2010-07-27 Paolo Carlini <paolo.carlini@oracle.com>
[official-gcc/alias-decl.git] / gcc / tree-ssa-pre.c
blob8293e978d67b113326fc1c2cfdbd093a6b3b48f7
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_bit_p (&set->expressions, i))
860 bitmap_clear_bit (&set->expressions, i);
861 bitmap_set_bit (&set->expressions, get_expression_id (expr));
862 return;
867 /* Return true if two bitmap sets are equal. */
869 static bool
870 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
872 return bitmap_equal_p (&a->values, &b->values);
875 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
876 and add it otherwise. */
878 static void
879 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
881 unsigned int val = get_expr_value_id (expr);
883 if (bitmap_set_contains_value (set, val))
884 bitmap_set_replace_value (set, val, expr);
885 else
886 bitmap_insert_into_set (set, expr);
889 /* Insert EXPR into SET if EXPR's value is not already present in
890 SET. */
892 static void
893 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
895 unsigned int val = get_expr_value_id (expr);
897 #ifdef ENABLE_CHECKING
898 gcc_assert (expr->id == get_or_alloc_expression_id (expr));
899 #endif
901 /* Constant values are always considered to be part of the set. */
902 if (value_id_constant_p (val))
903 return;
905 /* If the value membership changed, add the expression. */
906 if (bitmap_set_bit (&set->values, val))
907 bitmap_set_bit (&set->expressions, expr->id);
910 /* Print out EXPR to outfile. */
912 static void
913 print_pre_expr (FILE *outfile, const pre_expr expr)
915 switch (expr->kind)
917 case CONSTANT:
918 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
919 break;
920 case NAME:
921 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
922 break;
923 case NARY:
925 unsigned int i;
926 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
927 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
928 for (i = 0; i < nary->length; i++)
930 print_generic_expr (outfile, nary->op[i], 0);
931 if (i != (unsigned) nary->length - 1)
932 fprintf (outfile, ",");
934 fprintf (outfile, "}");
936 break;
938 case REFERENCE:
940 vn_reference_op_t vro;
941 unsigned int i;
942 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
943 fprintf (outfile, "{");
944 for (i = 0;
945 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
946 i++)
948 bool closebrace = false;
949 if (vro->opcode != SSA_NAME
950 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
952 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
953 if (vro->op0)
955 fprintf (outfile, "<");
956 closebrace = true;
959 if (vro->op0)
961 print_generic_expr (outfile, vro->op0, 0);
962 if (vro->op1)
964 fprintf (outfile, ",");
965 print_generic_expr (outfile, vro->op1, 0);
967 if (vro->op2)
969 fprintf (outfile, ",");
970 print_generic_expr (outfile, vro->op2, 0);
973 if (closebrace)
974 fprintf (outfile, ">");
975 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
976 fprintf (outfile, ",");
978 fprintf (outfile, "}");
979 if (ref->vuse)
981 fprintf (outfile, "@");
982 print_generic_expr (outfile, ref->vuse, 0);
985 break;
988 void debug_pre_expr (pre_expr);
990 /* Like print_pre_expr but always prints to stderr. */
991 DEBUG_FUNCTION void
992 debug_pre_expr (pre_expr e)
994 print_pre_expr (stderr, e);
995 fprintf (stderr, "\n");
998 /* Print out SET to OUTFILE. */
1000 static void
1001 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1002 const char *setname, int blockindex)
1004 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1005 if (set)
1007 bool first = true;
1008 unsigned i;
1009 bitmap_iterator bi;
1011 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1013 const pre_expr expr = expression_for_id (i);
1015 if (!first)
1016 fprintf (outfile, ", ");
1017 first = false;
1018 print_pre_expr (outfile, expr);
1020 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1023 fprintf (outfile, " }\n");
1026 void debug_bitmap_set (bitmap_set_t);
1028 DEBUG_FUNCTION void
1029 debug_bitmap_set (bitmap_set_t set)
1031 print_bitmap_set (stderr, set, "debug", 0);
1034 /* Print out the expressions that have VAL to OUTFILE. */
1036 void
1037 print_value_expressions (FILE *outfile, unsigned int val)
1039 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1040 if (set)
1042 char s[10];
1043 sprintf (s, "%04d", val);
1044 print_bitmap_set (outfile, set, s, 0);
1049 DEBUG_FUNCTION void
1050 debug_value_expressions (unsigned int val)
1052 print_value_expressions (stderr, val);
1055 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1056 represent it. */
1058 static pre_expr
1059 get_or_alloc_expr_for_constant (tree constant)
1061 unsigned int result_id;
1062 unsigned int value_id;
1063 struct pre_expr_d expr;
1064 pre_expr newexpr;
1066 expr.kind = CONSTANT;
1067 PRE_EXPR_CONSTANT (&expr) = constant;
1068 result_id = lookup_expression_id (&expr);
1069 if (result_id != 0)
1070 return expression_for_id (result_id);
1072 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1073 newexpr->kind = CONSTANT;
1074 PRE_EXPR_CONSTANT (newexpr) = constant;
1075 alloc_expression_id (newexpr);
1076 value_id = get_or_alloc_constant_value_id (constant);
1077 add_to_value (value_id, newexpr);
1078 return newexpr;
1081 /* Given a value id V, find the actual tree representing the constant
1082 value if there is one, and return it. Return NULL if we can't find
1083 a constant. */
1085 static tree
1086 get_constant_for_value_id (unsigned int v)
1088 if (value_id_constant_p (v))
1090 unsigned int i;
1091 bitmap_iterator bi;
1092 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1094 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1096 pre_expr expr = expression_for_id (i);
1097 if (expr->kind == CONSTANT)
1098 return PRE_EXPR_CONSTANT (expr);
1101 return NULL;
1104 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1105 Currently only supports constants and SSA_NAMES. */
1106 static pre_expr
1107 get_or_alloc_expr_for (tree t)
1109 if (TREE_CODE (t) == SSA_NAME)
1110 return get_or_alloc_expr_for_name (t);
1111 else if (is_gimple_min_invariant (t))
1112 return get_or_alloc_expr_for_constant (t);
1113 else
1115 /* More complex expressions can result from SCCVN expression
1116 simplification that inserts values for them. As they all
1117 do not have VOPs the get handled by the nary ops struct. */
1118 vn_nary_op_t result;
1119 unsigned int result_id;
1120 vn_nary_op_lookup (t, &result);
1121 if (result != NULL)
1123 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1124 e->kind = NARY;
1125 PRE_EXPR_NARY (e) = result;
1126 result_id = lookup_expression_id (e);
1127 if (result_id != 0)
1129 pool_free (pre_expr_pool, e);
1130 e = expression_for_id (result_id);
1131 return e;
1133 alloc_expression_id (e);
1134 return e;
1137 return NULL;
1140 /* Return the folded version of T if T, when folded, is a gimple
1141 min_invariant. Otherwise, return T. */
1143 static pre_expr
1144 fully_constant_expression (pre_expr e)
1146 switch (e->kind)
1148 case CONSTANT:
1149 return e;
1150 case NARY:
1152 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1153 switch (TREE_CODE_CLASS (nary->opcode))
1155 case tcc_expression:
1156 if (nary->opcode == TRUTH_NOT_EXPR)
1157 goto do_unary;
1158 if (nary->opcode != TRUTH_AND_EXPR
1159 && nary->opcode != TRUTH_OR_EXPR
1160 && nary->opcode != TRUTH_XOR_EXPR)
1161 return e;
1162 /* Fallthrough. */
1163 case tcc_binary:
1164 case tcc_comparison:
1166 /* We have to go from trees to pre exprs to value ids to
1167 constants. */
1168 tree naryop0 = nary->op[0];
1169 tree naryop1 = nary->op[1];
1170 tree result;
1171 if (!is_gimple_min_invariant (naryop0))
1173 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1174 unsigned int vrep0 = get_expr_value_id (rep0);
1175 tree const0 = get_constant_for_value_id (vrep0);
1176 if (const0)
1177 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1179 if (!is_gimple_min_invariant (naryop1))
1181 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1182 unsigned int vrep1 = get_expr_value_id (rep1);
1183 tree const1 = get_constant_for_value_id (vrep1);
1184 if (const1)
1185 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1187 result = fold_binary (nary->opcode, nary->type,
1188 naryop0, naryop1);
1189 if (result && is_gimple_min_invariant (result))
1190 return get_or_alloc_expr_for_constant (result);
1191 /* We might have simplified the expression to a
1192 SSA_NAME for example from x_1 * 1. But we cannot
1193 insert a PHI for x_1 unconditionally as x_1 might
1194 not be available readily. */
1195 return e;
1197 case tcc_reference:
1198 if (nary->opcode != REALPART_EXPR
1199 && nary->opcode != IMAGPART_EXPR
1200 && nary->opcode != VIEW_CONVERT_EXPR)
1201 return e;
1202 /* Fallthrough. */
1203 case tcc_unary:
1204 do_unary:
1206 /* We have to go from trees to pre exprs to value ids to
1207 constants. */
1208 tree naryop0 = nary->op[0];
1209 tree const0, result;
1210 if (is_gimple_min_invariant (naryop0))
1211 const0 = naryop0;
1212 else
1214 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1215 unsigned int vrep0 = get_expr_value_id (rep0);
1216 const0 = get_constant_for_value_id (vrep0);
1218 result = NULL;
1219 if (const0)
1221 tree type1 = TREE_TYPE (nary->op[0]);
1222 const0 = fold_convert (type1, const0);
1223 result = fold_unary (nary->opcode, nary->type, const0);
1225 if (result && is_gimple_min_invariant (result))
1226 return get_or_alloc_expr_for_constant (result);
1227 return e;
1229 default:
1230 return e;
1233 case REFERENCE:
1235 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1236 tree folded;
1237 if ((folded = fully_constant_vn_reference_p (ref)))
1238 return get_or_alloc_expr_for_constant (folded);
1239 return e;
1241 default:
1242 return e;
1244 return e;
1247 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1248 it has the value it would have in BLOCK. Set *SAME_VALID to true
1249 in case the new vuse doesn't change the value id of the OPERANDS. */
1251 static tree
1252 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1253 alias_set_type set, tree type, tree vuse,
1254 basic_block phiblock,
1255 basic_block block, bool *same_valid)
1257 gimple phi = SSA_NAME_DEF_STMT (vuse);
1258 ao_ref ref;
1259 edge e = NULL;
1260 bool use_oracle;
1262 *same_valid = true;
1264 if (gimple_bb (phi) != phiblock)
1265 return vuse;
1267 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1269 /* Use the alias-oracle to find either the PHI node in this block,
1270 the first VUSE used in this block that is equivalent to vuse or
1271 the first VUSE which definition in this block kills the value. */
1272 if (gimple_code (phi) == GIMPLE_PHI)
1273 e = find_edge (block, phiblock);
1274 else if (use_oracle)
1275 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1277 vuse = gimple_vuse (phi);
1278 phi = SSA_NAME_DEF_STMT (vuse);
1279 if (gimple_bb (phi) != phiblock)
1280 return vuse;
1281 if (gimple_code (phi) == GIMPLE_PHI)
1283 e = find_edge (block, phiblock);
1284 break;
1287 else
1288 return NULL_TREE;
1290 if (e)
1292 if (use_oracle)
1294 bitmap visited = NULL;
1295 /* Try to find a vuse that dominates this phi node by skipping
1296 non-clobbering statements. */
1297 vuse = get_continuation_for_phi (phi, &ref, &visited);
1298 if (visited)
1299 BITMAP_FREE (visited);
1301 else
1302 vuse = NULL_TREE;
1303 if (!vuse)
1305 /* If we didn't find any, the value ID can't stay the same,
1306 but return the translated vuse. */
1307 *same_valid = false;
1308 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1310 /* ??? We would like to return vuse here as this is the canonical
1311 upmost vdef that this reference is associated with. But during
1312 insertion of the references into the hash tables we only ever
1313 directly insert with their direct gimple_vuse, hence returning
1314 something else would make us not find the other expression. */
1315 return PHI_ARG_DEF (phi, e->dest_idx);
1318 return NULL_TREE;
1321 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1322 SET2. This is used to avoid making a set consisting of the union
1323 of PA_IN and ANTIC_IN during insert. */
1325 static inline pre_expr
1326 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1328 pre_expr result;
1330 result = bitmap_find_leader (set1, val, NULL);
1331 if (!result && set2)
1332 result = bitmap_find_leader (set2, val, NULL);
1333 return result;
1336 /* Get the tree type for our PRE expression e. */
1338 static tree
1339 get_expr_type (const pre_expr e)
1341 switch (e->kind)
1343 case NAME:
1344 return TREE_TYPE (PRE_EXPR_NAME (e));
1345 case CONSTANT:
1346 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1347 case REFERENCE:
1348 return PRE_EXPR_REFERENCE (e)->type;
1349 case NARY:
1350 return PRE_EXPR_NARY (e)->type;
1352 gcc_unreachable();
1355 /* Get a representative SSA_NAME for a given expression.
1356 Since all of our sub-expressions are treated as values, we require
1357 them to be SSA_NAME's for simplicity.
1358 Prior versions of GVNPRE used to use "value handles" here, so that
1359 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1360 either case, the operands are really values (IE we do not expect
1361 them to be usable without finding leaders). */
1363 static tree
1364 get_representative_for (const pre_expr e)
1366 tree exprtype;
1367 tree name;
1368 unsigned int value_id = get_expr_value_id (e);
1370 switch (e->kind)
1372 case NAME:
1373 return PRE_EXPR_NAME (e);
1374 case CONSTANT:
1375 return PRE_EXPR_CONSTANT (e);
1376 case NARY:
1377 case REFERENCE:
1379 /* Go through all of the expressions representing this value
1380 and pick out an SSA_NAME. */
1381 unsigned int i;
1382 bitmap_iterator bi;
1383 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1384 value_id);
1385 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1387 pre_expr rep = expression_for_id (i);
1388 if (rep->kind == NAME)
1389 return PRE_EXPR_NAME (rep);
1392 break;
1394 /* If we reached here we couldn't find an SSA_NAME. This can
1395 happen when we've discovered a value that has never appeared in
1396 the program as set to an SSA_NAME, most likely as the result of
1397 phi translation. */
1398 if (dump_file)
1400 fprintf (dump_file,
1401 "Could not find SSA_NAME representative for expression:");
1402 print_pre_expr (dump_file, e);
1403 fprintf (dump_file, "\n");
1406 exprtype = get_expr_type (e);
1408 /* Build and insert the assignment of the end result to the temporary
1409 that we will return. */
1410 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1412 pretemp = create_tmp_reg (exprtype, "pretmp");
1413 get_var_ann (pretemp);
1416 name = make_ssa_name (pretemp, gimple_build_nop ());
1417 VN_INFO_GET (name)->value_id = value_id;
1418 if (e->kind == CONSTANT)
1419 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1420 else
1421 VN_INFO (name)->valnum = name;
1423 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1424 if (dump_file)
1426 fprintf (dump_file, "Created SSA_NAME representative ");
1427 print_generic_expr (dump_file, name, 0);
1428 fprintf (dump_file, " for expression:");
1429 print_pre_expr (dump_file, e);
1430 fprintf (dump_file, "\n");
1433 return name;
1438 static pre_expr
1439 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1440 basic_block pred, basic_block phiblock);
1442 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1443 the phis in PRED. Return NULL if we can't find a leader for each part
1444 of the translated expression. */
1446 static pre_expr
1447 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1448 basic_block pred, basic_block phiblock)
1450 switch (expr->kind)
1452 case NARY:
1454 unsigned int i;
1455 bool changed = false;
1456 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1457 struct vn_nary_op_s newnary;
1458 /* The NARY structure is only guaranteed to have been
1459 allocated to the nary->length operands. */
1460 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1461 - sizeof (tree) * (4 - nary->length)));
1463 for (i = 0; i < newnary.length; i++)
1465 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1466 continue;
1467 else
1469 pre_expr leader, result;
1470 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1471 leader = find_leader_in_sets (op_val_id, set1, set2);
1472 result = phi_translate (leader, set1, set2, pred, phiblock);
1473 if (result && result != leader)
1475 tree name = get_representative_for (result);
1476 if (!name)
1477 return NULL;
1478 newnary.op[i] = name;
1480 else if (!result)
1481 return NULL;
1483 changed |= newnary.op[i] != nary->op[i];
1486 if (changed)
1488 pre_expr constant;
1489 unsigned int new_val_id;
1491 tree result = vn_nary_op_lookup_pieces (newnary.length,
1492 newnary.opcode,
1493 newnary.type,
1494 newnary.op[0],
1495 newnary.op[1],
1496 newnary.op[2],
1497 newnary.op[3],
1498 &nary);
1499 if (result && is_gimple_min_invariant (result))
1500 return get_or_alloc_expr_for_constant (result);
1502 expr = (pre_expr) pool_alloc (pre_expr_pool);
1503 expr->kind = NARY;
1504 expr->id = 0;
1505 if (nary)
1507 PRE_EXPR_NARY (expr) = nary;
1508 constant = fully_constant_expression (expr);
1509 if (constant != expr)
1510 return constant;
1512 new_val_id = nary->value_id;
1513 get_or_alloc_expression_id (expr);
1515 else
1517 new_val_id = get_next_value_id ();
1518 VEC_safe_grow_cleared (bitmap_set_t, heap,
1519 value_expressions,
1520 get_max_value_id() + 1);
1521 nary = vn_nary_op_insert_pieces (newnary.length,
1522 newnary.opcode,
1523 newnary.type,
1524 newnary.op[0],
1525 newnary.op[1],
1526 newnary.op[2],
1527 newnary.op[3],
1528 result, new_val_id);
1529 PRE_EXPR_NARY (expr) = nary;
1530 constant = fully_constant_expression (expr);
1531 if (constant != expr)
1532 return constant;
1533 get_or_alloc_expression_id (expr);
1535 add_to_value (new_val_id, expr);
1537 return expr;
1539 break;
1541 case REFERENCE:
1543 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1544 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1545 tree vuse = ref->vuse;
1546 tree newvuse = vuse;
1547 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1548 bool changed = false, same_valid = true;
1549 unsigned int i, j;
1550 vn_reference_op_t operand;
1551 vn_reference_t newref;
1553 for (i = 0, j = 0;
1554 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1556 pre_expr opresult;
1557 pre_expr leader;
1558 tree oldop0 = operand->op0;
1559 tree oldop1 = operand->op1;
1560 tree oldop2 = operand->op2;
1561 tree op0 = oldop0;
1562 tree op1 = oldop1;
1563 tree op2 = oldop2;
1564 tree type = operand->type;
1565 vn_reference_op_s newop = *operand;
1567 if (op0 && TREE_CODE (op0) == SSA_NAME)
1569 unsigned int op_val_id = VN_INFO (op0)->value_id;
1570 leader = find_leader_in_sets (op_val_id, set1, set2);
1571 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1572 if (opresult && opresult != leader)
1574 tree name = get_representative_for (opresult);
1575 if (!name)
1576 break;
1577 op0 = name;
1579 else if (!opresult)
1580 break;
1582 changed |= op0 != oldop0;
1584 if (op1 && TREE_CODE (op1) == SSA_NAME)
1586 unsigned int op_val_id = VN_INFO (op1)->value_id;
1587 leader = find_leader_in_sets (op_val_id, set1, set2);
1588 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1589 if (opresult && opresult != leader)
1591 tree name = get_representative_for (opresult);
1592 if (!name)
1593 break;
1594 op1 = name;
1596 else if (!opresult)
1597 break;
1599 /* We can't possibly insert these. */
1600 else if (op1 && !is_gimple_min_invariant (op1))
1601 break;
1602 changed |= op1 != oldop1;
1603 if (op2 && TREE_CODE (op2) == SSA_NAME)
1605 unsigned int op_val_id = VN_INFO (op2)->value_id;
1606 leader = find_leader_in_sets (op_val_id, set1, set2);
1607 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1608 if (opresult && opresult != leader)
1610 tree name = get_representative_for (opresult);
1611 if (!name)
1612 break;
1613 op2 = name;
1615 else if (!opresult)
1616 break;
1618 /* We can't possibly insert these. */
1619 else if (op2 && !is_gimple_min_invariant (op2))
1620 break;
1621 changed |= op2 != oldop2;
1623 if (!newoperands)
1624 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1625 /* We may have changed from an SSA_NAME to a constant */
1626 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1627 newop.opcode = TREE_CODE (op0);
1628 newop.type = type;
1629 newop.op0 = op0;
1630 newop.op1 = op1;
1631 newop.op2 = op2;
1632 /* If it transforms a non-constant ARRAY_REF into a constant
1633 one, adjust the constant offset. */
1634 if (newop.opcode == ARRAY_REF
1635 && newop.off == -1
1636 && TREE_CODE (op0) == INTEGER_CST
1637 && TREE_CODE (op1) == INTEGER_CST
1638 && TREE_CODE (op2) == INTEGER_CST)
1640 double_int off = tree_to_double_int (op0);
1641 off = double_int_add (off,
1642 double_int_neg
1643 (tree_to_double_int (op1)));
1644 off = double_int_mul (off, tree_to_double_int (op2));
1645 if (double_int_fits_in_shwi_p (off))
1646 newop.off = off.low;
1648 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1649 /* If it transforms from an SSA_NAME to an address, fold with
1650 a preceding indirect reference. */
1651 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1652 && VEC_index (vn_reference_op_s,
1653 newoperands, j - 1)->opcode == MEM_REF)
1654 vn_reference_fold_indirect (&newoperands, &j);
1656 if (i != VEC_length (vn_reference_op_s, operands))
1658 if (newoperands)
1659 VEC_free (vn_reference_op_s, heap, newoperands);
1660 return NULL;
1663 if (vuse)
1665 newvuse = translate_vuse_through_block (newoperands,
1666 ref->set, ref->type,
1667 vuse, phiblock, pred,
1668 &same_valid);
1669 if (newvuse == NULL_TREE)
1671 VEC_free (vn_reference_op_s, heap, newoperands);
1672 return NULL;
1676 if (changed || newvuse != vuse)
1678 unsigned int new_val_id;
1679 pre_expr constant;
1680 bool converted = false;
1682 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1683 ref->type,
1684 newoperands,
1685 &newref, true);
1686 if (result)
1687 VEC_free (vn_reference_op_s, heap, newoperands);
1689 if (result
1690 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1692 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1693 converted = true;
1696 if (result && is_gimple_min_invariant (result))
1698 gcc_assert (!newoperands);
1699 return get_or_alloc_expr_for_constant (result);
1702 expr = (pre_expr) pool_alloc (pre_expr_pool);
1703 expr->kind = REFERENCE;
1704 expr->id = 0;
1706 if (converted)
1708 vn_nary_op_t nary;
1709 tree nresult;
1711 gcc_assert (CONVERT_EXPR_P (result)
1712 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1714 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1715 TREE_TYPE (result),
1716 TREE_OPERAND (result, 0),
1717 NULL_TREE, NULL_TREE,
1718 NULL_TREE,
1719 &nary);
1720 if (nresult && is_gimple_min_invariant (nresult))
1721 return get_or_alloc_expr_for_constant (nresult);
1723 expr->kind = NARY;
1724 if (nary)
1726 PRE_EXPR_NARY (expr) = nary;
1727 constant = fully_constant_expression (expr);
1728 if (constant != expr)
1729 return constant;
1731 new_val_id = nary->value_id;
1732 get_or_alloc_expression_id (expr);
1734 else
1736 new_val_id = get_next_value_id ();
1737 VEC_safe_grow_cleared (bitmap_set_t, heap,
1738 value_expressions,
1739 get_max_value_id() + 1);
1740 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1741 TREE_TYPE (result),
1742 TREE_OPERAND (result, 0),
1743 NULL_TREE, NULL_TREE,
1744 NULL_TREE, NULL_TREE,
1745 new_val_id);
1746 PRE_EXPR_NARY (expr) = nary;
1747 constant = fully_constant_expression (expr);
1748 if (constant != expr)
1749 return constant;
1750 get_or_alloc_expression_id (expr);
1753 else if (newref)
1755 PRE_EXPR_REFERENCE (expr) = newref;
1756 constant = fully_constant_expression (expr);
1757 if (constant != expr)
1758 return constant;
1760 new_val_id = newref->value_id;
1761 get_or_alloc_expression_id (expr);
1763 else
1765 if (changed || !same_valid)
1767 new_val_id = get_next_value_id ();
1768 VEC_safe_grow_cleared (bitmap_set_t, heap,
1769 value_expressions,
1770 get_max_value_id() + 1);
1772 else
1773 new_val_id = ref->value_id;
1774 newref = vn_reference_insert_pieces (newvuse, ref->set,
1775 ref->type,
1776 newoperands,
1777 result, new_val_id);
1778 newoperands = NULL;
1779 PRE_EXPR_REFERENCE (expr) = newref;
1780 constant = fully_constant_expression (expr);
1781 if (constant != expr)
1782 return constant;
1783 get_or_alloc_expression_id (expr);
1785 add_to_value (new_val_id, expr);
1787 VEC_free (vn_reference_op_s, heap, newoperands);
1788 return expr;
1790 break;
1792 case NAME:
1794 gimple phi = NULL;
1795 edge e;
1796 gimple def_stmt;
1797 tree name = PRE_EXPR_NAME (expr);
1799 def_stmt = SSA_NAME_DEF_STMT (name);
1800 if (gimple_code (def_stmt) == GIMPLE_PHI
1801 && gimple_bb (def_stmt) == phiblock)
1802 phi = def_stmt;
1803 else
1804 return expr;
1806 e = find_edge (pred, gimple_bb (phi));
1807 if (e)
1809 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1810 pre_expr newexpr;
1812 if (TREE_CODE (def) == SSA_NAME)
1813 def = VN_INFO (def)->valnum;
1815 /* Handle constant. */
1816 if (is_gimple_min_invariant (def))
1817 return get_or_alloc_expr_for_constant (def);
1819 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1820 return NULL;
1822 newexpr = get_or_alloc_expr_for_name (def);
1823 return newexpr;
1826 return expr;
1828 default:
1829 gcc_unreachable ();
1833 /* Wrapper around phi_translate_1 providing caching functionality. */
1835 static pre_expr
1836 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1837 basic_block pred, basic_block phiblock)
1839 pre_expr phitrans;
1841 if (!expr)
1842 return NULL;
1844 /* Constants contain no values that need translation. */
1845 if (expr->kind == CONSTANT)
1846 return expr;
1848 if (value_id_constant_p (get_expr_value_id (expr)))
1849 return expr;
1851 if (expr->kind != NAME)
1853 phitrans = phi_trans_lookup (expr, pred);
1854 if (phitrans)
1855 return phitrans;
1858 /* Translate. */
1859 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1861 /* Don't add empty translations to the cache. Neither add
1862 translations of NAMEs as those are cheap to translate. */
1863 if (phitrans
1864 && expr->kind != NAME)
1865 phi_trans_add (expr, phitrans, pred);
1867 return phitrans;
1871 /* For each expression in SET, translate the values through phi nodes
1872 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1873 expressions in DEST. */
1875 static void
1876 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1877 basic_block phiblock)
1879 VEC (pre_expr, heap) *exprs;
1880 pre_expr expr;
1881 int i;
1883 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1885 bitmap_set_copy (dest, set);
1886 return;
1889 exprs = sorted_array_from_bitmap_set (set);
1890 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1892 pre_expr translated;
1893 translated = phi_translate (expr, set, NULL, pred, phiblock);
1894 if (!translated)
1895 continue;
1897 /* We might end up with multiple expressions from SET being
1898 translated to the same value. In this case we do not want
1899 to retain the NARY or REFERENCE expression but prefer a NAME
1900 which would be the leader. */
1901 if (translated->kind == NAME)
1902 bitmap_value_replace_in_set (dest, translated);
1903 else
1904 bitmap_value_insert_into_set (dest, translated);
1906 VEC_free (pre_expr, heap, exprs);
1909 /* Find the leader for a value (i.e., the name representing that
1910 value) in a given set, and return it. If STMT is non-NULL it
1911 makes sure the defining statement for the leader dominates it.
1912 Return NULL if no leader is found. */
1914 static pre_expr
1915 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1917 if (value_id_constant_p (val))
1919 unsigned int i;
1920 bitmap_iterator bi;
1921 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1923 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1925 pre_expr expr = expression_for_id (i);
1926 if (expr->kind == CONSTANT)
1927 return expr;
1930 if (bitmap_set_contains_value (set, val))
1932 /* Rather than walk the entire bitmap of expressions, and see
1933 whether any of them has the value we are looking for, we look
1934 at the reverse mapping, which tells us the set of expressions
1935 that have a given value (IE value->expressions with that
1936 value) and see if any of those expressions are in our set.
1937 The number of expressions per value is usually significantly
1938 less than the number of expressions in the set. In fact, for
1939 large testcases, doing it this way is roughly 5-10x faster
1940 than walking the bitmap.
1941 If this is somehow a significant lose for some cases, we can
1942 choose which set to walk based on which set is smaller. */
1943 unsigned int i;
1944 bitmap_iterator bi;
1945 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1947 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1948 &set->expressions, 0, i, bi)
1950 pre_expr val = expression_for_id (i);
1951 /* At the point where stmt is not null, there should always
1952 be an SSA_NAME first in the list of expressions. */
1953 if (stmt)
1955 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1956 if (gimple_code (def_stmt) != GIMPLE_PHI
1957 && gimple_bb (def_stmt) == gimple_bb (stmt)
1958 /* PRE insertions are at the end of the basic-block
1959 and have UID 0. */
1960 && (gimple_uid (def_stmt) == 0
1961 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1962 continue;
1964 return val;
1967 return NULL;
1970 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1971 BLOCK by seeing if it is not killed in the block. Note that we are
1972 only determining whether there is a store that kills it. Because
1973 of the order in which clean iterates over values, we are guaranteed
1974 that altered operands will have caused us to be eliminated from the
1975 ANTIC_IN set already. */
1977 static bool
1978 value_dies_in_block_x (pre_expr expr, basic_block block)
1980 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1981 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1982 gimple def;
1983 gimple_stmt_iterator gsi;
1984 unsigned id = get_expression_id (expr);
1985 bool res = false;
1986 ao_ref ref;
1988 if (!vuse)
1989 return false;
1991 /* Lookup a previously calculated result. */
1992 if (EXPR_DIES (block)
1993 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1994 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1996 /* A memory expression {e, VUSE} dies in the block if there is a
1997 statement that may clobber e. If, starting statement walk from the
1998 top of the basic block, a statement uses VUSE there can be no kill
1999 inbetween that use and the original statement that loaded {e, VUSE},
2000 so we can stop walking. */
2001 ref.base = NULL_TREE;
2002 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
2004 tree def_vuse, def_vdef;
2005 def = gsi_stmt (gsi);
2006 def_vuse = gimple_vuse (def);
2007 def_vdef = gimple_vdef (def);
2009 /* Not a memory statement. */
2010 if (!def_vuse)
2011 continue;
2013 /* Not a may-def. */
2014 if (!def_vdef)
2016 /* A load with the same VUSE, we're done. */
2017 if (def_vuse == vuse)
2018 break;
2020 continue;
2023 /* Init ref only if we really need it. */
2024 if (ref.base == NULL_TREE
2025 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
2026 refx->operands))
2028 res = true;
2029 break;
2031 /* If the statement may clobber expr, it dies. */
2032 if (stmt_may_clobber_ref_p_1 (def, &ref))
2034 res = true;
2035 break;
2039 /* Remember the result. */
2040 if (!EXPR_DIES (block))
2041 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2042 bitmap_set_bit (EXPR_DIES (block), id * 2);
2043 if (res)
2044 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2046 return res;
2050 #define union_contains_value(SET1, SET2, VAL) \
2051 (bitmap_set_contains_value ((SET1), (VAL)) \
2052 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
2054 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2056 static bool
2057 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2058 vn_reference_op_t vro)
2060 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2062 struct pre_expr_d temp;
2063 temp.kind = NAME;
2064 temp.id = 0;
2065 PRE_EXPR_NAME (&temp) = vro->op0;
2066 temp.id = lookup_expression_id (&temp);
2067 if (temp.id == 0)
2068 return false;
2069 if (!union_contains_value (set1, set2,
2070 get_expr_value_id (&temp)))
2071 return false;
2073 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2075 struct pre_expr_d temp;
2076 temp.kind = NAME;
2077 temp.id = 0;
2078 PRE_EXPR_NAME (&temp) = vro->op1;
2079 temp.id = lookup_expression_id (&temp);
2080 if (temp.id == 0)
2081 return false;
2082 if (!union_contains_value (set1, set2,
2083 get_expr_value_id (&temp)))
2084 return false;
2087 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2089 struct pre_expr_d temp;
2090 temp.kind = NAME;
2091 temp.id = 0;
2092 PRE_EXPR_NAME (&temp) = vro->op2;
2093 temp.id = lookup_expression_id (&temp);
2094 if (temp.id == 0)
2095 return false;
2096 if (!union_contains_value (set1, set2,
2097 get_expr_value_id (&temp)))
2098 return false;
2101 return true;
2104 /* Determine if the expression EXPR is valid in SET1 U SET2.
2105 ONLY SET2 CAN BE NULL.
2106 This means that we have a leader for each part of the expression
2107 (if it consists of values), or the expression is an SSA_NAME.
2108 For loads/calls, we also see if the vuse is killed in this block. */
2110 static bool
2111 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2112 basic_block block)
2114 switch (expr->kind)
2116 case NAME:
2117 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2118 case NARY:
2120 unsigned int i;
2121 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2122 for (i = 0; i < nary->length; i++)
2124 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2126 struct pre_expr_d temp;
2127 temp.kind = NAME;
2128 temp.id = 0;
2129 PRE_EXPR_NAME (&temp) = nary->op[i];
2130 temp.id = lookup_expression_id (&temp);
2131 if (temp.id == 0)
2132 return false;
2133 if (!union_contains_value (set1, set2,
2134 get_expr_value_id (&temp)))
2135 return false;
2138 /* If the NARY may trap make sure the block does not contain
2139 a possible exit point.
2140 ??? This is overly conservative if we translate AVAIL_OUT
2141 as the available expression might be after the exit point. */
2142 if (BB_MAY_NOTRETURN (block)
2143 && vn_nary_may_trap (nary))
2144 return false;
2145 return true;
2147 break;
2148 case REFERENCE:
2150 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2151 vn_reference_op_t vro;
2152 unsigned int i;
2154 for (i = 0; VEC_iterate (vn_reference_op_s, ref->operands, i, vro); i++)
2156 if (!vro_valid_in_sets (set1, set2, vro))
2157 return false;
2159 if (ref->vuse)
2161 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2162 if (!gimple_nop_p (def_stmt)
2163 && gimple_bb (def_stmt) != block
2164 && !dominated_by_p (CDI_DOMINATORS,
2165 block, gimple_bb (def_stmt)))
2166 return false;
2168 return !value_dies_in_block_x (expr, block);
2170 default:
2171 gcc_unreachable ();
2175 /* Clean the set of expressions that are no longer valid in SET1 or
2176 SET2. This means expressions that are made up of values we have no
2177 leaders for in SET1 or SET2. This version is used for partial
2178 anticipation, which means it is not valid in either ANTIC_IN or
2179 PA_IN. */
2181 static void
2182 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2184 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2185 pre_expr expr;
2186 int i;
2188 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2190 if (!valid_in_sets (set1, set2, expr, block))
2191 bitmap_remove_from_set (set1, expr);
2193 VEC_free (pre_expr, heap, exprs);
2196 /* Clean the set of expressions that are no longer valid in SET. This
2197 means expressions that are made up of values we have no leaders for
2198 in SET. */
2200 static void
2201 clean (bitmap_set_t set, basic_block block)
2203 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2204 pre_expr expr;
2205 int i;
2207 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2209 if (!valid_in_sets (set, NULL, expr, block))
2210 bitmap_remove_from_set (set, expr);
2212 VEC_free (pre_expr, heap, exprs);
2215 static sbitmap has_abnormal_preds;
2217 /* List of blocks that may have changed during ANTIC computation and
2218 thus need to be iterated over. */
2220 static sbitmap changed_blocks;
2222 /* Decide whether to defer a block for a later iteration, or PHI
2223 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2224 should defer the block, and true if we processed it. */
2226 static bool
2227 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2228 basic_block block, basic_block phiblock)
2230 if (!BB_VISITED (phiblock))
2232 SET_BIT (changed_blocks, block->index);
2233 BB_VISITED (block) = 0;
2234 BB_DEFERRED (block) = 1;
2235 return false;
2237 else
2238 phi_translate_set (dest, source, block, phiblock);
2239 return true;
2242 /* Compute the ANTIC set for BLOCK.
2244 If succs(BLOCK) > 1 then
2245 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2246 else if succs(BLOCK) == 1 then
2247 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2249 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2252 static bool
2253 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2255 bool changed = false;
2256 bitmap_set_t S, old, ANTIC_OUT;
2257 bitmap_iterator bi;
2258 unsigned int bii;
2259 edge e;
2260 edge_iterator ei;
2262 old = ANTIC_OUT = S = NULL;
2263 BB_VISITED (block) = 1;
2265 /* If any edges from predecessors are abnormal, antic_in is empty,
2266 so do nothing. */
2267 if (block_has_abnormal_pred_edge)
2268 goto maybe_dump_sets;
2270 old = ANTIC_IN (block);
2271 ANTIC_OUT = bitmap_set_new ();
2273 /* If the block has no successors, ANTIC_OUT is empty. */
2274 if (EDGE_COUNT (block->succs) == 0)
2276 /* If we have one successor, we could have some phi nodes to
2277 translate through. */
2278 else if (single_succ_p (block))
2280 basic_block succ_bb = single_succ (block);
2282 /* We trade iterations of the dataflow equations for having to
2283 phi translate the maximal set, which is incredibly slow
2284 (since the maximal set often has 300+ members, even when you
2285 have a small number of blocks).
2286 Basically, we defer the computation of ANTIC for this block
2287 until we have processed it's successor, which will inevitably
2288 have a *much* smaller set of values to phi translate once
2289 clean has been run on it.
2290 The cost of doing this is that we technically perform more
2291 iterations, however, they are lower cost iterations.
2293 Timings for PRE on tramp3d-v4:
2294 without maximal set fix: 11 seconds
2295 with maximal set fix/without deferring: 26 seconds
2296 with maximal set fix/with deferring: 11 seconds
2299 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2300 block, succ_bb))
2302 changed = true;
2303 goto maybe_dump_sets;
2306 /* If we have multiple successors, we take the intersection of all of
2307 them. Note that in the case of loop exit phi nodes, we may have
2308 phis to translate through. */
2309 else
2311 VEC(basic_block, heap) * worklist;
2312 size_t i;
2313 basic_block bprime, first = NULL;
2315 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2316 FOR_EACH_EDGE (e, ei, block->succs)
2318 if (!first
2319 && BB_VISITED (e->dest))
2320 first = e->dest;
2321 else if (BB_VISITED (e->dest))
2322 VEC_quick_push (basic_block, worklist, e->dest);
2325 /* Of multiple successors we have to have visited one already. */
2326 if (!first)
2328 SET_BIT (changed_blocks, block->index);
2329 BB_VISITED (block) = 0;
2330 BB_DEFERRED (block) = 1;
2331 changed = true;
2332 VEC_free (basic_block, heap, worklist);
2333 goto maybe_dump_sets;
2336 if (!gimple_seq_empty_p (phi_nodes (first)))
2337 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2338 else
2339 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2341 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2343 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2345 bitmap_set_t tmp = bitmap_set_new ();
2346 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2347 bitmap_set_and (ANTIC_OUT, tmp);
2348 bitmap_set_free (tmp);
2350 else
2351 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2353 VEC_free (basic_block, heap, worklist);
2356 /* Generate ANTIC_OUT - TMP_GEN. */
2357 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2359 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2360 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2361 TMP_GEN (block));
2363 /* Then union in the ANTIC_OUT - TMP_GEN values,
2364 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2365 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2366 bitmap_value_insert_into_set (ANTIC_IN (block),
2367 expression_for_id (bii));
2369 clean (ANTIC_IN (block), block);
2371 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2373 changed = true;
2374 SET_BIT (changed_blocks, block->index);
2375 FOR_EACH_EDGE (e, ei, block->preds)
2376 SET_BIT (changed_blocks, e->src->index);
2378 else
2379 RESET_BIT (changed_blocks, block->index);
2381 maybe_dump_sets:
2382 if (dump_file && (dump_flags & TDF_DETAILS))
2384 if (!BB_DEFERRED (block) || BB_VISITED (block))
2386 if (ANTIC_OUT)
2387 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2389 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2390 block->index);
2392 if (S)
2393 print_bitmap_set (dump_file, S, "S", block->index);
2395 else
2397 fprintf (dump_file,
2398 "Block %d was deferred for a future iteration.\n",
2399 block->index);
2402 if (old)
2403 bitmap_set_free (old);
2404 if (S)
2405 bitmap_set_free (S);
2406 if (ANTIC_OUT)
2407 bitmap_set_free (ANTIC_OUT);
2408 return changed;
2411 /* Compute PARTIAL_ANTIC for BLOCK.
2413 If succs(BLOCK) > 1 then
2414 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2415 in ANTIC_OUT for all succ(BLOCK)
2416 else if succs(BLOCK) == 1 then
2417 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2419 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2420 - ANTIC_IN[BLOCK])
2423 static bool
2424 compute_partial_antic_aux (basic_block block,
2425 bool block_has_abnormal_pred_edge)
2427 bool changed = false;
2428 bitmap_set_t old_PA_IN;
2429 bitmap_set_t PA_OUT;
2430 edge e;
2431 edge_iterator ei;
2432 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2434 old_PA_IN = PA_OUT = NULL;
2436 /* If any edges from predecessors are abnormal, antic_in is empty,
2437 so do nothing. */
2438 if (block_has_abnormal_pred_edge)
2439 goto maybe_dump_sets;
2441 /* If there are too many partially anticipatable values in the
2442 block, phi_translate_set can take an exponential time: stop
2443 before the translation starts. */
2444 if (max_pa
2445 && single_succ_p (block)
2446 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2447 goto maybe_dump_sets;
2449 old_PA_IN = PA_IN (block);
2450 PA_OUT = bitmap_set_new ();
2452 /* If the block has no successors, ANTIC_OUT is empty. */
2453 if (EDGE_COUNT (block->succs) == 0)
2455 /* If we have one successor, we could have some phi nodes to
2456 translate through. Note that we can't phi translate across DFS
2457 back edges in partial antic, because it uses a union operation on
2458 the successors. For recurrences like IV's, we will end up
2459 generating a new value in the set on each go around (i + 3 (VH.1)
2460 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2461 else if (single_succ_p (block))
2463 basic_block succ = single_succ (block);
2464 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2465 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2467 /* If we have multiple successors, we take the union of all of
2468 them. */
2469 else
2471 VEC(basic_block, heap) * worklist;
2472 size_t i;
2473 basic_block bprime;
2475 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2476 FOR_EACH_EDGE (e, ei, block->succs)
2478 if (e->flags & EDGE_DFS_BACK)
2479 continue;
2480 VEC_quick_push (basic_block, worklist, e->dest);
2482 if (VEC_length (basic_block, worklist) > 0)
2484 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2486 unsigned int i;
2487 bitmap_iterator bi;
2489 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2490 bitmap_value_insert_into_set (PA_OUT,
2491 expression_for_id (i));
2492 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2494 bitmap_set_t pa_in = bitmap_set_new ();
2495 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2496 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2497 bitmap_value_insert_into_set (PA_OUT,
2498 expression_for_id (i));
2499 bitmap_set_free (pa_in);
2501 else
2502 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2503 bitmap_value_insert_into_set (PA_OUT,
2504 expression_for_id (i));
2507 VEC_free (basic_block, heap, worklist);
2510 /* PA_IN starts with PA_OUT - TMP_GEN.
2511 Then we subtract things from ANTIC_IN. */
2512 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2514 /* For partial antic, we want to put back in the phi results, since
2515 we will properly avoid making them partially antic over backedges. */
2516 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2517 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2519 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2520 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2522 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2524 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2526 changed = true;
2527 SET_BIT (changed_blocks, block->index);
2528 FOR_EACH_EDGE (e, ei, block->preds)
2529 SET_BIT (changed_blocks, e->src->index);
2531 else
2532 RESET_BIT (changed_blocks, block->index);
2534 maybe_dump_sets:
2535 if (dump_file && (dump_flags & TDF_DETAILS))
2537 if (PA_OUT)
2538 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2540 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2542 if (old_PA_IN)
2543 bitmap_set_free (old_PA_IN);
2544 if (PA_OUT)
2545 bitmap_set_free (PA_OUT);
2546 return changed;
2549 /* Compute ANTIC and partial ANTIC sets. */
2551 static void
2552 compute_antic (void)
2554 bool changed = true;
2555 int num_iterations = 0;
2556 basic_block block;
2557 int i;
2559 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2560 We pre-build the map of blocks with incoming abnormal edges here. */
2561 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2562 sbitmap_zero (has_abnormal_preds);
2564 FOR_EACH_BB (block)
2566 edge_iterator ei;
2567 edge e;
2569 FOR_EACH_EDGE (e, ei, block->preds)
2571 e->flags &= ~EDGE_DFS_BACK;
2572 if (e->flags & EDGE_ABNORMAL)
2574 SET_BIT (has_abnormal_preds, block->index);
2575 break;
2579 BB_VISITED (block) = 0;
2580 BB_DEFERRED (block) = 0;
2582 /* While we are here, give empty ANTIC_IN sets to each block. */
2583 ANTIC_IN (block) = bitmap_set_new ();
2584 PA_IN (block) = bitmap_set_new ();
2587 /* At the exit block we anticipate nothing. */
2588 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2589 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2590 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2592 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2593 sbitmap_ones (changed_blocks);
2594 while (changed)
2596 if (dump_file && (dump_flags & TDF_DETAILS))
2597 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2598 num_iterations++;
2599 changed = false;
2600 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2602 if (TEST_BIT (changed_blocks, postorder[i]))
2604 basic_block block = BASIC_BLOCK (postorder[i]);
2605 changed |= compute_antic_aux (block,
2606 TEST_BIT (has_abnormal_preds,
2607 block->index));
2610 #ifdef ENABLE_CHECKING
2611 /* Theoretically possible, but *highly* unlikely. */
2612 gcc_assert (num_iterations < 500);
2613 #endif
2616 statistics_histogram_event (cfun, "compute_antic iterations",
2617 num_iterations);
2619 if (do_partial_partial)
2621 sbitmap_ones (changed_blocks);
2622 mark_dfs_back_edges ();
2623 num_iterations = 0;
2624 changed = true;
2625 while (changed)
2627 if (dump_file && (dump_flags & TDF_DETAILS))
2628 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2629 num_iterations++;
2630 changed = false;
2631 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2633 if (TEST_BIT (changed_blocks, postorder[i]))
2635 basic_block block = BASIC_BLOCK (postorder[i]);
2636 changed
2637 |= compute_partial_antic_aux (block,
2638 TEST_BIT (has_abnormal_preds,
2639 block->index));
2642 #ifdef ENABLE_CHECKING
2643 /* Theoretically possible, but *highly* unlikely. */
2644 gcc_assert (num_iterations < 500);
2645 #endif
2647 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2648 num_iterations);
2650 sbitmap_free (has_abnormal_preds);
2651 sbitmap_free (changed_blocks);
2654 /* Return true if we can value number the call in STMT. This is true
2655 if we have a pure or constant call. */
2657 static bool
2658 can_value_number_call (gimple stmt)
2660 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2661 return true;
2662 return false;
2665 /* Return true if OP is a tree which we can perform PRE on.
2666 This may not match the operations we can value number, but in
2667 a perfect world would. */
2669 static bool
2670 can_PRE_operation (tree op)
2672 return UNARY_CLASS_P (op)
2673 || BINARY_CLASS_P (op)
2674 || COMPARISON_CLASS_P (op)
2675 || TREE_CODE (op) == MEM_REF
2676 || TREE_CODE (op) == COMPONENT_REF
2677 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2678 || TREE_CODE (op) == CALL_EXPR
2679 || TREE_CODE (op) == ARRAY_REF;
2683 /* Inserted expressions are placed onto this worklist, which is used
2684 for performing quick dead code elimination of insertions we made
2685 that didn't turn out to be necessary. */
2686 static bitmap inserted_exprs;
2688 /* Pool allocated fake store expressions are placed onto this
2689 worklist, which, after performing dead code elimination, is walked
2690 to see which expressions need to be put into GC'able memory */
2691 static VEC(gimple, heap) *need_creation;
2693 /* The actual worker for create_component_ref_by_pieces. */
2695 static tree
2696 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2697 unsigned int *operand, gimple_seq *stmts,
2698 gimple domstmt)
2700 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2701 *operand);
2702 tree genop;
2703 ++*operand;
2704 switch (currop->opcode)
2706 case CALL_EXPR:
2708 tree folded, sc = NULL_TREE;
2709 unsigned int nargs = 0;
2710 tree fn, *args;
2711 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2712 fn = currop->op0;
2713 else
2715 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2716 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2717 if (!fn)
2718 return NULL_TREE;
2720 if (currop->op1)
2722 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2723 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2724 if (!sc)
2725 return NULL_TREE;
2727 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2728 ref->operands) - 1);
2729 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2731 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2732 operand, stmts,
2733 domstmt);
2734 if (!args[nargs])
2736 free (args);
2737 return NULL_TREE;
2739 nargs++;
2741 folded = build_call_array (currop->type,
2742 (TREE_CODE (fn) == FUNCTION_DECL
2743 ? build_fold_addr_expr (fn) : fn),
2744 nargs, args);
2745 free (args);
2746 if (sc)
2747 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2748 return folded;
2750 break;
2751 case MEM_REF:
2753 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2754 stmts, domstmt);
2755 tree offset = currop->op0;
2756 if (!baseop)
2757 return NULL_TREE;
2758 if (TREE_CODE (baseop) == ADDR_EXPR
2759 && handled_component_p (TREE_OPERAND (baseop, 0)))
2761 HOST_WIDE_INT off;
2762 tree base;
2763 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2764 &off);
2765 gcc_assert (base);
2766 offset = int_const_binop (PLUS_EXPR, offset,
2767 build_int_cst (TREE_TYPE (offset),
2768 off), 0);
2769 baseop = build_fold_addr_expr (base);
2771 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2773 break;
2774 case TARGET_MEM_REF:
2776 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2777 *operand);
2778 pre_expr op0expr;
2779 tree genop0 = NULL_TREE;
2780 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2781 stmts, domstmt);
2782 if (!baseop)
2783 return NULL_TREE;
2784 if (currop->op0)
2786 op0expr = get_or_alloc_expr_for (currop->op0);
2787 genop0 = find_or_generate_expression (block, op0expr,
2788 stmts, domstmt);
2789 if (!genop0)
2790 return NULL_TREE;
2792 if (DECL_P (baseop))
2793 return build6 (TARGET_MEM_REF, currop->type,
2794 baseop, NULL_TREE,
2795 genop0, currop->op1, currop->op2,
2796 unshare_expr (nextop->op1));
2797 else
2798 return build6 (TARGET_MEM_REF, currop->type,
2799 NULL_TREE, baseop,
2800 genop0, currop->op1, currop->op2,
2801 unshare_expr (nextop->op1));
2803 break;
2804 case ADDR_EXPR:
2805 if (currop->op0)
2807 gcc_assert (is_gimple_min_invariant (currop->op0));
2808 return currop->op0;
2810 /* Fallthrough. */
2811 case REALPART_EXPR:
2812 case IMAGPART_EXPR:
2813 case VIEW_CONVERT_EXPR:
2815 tree folded;
2816 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2817 operand,
2818 stmts, domstmt);
2819 if (!genop0)
2820 return NULL_TREE;
2821 folded = fold_build1 (currop->opcode, currop->type,
2822 genop0);
2823 return folded;
2825 break;
2826 case MISALIGNED_INDIRECT_REF:
2828 tree folded;
2829 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2830 operand,
2831 stmts, domstmt);
2832 if (!genop1)
2833 return NULL_TREE;
2834 genop1 = fold_convert (build_pointer_type (currop->type),
2835 genop1);
2837 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2838 folded = fold_build2 (currop->opcode, currop->type,
2839 genop1, currop->op1);
2840 else
2841 folded = fold_build1 (currop->opcode, currop->type,
2842 genop1);
2843 return folded;
2845 break;
2846 case BIT_FIELD_REF:
2848 tree folded;
2849 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2850 stmts, domstmt);
2851 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2852 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2853 tree genop1;
2854 tree genop2;
2856 if (!genop0)
2857 return NULL_TREE;
2858 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2859 if (!genop1)
2860 return NULL_TREE;
2861 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2862 if (!genop2)
2863 return NULL_TREE;
2864 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2865 genop2);
2866 return folded;
2869 /* For array ref vn_reference_op's, operand 1 of the array ref
2870 is op0 of the reference op and operand 3 of the array ref is
2871 op1. */
2872 case ARRAY_RANGE_REF:
2873 case ARRAY_REF:
2875 tree genop0;
2876 tree genop1 = currop->op0;
2877 pre_expr op1expr;
2878 tree genop2 = currop->op1;
2879 pre_expr op2expr;
2880 tree genop3 = currop->op2;
2881 pre_expr op3expr;
2882 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2883 stmts, domstmt);
2884 if (!genop0)
2885 return NULL_TREE;
2886 op1expr = get_or_alloc_expr_for (genop1);
2887 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2888 if (!genop1)
2889 return NULL_TREE;
2890 if (genop2)
2892 /* Drop zero minimum index. */
2893 if (tree_int_cst_equal (genop2, integer_zero_node))
2894 genop2 = NULL_TREE;
2895 else
2897 op2expr = get_or_alloc_expr_for (genop2);
2898 genop2 = find_or_generate_expression (block, op2expr, stmts,
2899 domstmt);
2900 if (!genop2)
2901 return NULL_TREE;
2904 if (genop3)
2906 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2907 /* We can't always put a size in units of the element alignment
2908 here as the element alignment may be not visible. See
2909 PR43783. Simply drop the element size for constant
2910 sizes. */
2911 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2912 genop3 = NULL_TREE;
2913 else
2915 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2916 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2917 op3expr = get_or_alloc_expr_for (genop3);
2918 genop3 = find_or_generate_expression (block, op3expr, stmts,
2919 domstmt);
2920 if (!genop3)
2921 return NULL_TREE;
2924 return build4 (currop->opcode, currop->type, genop0, genop1,
2925 genop2, genop3);
2927 case COMPONENT_REF:
2929 tree op0;
2930 tree op1;
2931 tree genop2 = currop->op1;
2932 pre_expr op2expr;
2933 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2934 stmts, domstmt);
2935 if (!op0)
2936 return NULL_TREE;
2937 /* op1 should be a FIELD_DECL, which are represented by
2938 themselves. */
2939 op1 = currop->op0;
2940 if (genop2)
2942 op2expr = get_or_alloc_expr_for (genop2);
2943 genop2 = find_or_generate_expression (block, op2expr, stmts,
2944 domstmt);
2945 if (!genop2)
2946 return NULL_TREE;
2949 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2950 genop2);
2952 break;
2953 case SSA_NAME:
2955 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2956 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2957 return genop;
2959 case STRING_CST:
2960 case INTEGER_CST:
2961 case COMPLEX_CST:
2962 case VECTOR_CST:
2963 case REAL_CST:
2964 case CONSTRUCTOR:
2965 case VAR_DECL:
2966 case PARM_DECL:
2967 case CONST_DECL:
2968 case RESULT_DECL:
2969 case FUNCTION_DECL:
2970 return currop->op0;
2972 default:
2973 gcc_unreachable ();
2977 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2978 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2979 trying to rename aggregates into ssa form directly, which is a no no.
2981 Thus, this routine doesn't create temporaries, it just builds a
2982 single access expression for the array, calling
2983 find_or_generate_expression to build the innermost pieces.
2985 This function is a subroutine of create_expression_by_pieces, and
2986 should not be called on it's own unless you really know what you
2987 are doing. */
2989 static tree
2990 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2991 gimple_seq *stmts, gimple domstmt)
2993 unsigned int op = 0;
2994 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2997 /* Find a leader for an expression, or generate one using
2998 create_expression_by_pieces if it's ANTIC but
2999 complex.
3000 BLOCK is the basic_block we are looking for leaders in.
3001 EXPR is the expression to find a leader or generate for.
3002 STMTS is the statement list to put the inserted expressions on.
3003 Returns the SSA_NAME of the LHS of the generated expression or the
3004 leader.
3005 DOMSTMT if non-NULL is a statement that should be dominated by
3006 all uses in the generated expression. If DOMSTMT is non-NULL this
3007 routine can fail and return NULL_TREE. Otherwise it will assert
3008 on failure. */
3010 static tree
3011 find_or_generate_expression (basic_block block, pre_expr expr,
3012 gimple_seq *stmts, gimple domstmt)
3014 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
3015 get_expr_value_id (expr), domstmt);
3016 tree genop = NULL;
3017 if (leader)
3019 if (leader->kind == NAME)
3020 genop = PRE_EXPR_NAME (leader);
3021 else if (leader->kind == CONSTANT)
3022 genop = PRE_EXPR_CONSTANT (leader);
3025 /* If it's still NULL, it must be a complex expression, so generate
3026 it recursively. Not so if inserting expressions for values generated
3027 by SCCVN. */
3028 if (genop == NULL
3029 && !domstmt)
3031 bitmap_set_t exprset;
3032 unsigned int lookfor = get_expr_value_id (expr);
3033 bool handled = false;
3034 bitmap_iterator bi;
3035 unsigned int i;
3037 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3038 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3040 pre_expr temp = expression_for_id (i);
3041 if (temp->kind != NAME)
3043 handled = true;
3044 genop = create_expression_by_pieces (block, temp, stmts,
3045 domstmt,
3046 get_expr_type (expr));
3047 break;
3050 if (!handled && domstmt)
3051 return NULL_TREE;
3053 gcc_assert (handled);
3055 return genop;
3058 #define NECESSARY GF_PLF_1
3060 /* Create an expression in pieces, so that we can handle very complex
3061 expressions that may be ANTIC, but not necessary GIMPLE.
3062 BLOCK is the basic block the expression will be inserted into,
3063 EXPR is the expression to insert (in value form)
3064 STMTS is a statement list to append the necessary insertions into.
3066 This function will die if we hit some value that shouldn't be
3067 ANTIC but is (IE there is no leader for it, or its components).
3068 This function may also generate expressions that are themselves
3069 partially or fully redundant. Those that are will be either made
3070 fully redundant during the next iteration of insert (for partially
3071 redundant ones), or eliminated by eliminate (for fully redundant
3072 ones).
3074 If DOMSTMT is non-NULL then we make sure that all uses in the
3075 expressions dominate that statement. In this case the function
3076 can return NULL_TREE to signal failure. */
3078 static tree
3079 create_expression_by_pieces (basic_block block, pre_expr expr,
3080 gimple_seq *stmts, gimple domstmt, tree type)
3082 tree temp, name;
3083 tree folded;
3084 gimple_seq forced_stmts = NULL;
3085 unsigned int value_id;
3086 gimple_stmt_iterator gsi;
3087 tree exprtype = type ? type : get_expr_type (expr);
3088 pre_expr nameexpr;
3089 gimple newstmt;
3091 switch (expr->kind)
3093 /* We may hit the NAME/CONSTANT case if we have to convert types
3094 that value numbering saw through. */
3095 case NAME:
3096 folded = PRE_EXPR_NAME (expr);
3097 break;
3098 case CONSTANT:
3099 folded = PRE_EXPR_CONSTANT (expr);
3100 break;
3101 case REFERENCE:
3103 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3104 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3106 break;
3107 case NARY:
3109 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3110 switch (nary->length)
3112 case 2:
3114 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3115 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
3116 tree genop1 = find_or_generate_expression (block, op1,
3117 stmts, domstmt);
3118 tree genop2 = find_or_generate_expression (block, op2,
3119 stmts, domstmt);
3120 if (!genop1 || !genop2)
3121 return NULL_TREE;
3122 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
3123 may be a constant with the wrong type. */
3124 if (nary->opcode == POINTER_PLUS_EXPR)
3126 genop1 = fold_convert (nary->type, genop1);
3127 genop2 = fold_convert (sizetype, genop2);
3129 else
3131 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3132 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
3135 folded = fold_build2 (nary->opcode, nary->type,
3136 genop1, genop2);
3138 break;
3139 case 1:
3141 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3142 tree genop1 = find_or_generate_expression (block, op1,
3143 stmts, domstmt);
3144 if (!genop1)
3145 return NULL_TREE;
3146 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3148 folded = fold_build1 (nary->opcode, nary->type,
3149 genop1);
3151 break;
3152 default:
3153 return NULL_TREE;
3156 break;
3157 default:
3158 return NULL_TREE;
3161 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3162 folded = fold_convert (exprtype, folded);
3164 /* Force the generated expression to be a sequence of GIMPLE
3165 statements.
3166 We have to call unshare_expr because force_gimple_operand may
3167 modify the tree we pass to it. */
3168 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3169 false, NULL);
3171 /* If we have any intermediate expressions to the value sets, add them
3172 to the value sets and chain them in the instruction stream. */
3173 if (forced_stmts)
3175 gsi = gsi_start (forced_stmts);
3176 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3178 gimple stmt = gsi_stmt (gsi);
3179 tree forcedname = gimple_get_lhs (stmt);
3180 pre_expr nameexpr;
3182 if (TREE_CODE (forcedname) == SSA_NAME)
3184 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3185 VN_INFO_GET (forcedname)->valnum = forcedname;
3186 VN_INFO (forcedname)->value_id = get_next_value_id ();
3187 nameexpr = get_or_alloc_expr_for_name (forcedname);
3188 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3189 if (!in_fre)
3190 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3191 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3193 mark_symbols_for_renaming (stmt);
3195 gimple_seq_add_seq (stmts, forced_stmts);
3198 /* Build and insert the assignment of the end result to the temporary
3199 that we will return. */
3200 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3202 pretemp = create_tmp_reg (exprtype, "pretmp");
3203 get_var_ann (pretemp);
3206 temp = pretemp;
3207 add_referenced_var (temp);
3209 newstmt = gimple_build_assign (temp, folded);
3210 name = make_ssa_name (temp, newstmt);
3211 gimple_assign_set_lhs (newstmt, name);
3212 gimple_set_plf (newstmt, NECESSARY, false);
3214 gimple_seq_add_stmt (stmts, newstmt);
3215 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3217 /* All the symbols in NEWEXPR should be put into SSA form. */
3218 mark_symbols_for_renaming (newstmt);
3220 /* Add a value number to the temporary.
3221 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3222 we are creating the expression by pieces, and this particular piece of
3223 the expression may have been represented. There is no harm in replacing
3224 here. */
3225 VN_INFO_GET (name)->valnum = name;
3226 value_id = get_expr_value_id (expr);
3227 VN_INFO (name)->value_id = value_id;
3228 nameexpr = get_or_alloc_expr_for_name (name);
3229 add_to_value (value_id, nameexpr);
3230 if (NEW_SETS (block))
3231 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3232 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3234 pre_stats.insertions++;
3235 if (dump_file && (dump_flags & TDF_DETAILS))
3237 fprintf (dump_file, "Inserted ");
3238 print_gimple_stmt (dump_file, newstmt, 0, 0);
3239 fprintf (dump_file, " in predecessor %d\n", block->index);
3242 return name;
3246 /* Returns true if we want to inhibit the insertions of PHI nodes
3247 for the given EXPR for basic block BB (a member of a loop).
3248 We want to do this, when we fear that the induction variable we
3249 create might inhibit vectorization. */
3251 static bool
3252 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3254 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3255 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3256 vn_reference_op_t op;
3257 unsigned i;
3259 /* If we aren't going to vectorize we don't inhibit anything. */
3260 if (!flag_tree_vectorize)
3261 return false;
3263 /* Otherwise we inhibit the insertion when the address of the
3264 memory reference is a simple induction variable. In other
3265 cases the vectorizer won't do anything anyway (either it's
3266 loop invariant or a complicated expression). */
3267 for (i = 0; VEC_iterate (vn_reference_op_s, ops, i, op); ++i)
3269 switch (op->opcode)
3271 case ARRAY_REF:
3272 case ARRAY_RANGE_REF:
3273 if (TREE_CODE (op->op0) != SSA_NAME)
3274 break;
3275 /* Fallthru. */
3276 case SSA_NAME:
3278 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3279 affine_iv iv;
3280 /* Default defs are loop invariant. */
3281 if (!defbb)
3282 break;
3283 /* Defined outside this loop, also loop invariant. */
3284 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3285 break;
3286 /* If it's a simple induction variable inhibit insertion,
3287 the vectorizer might be interested in this one. */
3288 if (simple_iv (bb->loop_father, bb->loop_father,
3289 op->op0, &iv, true))
3290 return true;
3291 /* No simple IV, vectorizer can't do anything, hence no
3292 reason to inhibit the transformation for this operand. */
3293 break;
3295 default:
3296 break;
3299 return false;
3302 /* Insert the to-be-made-available values of expression EXPRNUM for each
3303 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3304 merge the result with a phi node, given the same value number as
3305 NODE. Return true if we have inserted new stuff. */
3307 static bool
3308 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3309 pre_expr *avail)
3311 pre_expr expr = expression_for_id (exprnum);
3312 pre_expr newphi;
3313 unsigned int val = get_expr_value_id (expr);
3314 edge pred;
3315 bool insertions = false;
3316 bool nophi = false;
3317 basic_block bprime;
3318 pre_expr eprime;
3319 edge_iterator ei;
3320 tree type = get_expr_type (expr);
3321 tree temp;
3322 gimple phi;
3324 if (dump_file && (dump_flags & TDF_DETAILS))
3326 fprintf (dump_file, "Found partial redundancy for expression ");
3327 print_pre_expr (dump_file, expr);
3328 fprintf (dump_file, " (%04d)\n", val);
3331 /* Make sure we aren't creating an induction variable. */
3332 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3334 bool firstinsideloop = false;
3335 bool secondinsideloop = false;
3336 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3337 EDGE_PRED (block, 0)->src);
3338 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3339 EDGE_PRED (block, 1)->src);
3340 /* Induction variables only have one edge inside the loop. */
3341 if ((firstinsideloop ^ secondinsideloop)
3342 && (expr->kind != REFERENCE
3343 || inhibit_phi_insertion (block, expr)))
3345 if (dump_file && (dump_flags & TDF_DETAILS))
3346 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3347 nophi = true;
3351 /* Make the necessary insertions. */
3352 FOR_EACH_EDGE (pred, ei, block->preds)
3354 gimple_seq stmts = NULL;
3355 tree builtexpr;
3356 bprime = pred->src;
3357 eprime = avail[bprime->index];
3359 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3361 builtexpr = create_expression_by_pieces (bprime,
3362 eprime,
3363 &stmts, NULL,
3364 type);
3365 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3366 gsi_insert_seq_on_edge (pred, stmts);
3367 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3368 insertions = true;
3370 else if (eprime->kind == CONSTANT)
3372 /* Constants may not have the right type, fold_convert
3373 should give us back a constant with the right type.
3375 tree constant = PRE_EXPR_CONSTANT (eprime);
3376 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3378 tree builtexpr = fold_convert (type, constant);
3379 if (!is_gimple_min_invariant (builtexpr))
3381 tree forcedexpr = force_gimple_operand (builtexpr,
3382 &stmts, true,
3383 NULL);
3384 if (!is_gimple_min_invariant (forcedexpr))
3386 if (forcedexpr != builtexpr)
3388 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3389 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3391 if (stmts)
3393 gimple_stmt_iterator gsi;
3394 gsi = gsi_start (stmts);
3395 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3397 gimple stmt = gsi_stmt (gsi);
3398 tree lhs = gimple_get_lhs (stmt);
3399 if (TREE_CODE (lhs) == SSA_NAME)
3400 bitmap_set_bit (inserted_exprs,
3401 SSA_NAME_VERSION (lhs));
3402 gimple_set_plf (stmt, NECESSARY, false);
3404 gsi_insert_seq_on_edge (pred, stmts);
3406 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3409 else
3410 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3413 else if (eprime->kind == NAME)
3415 /* We may have to do a conversion because our value
3416 numbering can look through types in certain cases, but
3417 our IL requires all operands of a phi node have the same
3418 type. */
3419 tree name = PRE_EXPR_NAME (eprime);
3420 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3422 tree builtexpr;
3423 tree forcedexpr;
3424 builtexpr = fold_convert (type, name);
3425 forcedexpr = force_gimple_operand (builtexpr,
3426 &stmts, true,
3427 NULL);
3429 if (forcedexpr != name)
3431 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3432 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3435 if (stmts)
3437 gimple_stmt_iterator gsi;
3438 gsi = gsi_start (stmts);
3439 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3441 gimple stmt = gsi_stmt (gsi);
3442 tree lhs = gimple_get_lhs (stmt);
3443 if (TREE_CODE (lhs) == SSA_NAME)
3444 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3445 gimple_set_plf (stmt, NECESSARY, false);
3447 gsi_insert_seq_on_edge (pred, stmts);
3449 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3453 /* If we didn't want a phi node, and we made insertions, we still have
3454 inserted new stuff, and thus return true. If we didn't want a phi node,
3455 and didn't make insertions, we haven't added anything new, so return
3456 false. */
3457 if (nophi && insertions)
3458 return true;
3459 else if (nophi && !insertions)
3460 return false;
3462 /* Now build a phi for the new variable. */
3463 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3465 prephitemp = create_tmp_var (type, "prephitmp");
3466 get_var_ann (prephitemp);
3469 temp = prephitemp;
3470 add_referenced_var (temp);
3472 if (TREE_CODE (type) == COMPLEX_TYPE
3473 || TREE_CODE (type) == VECTOR_TYPE)
3474 DECL_GIMPLE_REG_P (temp) = 1;
3475 phi = create_phi_node (temp, block);
3477 gimple_set_plf (phi, NECESSARY, false);
3478 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3479 VN_INFO (gimple_phi_result (phi))->value_id = val;
3480 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3481 FOR_EACH_EDGE (pred, ei, block->preds)
3483 pre_expr ae = avail[pred->src->index];
3484 gcc_assert (get_expr_type (ae) == type
3485 || useless_type_conversion_p (type, get_expr_type (ae)));
3486 if (ae->kind == CONSTANT)
3487 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3488 else
3489 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3490 UNKNOWN_LOCATION);
3493 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3494 add_to_value (val, newphi);
3496 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3497 this insertion, since we test for the existence of this value in PHI_GEN
3498 before proceeding with the partial redundancy checks in insert_aux.
3500 The value may exist in AVAIL_OUT, in particular, it could be represented
3501 by the expression we are trying to eliminate, in which case we want the
3502 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3503 inserted there.
3505 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3506 this block, because if it did, it would have existed in our dominator's
3507 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3510 bitmap_insert_into_set (PHI_GEN (block), newphi);
3511 bitmap_value_replace_in_set (AVAIL_OUT (block),
3512 newphi);
3513 bitmap_insert_into_set (NEW_SETS (block),
3514 newphi);
3516 if (dump_file && (dump_flags & TDF_DETAILS))
3518 fprintf (dump_file, "Created phi ");
3519 print_gimple_stmt (dump_file, phi, 0, 0);
3520 fprintf (dump_file, " in block %d\n", block->index);
3522 pre_stats.phis++;
3523 return true;
3528 /* Perform insertion of partially redundant values.
3529 For BLOCK, do the following:
3530 1. Propagate the NEW_SETS of the dominator into the current block.
3531 If the block has multiple predecessors,
3532 2a. Iterate over the ANTIC expressions for the block to see if
3533 any of them are partially redundant.
3534 2b. If so, insert them into the necessary predecessors to make
3535 the expression fully redundant.
3536 2c. Insert a new PHI merging the values of the predecessors.
3537 2d. Insert the new PHI, and the new expressions, into the
3538 NEW_SETS set.
3539 3. Recursively call ourselves on the dominator children of BLOCK.
3541 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3542 do_regular_insertion and do_partial_insertion.
3546 static bool
3547 do_regular_insertion (basic_block block, basic_block dom)
3549 bool new_stuff = false;
3550 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3551 pre_expr expr;
3552 int i;
3554 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3556 if (expr->kind != NAME)
3558 pre_expr *avail;
3559 unsigned int val;
3560 bool by_some = false;
3561 bool cant_insert = false;
3562 bool all_same = true;
3563 pre_expr first_s = NULL;
3564 edge pred;
3565 basic_block bprime;
3566 pre_expr eprime = NULL;
3567 edge_iterator ei;
3568 pre_expr edoubleprime = NULL;
3569 bool do_insertion = false;
3571 val = get_expr_value_id (expr);
3572 if (bitmap_set_contains_value (PHI_GEN (block), val))
3573 continue;
3574 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3576 if (dump_file && (dump_flags & TDF_DETAILS))
3577 fprintf (dump_file, "Found fully redundant value\n");
3578 continue;
3581 avail = XCNEWVEC (pre_expr, last_basic_block);
3582 FOR_EACH_EDGE (pred, ei, block->preds)
3584 unsigned int vprime;
3586 /* We should never run insertion for the exit block
3587 and so not come across fake pred edges. */
3588 gcc_assert (!(pred->flags & EDGE_FAKE));
3589 bprime = pred->src;
3590 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3591 bprime, block);
3593 /* eprime will generally only be NULL if the
3594 value of the expression, translated
3595 through the PHI for this predecessor, is
3596 undefined. If that is the case, we can't
3597 make the expression fully redundant,
3598 because its value is undefined along a
3599 predecessor path. We can thus break out
3600 early because it doesn't matter what the
3601 rest of the results are. */
3602 if (eprime == NULL)
3604 cant_insert = true;
3605 break;
3608 eprime = fully_constant_expression (eprime);
3609 vprime = get_expr_value_id (eprime);
3610 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3611 vprime, NULL);
3612 if (edoubleprime == NULL)
3614 avail[bprime->index] = eprime;
3615 all_same = false;
3617 else
3619 avail[bprime->index] = edoubleprime;
3620 by_some = true;
3621 /* We want to perform insertions to remove a redundancy on
3622 a path in the CFG we want to optimize for speed. */
3623 if (optimize_edge_for_speed_p (pred))
3624 do_insertion = true;
3625 if (first_s == NULL)
3626 first_s = edoubleprime;
3627 else if (!pre_expr_eq (first_s, edoubleprime))
3628 all_same = false;
3631 /* If we can insert it, it's not the same value
3632 already existing along every predecessor, and
3633 it's defined by some predecessor, it is
3634 partially redundant. */
3635 if (!cant_insert && !all_same && by_some && do_insertion
3636 && dbg_cnt (treepre_insert))
3638 if (insert_into_preds_of_block (block, get_expression_id (expr),
3639 avail))
3640 new_stuff = true;
3642 /* If all edges produce the same value and that value is
3643 an invariant, then the PHI has the same value on all
3644 edges. Note this. */
3645 else if (!cant_insert && all_same && eprime
3646 && (edoubleprime->kind == CONSTANT
3647 || edoubleprime->kind == NAME)
3648 && !value_id_constant_p (val))
3650 unsigned int j;
3651 bitmap_iterator bi;
3652 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3653 value_expressions, val);
3655 unsigned int new_val = get_expr_value_id (edoubleprime);
3656 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3658 pre_expr expr = expression_for_id (j);
3660 if (expr->kind == NAME)
3662 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3663 /* Just reset the value id and valnum so it is
3664 the same as the constant we have discovered. */
3665 if (edoubleprime->kind == CONSTANT)
3667 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3668 pre_stats.constified++;
3670 else
3671 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3672 info->value_id = new_val;
3676 free (avail);
3680 VEC_free (pre_expr, heap, exprs);
3681 return new_stuff;
3685 /* Perform insertion for partially anticipatable expressions. There
3686 is only one case we will perform insertion for these. This case is
3687 if the expression is partially anticipatable, and fully available.
3688 In this case, we know that putting it earlier will enable us to
3689 remove the later computation. */
3692 static bool
3693 do_partial_partial_insertion (basic_block block, basic_block dom)
3695 bool new_stuff = false;
3696 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3697 pre_expr expr;
3698 int i;
3700 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3702 if (expr->kind != NAME)
3704 pre_expr *avail;
3705 unsigned int val;
3706 bool by_all = true;
3707 bool cant_insert = false;
3708 edge pred;
3709 basic_block bprime;
3710 pre_expr eprime = NULL;
3711 edge_iterator ei;
3713 val = get_expr_value_id (expr);
3714 if (bitmap_set_contains_value (PHI_GEN (block), val))
3715 continue;
3716 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3717 continue;
3719 avail = XCNEWVEC (pre_expr, last_basic_block);
3720 FOR_EACH_EDGE (pred, ei, block->preds)
3722 unsigned int vprime;
3723 pre_expr edoubleprime;
3725 /* We should never run insertion for the exit block
3726 and so not come across fake pred edges. */
3727 gcc_assert (!(pred->flags & EDGE_FAKE));
3728 bprime = pred->src;
3729 eprime = phi_translate (expr, ANTIC_IN (block),
3730 PA_IN (block),
3731 bprime, block);
3733 /* eprime will generally only be NULL if the
3734 value of the expression, translated
3735 through the PHI for this predecessor, is
3736 undefined. If that is the case, we can't
3737 make the expression fully redundant,
3738 because its value is undefined along a
3739 predecessor path. We can thus break out
3740 early because it doesn't matter what the
3741 rest of the results are. */
3742 if (eprime == NULL)
3744 cant_insert = true;
3745 break;
3748 eprime = fully_constant_expression (eprime);
3749 vprime = get_expr_value_id (eprime);
3750 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3751 vprime, NULL);
3752 if (edoubleprime == NULL)
3754 by_all = false;
3755 break;
3757 else
3758 avail[bprime->index] = edoubleprime;
3762 /* If we can insert it, it's not the same value
3763 already existing along every predecessor, and
3764 it's defined by some predecessor, it is
3765 partially redundant. */
3766 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3768 pre_stats.pa_insert++;
3769 if (insert_into_preds_of_block (block, get_expression_id (expr),
3770 avail))
3771 new_stuff = true;
3773 free (avail);
3777 VEC_free (pre_expr, heap, exprs);
3778 return new_stuff;
3781 static bool
3782 insert_aux (basic_block block)
3784 basic_block son;
3785 bool new_stuff = false;
3787 if (block)
3789 basic_block dom;
3790 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3791 if (dom)
3793 unsigned i;
3794 bitmap_iterator bi;
3795 bitmap_set_t newset = NEW_SETS (dom);
3796 if (newset)
3798 /* Note that we need to value_replace both NEW_SETS, and
3799 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3800 represented by some non-simple expression here that we want
3801 to replace it with. */
3802 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3804 pre_expr expr = expression_for_id (i);
3805 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3806 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3809 if (!single_pred_p (block))
3811 new_stuff |= do_regular_insertion (block, dom);
3812 if (do_partial_partial)
3813 new_stuff |= do_partial_partial_insertion (block, dom);
3817 for (son = first_dom_son (CDI_DOMINATORS, block);
3818 son;
3819 son = next_dom_son (CDI_DOMINATORS, son))
3821 new_stuff |= insert_aux (son);
3824 return new_stuff;
3827 /* Perform insertion of partially redundant values. */
3829 static void
3830 insert (void)
3832 bool new_stuff = true;
3833 basic_block bb;
3834 int num_iterations = 0;
3836 FOR_ALL_BB (bb)
3837 NEW_SETS (bb) = bitmap_set_new ();
3839 while (new_stuff)
3841 num_iterations++;
3842 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3844 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3848 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3850 static void
3851 add_to_exp_gen (basic_block block, tree op)
3853 if (!in_fre)
3855 pre_expr result;
3856 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3857 return;
3858 result = get_or_alloc_expr_for_name (op);
3859 bitmap_value_insert_into_set (EXP_GEN (block), result);
3863 /* Create value ids for PHI in BLOCK. */
3865 static void
3866 make_values_for_phi (gimple phi, basic_block block)
3868 tree result = gimple_phi_result (phi);
3870 /* We have no need for virtual phis, as they don't represent
3871 actual computations. */
3872 if (is_gimple_reg (result))
3874 pre_expr e = get_or_alloc_expr_for_name (result);
3875 add_to_value (get_expr_value_id (e), e);
3876 bitmap_insert_into_set (PHI_GEN (block), e);
3877 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3878 if (!in_fre)
3880 unsigned i;
3881 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3883 tree arg = gimple_phi_arg_def (phi, i);
3884 if (TREE_CODE (arg) == SSA_NAME)
3886 e = get_or_alloc_expr_for_name (arg);
3887 add_to_value (get_expr_value_id (e), e);
3894 /* Compute the AVAIL set for all basic blocks.
3896 This function performs value numbering of the statements in each basic
3897 block. The AVAIL sets are built from information we glean while doing
3898 this value numbering, since the AVAIL sets contain only one entry per
3899 value.
3901 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3902 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3904 static void
3905 compute_avail (void)
3908 basic_block block, son;
3909 basic_block *worklist;
3910 size_t sp = 0;
3911 unsigned i;
3913 /* We pretend that default definitions are defined in the entry block.
3914 This includes function arguments and the static chain decl. */
3915 for (i = 1; i < num_ssa_names; ++i)
3917 tree name = ssa_name (i);
3918 pre_expr e;
3919 if (!name
3920 || !SSA_NAME_IS_DEFAULT_DEF (name)
3921 || has_zero_uses (name)
3922 || !is_gimple_reg (name))
3923 continue;
3925 e = get_or_alloc_expr_for_name (name);
3926 add_to_value (get_expr_value_id (e), e);
3927 if (!in_fre)
3928 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3929 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3932 /* Allocate the worklist. */
3933 worklist = XNEWVEC (basic_block, n_basic_blocks);
3935 /* Seed the algorithm by putting the dominator children of the entry
3936 block on the worklist. */
3937 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3938 son;
3939 son = next_dom_son (CDI_DOMINATORS, son))
3940 worklist[sp++] = son;
3942 /* Loop until the worklist is empty. */
3943 while (sp)
3945 gimple_stmt_iterator gsi;
3946 gimple stmt;
3947 basic_block dom;
3948 unsigned int stmt_uid = 1;
3950 /* Pick a block from the worklist. */
3951 block = worklist[--sp];
3953 /* Initially, the set of available values in BLOCK is that of
3954 its immediate dominator. */
3955 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3956 if (dom)
3957 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3959 /* Generate values for PHI nodes. */
3960 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3961 make_values_for_phi (gsi_stmt (gsi), block);
3963 BB_MAY_NOTRETURN (block) = 0;
3965 /* Now compute value numbers and populate value sets with all
3966 the expressions computed in BLOCK. */
3967 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3969 ssa_op_iter iter;
3970 tree op;
3972 stmt = gsi_stmt (gsi);
3973 gimple_set_uid (stmt, stmt_uid++);
3975 /* Cache whether the basic-block has any non-visible side-effect
3976 or control flow.
3977 If this isn't a call or it is the last stmt in the
3978 basic-block then the CFG represents things correctly. */
3979 if (is_gimple_call (stmt)
3980 && !stmt_ends_bb_p (stmt))
3982 /* Non-looping const functions always return normally.
3983 Otherwise the call might not return or have side-effects
3984 that forbids hoisting possibly trapping expressions
3985 before it. */
3986 int flags = gimple_call_flags (stmt);
3987 if (!(flags & ECF_CONST)
3988 || (flags & ECF_LOOPING_CONST_OR_PURE))
3989 BB_MAY_NOTRETURN (block) = 1;
3992 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3994 pre_expr e = get_or_alloc_expr_for_name (op);
3996 add_to_value (get_expr_value_id (e), e);
3997 if (!in_fre)
3998 bitmap_insert_into_set (TMP_GEN (block), e);
3999 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
4002 if (gimple_has_volatile_ops (stmt)
4003 || stmt_could_throw_p (stmt))
4004 continue;
4006 switch (gimple_code (stmt))
4008 case GIMPLE_RETURN:
4009 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4010 add_to_exp_gen (block, op);
4011 continue;
4013 case GIMPLE_CALL:
4015 vn_reference_t ref;
4016 unsigned int i;
4017 vn_reference_op_t vro;
4018 pre_expr result = NULL;
4019 VEC(vn_reference_op_s, heap) *ops = NULL;
4021 if (!can_value_number_call (stmt))
4022 continue;
4024 copy_reference_ops_from_call (stmt, &ops);
4025 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4026 gimple_expr_type (stmt),
4027 ops, &ref, false);
4028 VEC_free (vn_reference_op_s, heap, ops);
4029 if (!ref)
4030 continue;
4032 for (i = 0; VEC_iterate (vn_reference_op_s,
4033 ref->operands, i,
4034 vro); i++)
4036 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4037 add_to_exp_gen (block, vro->op0);
4038 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4039 add_to_exp_gen (block, vro->op1);
4040 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4041 add_to_exp_gen (block, vro->op2);
4043 result = (pre_expr) pool_alloc (pre_expr_pool);
4044 result->kind = REFERENCE;
4045 result->id = 0;
4046 PRE_EXPR_REFERENCE (result) = ref;
4048 get_or_alloc_expression_id (result);
4049 add_to_value (get_expr_value_id (result), result);
4050 if (!in_fre)
4051 bitmap_value_insert_into_set (EXP_GEN (block), result);
4052 continue;
4055 case GIMPLE_ASSIGN:
4057 pre_expr result = NULL;
4058 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4060 case tcc_unary:
4061 case tcc_binary:
4062 case tcc_comparison:
4064 vn_nary_op_t nary;
4065 unsigned int i;
4067 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4068 gimple_assign_rhs_code (stmt),
4069 gimple_expr_type (stmt),
4070 gimple_assign_rhs1 (stmt),
4071 gimple_assign_rhs2 (stmt),
4072 NULL_TREE, NULL_TREE, &nary);
4074 if (!nary)
4075 continue;
4077 for (i = 0; i < nary->length; i++)
4078 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4079 add_to_exp_gen (block, nary->op[i]);
4081 result = (pre_expr) pool_alloc (pre_expr_pool);
4082 result->kind = NARY;
4083 result->id = 0;
4084 PRE_EXPR_NARY (result) = nary;
4085 break;
4088 case tcc_declaration:
4089 case tcc_reference:
4091 vn_reference_t ref;
4092 unsigned int i;
4093 vn_reference_op_t vro;
4095 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4096 gimple_vuse (stmt),
4097 true, &ref);
4098 if (!ref)
4099 continue;
4101 for (i = 0; VEC_iterate (vn_reference_op_s,
4102 ref->operands, i,
4103 vro); i++)
4105 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4106 add_to_exp_gen (block, vro->op0);
4107 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4108 add_to_exp_gen (block, vro->op1);
4109 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4110 add_to_exp_gen (block, vro->op2);
4112 result = (pre_expr) pool_alloc (pre_expr_pool);
4113 result->kind = REFERENCE;
4114 result->id = 0;
4115 PRE_EXPR_REFERENCE (result) = ref;
4116 break;
4119 default:
4120 /* For any other statement that we don't
4121 recognize, simply add all referenced
4122 SSA_NAMEs to EXP_GEN. */
4123 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4124 add_to_exp_gen (block, op);
4125 continue;
4128 get_or_alloc_expression_id (result);
4129 add_to_value (get_expr_value_id (result), result);
4130 if (!in_fre)
4131 bitmap_value_insert_into_set (EXP_GEN (block), result);
4133 continue;
4135 default:
4136 break;
4140 /* Put the dominator children of BLOCK on the worklist of blocks
4141 to compute available sets for. */
4142 for (son = first_dom_son (CDI_DOMINATORS, block);
4143 son;
4144 son = next_dom_son (CDI_DOMINATORS, son))
4145 worklist[sp++] = son;
4148 free (worklist);
4151 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4152 than the available expressions for it. The insertion point is
4153 right before the first use in STMT. Returns the SSA_NAME that should
4154 be used for replacement. */
4156 static tree
4157 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4159 basic_block bb = gimple_bb (stmt);
4160 gimple_stmt_iterator gsi;
4161 gimple_seq stmts = NULL;
4162 tree expr;
4163 pre_expr e;
4165 /* First create a value expression from the expression we want
4166 to insert and associate it with the value handle for SSA_VN. */
4167 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4168 if (e == NULL)
4169 return NULL_TREE;
4171 /* Then use create_expression_by_pieces to generate a valid
4172 expression to insert at this point of the IL stream. */
4173 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4174 if (expr == NULL_TREE)
4175 return NULL_TREE;
4176 gsi = gsi_for_stmt (stmt);
4177 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4179 return expr;
4182 /* Eliminate fully redundant computations. */
4184 static unsigned int
4185 eliminate (void)
4187 VEC (gimple, heap) *to_remove = NULL;
4188 basic_block b;
4189 unsigned int todo = 0;
4190 gimple_stmt_iterator gsi;
4191 gimple stmt;
4192 unsigned i;
4194 FOR_EACH_BB (b)
4196 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4198 stmt = gsi_stmt (gsi);
4200 /* Lookup the RHS of the expression, see if we have an
4201 available computation for it. If so, replace the RHS with
4202 the available computation. */
4203 if (gimple_has_lhs (stmt)
4204 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4205 && !gimple_assign_ssa_name_copy_p (stmt)
4206 && (!gimple_assign_single_p (stmt)
4207 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4208 && !gimple_has_volatile_ops (stmt)
4209 && !has_zero_uses (gimple_get_lhs (stmt)))
4211 tree lhs = gimple_get_lhs (stmt);
4212 tree rhs = NULL_TREE;
4213 tree sprime = NULL;
4214 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4215 pre_expr sprimeexpr;
4217 if (gimple_assign_single_p (stmt))
4218 rhs = gimple_assign_rhs1 (stmt);
4220 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4221 get_expr_value_id (lhsexpr),
4222 NULL);
4224 if (sprimeexpr)
4226 if (sprimeexpr->kind == CONSTANT)
4227 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4228 else if (sprimeexpr->kind == NAME)
4229 sprime = PRE_EXPR_NAME (sprimeexpr);
4230 else
4231 gcc_unreachable ();
4234 /* If there is no existing leader but SCCVN knows this
4235 value is constant, use that constant. */
4236 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4238 sprime = VN_INFO (lhs)->valnum;
4239 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4240 TREE_TYPE (sprime)))
4241 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4243 if (dump_file && (dump_flags & TDF_DETAILS))
4245 fprintf (dump_file, "Replaced ");
4246 print_gimple_expr (dump_file, stmt, 0, 0);
4247 fprintf (dump_file, " with ");
4248 print_generic_expr (dump_file, sprime, 0);
4249 fprintf (dump_file, " in ");
4250 print_gimple_stmt (dump_file, stmt, 0, 0);
4252 pre_stats.eliminations++;
4253 propagate_tree_value_into_stmt (&gsi, sprime);
4254 stmt = gsi_stmt (gsi);
4255 update_stmt (stmt);
4256 continue;
4259 /* If there is no existing usable leader but SCCVN thinks
4260 it has an expression it wants to use as replacement,
4261 insert that. */
4262 if (!sprime || sprime == lhs)
4264 tree val = VN_INFO (lhs)->valnum;
4265 if (val != VN_TOP
4266 && TREE_CODE (val) == SSA_NAME
4267 && VN_INFO (val)->needs_insertion
4268 && can_PRE_operation (vn_get_expr_for (val)))
4269 sprime = do_SCCVN_insertion (stmt, val);
4271 if (sprime
4272 && sprime != lhs
4273 && (rhs == NULL_TREE
4274 || TREE_CODE (rhs) != SSA_NAME
4275 || may_propagate_copy (rhs, sprime)))
4277 gcc_assert (sprime != rhs);
4279 if (dump_file && (dump_flags & TDF_DETAILS))
4281 fprintf (dump_file, "Replaced ");
4282 print_gimple_expr (dump_file, stmt, 0, 0);
4283 fprintf (dump_file, " with ");
4284 print_generic_expr (dump_file, sprime, 0);
4285 fprintf (dump_file, " in ");
4286 print_gimple_stmt (dump_file, stmt, 0, 0);
4289 if (TREE_CODE (sprime) == SSA_NAME)
4290 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4291 NECESSARY, true);
4292 /* We need to make sure the new and old types actually match,
4293 which may require adding a simple cast, which fold_convert
4294 will do for us. */
4295 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4296 && !useless_type_conversion_p (gimple_expr_type (stmt),
4297 TREE_TYPE (sprime)))
4298 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4300 pre_stats.eliminations++;
4301 propagate_tree_value_into_stmt (&gsi, sprime);
4302 stmt = gsi_stmt (gsi);
4303 update_stmt (stmt);
4305 /* If we removed EH side effects from the statement, clean
4306 its EH information. */
4307 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4309 bitmap_set_bit (need_eh_cleanup,
4310 gimple_bb (stmt)->index);
4311 if (dump_file && (dump_flags & TDF_DETAILS))
4312 fprintf (dump_file, " Removed EH side effects.\n");
4316 /* If the statement is a scalar store, see if the expression
4317 has the same value number as its rhs. If so, the store is
4318 dead. */
4319 else if (gimple_assign_single_p (stmt)
4320 && !is_gimple_reg (gimple_assign_lhs (stmt))
4321 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4322 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4324 tree rhs = gimple_assign_rhs1 (stmt);
4325 tree val;
4326 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4327 gimple_vuse (stmt), true, NULL);
4328 if (TREE_CODE (rhs) == SSA_NAME)
4329 rhs = VN_INFO (rhs)->valnum;
4330 if (val
4331 && operand_equal_p (val, rhs, 0))
4333 if (dump_file && (dump_flags & TDF_DETAILS))
4335 fprintf (dump_file, "Deleted redundant store ");
4336 print_gimple_stmt (dump_file, stmt, 0, 0);
4339 /* Queue stmt for removal. */
4340 VEC_safe_push (gimple, heap, to_remove, stmt);
4343 /* Visit COND_EXPRs and fold the comparison with the
4344 available value-numbers. */
4345 else if (gimple_code (stmt) == GIMPLE_COND)
4347 tree op0 = gimple_cond_lhs (stmt);
4348 tree op1 = gimple_cond_rhs (stmt);
4349 tree result;
4351 if (TREE_CODE (op0) == SSA_NAME)
4352 op0 = VN_INFO (op0)->valnum;
4353 if (TREE_CODE (op1) == SSA_NAME)
4354 op1 = VN_INFO (op1)->valnum;
4355 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4356 op0, op1);
4357 if (result && TREE_CODE (result) == INTEGER_CST)
4359 if (integer_zerop (result))
4360 gimple_cond_make_false (stmt);
4361 else
4362 gimple_cond_make_true (stmt);
4363 update_stmt (stmt);
4364 todo = TODO_cleanup_cfg;
4367 /* Visit indirect calls and turn them into direct calls if
4368 possible. */
4369 if (gimple_code (stmt) == GIMPLE_CALL
4370 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4372 tree fn = VN_INFO (gimple_call_fn (stmt))->valnum;
4373 if (TREE_CODE (fn) == ADDR_EXPR
4374 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
4376 if (dump_file && (dump_flags & TDF_DETAILS))
4378 fprintf (dump_file, "Replacing call target with ");
4379 print_generic_expr (dump_file, fn, 0);
4380 fprintf (dump_file, " in ");
4381 print_gimple_stmt (dump_file, stmt, 0, 0);
4384 gimple_call_set_fn (stmt, fn);
4385 update_stmt (stmt);
4386 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4388 bitmap_set_bit (need_eh_cleanup,
4389 gimple_bb (stmt)->index);
4390 if (dump_file && (dump_flags & TDF_DETAILS))
4391 fprintf (dump_file, " Removed EH side effects.\n");
4394 /* Changing an indirect call to a direct call may
4395 have exposed different semantics. This may
4396 require an SSA update. */
4397 todo |= TODO_update_ssa_only_virtuals;
4402 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4404 gimple stmt, phi = gsi_stmt (gsi);
4405 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4406 pre_expr sprimeexpr, resexpr;
4407 gimple_stmt_iterator gsi2;
4409 /* We want to perform redundant PHI elimination. Do so by
4410 replacing the PHI with a single copy if possible.
4411 Do not touch inserted, single-argument or virtual PHIs. */
4412 if (gimple_phi_num_args (phi) == 1
4413 || !is_gimple_reg (res))
4415 gsi_next (&gsi);
4416 continue;
4419 resexpr = get_or_alloc_expr_for_name (res);
4420 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4421 get_expr_value_id (resexpr), NULL);
4422 if (sprimeexpr)
4424 if (sprimeexpr->kind == CONSTANT)
4425 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4426 else if (sprimeexpr->kind == NAME)
4427 sprime = PRE_EXPR_NAME (sprimeexpr);
4428 else
4429 gcc_unreachable ();
4431 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4433 sprime = VN_INFO (res)->valnum;
4434 if (!useless_type_conversion_p (TREE_TYPE (res),
4435 TREE_TYPE (sprime)))
4436 sprime = fold_convert (TREE_TYPE (res), sprime);
4438 if (!sprime
4439 || sprime == res)
4441 gsi_next (&gsi);
4442 continue;
4445 if (dump_file && (dump_flags & TDF_DETAILS))
4447 fprintf (dump_file, "Replaced redundant PHI node defining ");
4448 print_generic_expr (dump_file, res, 0);
4449 fprintf (dump_file, " with ");
4450 print_generic_expr (dump_file, sprime, 0);
4451 fprintf (dump_file, "\n");
4454 remove_phi_node (&gsi, false);
4456 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4457 && TREE_CODE (sprime) == SSA_NAME)
4458 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4460 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4461 sprime = fold_convert (TREE_TYPE (res), sprime);
4462 stmt = gimple_build_assign (res, sprime);
4463 SSA_NAME_DEF_STMT (res) = stmt;
4464 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4466 gsi2 = gsi_after_labels (b);
4467 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4468 /* Queue the copy for eventual removal. */
4469 VEC_safe_push (gimple, heap, to_remove, stmt);
4470 /* If we inserted this PHI node ourself, it's not an elimination. */
4471 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4472 pre_stats.phis--;
4473 else
4474 pre_stats.eliminations++;
4478 /* We cannot remove stmts during BB walk, especially not release SSA
4479 names there as this confuses the VN machinery. The stmts ending
4480 up in to_remove are either stores or simple copies. */
4481 for (i = 0; VEC_iterate (gimple, to_remove, i, stmt); ++i)
4483 tree lhs = gimple_assign_lhs (stmt);
4484 tree rhs = gimple_assign_rhs1 (stmt);
4485 use_operand_p use_p;
4486 gimple use_stmt;
4488 /* If there is a single use only, propagate the equivalency
4489 instead of keeping the copy. */
4490 if (TREE_CODE (lhs) == SSA_NAME
4491 && TREE_CODE (rhs) == SSA_NAME
4492 && single_imm_use (lhs, &use_p, &use_stmt)
4493 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4495 SET_USE (use_p, rhs);
4496 update_stmt (use_stmt);
4497 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4498 && TREE_CODE (rhs) == SSA_NAME)
4499 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4502 /* If this is a store or a now unused copy, remove it. */
4503 if (TREE_CODE (lhs) != SSA_NAME
4504 || has_zero_uses (lhs))
4506 gsi = gsi_for_stmt (stmt);
4507 unlink_stmt_vdef (stmt);
4508 gsi_remove (&gsi, true);
4509 if (TREE_CODE (lhs) == SSA_NAME)
4510 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4511 release_defs (stmt);
4514 VEC_free (gimple, heap, to_remove);
4516 return todo;
4519 /* Borrow a bit of tree-ssa-dce.c for the moment.
4520 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4521 this may be a bit faster, and we may want critical edges kept split. */
4523 /* If OP's defining statement has not already been determined to be necessary,
4524 mark that statement necessary. Return the stmt, if it is newly
4525 necessary. */
4527 static inline gimple
4528 mark_operand_necessary (tree op)
4530 gimple stmt;
4532 gcc_assert (op);
4534 if (TREE_CODE (op) != SSA_NAME)
4535 return NULL;
4537 stmt = SSA_NAME_DEF_STMT (op);
4538 gcc_assert (stmt);
4540 if (gimple_plf (stmt, NECESSARY)
4541 || gimple_nop_p (stmt))
4542 return NULL;
4544 gimple_set_plf (stmt, NECESSARY, true);
4545 return stmt;
4548 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4549 to insert PHI nodes sometimes, and because value numbering of casts isn't
4550 perfect, we sometimes end up inserting dead code. This simple DCE-like
4551 pass removes any insertions we made that weren't actually used. */
4553 static void
4554 remove_dead_inserted_code (void)
4556 bitmap worklist;
4557 unsigned i;
4558 bitmap_iterator bi;
4559 gimple t;
4561 worklist = BITMAP_ALLOC (NULL);
4562 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4564 t = SSA_NAME_DEF_STMT (ssa_name (i));
4565 if (gimple_plf (t, NECESSARY))
4566 bitmap_set_bit (worklist, i);
4568 while (!bitmap_empty_p (worklist))
4570 i = bitmap_first_set_bit (worklist);
4571 bitmap_clear_bit (worklist, i);
4572 t = SSA_NAME_DEF_STMT (ssa_name (i));
4574 /* PHI nodes are somewhat special in that each PHI alternative has
4575 data and control dependencies. All the statements feeding the
4576 PHI node's arguments are always necessary. */
4577 if (gimple_code (t) == GIMPLE_PHI)
4579 unsigned k;
4581 for (k = 0; k < gimple_phi_num_args (t); k++)
4583 tree arg = PHI_ARG_DEF (t, k);
4584 if (TREE_CODE (arg) == SSA_NAME)
4586 gimple n = mark_operand_necessary (arg);
4587 if (n)
4588 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4592 else
4594 /* Propagate through the operands. Examine all the USE, VUSE and
4595 VDEF operands in this statement. Mark all the statements
4596 which feed this statement's uses as necessary. */
4597 ssa_op_iter iter;
4598 tree use;
4600 /* The operands of VDEF expressions are also needed as they
4601 represent potential definitions that may reach this
4602 statement (VDEF operands allow us to follow def-def
4603 links). */
4605 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4607 gimple n = mark_operand_necessary (use);
4608 if (n)
4609 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4614 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4616 t = SSA_NAME_DEF_STMT (ssa_name (i));
4617 if (!gimple_plf (t, NECESSARY))
4619 gimple_stmt_iterator gsi;
4621 if (dump_file && (dump_flags & TDF_DETAILS))
4623 fprintf (dump_file, "Removing unnecessary insertion:");
4624 print_gimple_stmt (dump_file, t, 0, 0);
4627 gsi = gsi_for_stmt (t);
4628 if (gimple_code (t) == GIMPLE_PHI)
4629 remove_phi_node (&gsi, true);
4630 else
4632 gsi_remove (&gsi, true);
4633 release_defs (t);
4637 BITMAP_FREE (worklist);
4640 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4641 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4642 the number of visited blocks. */
4644 static int
4645 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4647 edge_iterator *stack;
4648 int sp;
4649 int post_order_num = 0;
4650 sbitmap visited;
4652 if (include_entry_exit)
4653 post_order[post_order_num++] = EXIT_BLOCK;
4655 /* Allocate stack for back-tracking up CFG. */
4656 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4657 sp = 0;
4659 /* Allocate bitmap to track nodes that have been visited. */
4660 visited = sbitmap_alloc (last_basic_block);
4662 /* None of the nodes in the CFG have been visited yet. */
4663 sbitmap_zero (visited);
4665 /* Push the last edge on to the stack. */
4666 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4668 while (sp)
4670 edge_iterator ei;
4671 basic_block src;
4672 basic_block dest;
4674 /* Look at the edge on the top of the stack. */
4675 ei = stack[sp - 1];
4676 src = ei_edge (ei)->src;
4677 dest = ei_edge (ei)->dest;
4679 /* Check if the edge destination has been visited yet. */
4680 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4682 /* Mark that we have visited the destination. */
4683 SET_BIT (visited, src->index);
4685 if (EDGE_COUNT (src->preds) > 0)
4686 /* Since the DEST node has been visited for the first
4687 time, check its successors. */
4688 stack[sp++] = ei_start (src->preds);
4689 else
4690 post_order[post_order_num++] = src->index;
4692 else
4694 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4695 post_order[post_order_num++] = dest->index;
4697 if (!ei_one_before_end_p (ei))
4698 ei_next (&stack[sp - 1]);
4699 else
4700 sp--;
4704 if (include_entry_exit)
4705 post_order[post_order_num++] = ENTRY_BLOCK;
4707 free (stack);
4708 sbitmap_free (visited);
4709 return post_order_num;
4713 /* Initialize data structures used by PRE. */
4715 static void
4716 init_pre (bool do_fre)
4718 basic_block bb;
4720 next_expression_id = 1;
4721 expressions = NULL;
4722 VEC_safe_push (pre_expr, heap, expressions, NULL);
4723 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4724 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4725 get_max_value_id() + 1);
4726 name_to_id = NULL;
4728 in_fre = do_fre;
4730 inserted_exprs = BITMAP_ALLOC (NULL);
4731 need_creation = NULL;
4732 pretemp = NULL_TREE;
4733 storetemp = NULL_TREE;
4734 prephitemp = NULL_TREE;
4736 connect_infinite_loops_to_exit ();
4737 memset (&pre_stats, 0, sizeof (pre_stats));
4740 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4741 my_rev_post_order_compute (postorder, false);
4743 FOR_ALL_BB (bb)
4744 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4746 calculate_dominance_info (CDI_POST_DOMINATORS);
4747 calculate_dominance_info (CDI_DOMINATORS);
4749 bitmap_obstack_initialize (&grand_bitmap_obstack);
4750 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4751 expr_pred_trans_eq, free);
4752 expression_to_id = htab_create (num_ssa_names * 3,
4753 pre_expr_hash,
4754 pre_expr_eq, NULL);
4755 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4756 sizeof (struct bitmap_set), 30);
4757 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4758 sizeof (struct pre_expr_d), 30);
4759 FOR_ALL_BB (bb)
4761 EXP_GEN (bb) = bitmap_set_new ();
4762 PHI_GEN (bb) = bitmap_set_new ();
4763 TMP_GEN (bb) = bitmap_set_new ();
4764 AVAIL_OUT (bb) = bitmap_set_new ();
4767 need_eh_cleanup = BITMAP_ALLOC (NULL);
4771 /* Deallocate data structures used by PRE. */
4773 static void
4774 fini_pre (bool do_fre)
4776 basic_block bb;
4778 free (postorder);
4779 VEC_free (bitmap_set_t, heap, value_expressions);
4780 BITMAP_FREE (inserted_exprs);
4781 VEC_free (gimple, heap, need_creation);
4782 bitmap_obstack_release (&grand_bitmap_obstack);
4783 free_alloc_pool (bitmap_set_pool);
4784 free_alloc_pool (pre_expr_pool);
4785 htab_delete (phi_translate_table);
4786 htab_delete (expression_to_id);
4787 VEC_free (unsigned, heap, name_to_id);
4789 FOR_ALL_BB (bb)
4791 free (bb->aux);
4792 bb->aux = NULL;
4795 free_dominance_info (CDI_POST_DOMINATORS);
4797 if (!bitmap_empty_p (need_eh_cleanup))
4799 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4800 cleanup_tree_cfg ();
4803 BITMAP_FREE (need_eh_cleanup);
4805 if (!do_fre)
4806 loop_optimizer_finalize ();
4809 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4810 only wants to do full redundancy elimination. */
4812 static unsigned int
4813 execute_pre (bool do_fre)
4815 unsigned int todo = 0;
4817 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4819 /* This has to happen before SCCVN runs because
4820 loop_optimizer_init may create new phis, etc. */
4821 if (!do_fre)
4822 loop_optimizer_init (LOOPS_NORMAL);
4824 if (!run_scc_vn ())
4826 if (!do_fre)
4827 loop_optimizer_finalize ();
4829 return 0;
4832 init_pre (do_fre);
4833 scev_initialize ();
4835 /* Collect and value number expressions computed in each basic block. */
4836 compute_avail ();
4838 if (dump_file && (dump_flags & TDF_DETAILS))
4840 basic_block bb;
4842 FOR_ALL_BB (bb)
4844 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4845 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4846 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4847 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4851 /* Insert can get quite slow on an incredibly large number of basic
4852 blocks due to some quadratic behavior. Until this behavior is
4853 fixed, don't run it when he have an incredibly large number of
4854 bb's. If we aren't going to run insert, there is no point in
4855 computing ANTIC, either, even though it's plenty fast. */
4856 if (!do_fre && n_basic_blocks < 4000)
4858 compute_antic ();
4859 insert ();
4862 /* Make sure to remove fake edges before committing our inserts.
4863 This makes sure we don't end up with extra critical edges that
4864 we would need to split. */
4865 remove_fake_exit_edges ();
4866 gsi_commit_edge_inserts ();
4868 /* Remove all the redundant expressions. */
4869 todo |= eliminate ();
4871 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4872 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4873 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4874 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4875 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4877 clear_expression_ids ();
4878 free_scc_vn ();
4879 if (!do_fre)
4880 remove_dead_inserted_code ();
4882 scev_finalize ();
4883 fini_pre (do_fre);
4885 return todo;
4888 /* Gate and execute functions for PRE. */
4890 static unsigned int
4891 do_pre (void)
4893 return execute_pre (false);
4896 static bool
4897 gate_pre (void)
4899 return flag_tree_pre != 0;
4902 struct gimple_opt_pass pass_pre =
4905 GIMPLE_PASS,
4906 "pre", /* name */
4907 gate_pre, /* gate */
4908 do_pre, /* execute */
4909 NULL, /* sub */
4910 NULL, /* next */
4911 0, /* static_pass_number */
4912 TV_TREE_PRE, /* tv_id */
4913 PROP_no_crit_edges | PROP_cfg
4914 | PROP_ssa, /* properties_required */
4915 0, /* properties_provided */
4916 0, /* properties_destroyed */
4917 TODO_rebuild_alias, /* todo_flags_start */
4918 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4919 | TODO_verify_ssa /* todo_flags_finish */
4924 /* Gate and execute functions for FRE. */
4926 static unsigned int
4927 execute_fre (void)
4929 return execute_pre (true);
4932 static bool
4933 gate_fre (void)
4935 return flag_tree_fre != 0;
4938 struct gimple_opt_pass pass_fre =
4941 GIMPLE_PASS,
4942 "fre", /* name */
4943 gate_fre, /* gate */
4944 execute_fre, /* execute */
4945 NULL, /* sub */
4946 NULL, /* next */
4947 0, /* static_pass_number */
4948 TV_TREE_FRE, /* tv_id */
4949 PROP_cfg | PROP_ssa, /* properties_required */
4950 0, /* properties_provided */
4951 0, /* properties_destroyed */
4952 0, /* todo_flags_start */
4953 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */