pred_trans_eq (const void *p1, const void *p2)
531 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
532 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
533 basic_block b1 = ve1->pred;
534 basic_block b2 = ve2->pred;
536 /* If they are not translations for the same basic block, they can't
537 be equal. */
538 if (b1 != b2)
539 return false;
540 return pre_expr_eq (ve1->e, ve2->e);
543 /* Search in the phi translation table for the translation of
544 expression E in basic block PRED.
545 Return the translated value, if found, NULL otherwise. */
547 static inline pre_expr
548 phi_trans_lookup (pre_expr e, basic_block pred)
550 void **slot;
551 struct expr_pred_trans_d ept;
553 ept.e = e;
554 ept.pred = pred;
555 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
556 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
557 NO_INSERT);
558 if (!slot)
559 return NULL;
560 else
561 return ((expr_pred_trans_t) *slot)->v;
565 /* Add the tuple mapping from {expression E, basic block PRED} to
566 value V, to the phi translation table. */
568 static inline void
569 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
571 void **slot;
572 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
573 new_pair->e = e;
574 new_pair->pred = pred;
575 new_pair->v = v;
576 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
577 pred->index);
579 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
580 new_pair->hashcode, INSERT);
581 if (*slot)
582 free (*slot);
583 *slot = (void *) new_pair;
587 /* Add expression E to the expression set of value id V. */
589 void
590 add_to_value (unsigned int v, pre_expr e)
592 bitmap_set_t set;
594 gcc_assert (get_expr_value_id (e) == v);
596 if (v >= VEC_length (bitmap_set_t, value_expressions))
598 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
599 v + 1);
602 set = VEC_index (bitmap_set_t, value_expressions, v);
603 if (!set)
605 set = bitmap_set_new ();
606 VEC_replace (bitmap_set_t, value_expressions, v, set);
609 bitmap_insert_into_set_1 (set, e, v, true);
612 /* Create a new bitmap set and return it. */
614 static bitmap_set_t
615 bitmap_set_new (void)
617 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
618 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
619 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
620 return ret;
623 /* Return the value id for a PRE expression EXPR. */
625 static unsigned int
626 get_expr_value_id (pre_expr expr)
628 switch (expr->kind)
630 case CONSTANT:
632 unsigned int id;
633 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
634 if (id == 0)
636 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
637 add_to_value (id, expr);
639 return id;
641 case NAME:
642 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
643 case NARY:
644 return PRE_EXPR_NARY (expr)->value_id;
645 case REFERENCE:
646 return PRE_EXPR_REFERENCE (expr)->value_id;
647 default:
648 gcc_unreachable ();
652 /* Remove an expression EXPR from a bitmapped set. */
654 static void
655 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
657 unsigned int val = get_expr_value_id (expr);
658 if (!value_id_constant_p (val))
660 bitmap_clear_bit (&set->values, val);
661 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
665 static void
666 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
667 unsigned int val, bool allow_constants)
669 if (allow_constants || !value_id_constant_p (val))
671 /* We specifically expect this and only this function to be able to
672 insert constants into a set. */
673 bitmap_set_bit (&set->values, val);
674 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
678 /* Insert an expression EXPR into a bitmapped set. */
680 static void
681 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
683 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
686 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
688 static void
689 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
691 bitmap_copy (&dest->expressions, &orig->expressions);
692 bitmap_copy (&dest->values, &orig->values);
696 /* Free memory used up by SET. */
697 static void
698 bitmap_set_free (bitmap_set_t set)
700 bitmap_clear (&set->expressions);
701 bitmap_clear (&set->values);
705 /* Generate an topological-ordered array of bitmap set SET. */
707 static VEC(pre_expr, heap) *
708 sorted_array_from_bitmap_set (bitmap_set_t set)
710 unsigned int i, j;
711 bitmap_iterator bi, bj;
712 VEC(pre_expr, heap) *result;
714 /* Pre-allocate roughly enough space for the array. */
715 result = VEC_alloc (pre_expr, heap, bitmap_count_bits (&set->values));
717 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
719 /* The number of expressions having a given value is usually
720 relatively small. Thus, rather than making a vector of all
721 the expressions and sorting it by value-id, we walk the values
722 and check in the reverse mapping that tells us what expressions
723 have a given value, to filter those in our set. As a result,
724 the expressions are inserted in value-id order, which means
725 topological order.
727 If this is somehow a significant lose for some cases, we can
728 choose which set to walk based on the set size. */
729 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
730 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
732 if (bitmap_bit_p (&set->expressions, j))
733 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
737 return result;
740 /* Perform bitmapped set operation DEST &= ORIG. */
742 static void
743 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
745 bitmap_iterator bi;
746 unsigned int i;
748 if (dest != orig)
750 bitmap_head temp;
751 bitmap_initialize (&temp, &grand_bitmap_obstack);
753 bitmap_and_into (&dest->values, &orig->values);
754 bitmap_copy (&temp, &dest->expressions);
755 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
757 pre_expr expr = expression_for_id (i);
758 unsigned int value_id = get_expr_value_id (expr);
759 if (!bitmap_bit_p (&dest->values, value_id))
760 bitmap_clear_bit (&dest->expressions, i);
762 bitmap_clear (&temp);
766 /* Subtract all values and expressions contained in ORIG from DEST. */
768 static bitmap_set_t
769 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
771 bitmap_set_t result = bitmap_set_new ();
772 bitmap_iterator bi;
773 unsigned int i;
775 bitmap_and_compl (&result->expressions, &dest->expressions,
776 &orig->expressions);
778 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
780 pre_expr expr = expression_for_id (i);
781 unsigned int value_id = get_expr_value_id (expr);
782 bitmap_set_bit (&result->values, value_id);
785 return result;
788 /* Subtract all the values in bitmap set B from bitmap set A. */
790 static void
791 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
793 unsigned int i;
794 bitmap_iterator bi;
795 bitmap_head temp;
797 bitmap_initialize (&temp, &grand_bitmap_obstack);
799 bitmap_copy (&temp, &a->expressions);
800 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
802 pre_expr expr = expression_for_id (i);
803 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
804 bitmap_remove_from_set (a, expr);
806 bitmap_clear (&temp);
810 /* Return true if bitmapped set SET contains the value VALUE_ID. */
812 static bool
813 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
815 if (value_id_constant_p (value_id))
816 return true;
818 if (!set || bitmap_empty_p (&set->expressions))
819 return false;
821 return bitmap_bit_p (&set->values, value_id);
824 static inline bool
825 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
827 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
830 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
832 static void
833 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
834 const pre_expr expr)
836 bitmap_set_t exprset;
837 unsigned int i;
838 bitmap_iterator bi;
840 if (value_id_constant_p (lookfor))
841 return;
843 if (!bitmap_set_contains_value (set, lookfor))
844 return;
846 /* The number of expressions having a given value is usually
847 significantly less than the total number of expressions in SET.
848 Thus, rather than check, for each expression in SET, whether it
849 has the value LOOKFOR, we walk the reverse mapping that tells us
850 what expressions have a given value, and see if any of those
851 expressions are in our set. For large testcases, this is about
852 5-10x faster than walking the bitmap. If this is somehow a
853 significant lose for some cases, we can choose which set to walk
854 based on the set size. */
855 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
856 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
858 if (bitmap_clear_bit (&set->expressions, i))
860 bitmap_set_bit (&set->expressions, get_expression_id (expr));
861 return;
866 /* Return true if two bitmap sets are equal. */
868 static bool
869 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
871 return bitmap_equal_p (&a->values, &b->values);
874 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
875 and add it otherwise. */
877 static void
878 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
880 unsigned int val = get_expr_value_id (expr);
882 if (bitmap_set_contains_value (set, val))
883 bitmap_set_replace_value (set, val, expr);
884 else
885 bitmap_insert_into_set (set, expr);
888 /* Insert EXPR into SET if EXPR's value is not already present in
889 SET. */
891 static void
892 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
894 unsigned int val = get_expr_value_id (expr);
896 #ifdef ENABLE_CHECKING
897 gcc_assert (expr->id == get_or_alloc_expression_id (expr));
898 #endif
900 /* Constant values are always considered to be part of the set. */
901 if (value_id_constant_p (val))
902 return;
904 /* If the value membership changed, add the expression. */
905 if (bitmap_set_bit (&set->values, val))
906 bitmap_set_bit (&set->expressions, expr->id);
909 /* Print out EXPR to outfile. */
911 static void
912 print_pre_expr (FILE *outfile, const pre_expr expr)
914 switch (expr->kind)
916 case CONSTANT:
917 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
918 break;
919 case NAME:
920 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
921 break;
922 case NARY:
924 unsigned int i;
925 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
926 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
927 for (i = 0; i < nary->length; i++)
929 print_generic_expr (outfile, nary->op[i], 0);
930 if (i != (unsigned) nary->length - 1)
931 fprintf (outfile, ",");
933 fprintf (outfile, "}");
935 break;
937 case REFERENCE:
939 vn_reference_op_t vro;
940 unsigned int i;
941 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
942 fprintf (outfile, "{");
943 for (i = 0;
944 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
945 i++)
947 bool closebrace = false;
948 if (vro->opcode != SSA_NAME
949 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
951 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
952 if (vro->op0)
954 fprintf (outfile, "<");
955 closebrace = true;
958 if (vro->op0)
960 print_generic_expr (outfile, vro->op0, 0);
961 if (vro->op1)
963 fprintf (outfile, ",");
964 print_generic_expr (outfile, vro->op1, 0);
966 if (vro->op2)
968 fprintf (outfile, ",");
969 print_generic_expr (outfile, vro->op2, 0);
972 if (closebrace)
973 fprintf (outfile, ">");
974 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
975 fprintf (outfile, ",");
977 fprintf (outfile, "}");
978 if (ref->vuse)
980 fprintf (outfile, "@");
981 print_generic_expr (outfile, ref->vuse, 0);
984 break;
987 void debug_pre_expr (pre_expr);
989 /* Like print_pre_expr but always prints to stderr. */
990 DEBUG_FUNCTION void
991 debug_pre_expr (pre_expr e)
993 print_pre_expr (stderr, e);
994 fprintf (stderr, "\n");
997 /* Print out SET to OUTFILE. */
999 static void
1000 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1001 const char *setname, int blockindex)
1003 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1004 if (set)
1006 bool first = true;
1007 unsigned i;
1008 bitmap_iterator bi;
1010 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1012 const pre_expr expr = expression_for_id (i);
1014 if (!first)
1015 fprintf (outfile, ", ");
1016 first = false;
1017 print_pre_expr (outfile, expr);
1019 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1022 fprintf (outfile, " }\n");
1025 void debug_bitmap_set (bitmap_set_t);
1027 DEBUG_FUNCTION void
1028 debug_bitmap_set (bitmap_set_t set)
1030 print_bitmap_set (stderr, set, "debug", 0);
1033 /* Print out the expressions that have VAL to OUTFILE. */
1035 void
1036 print_value_expressions (FILE *outfile, unsigned int val)
1038 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1039 if (set)
1041 char s[10];
1042 sprintf (s, "%04d", val);
1043 print_bitmap_set (outfile, set, s, 0);
1048 DEBUG_FUNCTION void
1049 debug_value_expressions (unsigned int val)
1051 print_value_expressions (stderr, val);
1054 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1055 represent it. */
1057 static pre_expr
1058 get_or_alloc_expr_for_constant (tree constant)
1060 unsigned int result_id;
1061 unsigned int value_id;
1062 struct pre_expr_d expr;
1063 pre_expr newexpr;
1065 expr.kind = CONSTANT;
1066 PRE_EXPR_CONSTANT (&expr) = constant;
1067 result_id = lookup_expression_id (&expr);
1068 if (result_id != 0)
1069 return expression_for_id (result_id);
1071 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1072 newexpr->kind = CONSTANT;
1073 PRE_EXPR_CONSTANT (newexpr) = constant;
1074 alloc_expression_id (newexpr);
1075 value_id = get_or_alloc_constant_value_id (constant);
1076 add_to_value (value_id, newexpr);
1077 return newexpr;
1080 /* Given a value id V, find the actual tree representing the constant
1081 value if there is one, and return it. Return NULL if we can't find
1082 a constant. */
1084 static tree
1085 get_constant_for_value_id (unsigned int v)
1087 if (value_id_constant_p (v))
1089 unsigned int i;
1090 bitmap_iterator bi;
1091 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1093 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1095 pre_expr expr = expression_for_id (i);
1096 if (expr->kind == CONSTANT)
1097 return PRE_EXPR_CONSTANT (expr);
1100 return NULL;
1103 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1104 Currently only supports constants and SSA_NAMES. */
1105 static pre_expr
1106 get_or_alloc_expr_for (tree t)
1108 if (TREE_CODE (t) == SSA_NAME)
1109 return get_or_alloc_expr_for_name (t);
1110 else if (is_gimple_min_invariant (t))
1111 return get_or_alloc_expr_for_constant (t);
1112 else
1114 /* More complex expressions can result from SCCVN expression
1115 simplification that inserts values for them. As they all
1116 do not have VOPs the get handled by the nary ops struct. */
1117 vn_nary_op_t result;
1118 unsigned int result_id;
1119 vn_nary_op_lookup (t, &result);
1120 if (result != NULL)
1122 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1123 e->kind = NARY;
1124 PRE_EXPR_NARY (e) = result;
1125 result_id = lookup_expression_id (e);
1126 if (result_id != 0)
1128 pool_free (pre_expr_pool, e);
1129 e = expression_for_id (result_id);
1130 return e;
1132 alloc_expression_id (e);
1133 return e;
1136 return NULL;
1139 /* Return the folded version of T if T, when folded, is a gimple
1140 min_invariant. Otherwise, return T. */
1142 static pre_expr
1143 fully_constant_expression (pre_expr e)
1145 switch (e->kind)
1147 case CONSTANT:
1148 return e;
1149 case NARY:
1151 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1152 switch (TREE_CODE_CLASS (nary->opcode))
1154 case tcc_expression:
1155 if (nary->opcode == TRUTH_NOT_EXPR)
1156 goto do_unary;
1157 if (nary->opcode != TRUTH_AND_EXPR
1158 && nary->opcode != TRUTH_OR_EXPR
1159 && nary->opcode != TRUTH_XOR_EXPR)
1160 return e;
1161 /* Fallthrough. */
1162 case tcc_binary:
1163 case tcc_comparison:
1165 /* We have to go from trees to pre exprs to value ids to
1166 constants. */
1167 tree naryop0 = nary->op[0];
1168 tree naryop1 = nary->op[1];
1169 tree result;
1170 if (!is_gimple_min_invariant (naryop0))
1172 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1173 unsigned int vrep0 = get_expr_value_id (rep0);
1174 tree const0 = get_constant_for_value_id (vrep0);
1175 if (const0)
1176 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1178 if (!is_gimple_min_invariant (naryop1))
1180 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1181 unsigned int vrep1 = get_expr_value_id (rep1);
1182 tree const1 = get_constant_for_value_id (vrep1);
1183 if (const1)
1184 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1186 result = fold_binary (nary->opcode, nary->type,
1187 naryop0, naryop1);
1188 if (result && is_gimple_min_invariant (result))
1189 return get_or_alloc_expr_for_constant (result);
1190 /* We might have simplified the expression to a
1191 SSA_NAME for example from x_1 * 1. But we cannot
1192 insert a PHI for x_1 unconditionally as x_1 might
1193 not be available readily. */
1194 return e;
1196 case tcc_reference:
1197 if (nary->opcode != REALPART_EXPR
1198 && nary->opcode != IMAGPART_EXPR
1199 && nary->opcode != VIEW_CONVERT_EXPR)
1200 return e;
1201 /* Fallthrough. */
1202 case tcc_unary:
1203 do_unary:
1205 /* We have to go from trees to pre exprs to value ids to
1206 constants. */
1207 tree naryop0 = nary->op[0];
1208 tree const0, result;
1209 if (is_gimple_min_invariant (naryop0))
1210 const0 = naryop0;
1211 else
1213 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1214 unsigned int vrep0 = get_expr_value_id (rep0);
1215 const0 = get_constant_for_value_id (vrep0);
1217 result = NULL;
1218 if (const0)
1220 tree type1 = TREE_TYPE (nary->op[0]);
1221 const0 = fold_convert (type1, const0);
1222 result = fold_unary (nary->opcode, nary->type, const0);
1224 if (result && is_gimple_min_invariant (result))
1225 return get_or_alloc_expr_for_constant (result);
1226 return e;
1228 default:
1229 return e;
1232 case REFERENCE:
1234 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1235 tree folded;
1236 if ((folded = fully_constant_vn_reference_p (ref)))
1237 return get_or_alloc_expr_for_constant (folded);
1238 return e;
1240 default:
1241 return e;
1243 return e;
1246 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1247 it has the value it would have in BLOCK. Set *SAME_VALID to true
1248 in case the new vuse doesn't change the value id of the OPERANDS. */
1250 static tree
1251 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1252 alias_set_type set, tree type, tree vuse,
1253 basic_block phiblock,
1254 basic_block block, bool *same_valid)
1256 gimple phi = SSA_NAME_DEF_STMT (vuse);
1257 ao_ref ref;
1258 edge e = NULL;
1259 bool use_oracle;
1261 *same_valid = true;
1263 if (gimple_bb (phi) != phiblock)
1264 return vuse;
1266 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1268 /* Use the alias-oracle to find either the PHI node in this block,
1269 the first VUSE used in this block that is equivalent to vuse or
1270 the first VUSE which definition in this block kills the value. */
1271 if (gimple_code (phi) == GIMPLE_PHI)
1272 e = find_edge (block, phiblock);
1273 else if (use_oracle)
1274 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1276 vuse = gimple_vuse (phi);
1277 phi = SSA_NAME_DEF_STMT (vuse);
1278 if (gimple_bb (phi) != phiblock)
1279 return vuse;
1280 if (gimple_code (phi) == GIMPLE_PHI)
1282 e = find_edge (block, phiblock);
1283 break;
1286 else
1287 return NULL_TREE;
1289 if (e)
1291 if (use_oracle)
1293 bitmap visited = NULL;
1294 /* Try to find a vuse that dominates this phi node by skipping
1295 non-clobbering statements. */
1296 vuse = get_continuation_for_phi (phi, &ref, &visited);
1297 if (visited)
1298 BITMAP_FREE (visited);
1300 else
1301 vuse = NULL_TREE;
1302 if (!vuse)
1304 /* If we didn't find any, the value ID can't stay the same,
1305 but return the translated vuse. */
1306 *same_valid = false;
1307 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1309 /* ??? We would like to return vuse here as this is the canonical
1310 upmost vdef that this reference is associated with. But during
1311 insertion of the references into the hash tables we only ever
1312 directly insert with their direct gimple_vuse, hence returning
1313 something else would make us not find the other expression. */
1314 return PHI_ARG_DEF (phi, e->dest_idx);
1317 return NULL_TREE;
1320 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1321 SET2. This is used to avoid making a set consisting of the union
1322 of PA_IN and ANTIC_IN during insert. */
1324 static inline pre_expr
1325 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1327 pre_expr result;
1329 result = bitmap_find_leader (set1, val, NULL);
1330 if (!result && set2)
1331 result = bitmap_find_leader (set2, val, NULL);
1332 return result;
1335 /* Get the tree type for our PRE expression e. */
1337 static tree
1338 get_expr_type (const pre_expr e)
1340 switch (e->kind)
1342 case NAME:
1343 return TREE_TYPE (PRE_EXPR_NAME (e));
1344 case CONSTANT:
1345 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1346 case REFERENCE:
1347 return PRE_EXPR_REFERENCE (e)->type;
1348 case NARY:
1349 return PRE_EXPR_NARY (e)->type;
1351 gcc_unreachable();
1354 /* Get a representative SSA_NAME for a given expression.
1355 Since all of our sub-expressions are treated as values, we require
1356 them to be SSA_NAME's for simplicity.
1357 Prior versions of GVNPRE used to use "value handles" here, so that
1358 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1359 either case, the operands are really values (IE we do not expect
1360 them to be usable without finding leaders). */
1362 static tree
1363 get_representative_for (const pre_expr e)
1365 tree exprtype;
1366 tree name;
1367 unsigned int value_id = get_expr_value_id (e);
1369 switch (e->kind)
1371 case NAME:
1372 return PRE_EXPR_NAME (e);
1373 case CONSTANT:
1374 return PRE_EXPR_CONSTANT (e);
1375 case NARY:
1376 case REFERENCE:
1378 /* Go through all of the expressions representing this value
1379 and pick out an SSA_NAME. */
1380 unsigned int i;
1381 bitmap_iterator bi;
1382 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1383 value_id);
1384 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1386 pre_expr rep = expression_for_id (i);
1387 if (rep->kind == NAME)
1388 return PRE_EXPR_NAME (rep);
1391 break;
1393 /* If we reached here we couldn't find an SSA_NAME. This can
1394 happen when we've discovered a value that has never appeared in
1395 the program as set to an SSA_NAME, most likely as the result of
1396 phi translation. */
1397 if (dump_file)
1399 fprintf (dump_file,
1400 "Could not find SSA_NAME representative for expression:");
1401 print_pre_expr (dump_file, e);
1402 fprintf (dump_file, "\n");
1405 exprtype = get_expr_type (e);
1407 /* Build and insert the assignment of the end result to the temporary
1408 that we will return. */
1409 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1411 pretemp = create_tmp_reg (exprtype, "pretmp");
1412 get_var_ann (pretemp);
1415 name = make_ssa_name (pretemp, gimple_build_nop ());
1416 VN_INFO_GET (name)->value_id = value_id;
1417 if (e->kind == CONSTANT)
1418 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1419 else
1420 VN_INFO (name)->valnum = name;
1422 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1423 if (dump_file)
1425 fprintf (dump_file, "Created SSA_NAME representative ");
1426 print_generic_expr (dump_file, name, 0);
1427 fprintf (dump_file, " for expression:");
1428 print_pre_expr (dump_file, e);
1429 fprintf (dump_file, "\n");
1432 return name;
1437 static pre_expr
1438 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1439 basic_block pred, basic_block phiblock);
1441 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1442 the phis in PRED. Return NULL if we can't find a leader for each part
1443 of the translated expression. */
1445 static pre_expr
1446 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1447 basic_block pred, basic_block phiblock)
1449 switch (expr->kind)
1451 case NARY:
1453 unsigned int i;
1454 bool changed = false;
1455 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1456 struct vn_nary_op_s newnary;
1457 /* The NARY structure is only guaranteed to have been
1458 allocated to the nary->length operands. */
1459 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1460 - sizeof (tree) * (4 - nary->length)));
1462 for (i = 0; i < newnary.length; i++)
1464 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1465 continue;
1466 else
1468 pre_expr leader, result;
1469 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1470 leader = find_leader_in_sets (op_val_id, set1, set2);
1471 result = phi_translate (leader, set1, set2, pred, phiblock);
1472 if (result && result != leader)
1474 tree name = get_representative_for (result);
1475 if (!name)
1476 return NULL;
1477 newnary.op[i] = name;
1479 else if (!result)
1480 return NULL;
1482 changed |= newnary.op[i] != nary->op[i];
1485 if (changed)
1487 pre_expr constant;
1488 unsigned int new_val_id;
1490 tree result = vn_nary_op_lookup_pieces (newnary.length,
1491 newnary.opcode,
1492 newnary.type,
1493 newnary.op[0],
1494 newnary.op[1],
1495 newnary.op[2],
1496 newnary.op[3],
1497 &nary);
1498 if (result && is_gimple_min_invariant (result))
1499 return get_or_alloc_expr_for_constant (result);
1501 expr = (pre_expr) pool_alloc (pre_expr_pool);
1502 expr->kind = NARY;
1503 expr->id = 0;
1504 if (nary)
1506 PRE_EXPR_NARY (expr) = nary;
1507 constant = fully_constant_expression (expr);
1508 if (constant != expr)
1509 return constant;
1511 new_val_id = nary->value_id;
1512 get_or_alloc_expression_id (expr);
1514 else
1516 new_val_id = get_next_value_id ();
1517 VEC_safe_grow_cleared (bitmap_set_t, heap,
1518 value_expressions,
1519 get_max_value_id() + 1);
1520 nary = vn_nary_op_insert_pieces (newnary.length,
1521 newnary.opcode,
1522 newnary.type,
1523 newnary.op[0],
1524 newnary.op[1],
1525 newnary.op[2],
1526 newnary.op[3],
1527 result, new_val_id);
1528 PRE_EXPR_NARY (expr) = nary;
1529 constant = fully_constant_expression (expr);
1530 if (constant != expr)
1531 return constant;
1532 get_or_alloc_expression_id (expr);
1534 add_to_value (new_val_id, expr);
1536 return expr;
1538 break;
1540 case REFERENCE:
1542 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1543 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1544 tree vuse = ref->vuse;
1545 tree newvuse = vuse;
1546 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1547 bool changed = false, same_valid = true;
1548 unsigned int i, j;
1549 vn_reference_op_t operand;
1550 vn_reference_t newref;
1552 for (i = 0, j = 0;
1553 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1555 pre_expr opresult;
1556 pre_expr leader;
1557 tree oldop0 = operand->op0;
1558 tree oldop1 = operand->op1;
1559 tree oldop2 = operand->op2;
1560 tree op0 = oldop0;
1561 tree op1 = oldop1;
1562 tree op2 = oldop2;
1563 tree type = operand->type;
1564 vn_reference_op_s newop = *operand;
1566 if (op0 && TREE_CODE (op0) == SSA_NAME)
1568 unsigned int op_val_id = VN_INFO (op0)->value_id;
1569 leader = find_leader_in_sets (op_val_id, set1, set2);
1570 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1571 if (opresult && opresult != leader)
1573 tree name = get_representative_for (opresult);
1574 if (!name)
1575 break;
1576 op0 = name;
1578 else if (!opresult)
1579 break;
1581 changed |= op0 != oldop0;
1583 if (op1 && TREE_CODE (op1) == SSA_NAME)
1585 unsigned int op_val_id = VN_INFO (op1)->value_id;
1586 leader = find_leader_in_sets (op_val_id, set1, set2);
1587 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1588 if (opresult && opresult != leader)
1590 tree name = get_representative_for (opresult);
1591 if (!name)
1592 break;
1593 op1 = name;
1595 else if (!opresult)
1596 break;
1598 /* We can't possibly insert these. */
1599 else if (op1 && !is_gimple_min_invariant (op1))
1600 break;
1601 changed |= op1 != oldop1;
1602 if (op2 && TREE_CODE (op2) == SSA_NAME)
1604 unsigned int op_val_id = VN_INFO (op2)->value_id;
1605 leader = find_leader_in_sets (op_val_id, set1, set2);
1606 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1607 if (opresult && opresult != leader)
1609 tree name = get_representative_for (opresult);
1610 if (!name)
1611 break;
1612 op2 = name;
1614 else if (!opresult)
1615 break;
1617 /* We can't possibly insert these. */
1618 else if (op2 && !is_gimple_min_invariant (op2))
1619 break;
1620 changed |= op2 != oldop2;
1622 if (!newoperands)
1623 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1624 /* We may have changed from an SSA_NAME to a constant */
1625 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1626 newop.opcode = TREE_CODE (op0);
1627 newop.type = type;
1628 newop.op0 = op0;
1629 newop.op1 = op1;
1630 newop.op2 = op2;
1631 /* If it transforms a non-constant ARRAY_REF into a constant
1632 one, adjust the constant offset. */
1633 if (newop.opcode == ARRAY_REF
1634 && newop.off == -1
1635 && TREE_CODE (op0) == INTEGER_CST
1636 && TREE_CODE (op1) == INTEGER_CST
1637 && TREE_CODE (op2) == INTEGER_CST)
1639 double_int off = tree_to_double_int (op0);
1640 off = double_int_add (off,
1641 double_int_neg
1642 (tree_to_double_int (op1)));
1643 off = double_int_mul (off, tree_to_double_int (op2));
1644 if (double_int_fits_in_shwi_p (off))
1645 newop.off = off.low;
1647 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1648 /* If it transforms from an SSA_NAME to an address, fold with
1649 a preceding indirect reference. */
1650 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1651 && VEC_index (vn_reference_op_s,
1652 newoperands, j - 1)->opcode == MEM_REF)
1653 vn_reference_fold_indirect (&newoperands, &j);
1655 if (i != VEC_length (vn_reference_op_s, operands))
1657 if (newoperands)
1658 VEC_free (vn_reference_op_s, heap, newoperands);
1659 return NULL;
1662 if (vuse)
1664 newvuse = translate_vuse_through_block (newoperands,
1665 ref->set, ref->type,
1666 vuse, phiblock, pred,
1667 &same_valid);
1668 if (newvuse == NULL_TREE)
1670 VEC_free (vn_reference_op_s, heap, newoperands);
1671 return NULL;
1675 if (changed || newvuse != vuse)
1677 unsigned int new_val_id;
1678 pre_expr constant;
1679 bool converted = false;
1681 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1682 ref->type,
1683 newoperands,
1684 &newref, true);
1685 if (result)
1686 VEC_free (vn_reference_op_s, heap, newoperands);
1688 if (result
1689 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1691 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1692 converted = true;
1695 if (result && is_gimple_min_invariant (result))
1697 gcc_assert (!newoperands);
1698 return get_or_alloc_expr_for_constant (result);
1701 expr = (pre_expr) pool_alloc (pre_expr_pool);
1702 expr->kind = REFERENCE;
1703 expr->id = 0;
1705 if (converted)
1707 vn_nary_op_t nary;
1708 tree nresult;
1710 gcc_assert (CONVERT_EXPR_P (result)
1711 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1713 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1714 TREE_TYPE (result),
1715 TREE_OPERAND (result, 0),
1716 NULL_TREE, NULL_TREE,
1717 NULL_TREE,
1718 &nary);
1719 if (nresult && is_gimple_min_invariant (nresult))
1720 return get_or_alloc_expr_for_constant (nresult);
1722 expr->kind = NARY;
1723 if (nary)
1725 PRE_EXPR_NARY (expr) = nary;
1726 constant = fully_constant_expression (expr);
1727 if (constant != expr)
1728 return constant;
1730 new_val_id = nary->value_id;
1731 get_or_alloc_expression_id (expr);
1733 else
1735 new_val_id = get_next_value_id ();
1736 VEC_safe_grow_cleared (bitmap_set_t, heap,
1737 value_expressions,
1738 get_max_value_id() + 1);
1739 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1740 TREE_TYPE (result),
1741 TREE_OPERAND (result, 0),
1742 NULL_TREE, NULL_TREE,
1743 NULL_TREE, NULL_TREE,
1744 new_val_id);
1745 PRE_EXPR_NARY (expr) = nary;
1746 constant = fully_constant_expression (expr);
1747 if (constant != expr)
1748 return constant;
1749 get_or_alloc_expression_id (expr);
1752 else if (newref)
1754 PRE_EXPR_REFERENCE (expr) = newref;
1755 constant = fully_constant_expression (expr);
1756 if (constant != expr)
1757 return constant;
1759 new_val_id = newref->value_id;
1760 get_or_alloc_expression_id (expr);
1762 else
1764 if (changed || !same_valid)
1766 new_val_id = get_next_value_id ();
1767 VEC_safe_grow_cleared (bitmap_set_t, heap,
1768 value_expressions,
1769 get_max_value_id() + 1);
1771 else
1772 new_val_id = ref->value_id;
1773 newref = vn_reference_insert_pieces (newvuse, ref->set,
1774 ref->type,
1775 newoperands,
1776 result, new_val_id);
1777 newoperands = NULL;
1778 PRE_EXPR_REFERENCE (expr) = newref;
1779 constant = fully_constant_expression (expr);
1780 if (constant != expr)
1781 return constant;
1782 get_or_alloc_expression_id (expr);
1784 add_to_value (new_val_id, expr);
1786 VEC_free (vn_reference_op_s, heap, newoperands);
1787 return expr;
1789 break;
1791 case NAME:
1793 gimple phi = NULL;
1794 edge e;
1795 gimple def_stmt;
1796 tree name = PRE_EXPR_NAME (expr);
1798 def_stmt = SSA_NAME_DEF_STMT (name);
1799 if (gimple_code (def_stmt) == GIMPLE_PHI
1800 && gimple_bb (def_stmt) == phiblock)
1801 phi = def_stmt;
1802 else
1803 return expr;
1805 e = find_edge (pred, gimple_bb (phi));
1806 if (e)
1808 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1809 pre_expr newexpr;
1811 if (TREE_CODE (def) == SSA_NAME)
1812 def = VN_INFO (def)->valnum;
1814 /* Handle constant. */
1815 if (is_gimple_min_invariant (def))
1816 return get_or_alloc_expr_for_constant (def);
1818 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1819 return NULL;
1821 newexpr = get_or_alloc_expr_for_name (def);
1822 return newexpr;
1825 return expr;
1827 default:
1828 gcc_unreachable ();
1832 /* Wrapper around phi_translate_1 providing caching functionality. */
1834 static pre_expr
1835 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1836 basic_block pred, basic_block phiblock)
1838 pre_expr phitrans;
1840 if (!expr)
1841 return NULL;
1843 /* Constants contain no values that need translation. */
1844 if (expr->kind == CONSTANT)
1845 return expr;
1847 if (value_id_constant_p (get_expr_value_id (expr)))
1848 return expr;
1850 if (expr->kind != NAME)
1852 phitrans = phi_trans_lookup (expr, pred);
1853 if (phitrans)
1854 return phitrans;
1857 /* Translate. */
1858 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1860 /* Don't add empty translations to the cache. Neither add
1861 translations of NAMEs as those are cheap to translate. */
1862 if (phitrans
1863 && expr->kind != NAME)
1864 phi_trans_add (expr, phitrans, pred);
1866 return phitrans;
1870 /* For each expression in SET, translate the values through phi nodes
1871 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1872 expressions in DEST. */
1874 static void
1875 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1876 basic_block phiblock)
1878 VEC (pre_expr, heap) *exprs;
1879 pre_expr expr;
1880 int i;
1882 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1884 bitmap_set_copy (dest, set);
1885 return;
1888 exprs = sorted_array_from_bitmap_set (set);
1889 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1891 pre_expr translated;
1892 translated = phi_translate (expr, set, NULL, pred, phiblock);
1893 if (!translated)
1894 continue;
1896 /* We might end up with multiple expressions from SET being
1897 translated to the same value. In this case we do not want
1898 to retain the NARY or REFERENCE expression but prefer a NAME
1899 which would be the leader. */
1900 if (translated->kind == NAME)
1901 bitmap_value_replace_in_set (dest, translated);
1902 else
1903 bitmap_value_insert_into_set (dest, translated);
1905 VEC_free (pre_expr, heap, exprs);
1908 /* Find the leader for a value (i.e., the name representing that
1909 value) in a given set, and return it. If STMT is non-NULL it
1910 makes sure the defining statement for the leader dominates it.
1911 Return NULL if no leader is found. */
1913 static pre_expr
1914 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1916 if (value_id_constant_p (val))
1918 unsigned int i;
1919 bitmap_iterator bi;
1920 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1922 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1924 pre_expr expr = expression_for_id (i);
1925 if (expr->kind == CONSTANT)
1926 return expr;
1929 if (bitmap_set_contains_value (set, val))
1931 /* Rather than walk the entire bitmap of expressions, and see
1932 whether any of them has the value we are looking for, we look
1933 at the reverse mapping, which tells us the set of expressions
1934 that have a given value (IE value->expressions with that
1935 value) and see if any of those expressions are in our set.
1936 The number of expressions per value is usually significantly
1937 less than the number of expressions in the set. In fact, for
1938 large testcases, doing it this way is roughly 5-10x faster
1939 than walking the bitmap.
1940 If this is somehow a significant lose for some cases, we can
1941 choose which set to walk based on which set is smaller. */
1942 unsigned int i;
1943 bitmap_iterator bi;
1944 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1946 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1947 &set->expressions, 0, i, bi)
1949 pre_expr val = expression_for_id (i);
1950 /* At the point where stmt is not null, there should always
1951 be an SSA_NAME first in the list of expressions. */
1952 if (stmt)
1954 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1955 if (gimple_code (def_stmt) != GIMPLE_PHI
1956 && gimple_bb (def_stmt) == gimple_bb (stmt)
1957 /* PRE insertions are at the end of the basic-block
1958 and have UID 0. */
1959 && (gimple_uid (def_stmt) == 0
1960 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1961 continue;
1963 return val;
1966 return NULL;
1969 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1970 BLOCK by seeing if it is not killed in the block. Note that we are
1971 only determining whether there is a store that kills it. Because
1972 of the order in which clean iterates over values, we are guaranteed
1973 that altered operands will have caused us to be eliminated from the
1974 ANTIC_IN set already. */
1976 static bool
1977 value_dies_in_block_x (pre_expr expr, basic_block block)
1979 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1980 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1981 gimple def;
1982 gimple_stmt_iterator gsi;
1983 unsigned id = get_expression_id (expr);
1984 bool res = false;
1985 ao_ref ref;
1987 if (!vuse)
1988 return false;
1990 /* Lookup a previously calculated result. */
1991 if (EXPR_DIES (block)
1992 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1993 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1995 /* A memory expression {e, VUSE} dies in the block if there is a
1996 statement that may clobber e. If, starting statement walk from the
1997 top of the basic block, a statement uses VUSE there can be no kill
1998 inbetween that use and the original statement that loaded {e, VUSE},
1999 so we can stop walking. */
2000 ref.base = NULL_TREE;
2001 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
2003 tree def_vuse, def_vdef;
2004 def = gsi_stmt (gsi);
2005 def_vuse = gimple_vuse (def);
2006 def_vdef = gimple_vdef (def);
2008 /* Not a memory statement. */
2009 if (!def_vuse)
2010 continue;
2012 /* Not a may-def. */
2013 if (!def_vdef)
2015 /* A load with the same VUSE, we're done. */
2016 if (def_vuse == vuse)
2017 break;
2019 continue;
2022 /* Init ref only if we really need it. */
2023 if (ref.base == NULL_TREE
2024 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
2025 refx->operands))
2027 res = true;
2028 break;
2030 /* If the statement may clobber expr, it dies. */
2031 if (stmt_may_clobber_ref_p_1 (def, &ref))
2033 res = true;
2034 break;
2038 /* Remember the result. */
2039 if (!EXPR_DIES (block))
2040 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2041 bitmap_set_bit (EXPR_DIES (block), id * 2);
2042 if (res)
2043 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2045 return res;
2049 #define union_contains_value(SET1, SET2, VAL) \
2050 (bitmap_set_contains_value ((SET1), (VAL)) \
2051 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
2053 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2055 static bool
2056 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2057 vn_reference_op_t vro)
2059 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2061 struct pre_expr_d temp;
2062 temp.kind = NAME;
2063 temp.id = 0;
2064 PRE_EXPR_NAME (&temp) = vro->op0;
2065 temp.id = lookup_expression_id (&temp);
2066 if (temp.id == 0)
2067 return false;
2068 if (!union_contains_value (set1, set2,
2069 get_expr_value_id (&temp)))
2070 return false;
2072 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2074 struct pre_expr_d temp;
2075 temp.kind = NAME;
2076 temp.id = 0;
2077 PRE_EXPR_NAME (&temp) = vro->op1;
2078 temp.id = lookup_expression_id (&temp);
2079 if (temp.id == 0)
2080 return false;
2081 if (!union_contains_value (set1, set2,
2082 get_expr_value_id (&temp)))
2083 return false;
2086 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2088 struct pre_expr_d temp;
2089 temp.kind = NAME;
2090 temp.id = 0;
2091 PRE_EXPR_NAME (&temp) = vro->op2;
2092 temp.id = lookup_expression_id (&temp);
2093 if (temp.id == 0)
2094 return false;
2095 if (!union_contains_value (set1, set2,
2096 get_expr_value_id (&temp)))
2097 return false;
2100 return true;
2103 /* Determine if the expression EXPR is valid in SET1 U SET2.
2104 ONLY SET2 CAN BE NULL.
2105 This means that we have a leader for each part of the expression
2106 (if it consists of values), or the expression is an SSA_NAME.
2107 For loads/calls, we also see if the vuse is killed in this block. */
2109 static bool
2110 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2111 basic_block block)
2113 switch (expr->kind)
2115 case NAME:
2116 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2117 case NARY:
2119 unsigned int i;
2120 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2121 for (i = 0; i < nary->length; i++)
2123 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2125 struct pre_expr_d temp;
2126 temp.kind = NAME;
2127 temp.id = 0;
2128 PRE_EXPR_NAME (&temp) = nary->op[i];
2129 temp.id = lookup_expression_id (&temp);
2130 if (temp.id == 0)
2131 return false;
2132 if (!union_contains_value (set1, set2,
2133 get_expr_value_id (&temp)))
2134 return false;
2137 /* If the NARY may trap make sure the block does not contain
2138 a possible exit point.
2139 ??? This is overly conservative if we translate AVAIL_OUT
2140 as the available expression might be after the exit point. */
2141 if (BB_MAY_NOTRETURN (block)
2142 && vn_nary_may_trap (nary))
2143 return false;
2144 return true;
2146 break;
2147 case REFERENCE:
2149 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2150 vn_reference_op_t vro;
2151 unsigned int i;
2153 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2155 if (!vro_valid_in_sets (set1, set2, vro))
2156 return false;
2158 if (ref->vuse)
2160 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2161 if (!gimple_nop_p (def_stmt)
2162 && gimple_bb (def_stmt) != block
2163 && !dominated_by_p (CDI_DOMINATORS,
2164 block, gimple_bb (def_stmt)))
2165 return false;
2167 return !value_dies_in_block_x (expr, block);
2169 default:
2170 gcc_unreachable ();
2174 /* Clean the set of expressions that are no longer valid in SET1 or
2175 SET2. This means expressions that are made up of values we have no
2176 leaders for in SET1 or SET2. This version is used for partial
2177 anticipation, which means it is not valid in either ANTIC_IN or
2178 PA_IN. */
2180 static void
2181 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2183 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2184 pre_expr expr;
2185 int i;
2187 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2189 if (!valid_in_sets (set1, set2, expr, block))
2190 bitmap_remove_from_set (set1, expr);
2192 VEC_free (pre_expr, heap, exprs);
2195 /* Clean the set of expressions that are no longer valid in SET. This
2196 means expressions that are made up of values we have no leaders for
2197 in SET. */
2199 static void
2200 clean (bitmap_set_t set, basic_block block)
2202 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2203 pre_expr expr;
2204 int i;
2206 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2208 if (!valid_in_sets (set, NULL, expr, block))
2209 bitmap_remove_from_set (set, expr);
2211 VEC_free (pre_expr, heap, exprs);
2214 static sbitmap has_abnormal_preds;
2216 /* List of blocks that may have changed during ANTIC computation and
2217 thus need to be iterated over. */
2219 static sbitmap changed_blocks;
2221 /* Decide whether to defer a block for a later iteration, or PHI
2222 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2223 should defer the block, and true if we processed it. */
2225 static bool
2226 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2227 basic_block block, basic_block phiblock)
2229 if (!BB_VISITED (phiblock))
2231 SET_BIT (changed_blocks, block->index);
2232 BB_VISITED (block) = 0;
2233 BB_DEFERRED (block) = 1;
2234 return false;
2236 else
2237 phi_translate_set (dest, source, block, phiblock);
2238 return true;
2241 /* Compute the ANTIC set for BLOCK.
2243 If succs(BLOCK) > 1 then
2244 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2245 else if succs(BLOCK) == 1 then
2246 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2248 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2251 static bool
2252 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2254 bool changed = false;
2255 bitmap_set_t S, old, ANTIC_OUT;
2256 bitmap_iterator bi;
2257 unsigned int bii;
2258 edge e;
2259 edge_iterator ei;
2261 old = ANTIC_OUT = S = NULL;
2262 BB_VISITED (block) = 1;
2264 /* If any edges from predecessors are abnormal, antic_in is empty,
2265 so do nothing. */
2266 if (block_has_abnormal_pred_edge)
2267 goto maybe_dump_sets;
2269 old = ANTIC_IN (block);
2270 ANTIC_OUT = bitmap_set_new ();
2272 /* If the block has no successors, ANTIC_OUT is empty. */
2273 if (EDGE_COUNT (block->succs) == 0)
2275 /* If we have one successor, we could have some phi nodes to
2276 translate through. */
2277 else if (single_succ_p (block))
2279 basic_block succ_bb = single_succ (block);
2281 /* We trade iterations of the dataflow equations for having to
2282 phi translate the maximal set, which is incredibly slow
2283 (since the maximal set often has 300+ members, even when you
2284 have a small number of blocks).
2285 Basically, we defer the computation of ANTIC for this block
2286 until we have processed it's successor, which will inevitably
2287 have a *much* smaller set of values to phi translate once
2288 clean has been run on it.
2289 The cost of doing this is that we technically perform more
2290 iterations, however, they are lower cost iterations.
2292 Timings for PRE on tramp3d-v4:
2293 without maximal set fix: 11 seconds
2294 with maximal set fix/without deferring: 26 seconds
2295 with maximal set fix/with deferring: 11 seconds
2298 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2299 block, succ_bb))
2301 changed = true;
2302 goto maybe_dump_sets;
2305 /* If we have multiple successors, we take the intersection of all of
2306 them. Note that in the case of loop exit phi nodes, we may have
2307 phis to translate through. */
2308 else
2310 VEC(basic_block, heap) * worklist;
2311 size_t i;
2312 basic_block bprime, first = NULL;
2314 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2315 FOR_EACH_EDGE (e, ei, block->succs)
2317 if (!first
2318 && BB_VISITED (e->dest))
2319 first = e->dest;
2320 else if (BB_VISITED (e->dest))
2321 VEC_quick_push (basic_block, worklist, e->dest);
2324 /* Of multiple successors we have to have visited one already. */
2325 if (!first)
2327 SET_BIT (changed_blocks, block->index);
2328 BB_VISITED (block) = 0;
2329 BB_DEFERRED (block) = 1;
2330 changed = true;
2331 VEC_free (basic_block, heap, worklist);
2332 goto maybe_dump_sets;
2335 if (!gimple_seq_empty_p (phi_nodes (first)))
2336 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2337 else
2338 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2340 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2342 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2344 bitmap_set_t tmp = bitmap_set_new ();
2345 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2346 bitmap_set_and (ANTIC_OUT, tmp);
2347 bitmap_set_free (tmp);
2349 else
2350 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2352 VEC_free (basic_block, heap, worklist);
2355 /* Generate ANTIC_OUT - TMP_GEN. */
2356 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2358 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2359 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2360 TMP_GEN (block));
2362 /* Then union in the ANTIC_OUT - TMP_GEN values,
2363 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2364 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2365 bitmap_value_insert_into_set (ANTIC_IN (block),
2366 expression_for_id (bii));
2368 clean (ANTIC_IN (block), block);
2370 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2372 changed = true;
2373 SET_BIT (changed_blocks, block->index);
2374 FOR_EACH_EDGE (e, ei, block->preds)
2375 SET_BIT (changed_blocks, e->src->index);
2377 else
2378 RESET_BIT (changed_blocks, block->index);
2380 maybe_dump_sets:
2381 if (dump_file && (dump_flags & TDF_DETAILS))
2383 if (!BB_DEFERRED (block) || BB_VISITED (block))
2385 if (ANTIC_OUT)
2386 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2388 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2389 block->index);
2391 if (S)
2392 print_bitmap_set (dump_file, S, "S", block->index);
2394 else
2396 fprintf (dump_file,
2397 "Block %d was deferred for a future iteration.\n",
2398 block->index);
2401 if (old)
2402 bitmap_set_free (old);
2403 if (S)
2404 bitmap_set_free (S);
2405 if (ANTIC_OUT)
2406 bitmap_set_free (ANTIC_OUT);
2407 return changed;
2410 /* Compute PARTIAL_ANTIC for BLOCK.
2412 If succs(BLOCK) > 1 then
2413 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2414 in ANTIC_OUT for all succ(BLOCK)
2415 else if succs(BLOCK) == 1 then
2416 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2418 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2419 - ANTIC_IN[BLOCK])
2422 static bool
2423 compute_partial_antic_aux (basic_block block,
2424 bool block_has_abnormal_pred_edge)
2426 bool changed = false;
2427 bitmap_set_t old_PA_IN;
2428 bitmap_set_t PA_OUT;
2429 edge e;
2430 edge_iterator ei;
2431 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2433 old_PA_IN = PA_OUT = NULL;
2435 /* If any edges from predecessors are abnormal, antic_in is empty,
2436 so do nothing. */
2437 if (block_has_abnormal_pred_edge)
2438 goto maybe_dump_sets;
2440 /* If there are too many partially anticipatable values in the
2441 block, phi_translate_set can take an exponential time: stop
2442 before the translation starts. */
2443 if (max_pa
2444 && single_succ_p (block)
2445 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2446 goto maybe_dump_sets;
2448 old_PA_IN = PA_IN (block);
2449 PA_OUT = bitmap_set_new ();
2451 /* If the block has no successors, ANTIC_OUT is empty. */
2452 if (EDGE_COUNT (block->succs) == 0)
2454 /* If we have one successor, we could have some phi nodes to
2455 translate through. Note that we can't phi translate across DFS
2456 back edges in partial antic, because it uses a union operation on
2457 the successors. For recurrences like IV's, we will end up
2458 generating a new value in the set on each go around (i + 3 (VH.1)
2459 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2460 else if (single_succ_p (block))
2462 basic_block succ = single_succ (block);
2463 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2464 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2466 /* If we have multiple successors, we take the union of all of
2467 them. */
2468 else
2470 VEC(basic_block, heap) * worklist;
2471 size_t i;
2472 basic_block bprime;
2474 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2475 FOR_EACH_EDGE (e, ei, block->succs)
2477 if (e->flags & EDGE_DFS_BACK)
2478 continue;
2479 VEC_quick_push (basic_block, worklist, e->dest);
2481 if (VEC_length (basic_block, worklist) > 0)
2483 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2485 unsigned int i;
2486 bitmap_iterator bi;
2488 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2489 bitmap_value_insert_into_set (PA_OUT,
2490 expression_for_id (i));
2491 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2493 bitmap_set_t pa_in = bitmap_set_new ();
2494 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2495 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2496 bitmap_value_insert_into_set (PA_OUT,
2497 expression_for_id (i));
2498 bitmap_set_free (pa_in);
2500 else
2501 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2502 bitmap_value_insert_into_set (PA_OUT,
2503 expression_for_id (i));
2506 VEC_free (basic_block, heap, worklist);
2509 /* PA_IN starts with PA_OUT - TMP_GEN.
2510 Then we subtract things from ANTIC_IN. */
2511 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2513 /* For partial antic, we want to put back in the phi results, since
2514 we will properly avoid making them partially antic over backedges. */
2515 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2516 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2518 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2519 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2521 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2523 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2525 changed = true;
2526 SET_BIT (changed_blocks, block->index);
2527 FOR_EACH_EDGE (e, ei, block->preds)
2528 SET_BIT (changed_blocks, e->src->index);
2530 else
2531 RESET_BIT (changed_blocks, block->index);
2533 maybe_dump_sets:
2534 if (dump_file && (dump_flags & TDF_DETAILS))
2536 if (PA_OUT)
2537 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2539 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2541 if (old_PA_IN)
2542 bitmap_set_free (old_PA_IN);
2543 if (PA_OUT)
2544 bitmap_set_free (PA_OUT);
2545 return changed;
2548 /* Compute ANTIC and partial ANTIC sets. */
2550 static void
2551 compute_antic (void)
2553 bool changed = true;
2554 int num_iterations = 0;
2555 basic_block block;
2556 int i;
2558 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2559 We pre-build the map of blocks with incoming abnormal edges here. */
2560 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2561 sbitmap_zero (has_abnormal_preds);
2563 FOR_EACH_BB (block)
2565 edge_iterator ei;
2566 edge e;
2568 FOR_EACH_EDGE (e, ei, block->preds)
2570 e->flags &= ~EDGE_DFS_BACK;
2571 if (e->flags & EDGE_ABNORMAL)
2573 SET_BIT (has_abnormal_preds, block->index);
2574 break;
2578 BB_VISITED (block) = 0;
2579 BB_DEFERRED (block) = 0;
2581 /* While we are here, give empty ANTIC_IN sets to each block. */
2582 ANTIC_IN (block) = bitmap_set_new ();
2583 PA_IN (block) = bitmap_set_new ();
2586 /* At the exit block we anticipate nothing. */
2587 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2588 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2589 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2591 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2592 sbitmap_ones (changed_blocks);
2593 while (changed)
2595 if (dump_file && (dump_flags & TDF_DETAILS))
2596 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2597 num_iterations++;
2598 changed = false;
2599 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2601 if (TEST_BIT (changed_blocks, postorder[i]))
2603 basic_block block = BASIC_BLOCK (postorder[i]);
2604 changed |= compute_antic_aux (block,
2605 TEST_BIT (has_abnormal_preds,
2606 block->index));
2609 #ifdef ENABLE_CHECKING
2610 /* Theoretically possible, but *highly* unlikely. */
2611 gcc_assert (num_iterations < 500);
2612 #endif
2615 statistics_histogram_event (cfun, "compute_antic iterations",
2616 num_iterations);
2618 if (do_partial_partial)
2620 sbitmap_ones (changed_blocks);
2621 mark_dfs_back_edges ();
2622 num_iterations = 0;
2623 changed = true;
2624 while (changed)
2626 if (dump_file && (dump_flags & TDF_DETAILS))
2627 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2628 num_iterations++;
2629 changed = false;
2630 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2632 if (TEST_BIT (changed_blocks, postorder[i]))
2634 basic_block block = BASIC_BLOCK (postorder[i]);
2635 changed
2636 |= compute_partial_antic_aux (block,
2637 TEST_BIT (has_abnormal_preds,
2638 block->index));
2641 #ifdef ENABLE_CHECKING
2642 /* Theoretically possible, but *highly* unlikely. */
2643 gcc_assert (num_iterations < 500);
2644 #endif
2646 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2647 num_iterations);
2649 sbitmap_free (has_abnormal_preds);
2650 sbitmap_free (changed_blocks);
2653 /* Return true if we can value number the call in STMT. This is true
2654 if we have a pure or constant call. */
2656 static bool
2657 can_value_number_call (gimple stmt)
2659 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2660 return true;
2661 return false;
2664 /* Return true if OP is a tree which we can perform PRE on.
2665 This may not match the operations we can value number, but in
2666 a perfect world would. */
2668 static bool
2669 can_PRE_operation (tree op)
2671 return UNARY_CLASS_P (op)
2672 || BINARY_CLASS_P (op)
2673 || COMPARISON_CLASS_P (op)
2674 || TREE_CODE (op) == MEM_REF
2675 || TREE_CODE (op) == COMPONENT_REF
2676 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2677 || TREE_CODE (op) == CALL_EXPR
2678 || TREE_CODE (op) == ARRAY_REF;
2682 /* Inserted expressions are placed onto this worklist, which is used
2683 for performing quick dead code elimination of insertions we made
2684 that didn't turn out to be necessary. */
2685 static bitmap inserted_exprs;
2687 /* Pool allocated fake store expressions are placed onto this
2688 worklist, which, after performing dead code elimination, is walked
2689 to see which expressions need to be put into GC'able memory */
2690 static VEC(gimple, heap) *need_creation;
2692 /* The actual worker for create_component_ref_by_pieces. */
2694 static tree
2695 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2696 unsigned int *operand, gimple_seq *stmts,
2697 gimple domstmt)
2699 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2700 *operand);
2701 tree genop;
2702 ++*operand;
2703 switch (currop->opcode)
2705 case CALL_EXPR:
2707 tree folded, sc = NULL_TREE;
2708 unsigned int nargs = 0;
2709 tree fn, *args;
2710 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2711 fn = currop->op0;
2712 else
2714 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2715 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2716 if (!fn)
2717 return NULL_TREE;
2719 if (currop->op1)
2721 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2722 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2723 if (!sc)
2724 return NULL_TREE;
2726 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2727 ref->operands) - 1);
2728 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2730 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2731 operand, stmts,
2732 domstmt);
2733 if (!args[nargs])
2735 free (args);
2736 return NULL_TREE;
2738 nargs++;
2740 folded = build_call_array (currop->type,
2741 (TREE_CODE (fn) == FUNCTION_DECL
2742 ? build_fold_addr_expr (fn) : fn),
2743 nargs, args);
2744 free (args);
2745 if (sc)
2746 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2747 return folded;
2749 break;
2750 case MEM_REF:
2752 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2753 stmts, domstmt);
2754 tree offset = currop->op0;
2755 if (!baseop)
2756 return NULL_TREE;
2757 if (TREE_CODE (baseop) == ADDR_EXPR
2758 && handled_component_p (TREE_OPERAND (baseop, 0)))
2760 HOST_WIDE_INT off;
2761 tree base;
2762 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2763 &off);
2764 gcc_assert (base);
2765 offset = int_const_binop (PLUS_EXPR, offset,
2766 build_int_cst (TREE_TYPE (offset),
2767 off), 0);
2768 baseop = build_fold_addr_expr (base);
2770 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2772 break;
2773 case TARGET_MEM_REF:
2775 pre_expr op0expr, op1expr;
2776 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2777 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2778 ++*operand);
2779 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2780 stmts, domstmt);
2781 if (!baseop)
2782 return NULL_TREE;
2783 if (currop->op0)
2785 op0expr = get_or_alloc_expr_for (currop->op0);
2786 genop0 = find_or_generate_expression (block, op0expr,
2787 stmts, domstmt);
2788 if (!genop0)
2789 return NULL_TREE;
2791 if (nextop->op0)
2793 op1expr = get_or_alloc_expr_for (nextop->op0);
2794 genop1 = find_or_generate_expression (block, op1expr,
2795 stmts, domstmt);
2796 if (!genop1)
2797 return NULL_TREE;
2799 return build5 (TARGET_MEM_REF, currop->type,
2800 baseop, currop->op2, genop0, currop->op1, genop1);
2802 break;
2803 case ADDR_EXPR:
2804 if (currop->op0)
2806 gcc_assert (is_gimple_min_invariant (currop->op0));
2807 return currop->op0;
2809 /* Fallthrough. */
2810 case REALPART_EXPR:
2811 case IMAGPART_EXPR:
2812 case VIEW_CONVERT_EXPR:
2814 tree folded;
2815 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2816 operand,
2817 stmts, domstmt);
2818 if (!genop0)
2819 return NULL_TREE;
2820 folded = fold_build1 (currop->opcode, currop->type,
2821 genop0);
2822 return folded;
2824 break;
2825 case BIT_FIELD_REF:
2827 tree folded;
2828 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2829 stmts, domstmt);
2830 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2831 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2832 tree genop1;
2833 tree genop2;
2835 if (!genop0)
2836 return NULL_TREE;
2837 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2838 if (!genop1)
2839 return NULL_TREE;
2840 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2841 if (!genop2)
2842 return NULL_TREE;
2843 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2844 genop2);
2845 return folded;
2848 /* For array ref vn_reference_op's, operand 1 of the array ref
2849 is op0 of the reference op and operand 3 of the array ref is
2850 op1. */
2851 case ARRAY_RANGE_REF:
2852 case ARRAY_REF:
2854 tree genop0;
2855 tree genop1 = currop->op0;
2856 pre_expr op1expr;
2857 tree genop2 = currop->op1;
2858 pre_expr op2expr;
2859 tree genop3 = currop->op2;
2860 pre_expr op3expr;
2861 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2862 stmts, domstmt);
2863 if (!genop0)
2864 return NULL_TREE;
2865 op1expr = get_or_alloc_expr_for (genop1);
2866 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2867 if (!genop1)
2868 return NULL_TREE;
2869 if (genop2)
2871 /* Drop zero minimum index. */
2872 if (tree_int_cst_equal (genop2, integer_zero_node))
2873 genop2 = NULL_TREE;
2874 else
2876 op2expr = get_or_alloc_expr_for (genop2);
2877 genop2 = find_or_generate_expression (block, op2expr, stmts,
2878 domstmt);
2879 if (!genop2)
2880 return NULL_TREE;
2883 if (genop3)
2885 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2886 /* We can't always put a size in units of the element alignment
2887 here as the element alignment may be not visible. See
2888 PR43783. Simply drop the element size for constant
2889 sizes. */
2890 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2891 genop3 = NULL_TREE;
2892 else
2894 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2895 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2896 op3expr = get_or_alloc_expr_for (genop3);
2897 genop3 = find_or_generate_expression (block, op3expr, stmts,
2898 domstmt);
2899 if (!genop3)
2900 return NULL_TREE;
2903 return build4 (currop->opcode, currop->type, genop0, genop1,
2904 genop2, genop3);
2906 case COMPONENT_REF:
2908 tree op0;
2909 tree op1;
2910 tree genop2 = currop->op1;
2911 pre_expr op2expr;
2912 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2913 stmts, domstmt);
2914 if (!op0)
2915 return NULL_TREE;
2916 /* op1 should be a FIELD_DECL, which are represented by
2917 themselves. */
2918 op1 = currop->op0;
2919 if (genop2)
2921 op2expr = get_or_alloc_expr_for (genop2);
2922 genop2 = find_or_generate_expression (block, op2expr, stmts,
2923 domstmt);
2924 if (!genop2)
2925 return NULL_TREE;
2928 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2929 genop2);
2931 break;
2932 case SSA_NAME:
2934 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2935 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2936 return genop;
2938 case STRING_CST:
2939 case INTEGER_CST:
2940 case COMPLEX_CST:
2941 case VECTOR_CST:
2942 case REAL_CST:
2943 case CONSTRUCTOR:
2944 case VAR_DECL:
2945 case PARM_DECL:
2946 case CONST_DECL:
2947 case RESULT_DECL:
2948 case FUNCTION_DECL:
2949 return currop->op0;
2951 default:
2952 gcc_unreachable ();
2956 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2957 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2958 trying to rename aggregates into ssa form directly, which is a no no.
2960 Thus, this routine doesn't create temporaries, it just builds a
2961 single access expression for the array, calling
2962 find_or_generate_expression to build the innermost pieces.
2964 This function is a subroutine of create_expression_by_pieces, and
2965 should not be called on it's own unless you really know what you
2966 are doing. */
2968 static tree
2969 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2970 gimple_seq *stmts, gimple domstmt)
2972 unsigned int op = 0;
2973 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2976 /* Find a leader for an expression, or generate one using
2977 create_expression_by_pieces if it's ANTIC but
2978 complex.
2979 BLOCK is the basic_block we are looking for leaders in.
2980 EXPR is the expression to find a leader or generate for.
2981 STMTS is the statement list to put the inserted expressions on.
2982 Returns the SSA_NAME of the LHS of the generated expression or the
2983 leader.
2984 DOMSTMT if non-NULL is a statement that should be dominated by
2985 all uses in the generated expression. If DOMSTMT is non-NULL this
2986 routine can fail and return NULL_TREE. Otherwise it will assert
2987 on failure. */
2989 static tree
2990 find_or_generate_expression (basic_block block, pre_expr expr,
2991 gimple_seq *stmts, gimple domstmt)
2993 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2994 get_expr_value_id (expr), domstmt);
2995 tree genop = NULL;
2996 if (leader)
2998 if (leader->kind == NAME)
2999 genop = PRE_EXPR_NAME (leader);
3000 else if (leader->kind == CONSTANT)
3001 genop = PRE_EXPR_CONSTANT (leader);
3004 /* If it's still NULL, it must be a complex expression, so generate
3005 it recursively. Not so if inserting expressions for values generated
3006 by SCCVN. */
3007 if (genop == NULL
3008 && !domstmt)
3010 bitmap_set_t exprset;
3011 unsigned int lookfor = get_expr_value_id (expr);
3012 bool handled = false;
3013 bitmap_iterator bi;
3014 unsigned int i;
3016 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3017 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3019 pre_expr temp = expression_for_id (i);
3020 if (temp->kind != NAME)
3022 handled = true;
3023 genop = create_expression_by_pieces (block, temp, stmts,
3024 domstmt,
3025 get_expr_type (expr));
3026 break;
3029 if (!handled && domstmt)
3030 return NULL_TREE;
3032 gcc_assert (handled);
3034 return genop;
3037 #define NECESSARY GF_PLF_1
3039 /* Create an expression in pieces, so that we can handle very complex
3040 expressions that may be ANTIC, but not necessary GIMPLE.
3041 BLOCK is the basic block the expression will be inserted into,
3042 EXPR is the expression to insert (in value form)
3043 STMTS is a statement list to append the necessary insertions into.
3045 This function will die if we hit some value that shouldn't be
3046 ANTIC but is (IE there is no leader for it, or its components).
3047 This function may also generate expressions that are themselves
3048 partially or fully redundant. Those that are will be either made
3049 fully redundant during the next iteration of insert (for partially
3050 redundant ones), or eliminated by eliminate (for fully redundant
3051 ones).
3053 If DOMSTMT is non-NULL then we make sure that all uses in the
3054 expressions dominate that statement. In this case the function
3055 can return NULL_TREE to signal failure. */
3057 static tree
3058 create_expression_by_pieces (basic_block block, pre_expr expr,
3059 gimple_seq *stmts, gimple domstmt, tree type)
3061 tree temp, name;
3062 tree folded;
3063 gimple_seq forced_stmts = NULL;
3064 unsigned int value_id;
3065 gimple_stmt_iterator gsi;
3066 tree exprtype = type ? type : get_expr_type (expr);
3067 pre_expr nameexpr;
3068 gimple newstmt;
3070 switch (expr->kind)
3072 /* We may hit the NAME/CONSTANT case if we have to convert types
3073 that value numbering saw through. */
3074 case NAME:
3075 folded = PRE_EXPR_NAME (expr);
3076 break;
3077 case CONSTANT:
3078 folded = PRE_EXPR_CONSTANT (expr);
3079 break;
3080 case REFERENCE:
3082 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3083 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3085 break;
3086 case NARY:
3088 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3089 switch (nary->length)
3091 case 2:
3093 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3094 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
3095 tree genop1 = find_or_generate_expression (block, op1,
3096 stmts, domstmt);
3097 tree genop2 = find_or_generate_expression (block, op2,
3098 stmts, domstmt);
3099 if (!genop1 || !genop2)
3100 return NULL_TREE;
3101 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
3102 may be a constant with the wrong type. */
3103 if (nary->opcode == POINTER_PLUS_EXPR)
3105 genop1 = fold_convert (nary->type, genop1);
3106 genop2 = fold_convert (sizetype, genop2);
3108 else
3110 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3111 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
3114 folded = fold_build2 (nary->opcode, nary->type,
3115 genop1, genop2);
3117 break;
3118 case 1:
3120 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3121 tree genop1 = find_or_generate_expression (block, op1,
3122 stmts, domstmt);
3123 if (!genop1)
3124 return NULL_TREE;
3125 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3127 folded = fold_build1 (nary->opcode, nary->type,
3128 genop1);
3130 break;
3131 default:
3132 return NULL_TREE;
3135 break;
3136 default:
3137 return NULL_TREE;
3140 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3141 folded = fold_convert (exprtype, folded);
3143 /* Force the generated expression to be a sequence of GIMPLE
3144 statements.
3145 We have to call unshare_expr because force_gimple_operand may
3146 modify the tree we pass to it. */
3147 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3148 false, NULL);
3150 /* If we have any intermediate expressions to the value sets, add them
3151 to the value sets and chain them in the instruction stream. */
3152 if (forced_stmts)
3154 gsi = gsi_start (forced_stmts);
3155 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3157 gimple stmt = gsi_stmt (gsi);
3158 tree forcedname = gimple_get_lhs (stmt);
3159 pre_expr nameexpr;
3161 if (TREE_CODE (forcedname) == SSA_NAME)
3163 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3164 VN_INFO_GET (forcedname)->valnum = forcedname;
3165 VN_INFO (forcedname)->value_id = get_next_value_id ();
3166 nameexpr = get_or_alloc_expr_for_name (forcedname);
3167 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3168 if (!in_fre)
3169 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3170 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3172 mark_symbols_for_renaming (stmt);
3174 gimple_seq_add_seq (stmts, forced_stmts);
3177 /* Build and insert the assignment of the end result to the temporary
3178 that we will return. */
3179 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3181 pretemp = create_tmp_reg (exprtype, "pretmp");
3182 get_var_ann (pretemp);
3185 temp = pretemp;
3186 add_referenced_var (temp);
3188 newstmt = gimple_build_assign (temp, folded);
3189 name = make_ssa_name (temp, newstmt);
3190 gimple_assign_set_lhs (newstmt, name);
3191 gimple_set_plf (newstmt, NECESSARY, false);
3193 gimple_seq_add_stmt (stmts, newstmt);
3194 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3196 /* All the symbols in NEWEXPR should be put into SSA form. */
3197 mark_symbols_for_renaming (newstmt);
3199 /* Add a value number to the temporary.
3200 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3201 we are creating the expression by pieces, and this particular piece of
3202 the expression may have been represented. There is no harm in replacing
3203 here. */
3204 VN_INFO_GET (name)->valnum = name;
3205 value_id = get_expr_value_id (expr);
3206 VN_INFO (name)->value_id = value_id;
3207 nameexpr = get_or_alloc_expr_for_name (name);
3208 add_to_value (value_id, nameexpr);
3209 if (NEW_SETS (block))
3210 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3211 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3213 pre_stats.insertions++;
3214 if (dump_file && (dump_flags & TDF_DETAILS))
3216 fprintf (dump_file, "Inserted ");
3217 print_gimple_stmt (dump_file, newstmt, 0, 0);
3218 fprintf (dump_file, " in predecessor %d\n", block->index);
3221 return name;
3225 /* Returns true if we want to inhibit the insertions of PHI nodes
3226 for the given EXPR for basic block BB (a member of a loop).
3227 We want to do this, when we fear that the induction variable we
3228 create might inhibit vectorization. */
3230 static bool
3231 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3233 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3234 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3235 vn_reference_op_t op;
3236 unsigned i;
3238 /* If we aren't going to vectorize we don't inhibit anything. */
3239 if (!flag_tree_vectorize)
3240 return false;
3242 /* Otherwise we inhibit the insertion when the address of the
3243 memory reference is a simple induction variable. In other
3244 cases the vectorizer won't do anything anyway (either it's
3245 loop invariant or a complicated expression). */
3246 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3248 switch (op->opcode)
3250 case ARRAY_REF:
3251 case ARRAY_RANGE_REF:
3252 if (TREE_CODE (op->op0) != SSA_NAME)
3253 break;
3254 /* Fallthru. */
3255 case SSA_NAME:
3257 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3258 affine_iv iv;
3259 /* Default defs are loop invariant. */
3260 if (!defbb)
3261 break;
3262 /* Defined outside this loop, also loop invariant. */
3263 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3264 break;
3265 /* If it's a simple induction variable inhibit insertion,
3266 the vectorizer might be interested in this one. */
3267 if (simple_iv (bb->loop_father, bb->loop_father,
3268 op->op0, &iv, true))
3269 return true;
3270 /* No simple IV, vectorizer can't do anything, hence no
3271 reason to inhibit the transformation for this operand. */
3272 break;
3274 default:
3275 break;
3278 return false;
3281 /* Insert the to-be-made-available values of expression EXPRNUM for each
3282 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3283 merge the result with a phi node, given the same value number as
3284 NODE. Return true if we have inserted new stuff. */
3286 static bool
3287 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3288 pre_expr *avail)
3290 pre_expr expr = expression_for_id (exprnum);
3291 pre_expr newphi;
3292 unsigned int val = get_expr_value_id (expr);
3293 edge pred;
3294 bool insertions = false;
3295 bool nophi = false;
3296 basic_block bprime;
3297 pre_expr eprime;
3298 edge_iterator ei;
3299 tree type = get_expr_type (expr);
3300 tree temp;
3301 gimple phi;
3303 if (dump_file && (dump_flags & TDF_DETAILS))
3305 fprintf (dump_file, "Found partial redundancy for expression ");
3306 print_pre_expr (dump_file, expr);
3307 fprintf (dump_file, " (%04d)\n", val);
3310 /* Make sure we aren't creating an induction variable. */
3311 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3313 bool firstinsideloop = false;
3314 bool secondinsideloop = false;
3315 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3316 EDGE_PRED (block, 0)->src);
3317 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3318 EDGE_PRED (block, 1)->src);
3319 /* Induction variables only have one edge inside the loop. */
3320 if ((firstinsideloop ^ secondinsideloop)
3321 && (expr->kind != REFERENCE
3322 || inhibit_phi_insertion (block, expr)))
3324 if (dump_file && (dump_flags & TDF_DETAILS))
3325 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3326 nophi = true;
3330 /* Make the necessary insertions. */
3331 FOR_EACH_EDGE (pred, ei, block->preds)
3333 gimple_seq stmts = NULL;
3334 tree builtexpr;
3335 bprime = pred->src;
3336 eprime = avail[bprime->index];
3338 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3340 builtexpr = create_expression_by_pieces (bprime,
3341 eprime,
3342 &stmts, NULL,
3343 type);
3344 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3345 gsi_insert_seq_on_edge (pred, stmts);
3346 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3347 insertions = true;
3349 else if (eprime->kind == CONSTANT)
3351 /* Constants may not have the right type, fold_convert
3352 should give us back a constant with the right type.
3354 tree constant = PRE_EXPR_CONSTANT (eprime);
3355 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3357 tree builtexpr = fold_convert (type, constant);
3358 if (!is_gimple_min_invariant (builtexpr))
3360 tree forcedexpr = force_gimple_operand (builtexpr,
3361 &stmts, true,
3362 NULL);
3363 if (!is_gimple_min_invariant (forcedexpr))
3365 if (forcedexpr != builtexpr)
3367 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3368 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3370 if (stmts)
3372 gimple_stmt_iterator gsi;
3373 gsi = gsi_start (stmts);
3374 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3376 gimple stmt = gsi_stmt (gsi);
3377 tree lhs = gimple_get_lhs (stmt);
3378 if (TREE_CODE (lhs) == SSA_NAME)
3379 bitmap_set_bit (inserted_exprs,
3380 SSA_NAME_VERSION (lhs));
3381 gimple_set_plf (stmt, NECESSARY, false);
3383 gsi_insert_seq_on_edge (pred, stmts);
3385 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3388 else
3389 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3392 else if (eprime->kind == NAME)
3394 /* We may have to do a conversion because our value
3395 numbering can look through types in certain cases, but
3396 our IL requires all operands of a phi node have the same
3397 type. */
3398 tree name = PRE_EXPR_NAME (eprime);
3399 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3401 tree builtexpr;
3402 tree forcedexpr;
3403 builtexpr = fold_convert (type, name);
3404 forcedexpr = force_gimple_operand (builtexpr,
3405 &stmts, true,
3406 NULL);
3408 if (forcedexpr != name)
3410 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3411 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3414 if (stmts)
3416 gimple_stmt_iterator gsi;
3417 gsi = gsi_start (stmts);
3418 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3420 gimple stmt = gsi_stmt (gsi);
3421 tree lhs = gimple_get_lhs (stmt);
3422 if (TREE_CODE (lhs) == SSA_NAME)
3423 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3424 gimple_set_plf (stmt, NECESSARY, false);
3426 gsi_insert_seq_on_edge (pred, stmts);
3428 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3432 /* If we didn't want a phi node, and we made insertions, we still have
3433 inserted new stuff, and thus return true. If we didn't want a phi node,
3434 and didn't make insertions, we haven't added anything new, so return
3435 false. */
3436 if (nophi && insertions)
3437 return true;
3438 else if (nophi && !insertions)
3439 return false;
3441 /* Now build a phi for the new variable. */
3442 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3444 prephitemp = create_tmp_var (type, "prephitmp");
3445 get_var_ann (prephitemp);
3448 temp = prephitemp;
3449 add_referenced_var (temp);
3451 if (TREE_CODE (type) == COMPLEX_TYPE
3452 || TREE_CODE (type) == VECTOR_TYPE)
3453 DECL_GIMPLE_REG_P (temp) = 1;
3454 phi = create_phi_node (temp, block);
3456 gimple_set_plf (phi, NECESSARY, false);
3457 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3458 VN_INFO (gimple_phi_result (phi))->value_id = val;
3459 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3460 FOR_EACH_EDGE (pred, ei, block->preds)
3462 pre_expr ae = avail[pred->src->index];
3463 gcc_assert (get_expr_type (ae) == type
3464 || useless_type_conversion_p (type, get_expr_type (ae)));
3465 if (ae->kind == CONSTANT)
3466 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3467 else
3468 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3469 UNKNOWN_LOCATION);
3472 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3473 add_to_value (val, newphi);
3475 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3476 this insertion, since we test for the existence of this value in PHI_GEN
3477 before proceeding with the partial redundancy checks in insert_aux.
3479 The value may exist in AVAIL_OUT, in particular, it could be represented
3480 by the expression we are trying to eliminate, in which case we want the
3481 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3482 inserted there.
3484 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3485 this block, because if it did, it would have existed in our dominator's
3486 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3489 bitmap_insert_into_set (PHI_GEN (block), newphi);
3490 bitmap_value_replace_in_set (AVAIL_OUT (block),
3491 newphi);
3492 bitmap_insert_into_set (NEW_SETS (block),
3493 newphi);
3495 if (dump_file && (dump_flags & TDF_DETAILS))
3497 fprintf (dump_file, "Created phi ");
3498 print_gimple_stmt (dump_file, phi, 0, 0);
3499 fprintf (dump_file, " in block %d\n", block->index);
3501 pre_stats.phis++;
3502 return true;
3507 /* Perform insertion of partially redundant values.
3508 For BLOCK, do the following:
3509 1. Propagate the NEW_SETS of the dominator into the current block.
3510 If the block has multiple predecessors,
3511 2a. Iterate over the ANTIC expressions for the block to see if
3512 any of them are partially redundant.
3513 2b. If so, insert them into the necessary predecessors to make
3514 the expression fully redundant.
3515 2c. Insert a new PHI merging the values of the predecessors.
3516 2d. Insert the new PHI, and the new expressions, into the
3517 NEW_SETS set.
3518 3. Recursively call ourselves on the dominator children of BLOCK.
3520 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3521 do_regular_insertion and do_partial_insertion.
3525 static bool
3526 do_regular_insertion (basic_block block, basic_block dom)
3528 bool new_stuff = false;
3529 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3530 pre_expr expr;
3531 int i;
3533 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3535 if (expr->kind != NAME)
3537 pre_expr *avail;
3538 unsigned int val;
3539 bool by_some = false;
3540 bool cant_insert = false;
3541 bool all_same = true;
3542 pre_expr first_s = NULL;
3543 edge pred;
3544 basic_block bprime;
3545 pre_expr eprime = NULL;
3546 edge_iterator ei;
3547 pre_expr edoubleprime = NULL;
3548 bool do_insertion = false;
3550 val = get_expr_value_id (expr);
3551 if (bitmap_set_contains_value (PHI_GEN (block), val))
3552 continue;
3553 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3555 if (dump_file && (dump_flags & TDF_DETAILS))
3556 fprintf (dump_file, "Found fully redundant value\n");
3557 continue;
3560 avail = XCNEWVEC (pre_expr, last_basic_block);
3561 FOR_EACH_EDGE (pred, ei, block->preds)
3563 unsigned int vprime;
3565 /* We should never run insertion for the exit block
3566 and so not come across fake pred edges. */
3567 gcc_assert (!(pred->flags & EDGE_FAKE));
3568 bprime = pred->src;
3569 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3570 bprime, block);
3572 /* eprime will generally only be NULL if the
3573 value of the expression, translated
3574 through the PHI for this predecessor, is
3575 undefined. If that is the case, we can't
3576 make the expression fully redundant,
3577 because its value is undefined along a
3578 predecessor path. We can thus break out
3579 early because it doesn't matter what the
3580 rest of the results are. */
3581 if (eprime == NULL)
3583 cant_insert = true;
3584 break;
3587 eprime = fully_constant_expression (eprime);
3588 vprime = get_expr_value_id (eprime);
3589 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3590 vprime, NULL);
3591 if (edoubleprime == NULL)
3593 avail[bprime->index] = eprime;
3594 all_same = false;
3596 else
3598 avail[bprime->index] = edoubleprime;
3599 by_some = true;
3600 /* We want to perform insertions to remove a redundancy on
3601 a path in the CFG we want to optimize for speed. */
3602 if (optimize_edge_for_speed_p (pred))
3603 do_insertion = true;
3604 if (first_s == NULL)
3605 first_s = edoubleprime;
3606 else if (!pre_expr_eq (first_s, edoubleprime))
3607 all_same = false;
3610 /* If we can insert it, it's not the same value
3611 already existing along every predecessor, and
3612 it's defined by some predecessor, it is
3613 partially redundant. */
3614 if (!cant_insert && !all_same && by_some && do_insertion
3615 && dbg_cnt (treepre_insert))
3617 if (insert_into_preds_of_block (block, get_expression_id (expr),
3618 avail))
3619 new_stuff = true;
3621 /* If all edges produce the same value and that value is
3622 an invariant, then the PHI has the same value on all
3623 edges. Note this. */
3624 else if (!cant_insert && all_same && eprime
3625 && (edoubleprime->kind == CONSTANT
3626 || edoubleprime->kind == NAME)
3627 && !value_id_constant_p (val))
3629 unsigned int j;
3630 bitmap_iterator bi;
3631 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3632 value_expressions, val);
3634 unsigned int new_val = get_expr_value_id (edoubleprime);
3635 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3637 pre_expr expr = expression_for_id (j);
3639 if (expr->kind == NAME)
3641 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3642 /* Just reset the value id and valnum so it is
3643 the same as the constant we have discovered. */
3644 if (edoubleprime->kind == CONSTANT)
3646 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3647 pre_stats.constified++;
3649 else
3650 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3651 info->value_id = new_val;
3655 free (avail);
3659 VEC_free (pre_expr, heap, exprs);
3660 return new_stuff;
3664 /* Perform insertion for partially anticipatable expressions. There
3665 is only one case we will perform insertion for these. This case is
3666 if the expression is partially anticipatable, and fully available.
3667 In this case, we know that putting it earlier will enable us to
3668 remove the later computation. */
3671 static bool
3672 do_partial_partial_insertion (basic_block block, basic_block dom)
3674 bool new_stuff = false;
3675 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3676 pre_expr expr;
3677 int i;
3679 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3681 if (expr->kind != NAME)
3683 pre_expr *avail;
3684 unsigned int val;
3685 bool by_all = true;
3686 bool cant_insert = false;
3687 edge pred;
3688 basic_block bprime;
3689 pre_expr eprime = NULL;
3690 edge_iterator ei;
3692 val = get_expr_value_id (expr);
3693 if (bitmap_set_contains_value (PHI_GEN (block), val))
3694 continue;
3695 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3696 continue;
3698 avail = XCNEWVEC (pre_expr, last_basic_block);
3699 FOR_EACH_EDGE (pred, ei, block->preds)
3701 unsigned int vprime;
3702 pre_expr edoubleprime;
3704 /* We should never run insertion for the exit block
3705 and so not come across fake pred edges. */
3706 gcc_assert (!(pred->flags & EDGE_FAKE));
3707 bprime = pred->src;
3708 eprime = phi_translate (expr, ANTIC_IN (block),
3709 PA_IN (block),
3710 bprime, block);
3712 /* eprime will generally only be NULL if the
3713 value of the expression, translated
3714 through the PHI for this predecessor, is
3715 undefined. If that is the case, we can't
3716 make the expression fully redundant,
3717 because its value is undefined along a
3718 predecessor path. We can thus break out
3719 early because it doesn't matter what the
3720 rest of the results are. */
3721 if (eprime == NULL)
3723 cant_insert = true;
3724 break;
3727 eprime = fully_constant_expression (eprime);
3728 vprime = get_expr_value_id (eprime);
3729 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3730 vprime, NULL);
3731 if (edoubleprime == NULL)
3733 by_all = false;
3734 break;
3736 else
3737 avail[bprime->index] = edoubleprime;
3741 /* If we can insert it, it's not the same value
3742 already existing along every predecessor, and
3743 it's defined by some predecessor, it is
3744 partially redundant. */
3745 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3747 pre_stats.pa_insert++;
3748 if (insert_into_preds_of_block (block, get_expression_id (expr),
3749 avail))
3750 new_stuff = true;
3752 free (avail);
3756 VEC_free (pre_expr, heap, exprs);
3757 return new_stuff;
3760 static bool
3761 insert_aux (basic_block block)
3763 basic_block son;
3764 bool new_stuff = false;
3766 if (block)
3768 basic_block dom;
3769 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3770 if (dom)
3772 unsigned i;
3773 bitmap_iterator bi;
3774 bitmap_set_t newset = NEW_SETS (dom);
3775 if (newset)
3777 /* Note that we need to value_replace both NEW_SETS, and
3778 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3779 represented by some non-simple expression here that we want
3780 to replace it with. */
3781 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3783 pre_expr expr = expression_for_id (i);
3784 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3785 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3788 if (!single_pred_p (block))
3790 new_stuff |= do_regular_insertion (block, dom);
3791 if (do_partial_partial)
3792 new_stuff |= do_partial_partial_insertion (block, dom);
3796 for (son = first_dom_son (CDI_DOMINATORS, block);
3797 son;
3798 son = next_dom_son (CDI_DOMINATORS, son))
3800 new_stuff |= insert_aux (son);
3803 return new_stuff;
3806 /* Perform insertion of partially redundant values. */
3808 static void
3809 insert (void)
3811 bool new_stuff = true;
3812 basic_block bb;
3813 int num_iterations = 0;
3815 FOR_ALL_BB (bb)
3816 NEW_SETS (bb) = bitmap_set_new ();
3818 while (new_stuff)
3820 num_iterations++;
3821 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3823 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3827 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3829 static void
3830 add_to_exp_gen (basic_block block, tree op)
3832 if (!in_fre)
3834 pre_expr result;
3835 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3836 return;
3837 result = get_or_alloc_expr_for_name (op);
3838 bitmap_value_insert_into_set (EXP_GEN (block), result);
3842 /* Create value ids for PHI in BLOCK. */
3844 static void
3845 make_values_for_phi (gimple phi, basic_block block)
3847 tree result = gimple_phi_result (phi);
3849 /* We have no need for virtual phis, as they don't represent
3850 actual computations. */
3851 if (is_gimple_reg (result))
3853 pre_expr e = get_or_alloc_expr_for_name (result);
3854 add_to_value (get_expr_value_id (e), e);
3855 bitmap_insert_into_set (PHI_GEN (block), e);
3856 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3857 if (!in_fre)
3859 unsigned i;
3860 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3862 tree arg = gimple_phi_arg_def (phi, i);
3863 if (TREE_CODE (arg) == SSA_NAME)
3865 e = get_or_alloc_expr_for_name (arg);
3866 add_to_value (get_expr_value_id (e), e);
3873 /* Compute the AVAIL set for all basic blocks.
3875 This function performs value numbering of the statements in each basic
3876 block. The AVAIL sets are built from information we glean while doing
3877 this value numbering, since the AVAIL sets contain only one entry per
3878 value.
3880 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3881 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3883 static void
3884 compute_avail (void)
3887 basic_block block, son;
3888 basic_block *worklist;
3889 size_t sp = 0;
3890 unsigned i;
3892 /* We pretend that default definitions are defined in the entry block.
3893 This includes function arguments and the static chain decl. */
3894 for (i = 1; i < num_ssa_names; ++i)
3896 tree name = ssa_name (i);
3897 pre_expr e;
3898 if (!name
3899 || !SSA_NAME_IS_DEFAULT_DEF (name)
3900 || has_zero_uses (name)
3901 || !is_gimple_reg (name))
3902 continue;
3904 e = get_or_alloc_expr_for_name (name);
3905 add_to_value (get_expr_value_id (e), e);
3906 if (!in_fre)
3907 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3908 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3911 /* Allocate the worklist. */
3912 worklist = XNEWVEC (basic_block, n_basic_blocks);
3914 /* Seed the algorithm by putting the dominator children of the entry
3915 block on the worklist. */
3916 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3917 son;
3918 son = next_dom_son (CDI_DOMINATORS, son))
3919 worklist[sp++] = son;
3921 /* Loop until the worklist is empty. */
3922 while (sp)
3924 gimple_stmt_iterator gsi;
3925 gimple stmt;
3926 basic_block dom;
3927 unsigned int stmt_uid = 1;
3929 /* Pick a block from the worklist. */
3930 block = worklist[--sp];
3932 /* Initially, the set of available values in BLOCK is that of
3933 its immediate dominator. */
3934 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3935 if (dom)
3936 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3938 /* Generate values for PHI nodes. */
3939 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3940 make_values_for_phi (gsi_stmt (gsi), block);
3942 BB_MAY_NOTRETURN (block) = 0;
3944 /* Now compute value numbers and populate value sets with all
3945 the expressions computed in BLOCK. */
3946 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3948 ssa_op_iter iter;
3949 tree op;
3951 stmt = gsi_stmt (gsi);
3952 gimple_set_uid (stmt, stmt_uid++);
3954 /* Cache whether the basic-block has any non-visible side-effect
3955 or control flow.
3956 If this isn't a call or it is the last stmt in the
3957 basic-block then the CFG represents things correctly. */
3958 if (is_gimple_call (stmt)
3959 && !stmt_ends_bb_p (stmt))
3961 /* Non-looping const functions always return normally.
3962 Otherwise the call might not return or have side-effects
3963 that forbids hoisting possibly trapping expressions
3964 before it. */
3965 int flags = gimple_call_flags (stmt);
3966 if (!(flags & ECF_CONST)
3967 || (flags & ECF_LOOPING_CONST_OR_PURE))
3968 BB_MAY_NOTRETURN (block) = 1;
3971 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3973 pre_expr e = get_or_alloc_expr_for_name (op);
3975 add_to_value (get_expr_value_id (e), e);
3976 if (!in_fre)
3977 bitmap_insert_into_set (TMP_GEN (block), e);
3978 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3981 if (gimple_has_volatile_ops (stmt)
3982 || stmt_could_throw_p (stmt))
3983 continue;
3985 switch (gimple_code (stmt))
3987 case GIMPLE_RETURN:
3988 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3989 add_to_exp_gen (block, op);
3990 continue;
3992 case GIMPLE_CALL:
3994 vn_reference_t ref;
3995 unsigned int i;
3996 vn_reference_op_t vro;
3997 pre_expr result = NULL;
3998 VEC(vn_reference_op_s, heap) *ops = NULL;
4000 if (!can_value_number_call (stmt))
4001 continue;
4003 copy_reference_ops_from_call (stmt, &ops);
4004 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4005 gimple_expr_type (stmt),
4006 ops, &ref, false);
4007 VEC_free (vn_reference_op_s, heap, ops);
4008 if (!ref)
4009 continue;
4011 for (i = 0; VEC_iterate (vn_reference_op_s,
4012 ref->operands, i,
4013 vro); i++)
4015 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4016 add_to_exp_gen (block, vro->op0);
4017 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4018 add_to_exp_gen (block, vro->op1);
4019 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4020 add_to_exp_gen (block, vro->op2);
4022 result = (pre_expr) pool_alloc (pre_expr_pool);
4023 result->kind = REFERENCE;
4024 result->id = 0;
4025 PRE_EXPR_REFERENCE (result) = ref;
4027 get_or_alloc_expression_id (result);
4028 add_to_value (get_expr_value_id (result), result);
4029 if (!in_fre)
4030 bitmap_value_insert_into_set (EXP_GEN (block), result);
4031 continue;
4034 case GIMPLE_ASSIGN:
4036 pre_expr result = NULL;
4037 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4039 case tcc_unary:
4040 case tcc_binary:
4041 case tcc_comparison:
4043 vn_nary_op_t nary;
4044 unsigned int i;
4046 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4047 gimple_assign_rhs_code (stmt),
4048 gimple_expr_type (stmt),
4049 gimple_assign_rhs1 (stmt),
4050 gimple_assign_rhs2 (stmt),
4051 NULL_TREE, NULL_TREE, &nary);
4053 if (!nary)
4054 continue;
4056 for (i = 0; i < nary->length; i++)
4057 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4058 add_to_exp_gen (block, nary->op[i]);
4060 result = (pre_expr) pool_alloc (pre_expr_pool);
4061 result->kind = NARY;
4062 result->id = 0;
4063 PRE_EXPR_NARY (result) = nary;
4064 break;
4067 case tcc_declaration:
4068 case tcc_reference:
4070 vn_reference_t ref;
4071 unsigned int i;
4072 vn_reference_op_t vro;
4074 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4075 gimple_vuse (stmt),
4076 true, &ref);
4077 if (!ref)
4078 continue;
4080 for (i = 0; VEC_iterate (vn_reference_op_s,
4081 ref->operands, i,
4082 vro); i++)
4084 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4085 add_to_exp_gen (block, vro->op0);
4086 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4087 add_to_exp_gen (block, vro->op1);
4088 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4089 add_to_exp_gen (block, vro->op2);
4091 result = (pre_expr) pool_alloc (pre_expr_pool);
4092 result->kind = REFERENCE;
4093 result->id = 0;
4094 PRE_EXPR_REFERENCE (result) = ref;
4095 break;
4098 default:
4099 /* For any other statement that we don't
4100 recognize, simply add all referenced
4101 SSA_NAMEs to EXP_GEN. */
4102 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4103 add_to_exp_gen (block, op);
4104 continue;
4107 get_or_alloc_expression_id (result);
4108 add_to_value (get_expr_value_id (result), result);
4109 if (!in_fre)
4110 bitmap_value_insert_into_set (EXP_GEN (block), result);
4112 continue;
4114 default:
4115 break;
4119 /* Put the dominator children of BLOCK on the worklist of blocks
4120 to compute available sets for. */
4121 for (son = first_dom_son (CDI_DOMINATORS, block);
4122 son;
4123 son = next_dom_son (CDI_DOMINATORS, son))
4124 worklist[sp++] = son;
4127 free (worklist);
4130 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4131 than the available expressions for it. The insertion point is
4132 right before the first use in STMT. Returns the SSA_NAME that should
4133 be used for replacement. */
4135 static tree
4136 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4138 basic_block bb = gimple_bb (stmt);
4139 gimple_stmt_iterator gsi;
4140 gimple_seq stmts = NULL;
4141 tree expr;
4142 pre_expr e;
4144 /* First create a value expression from the expression we want
4145 to insert and associate it with the value handle for SSA_VN. */
4146 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4147 if (e == NULL)
4148 return NULL_TREE;
4150 /* Then use create_expression_by_pieces to generate a valid
4151 expression to insert at this point of the IL stream. */
4152 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4153 if (expr == NULL_TREE)
4154 return NULL_TREE;
4155 gsi = gsi_for_stmt (stmt);
4156 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4158 return expr;
4161 /* Eliminate fully redundant computations. */
4163 static unsigned int
4164 eliminate (void)
4166 VEC (gimple, heap) *to_remove = NULL;
4167 basic_block b;
4168 unsigned int todo = 0;
4169 gimple_stmt_iterator gsi;
4170 gimple stmt;
4171 unsigned i;
4173 FOR_EACH_BB (b)
4175 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4177 stmt = gsi_stmt (gsi);
4179 /* Lookup the RHS of the expression, see if we have an
4180 available computation for it. If so, replace the RHS with
4181 the available computation. */
4182 if (gimple_has_lhs (stmt)
4183 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4184 && !gimple_assign_ssa_name_copy_p (stmt)
4185 && (!gimple_assign_single_p (stmt)
4186 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4187 && !gimple_has_volatile_ops (stmt)
4188 && !has_zero_uses (gimple_get_lhs (stmt)))
4190 tree lhs = gimple_get_lhs (stmt);
4191 tree rhs = NULL_TREE;
4192 tree sprime = NULL;
4193 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4194 pre_expr sprimeexpr;
4196 if (gimple_assign_single_p (stmt))
4197 rhs = gimple_assign_rhs1 (stmt);
4199 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4200 get_expr_value_id (lhsexpr),
4201 NULL);
4203 if (sprimeexpr)
4205 if (sprimeexpr->kind == CONSTANT)
4206 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4207 else if (sprimeexpr->kind == NAME)
4208 sprime = PRE_EXPR_NAME (sprimeexpr);
4209 else
4210 gcc_unreachable ();
4213 /* If there is no existing leader but SCCVN knows this
4214 value is constant, use that constant. */
4215 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4217 sprime = VN_INFO (lhs)->valnum;
4218 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4219 TREE_TYPE (sprime)))
4220 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4222 if (dump_file && (dump_flags & TDF_DETAILS))
4224 fprintf (dump_file, "Replaced ");
4225 print_gimple_expr (dump_file, stmt, 0, 0);
4226 fprintf (dump_file, " with ");
4227 print_generic_expr (dump_file, sprime, 0);
4228 fprintf (dump_file, " in ");
4229 print_gimple_stmt (dump_file, stmt, 0, 0);
4231 pre_stats.eliminations++;
4232 propagate_tree_value_into_stmt (&gsi, sprime);
4233 stmt = gsi_stmt (gsi);
4234 update_stmt (stmt);
4235 continue;
4238 /* If there is no existing usable leader but SCCVN thinks
4239 it has an expression it wants to use as replacement,
4240 insert that. */
4241 if (!sprime || sprime == lhs)
4243 tree val = VN_INFO (lhs)->valnum;
4244 if (val != VN_TOP
4245 && TREE_CODE (val) == SSA_NAME
4246 && VN_INFO (val)->needs_insertion
4247 && can_PRE_operation (vn_get_expr_for (val)))
4248 sprime = do_SCCVN_insertion (stmt, val);
4250 if (sprime
4251 && sprime != lhs
4252 && (rhs == NULL_TREE
4253 || TREE_CODE (rhs) != SSA_NAME
4254 || may_propagate_copy (rhs, sprime)))
4256 gcc_assert (sprime != rhs);
4258 if (dump_file && (dump_flags & TDF_DETAILS))
4260 fprintf (dump_file, "Replaced ");
4261 print_gimple_expr (dump_file, stmt, 0, 0);
4262 fprintf (dump_file, " with ");
4263 print_generic_expr (dump_file, sprime, 0);
4264 fprintf (dump_file, " in ");
4265 print_gimple_stmt (dump_file, stmt, 0, 0);
4268 if (TREE_CODE (sprime) == SSA_NAME)
4269 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4270 NECESSARY, true);
4271 /* We need to make sure the new and old types actually match,
4272 which may require adding a simple cast, which fold_convert
4273 will do for us. */
4274 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4275 && !useless_type_conversion_p (gimple_expr_type (stmt),
4276 TREE_TYPE (sprime)))
4277 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4279 pre_stats.eliminations++;
4280 propagate_tree_value_into_stmt (&gsi, sprime);
4281 stmt = gsi_stmt (gsi);
4282 update_stmt (stmt);
4284 /* If we removed EH side effects from the statement, clean
4285 its EH information. */
4286 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4288 bitmap_set_bit (need_eh_cleanup,
4289 gimple_bb (stmt)->index);
4290 if (dump_file && (dump_flags & TDF_DETAILS))
4291 fprintf (dump_file, " Removed EH side effects.\n");
4295 /* If the statement is a scalar store, see if the expression
4296 has the same value number as its rhs. If so, the store is
4297 dead. */
4298 else if (gimple_assign_single_p (stmt)
4299 && !is_gimple_reg (gimple_assign_lhs (stmt))
4300 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4301 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4303 tree rhs = gimple_assign_rhs1 (stmt);
4304 tree val;
4305 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4306 gimple_vuse (stmt), true, NULL);
4307 if (TREE_CODE (rhs) == SSA_NAME)
4308 rhs = VN_INFO (rhs)->valnum;
4309 if (val
4310 && operand_equal_p (val, rhs, 0))
4312 if (dump_file && (dump_flags & TDF_DETAILS))
4314 fprintf (dump_file, "Deleted redundant store ");
4315 print_gimple_stmt (dump_file, stmt, 0, 0);
4318 /* Queue stmt for removal. */
4319 VEC_safe_push (gimple, heap, to_remove, stmt);
4322 /* Visit COND_EXPRs and fold the comparison with the
4323 available value-numbers. */
4324 else if (gimple_code (stmt) == GIMPLE_COND)
4326 tree op0 = gimple_cond_lhs (stmt);
4327 tree op1 = gimple_cond_rhs (stmt);
4328 tree result;
4330 if (TREE_CODE (op0) == SSA_NAME)
4331 op0 = VN_INFO (op0)->valnum;
4332 if (TREE_CODE (op1) == SSA_NAME)
4333 op1 = VN_INFO (op1)->valnum;
4334 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4335 op0, op1);
4336 if (result && TREE_CODE (result) == INTEGER_CST)
4338 if (integer_zerop (result))
4339 gimple_cond_make_false (stmt);
4340 else
4341 gimple_cond_make_true (stmt);
4342 update_stmt (stmt);
4343 todo = TODO_cleanup_cfg;
4346 /* Visit indirect calls and turn them into direct calls if
4347 possible. */
4348 if (gimple_code (stmt) == GIMPLE_CALL
4349 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4351 tree fn = VN_INFO (gimple_call_fn (stmt))->valnum;
4352 if (TREE_CODE (fn) == ADDR_EXPR
4353 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
4355 if (dump_file && (dump_flags & TDF_DETAILS))
4357 fprintf (dump_file, "Replacing call target with ");
4358 print_generic_expr (dump_file, fn, 0);
4359 fprintf (dump_file, " in ");
4360 print_gimple_stmt (dump_file, stmt, 0, 0);
4363 gimple_call_set_fn (stmt, fn);
4364 update_stmt (stmt);
4365 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4367 bitmap_set_bit (need_eh_cleanup,
4368 gimple_bb (stmt)->index);
4369 if (dump_file && (dump_flags & TDF_DETAILS))
4370 fprintf (dump_file, " Removed EH side effects.\n");
4373 /* Changing an indirect call to a direct call may
4374 have exposed different semantics. This may
4375 require an SSA update. */
4376 todo |= TODO_update_ssa_only_virtuals;
4381 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4383 gimple stmt, phi = gsi_stmt (gsi);
4384 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4385 pre_expr sprimeexpr, resexpr;
4386 gimple_stmt_iterator gsi2;
4388 /* We want to perform redundant PHI elimination. Do so by
4389 replacing the PHI with a single copy if possible.
4390 Do not touch inserted, single-argument or virtual PHIs. */
4391 if (gimple_phi_num_args (phi) == 1
4392 || !is_gimple_reg (res))
4394 gsi_next (&gsi);
4395 continue;
4398 resexpr = get_or_alloc_expr_for_name (res);
4399 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4400 get_expr_value_id (resexpr), NULL);
4401 if (sprimeexpr)
4403 if (sprimeexpr->kind == CONSTANT)
4404 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4405 else if (sprimeexpr->kind == NAME)
4406 sprime = PRE_EXPR_NAME (sprimeexpr);
4407 else
4408 gcc_unreachable ();
4410 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4412 sprime = VN_INFO (res)->valnum;
4413 if (!useless_type_conversion_p (TREE_TYPE (res),
4414 TREE_TYPE (sprime)))
4415 sprime = fold_convert (TREE_TYPE (res), sprime);
4417 if (!sprime
4418 || sprime == res)
4420 gsi_next (&gsi);
4421 continue;
4424 if (dump_file && (dump_flags & TDF_DETAILS))
4426 fprintf (dump_file, "Replaced redundant PHI node defining ");
4427 print_generic_expr (dump_file, res, 0);
4428 fprintf (dump_file, " with ");
4429 print_generic_expr (dump_file, sprime, 0);
4430 fprintf (dump_file, "\n");
4433 remove_phi_node (&gsi, false);
4435 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4436 && TREE_CODE (sprime) == SSA_NAME)
4437 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4439 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4440 sprime = fold_convert (TREE_TYPE (res), sprime);
4441 stmt = gimple_build_assign (res, sprime);
4442 SSA_NAME_DEF_STMT (res) = stmt;
4443 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4445 gsi2 = gsi_after_labels (b);
4446 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4447 /* Queue the copy for eventual removal. */
4448 VEC_safe_push (gimple, heap, to_remove, stmt);
4449 /* If we inserted this PHI node ourself, it's not an elimination. */
4450 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4451 pre_stats.phis--;
4452 else
4453 pre_stats.eliminations++;
4457 /* We cannot remove stmts during BB walk, especially not release SSA
4458 names there as this confuses the VN machinery. The stmts ending
4459 up in to_remove are either stores or simple copies. */
4460 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4462 tree lhs = gimple_assign_lhs (stmt);
4463 tree rhs = gimple_assign_rhs1 (stmt);
4464 use_operand_p use_p;
4465 gimple use_stmt;
4467 /* If there is a single use only, propagate the equivalency
4468 instead of keeping the copy. */
4469 if (TREE_CODE (lhs) == SSA_NAME
4470 && TREE_CODE (rhs) == SSA_NAME
4471 && single_imm_use (lhs, &use_p, &use_stmt)
4472 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4474 SET_USE (use_p, rhs);
4475 update_stmt (use_stmt);
4476 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4477 && TREE_CODE (rhs) == SSA_NAME)
4478 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4481 /* If this is a store or a now unused copy, remove it. */
4482 if (TREE_CODE (lhs) != SSA_NAME
4483 || has_zero_uses (lhs))
4485 basic_block bb = gimple_bb (stmt);
4486 gsi = gsi_for_stmt (stmt);
4487 unlink_stmt_vdef (stmt);
4488 gsi_remove (&gsi, true);
4489 if (gimple_purge_dead_eh_edges (bb))
4490 todo |= TODO_cleanup_cfg;
4491 if (TREE_CODE (lhs) == SSA_NAME)
4492 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4493 release_defs (stmt);
4496 VEC_free (gimple, heap, to_remove);
4498 return todo;
4501 /* Borrow a bit of tree-ssa-dce.c for the moment.
4502 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4503 this may be a bit faster, and we may want critical edges kept split. */
4505 /* If OP's defining statement has not already been determined to be necessary,
4506 mark that statement necessary. Return the stmt, if it is newly
4507 necessary. */
4509 static inline gimple
4510 mark_operand_necessary (tree op)
4512 gimple stmt;
4514 gcc_assert (op);
4516 if (TREE_CODE (op) != SSA_NAME)
4517 return NULL;
4519 stmt = SSA_NAME_DEF_STMT (op);
4520 gcc_assert (stmt);
4522 if (gimple_plf (stmt, NECESSARY)
4523 || gimple_nop_p (stmt))
4524 return NULL;
4526 gimple_set_plf (stmt, NECESSARY, true);
4527 return stmt;
4530 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4531 to insert PHI nodes sometimes, and because value numbering of casts isn't
4532 perfect, we sometimes end up inserting dead code. This simple DCE-like
4533 pass removes any insertions we made that weren't actually used. */
4535 static void
4536 remove_dead_inserted_code (void)
4538 bitmap worklist;
4539 unsigned i;
4540 bitmap_iterator bi;
4541 gimple t;
4543 worklist = BITMAP_ALLOC (NULL);
4544 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4546 t = SSA_NAME_DEF_STMT (ssa_name (i));
4547 if (gimple_plf (t, NECESSARY))
4548 bitmap_set_bit (worklist, i);
4550 while (!bitmap_empty_p (worklist))
4552 i = bitmap_first_set_bit (worklist);
4553 bitmap_clear_bit (worklist, i);
4554 t = SSA_NAME_DEF_STMT (ssa_name (i));
4556 /* PHI nodes are somewhat special in that each PHI alternative has
4557 data and control dependencies. All the statements feeding the
4558 PHI node's arguments are always necessary. */
4559 if (gimple_code (t) == GIMPLE_PHI)
4561 unsigned k;
4563 for (k = 0; k < gimple_phi_num_args (t); k++)
4565 tree arg = PHI_ARG_DEF (t, k);
4566 if (TREE_CODE (arg) == SSA_NAME)
4568 gimple n = mark_operand_necessary (arg);
4569 if (n)
4570 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4574 else
4576 /* Propagate through the operands. Examine all the USE, VUSE and
4577 VDEF operands in this statement. Mark all the statements
4578 which feed this statement's uses as necessary. */
4579 ssa_op_iter iter;
4580 tree use;
4582 /* The operands of VDEF expressions are also needed as they
4583 represent potential definitions that may reach this
4584 statement (VDEF operands allow us to follow def-def
4585 links). */
4587 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4589 gimple n = mark_operand_necessary (use);
4590 if (n)
4591 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4596 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4598 t = SSA_NAME_DEF_STMT (ssa_name (i));
4599 if (!gimple_plf (t, NECESSARY))
4601 gimple_stmt_iterator gsi;
4603 if (dump_file && (dump_flags & TDF_DETAILS))
4605 fprintf (dump_file, "Removing unnecessary insertion:");
4606 print_gimple_stmt (dump_file, t, 0, 0);
4609 gsi = gsi_for_stmt (t);
4610 if (gimple_code (t) == GIMPLE_PHI)
4611 remove_phi_node (&gsi, true);
4612 else
4614 gsi_remove (&gsi, true);
4615 release_defs (t);
4619 BITMAP_FREE (worklist);
4622 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4623 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4624 the number of visited blocks. */
4626 static int
4627 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4629 edge_iterator *stack;
4630 int sp;
4631 int post_order_num = 0;
4632 sbitmap visited;
4634 if (include_entry_exit)
4635 post_order[post_order_num++] = EXIT_BLOCK;
4637 /* Allocate stack for back-tracking up CFG. */
4638 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4639 sp = 0;
4641 /* Allocate bitmap to track nodes that have been visited. */
4642 visited = sbitmap_alloc (last_basic_block);
4644 /* None of the nodes in the CFG have been visited yet. */
4645 sbitmap_zero (visited);
4647 /* Push the last edge on to the stack. */
4648 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4650 while (sp)
4652 edge_iterator ei;
4653 basic_block src;
4654 basic_block dest;
4656 /* Look at the edge on the top of the stack. */
4657 ei = stack[sp - 1];
4658 src = ei_edge (ei)->src;
4659 dest = ei_edge (ei)->dest;
4661 /* Check if the edge destination has been visited yet. */
4662 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4664 /* Mark that we have visited the destination. */
4665 SET_BIT (visited, src->index);
4667 if (EDGE_COUNT (src->preds) > 0)
4668 /* Since the DEST node has been visited for the first
4669 time, check its successors. */
4670 stack[sp++] = ei_start (src->preds);
4671 else
4672 post_order[post_order_num++] = src->index;
4674 else
4676 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4677 post_order[post_order_num++] = dest->index;
4679 if (!ei_one_before_end_p (ei))
4680 ei_next (&stack[sp - 1]);
4681 else
4682 sp--;
4686 if (include_entry_exit)
4687 post_order[post_order_num++] = ENTRY_BLOCK;
4689 free (stack);
4690 sbitmap_free (visited);
4691 return post_order_num;
4695 /* Initialize data structures used by PRE. */
4697 static void
4698 init_pre (bool do_fre)
4700 basic_block bb;
4702 next_expression_id = 1;
4703 expressions = NULL;
4704 VEC_safe_push (pre_expr, heap, expressions, NULL);
4705 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4706 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4707 get_max_value_id() + 1);
4708 name_to_id = NULL;
4710 in_fre = do_fre;
4712 inserted_exprs = BITMAP_ALLOC (NULL);
4713 need_creation = NULL;
4714 pretemp = NULL_TREE;
4715 storetemp = NULL_TREE;
4716 prephitemp = NULL_TREE;
4718 connect_infinite_loops_to_exit ();
4719 memset (&pre_stats, 0, sizeof (pre_stats));
4722 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4723 my_rev_post_order_compute (postorder, false);
4725 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4727 calculate_dominance_info (CDI_POST_DOMINATORS);
4728 calculate_dominance_info (CDI_DOMINATORS);
4730 bitmap_obstack_initialize (&grand_bitmap_obstack);
4731 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4732 expr_pred_trans_eq, free);
4733 expression_to_id = htab_create (num_ssa_names * 3,
4734 pre_expr_hash,
4735 pre_expr_eq, NULL);
4736 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4737 sizeof (struct bitmap_set), 30);
4738 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4739 sizeof (struct pre_expr_d), 30);
4740 FOR_ALL_BB (bb)
4742 EXP_GEN (bb) = bitmap_set_new ();
4743 PHI_GEN (bb) = bitmap_set_new ();
4744 TMP_GEN (bb) = bitmap_set_new ();
4745 AVAIL_OUT (bb) = bitmap_set_new ();
4748 need_eh_cleanup = BITMAP_ALLOC (NULL);
4752 /* Deallocate data structures used by PRE. */
4754 static void
4755 fini_pre (bool do_fre)
4757 free (postorder);
4758 VEC_free (bitmap_set_t, heap, value_expressions);
4759 BITMAP_FREE (inserted_exprs);
4760 VEC_free (gimple, heap, need_creation);
4761 bitmap_obstack_release (&grand_bitmap_obstack);
4762 free_alloc_pool (bitmap_set_pool);
4763 free_alloc_pool (pre_expr_pool);
4764 htab_delete (phi_translate_table);
4765 htab_delete (expression_to_id);
4766 VEC_free (unsigned, heap, name_to_id);
4768 free_aux_for_blocks ();
4770 free_dominance_info (CDI_POST_DOMINATORS);
4772 if (!bitmap_empty_p (need_eh_cleanup))
4774 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4775 cleanup_tree_cfg ();
4778 BITMAP_FREE (need_eh_cleanup);
4780 if (!do_fre)
4781 loop_optimizer_finalize ();
4784 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4785 only wants to do full redundancy elimination. */
4787 static unsigned int
4788 execute_pre (bool do_fre)
4790 unsigned int todo = 0;
4792 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4794 /* This has to happen before SCCVN runs because
4795 loop_optimizer_init may create new phis, etc. */
4796 if (!do_fre)
4797 loop_optimizer_init (LOOPS_NORMAL);
4799 if (!run_scc_vn ())
4801 if (!do_fre)
4802 loop_optimizer_finalize ();
4804 return 0;
4807 init_pre (do_fre);
4808 scev_initialize ();
4810 /* Collect and value number expressions computed in each basic block. */
4811 compute_avail ();
4813 if (dump_file && (dump_flags & TDF_DETAILS))
4815 basic_block bb;
4817 FOR_ALL_BB (bb)
4819 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4820 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4821 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4822 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4826 /* Insert can get quite slow on an incredibly large number of basic
4827 blocks due to some quadratic behavior. Until this behavior is
4828 fixed, don't run it when he have an incredibly large number of
4829 bb's. If we aren't going to run insert, there is no point in
4830 computing ANTIC, either, even though it's plenty fast. */
4831 if (!do_fre && n_basic_blocks < 4000)
4833 compute_antic ();
4834 insert ();
4837 /* Make sure to remove fake edges before committing our inserts.
4838 This makes sure we don't end up with extra critical edges that
4839 we would need to split. */
4840 remove_fake_exit_edges ();
4841 gsi_commit_edge_inserts ();
4843 /* Remove all the redundant expressions. */
4844 todo |= eliminate ();
4846 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4847 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4848 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4849 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4850 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4852 clear_expression_ids ();
4853 free_scc_vn ();
4854 if (!do_fre)
4855 remove_dead_inserted_code ();
4857 scev_finalize ();
4858 fini_pre (do_fre);
4860 return todo;
4863 /* Gate and execute functions for PRE. */
4865 static unsigned int
4866 do_pre (void)
4868 return execute_pre (false);
4871 static bool
4872 gate_pre (void)
4874 return flag_tree_pre != 0;
4877 struct gimple_opt_pass pass_pre =
4880 GIMPLE_PASS,
4881 "pre", /* name */
4882 gate_pre, /* gate */
4883 do_pre, /* execute */
4884 NULL, /* sub */
4885 NULL, /* next */
4886 0, /* static_pass_number */
4887 TV_TREE_PRE, /* tv_id */
4888 PROP_no_crit_edges | PROP_cfg
4889 | PROP_ssa, /* properties_required */
4890 0, /* properties_provided */
4891 0, /* properties_destroyed */
4892 TODO_rebuild_alias, /* todo_flags_start */
4893 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4894 | TODO_verify_ssa /* todo_flags_finish */
4899 /* Gate and execute functions for FRE. */
4901 static unsigned int
4902 execute_fre (void)
4904 return execute_pre (true);
4907 static bool
4908 gate_fre (void)
4910 return flag_tree_fre != 0;
4913 struct gimple_opt_pass pass_fre =
4916 GIMPLE_PASS,
4917 "fre", /* name */
4918 gate_fre, /* gate */
4919 execute_fre, /* execute */
4920 NULL, /* sub */
4921 NULL, /* next */
4922 0, /* static_pass_number */
4923 TV_TREE_FRE, /* tv_id */
4924 PROP_cfg | PROP_ssa, /* properties_required */
4925 0, /* properties_provided */
4926 0, /* properties_destroyed */
4927 0, /* todo_flags_start */
4928 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */