Merge from mainline (154736:156693)
[official-gcc/graphite-test-results.git] / gcc / tree-ssa-pre.c
blob285b2c8b0b3dd7b3c4013f07b325146d536a6b9e
1 /* SSA-PRE for trees.
2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
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 "ggc.h"
28 #include "tree.h"
29 #include "basic-block.h"
30 #include "diagnostic.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 "real.h"
40 #include "alloc-pool.h"
41 #include "obstack.h"
42 #include "tree-pass.h"
43 #include "flags.h"
44 #include "bitmap.h"
45 #include "langhooks.h"
46 #include "cfgloop.h"
47 #include "tree-ssa-sccvn.h"
48 #include "tree-scalar-evolution.h"
49 #include "params.h"
50 #include "dbgcnt.h"
52 /* TODO:
54 1. Avail sets can be shared by making an avail_find_leader that
55 walks up the dominator tree and looks in those avail sets.
56 This might affect code optimality, it's unclear right now.
57 2. Strength reduction can be performed by anticipating expressions
58 we can repair later on.
59 3. We can do back-substitution or smarter value numbering to catch
60 commutative expressions split up over multiple statements.
63 /* For ease of terminology, "expression node" in the below refers to
64 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
65 represent the actual statement containing the expressions we care about,
66 and we cache the value number by putting it in the expression. */
68 /* Basic algorithm
70 First we walk the statements to generate the AVAIL sets, the
71 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
72 generation of values/expressions by a given block. We use them
73 when computing the ANTIC sets. The AVAIL sets consist of
74 SSA_NAME's that represent values, so we know what values are
75 available in what blocks. AVAIL is a forward dataflow problem. In
76 SSA, values are never killed, so we don't need a kill set, or a
77 fixpoint iteration, in order to calculate the AVAIL sets. In
78 traditional parlance, AVAIL sets tell us the downsafety of the
79 expressions/values.
81 Next, we generate the ANTIC sets. These sets represent the
82 anticipatable expressions. ANTIC is a backwards dataflow
83 problem. An expression is anticipatable in a given block if it could
84 be generated in that block. This means that if we had to perform
85 an insertion in that block, of the value of that expression, we
86 could. Calculating the ANTIC sets requires phi translation of
87 expressions, because the flow goes backwards through phis. We must
88 iterate to a fixpoint of the ANTIC sets, because we have a kill
89 set. Even in SSA form, values are not live over the entire
90 function, only from their definition point onwards. So we have to
91 remove values from the ANTIC set once we go past the definition
92 point of the leaders that make them up.
93 compute_antic/compute_antic_aux performs this computation.
95 Third, we perform insertions to make partially redundant
96 expressions fully redundant.
98 An expression is partially redundant (excluding partial
99 anticipation) if:
101 1. It is AVAIL in some, but not all, of the predecessors of a
102 given block.
103 2. It is ANTIC in all the predecessors.
105 In order to make it fully redundant, we insert the expression into
106 the predecessors where it is not available, but is ANTIC.
108 For the partial anticipation case, we only perform insertion if it
109 is partially anticipated in some block, and fully available in all
110 of the predecessors.
112 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
113 performs these steps.
115 Fourth, we eliminate fully redundant expressions.
116 This is a simple statement walk that replaces redundant
117 calculations with the now available values. */
119 /* Representations of value numbers:
121 Value numbers are represented by a representative SSA_NAME. We
122 will create fake SSA_NAME's in situations where we need a
123 representative but do not have one (because it is a complex
124 expression). In order to facilitate storing the value numbers in
125 bitmaps, and keep the number of wasted SSA_NAME's down, we also
126 associate a value_id with each value number, and create full blown
127 ssa_name's only where we actually need them (IE in operands of
128 existing expressions).
130 Theoretically you could replace all the value_id's with
131 SSA_NAME_VERSION, but this would allocate a large number of
132 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
133 It would also require an additional indirection at each point we
134 use the value id. */
136 /* Representation of expressions on value numbers:
138 Expressions consisting of value numbers are represented the same
139 way as our VN internally represents them, with an additional
140 "pre_expr" wrapping around them in order to facilitate storing all
141 of the expressions in the same sets. */
143 /* Representation of sets:
145 The dataflow sets do not need to be sorted in any particular order
146 for the majority of their lifetime, are simply represented as two
147 bitmaps, one that keeps track of values present in the set, and one
148 that keeps track of expressions present in the set.
150 When we need them in topological order, we produce it on demand by
151 transforming the bitmap into an array and sorting it into topo
152 order. */
154 /* Type of expression, used to know which member of the PRE_EXPR union
155 is valid. */
157 enum pre_expr_kind
159 NAME,
160 NARY,
161 REFERENCE,
162 CONSTANT
165 typedef union pre_expr_union_d
167 tree name;
168 tree constant;
169 vn_nary_op_t nary;
170 vn_reference_t reference;
171 } pre_expr_union;
173 typedef struct pre_expr_d
175 enum pre_expr_kind kind;
176 unsigned int id;
177 pre_expr_union u;
178 } *pre_expr;
180 #define PRE_EXPR_NAME(e) (e)->u.name
181 #define PRE_EXPR_NARY(e) (e)->u.nary
182 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
183 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
185 static int
186 pre_expr_eq (const void *p1, const void *p2)
188 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
189 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
191 if (e1->kind != e2->kind)
192 return false;
194 switch (e1->kind)
196 case CONSTANT:
197 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
198 PRE_EXPR_CONSTANT (e2));
199 case NAME:
200 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
201 case NARY:
202 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
203 case REFERENCE:
204 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
205 PRE_EXPR_REFERENCE (e2));
206 default:
207 gcc_unreachable ();
211 static hashval_t
212 pre_expr_hash (const void *p1)
214 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
215 switch (e->kind)
217 case CONSTANT:
218 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
219 case NAME:
220 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
221 case NARY:
222 return PRE_EXPR_NARY (e)->hashcode;
223 case REFERENCE:
224 return PRE_EXPR_REFERENCE (e)->hashcode;
225 default:
226 gcc_unreachable ();
231 /* Next global expression id number. */
232 static unsigned int next_expression_id;
234 /* Mapping from expression to id number we can use in bitmap sets. */
235 DEF_VEC_P (pre_expr);
236 DEF_VEC_ALLOC_P (pre_expr, heap);
237 static VEC(pre_expr, heap) *expressions;
238 static htab_t expression_to_id;
239 static VEC(unsigned, heap) *name_to_id;
241 /* Allocate an expression id for EXPR. */
243 static inline unsigned int
244 alloc_expression_id (pre_expr expr)
246 void **slot;
247 /* Make sure we won't overflow. */
248 gcc_assert (next_expression_id + 1 > next_expression_id);
249 expr->id = next_expression_id++;
250 VEC_safe_push (pre_expr, heap, expressions, expr);
251 if (expr->kind == NAME)
253 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
254 /* VEC_safe_grow_cleared allocates no headroom. Avoid frequent
255 re-allocations by using VEC_reserve upfront. There is no
256 VEC_quick_grow_cleared unfortunately. */
257 VEC_reserve (unsigned, heap, name_to_id, num_ssa_names);
258 VEC_safe_grow_cleared (unsigned, heap, name_to_id, num_ssa_names);
259 gcc_assert (VEC_index (unsigned, name_to_id, version) == 0);
260 VEC_replace (unsigned, name_to_id, version, expr->id);
262 else
264 slot = htab_find_slot (expression_to_id, expr, INSERT);
265 gcc_assert (!*slot);
266 *slot = expr;
268 return next_expression_id - 1;
271 /* Return the expression id for tree EXPR. */
273 static inline unsigned int
274 get_expression_id (const pre_expr expr)
276 return expr->id;
279 static inline unsigned int
280 lookup_expression_id (const pre_expr expr)
282 void **slot;
284 if (expr->kind == NAME)
286 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
287 if (VEC_length (unsigned, name_to_id) <= version)
288 return 0;
289 return VEC_index (unsigned, name_to_id, version);
291 else
293 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
294 if (!slot)
295 return 0;
296 return ((pre_expr)*slot)->id;
300 /* Return the existing expression id for EXPR, or create one if one
301 does not exist yet. */
303 static inline unsigned int
304 get_or_alloc_expression_id (pre_expr expr)
306 unsigned int id = lookup_expression_id (expr);
307 if (id == 0)
308 return alloc_expression_id (expr);
309 return expr->id = id;
312 /* Return the expression that has expression id ID */
314 static inline pre_expr
315 expression_for_id (unsigned int id)
317 return VEC_index (pre_expr, expressions, id);
320 /* Free the expression id field in all of our expressions,
321 and then destroy the expressions array. */
323 static void
324 clear_expression_ids (void)
326 VEC_free (pre_expr, heap, expressions);
329 static alloc_pool pre_expr_pool;
331 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
333 static pre_expr
334 get_or_alloc_expr_for_name (tree name)
336 struct pre_expr_d expr;
337 pre_expr result;
338 unsigned int result_id;
340 expr.kind = NAME;
341 expr.id = 0;
342 PRE_EXPR_NAME (&expr) = name;
343 result_id = lookup_expression_id (&expr);
344 if (result_id != 0)
345 return expression_for_id (result_id);
347 result = (pre_expr) pool_alloc (pre_expr_pool);
348 result->kind = NAME;
349 PRE_EXPR_NAME (result) = name;
350 alloc_expression_id (result);
351 return result;
354 static bool in_fre = false;
356 /* An unordered bitmap set. One bitmap tracks values, the other,
357 expressions. */
358 typedef struct bitmap_set
360 bitmap expressions;
361 bitmap values;
362 } *bitmap_set_t;
364 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
365 EXECUTE_IF_SET_IN_BITMAP((set)->expressions, 0, (id), (bi))
367 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
368 EXECUTE_IF_SET_IN_BITMAP((set)->values, 0, (id), (bi))
370 /* Mapping from value id to expressions with that value_id. */
371 DEF_VEC_P (bitmap_set_t);
372 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
373 static VEC(bitmap_set_t, heap) *value_expressions;
375 /* Sets that we need to keep track of. */
376 typedef struct bb_bitmap_sets
378 /* The EXP_GEN set, which represents expressions/values generated in
379 a basic block. */
380 bitmap_set_t exp_gen;
382 /* The PHI_GEN set, which represents PHI results generated in a
383 basic block. */
384 bitmap_set_t phi_gen;
386 /* The TMP_GEN set, which represents results/temporaries generated
387 in a basic block. IE the LHS of an expression. */
388 bitmap_set_t tmp_gen;
390 /* The AVAIL_OUT set, which represents which values are available in
391 a given basic block. */
392 bitmap_set_t avail_out;
394 /* The ANTIC_IN set, which represents which values are anticipatable
395 in a given basic block. */
396 bitmap_set_t antic_in;
398 /* The PA_IN set, which represents which values are
399 partially anticipatable in a given basic block. */
400 bitmap_set_t pa_in;
402 /* The NEW_SETS set, which is used during insertion to augment the
403 AVAIL_OUT set of blocks with the new insertions performed during
404 the current iteration. */
405 bitmap_set_t new_sets;
407 /* A cache for value_dies_in_block_x. */
408 bitmap expr_dies;
410 /* True if we have visited this block during ANTIC calculation. */
411 unsigned int visited : 1;
413 /* True we have deferred processing this block during ANTIC
414 calculation until its successor is processed. */
415 unsigned int deferred : 1;
417 /* True when the block contains a call that might not return. */
418 unsigned int contains_may_not_return_call : 1;
419 } *bb_value_sets_t;
421 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
422 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
423 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
424 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
425 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
426 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
427 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
428 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
429 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
430 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
431 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
434 /* Basic block list in postorder. */
435 static int *postorder;
437 /* This structure is used to keep track of statistics on what
438 optimization PRE was able to perform. */
439 static struct
441 /* The number of RHS computations eliminated by PRE. */
442 int eliminations;
444 /* The number of new expressions/temporaries generated by PRE. */
445 int insertions;
447 /* The number of inserts found due to partial anticipation */
448 int pa_insert;
450 /* The number of new PHI nodes added by PRE. */
451 int phis;
453 /* The number of values found constant. */
454 int constified;
456 } pre_stats;
458 static bool do_partial_partial;
459 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
460 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
461 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
462 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
463 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
464 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
465 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
466 unsigned int, bool);
467 static bitmap_set_t bitmap_set_new (void);
468 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
469 gimple, tree);
470 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
471 gimple);
472 static unsigned int get_expr_value_id (pre_expr);
474 /* We can add and remove elements and entries to and from sets
475 and hash tables, so we use alloc pools for them. */
477 static alloc_pool bitmap_set_pool;
478 static bitmap_obstack grand_bitmap_obstack;
480 /* To avoid adding 300 temporary variables when we only need one, we
481 only create one temporary variable, on demand, and build ssa names
482 off that. We do have to change the variable if the types don't
483 match the current variable's type. */
484 static tree pretemp;
485 static tree storetemp;
486 static tree prephitemp;
488 /* Set of blocks with statements that have had its EH information
489 cleaned up. */
490 static bitmap need_eh_cleanup;
492 /* The phi_translate_table caches phi translations for a given
493 expression and predecessor. */
495 static htab_t phi_translate_table;
497 /* A three tuple {e, pred, v} used to cache phi translations in the
498 phi_translate_table. */
500 typedef struct expr_pred_trans_d
502 /* The expression. */
503 pre_expr e;
505 /* The predecessor block along which we translated the expression. */
506 basic_block pred;
508 /* The value that resulted from the translation. */
509 pre_expr v;
511 /* The hashcode for the expression, pred pair. This is cached for
512 speed reasons. */
513 hashval_t hashcode;
514 } *expr_pred_trans_t;
515 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
517 /* Return the hash value for a phi translation table entry. */
519 static hashval_t
520 expr_pred_trans_hash (const void *p)
522 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
523 return ve->hashcode;
526 /* Return true if two phi translation table entries are the same.
527 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
529 static int
530 expr_pred_trans_eq (const void *p1, const void *p2)
532 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
533 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
534 basic_block b1 = ve1->pred;
535 basic_block b2 = ve2->pred;
537 /* If they are not translations for the same basic block, they can't
538 be equal. */
539 if (b1 != b2)
540 return false;
541 return pre_expr_eq (ve1->e, ve2->e);
544 /* Search in the phi translation table for the translation of
545 expression E in basic block PRED.
546 Return the translated value, if found, NULL otherwise. */
548 static inline pre_expr
549 phi_trans_lookup (pre_expr e, basic_block pred)
551 void **slot;
552 struct expr_pred_trans_d ept;
554 ept.e = e;
555 ept.pred = pred;
556 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
557 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
558 NO_INSERT);
559 if (!slot)
560 return NULL;
561 else
562 return ((expr_pred_trans_t) *slot)->v;
566 /* Add the tuple mapping from {expression E, basic block PRED} to
567 value V, to the phi translation table. */
569 static inline void
570 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
572 void **slot;
573 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
574 new_pair->e = e;
575 new_pair->pred = pred;
576 new_pair->v = v;
577 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
578 pred->index);
580 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
581 new_pair->hashcode, INSERT);
582 if (*slot)
583 free (*slot);
584 *slot = (void *) new_pair;
588 /* Add expression E to the expression set of value id V. */
590 void
591 add_to_value (unsigned int v, pre_expr e)
593 bitmap_set_t set;
595 gcc_assert (get_expr_value_id (e) == v);
597 if (v >= VEC_length (bitmap_set_t, value_expressions))
599 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
600 v + 1);
603 set = VEC_index (bitmap_set_t, value_expressions, v);
604 if (!set)
606 set = bitmap_set_new ();
607 VEC_replace (bitmap_set_t, value_expressions, v, set);
610 bitmap_insert_into_set_1 (set, e, v, true);
613 /* Create a new bitmap set and return it. */
615 static bitmap_set_t
616 bitmap_set_new (void)
618 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
619 ret->expressions = BITMAP_ALLOC (&grand_bitmap_obstack);
620 ret->values = BITMAP_ALLOC (&grand_bitmap_obstack);
621 return ret;
624 /* Return the value id for a PRE expression EXPR. */
626 static unsigned int
627 get_expr_value_id (pre_expr expr)
629 switch (expr->kind)
631 case CONSTANT:
633 unsigned int id;
634 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
635 if (id == 0)
637 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
638 add_to_value (id, expr);
640 return id;
642 case NAME:
643 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
644 case NARY:
645 return PRE_EXPR_NARY (expr)->value_id;
646 case REFERENCE:
647 return PRE_EXPR_REFERENCE (expr)->value_id;
648 default:
649 gcc_unreachable ();
653 /* Remove an expression EXPR from a bitmapped set. */
655 static void
656 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
658 unsigned int val = get_expr_value_id (expr);
659 if (!value_id_constant_p (val))
661 bitmap_clear_bit (set->values, val);
662 bitmap_clear_bit (set->expressions, get_expression_id (expr));
666 static void
667 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
668 unsigned int val, bool allow_constants)
670 if (allow_constants || !value_id_constant_p (val))
672 /* We specifically expect this and only this function to be able to
673 insert constants into a set. */
674 bitmap_set_bit (set->values, val);
675 bitmap_set_bit (set->expressions, get_or_alloc_expression_id (expr));
679 /* Insert an expression EXPR into a bitmapped set. */
681 static void
682 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
684 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
687 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
689 static void
690 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
692 bitmap_copy (dest->expressions, orig->expressions);
693 bitmap_copy (dest->values, orig->values);
697 /* Free memory used up by SET. */
698 static void
699 bitmap_set_free (bitmap_set_t set)
701 BITMAP_FREE (set->expressions);
702 BITMAP_FREE (set->values);
706 /* Generate an topological-ordered array of bitmap set SET. */
708 static VEC(pre_expr, heap) *
709 sorted_array_from_bitmap_set (bitmap_set_t set)
711 unsigned int i, j;
712 bitmap_iterator bi, bj;
713 VEC(pre_expr, heap) *result;
715 /* Pre-allocate roughly enough space for the array. */
716 result = VEC_alloc (pre_expr, heap, bitmap_count_bits (set->values));
718 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
720 /* The number of expressions having a given value is usually
721 relatively small. Thus, rather than making a vector of all
722 the expressions and sorting it by value-id, we walk the values
723 and check in the reverse mapping that tells us what expressions
724 have a given value, to filter those in our set. As a result,
725 the expressions are inserted in value-id order, which means
726 topological order.
728 If this is somehow a significant lose for some cases, we can
729 choose which set to walk based on the set size. */
730 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
731 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
733 if (bitmap_bit_p (set->expressions, j))
734 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
738 return result;
741 /* Perform bitmapped set operation DEST &= ORIG. */
743 static void
744 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
746 bitmap_iterator bi;
747 unsigned int i;
749 if (dest != orig)
751 bitmap temp = BITMAP_ALLOC (&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_FREE (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 temp = BITMAP_ALLOC (&grand_bitmap_obstack);
797 bitmap_copy (temp, a->expressions);
798 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
800 pre_expr expr = expression_for_id (i);
801 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
802 bitmap_remove_from_set (a, expr);
804 BITMAP_FREE (temp);
808 /* Return true if bitmapped set SET contains the value VALUE_ID. */
810 static bool
811 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
813 if (value_id_constant_p (value_id))
814 return true;
816 if (!set || bitmap_empty_p (set->expressions))
817 return false;
819 return bitmap_bit_p (set->values, value_id);
822 static inline bool
823 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
825 return bitmap_bit_p (set->expressions, get_expression_id (expr));
828 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
830 static void
831 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
832 const pre_expr expr)
834 bitmap_set_t exprset;
835 unsigned int i;
836 bitmap_iterator bi;
838 if (value_id_constant_p (lookfor))
839 return;
841 if (!bitmap_set_contains_value (set, lookfor))
842 return;
844 /* The number of expressions having a given value is usually
845 significantly less than the total number of expressions in SET.
846 Thus, rather than check, for each expression in SET, whether it
847 has the value LOOKFOR, we walk the reverse mapping that tells us
848 what expressions have a given value, and see if any of those
849 expressions are in our set. For large testcases, this is about
850 5-10x faster than walking the bitmap. If this is somehow a
851 significant lose for some cases, we can choose which set to walk
852 based on the set size. */
853 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
854 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
856 if (bitmap_bit_p (set->expressions, i))
858 bitmap_clear_bit (set->expressions, i);
859 bitmap_set_bit (set->expressions, get_expression_id (expr));
860 return;
865 /* Return true if two bitmap sets are equal. */
867 static bool
868 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
870 return bitmap_equal_p (a->values, b->values);
873 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
874 and add it otherwise. */
876 static void
877 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
879 unsigned int val = get_expr_value_id (expr);
881 if (bitmap_set_contains_value (set, val))
882 bitmap_set_replace_value (set, val, expr);
883 else
884 bitmap_insert_into_set (set, expr);
887 /* Insert EXPR into SET if EXPR's value is not already present in
888 SET. */
890 static void
891 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
893 unsigned int val = get_expr_value_id (expr);
895 #ifdef ENABLE_CHECKING
896 gcc_assert (expr->id == get_or_alloc_expression_id (expr));
897 #endif
899 /* Constant values are always considered to be part of the set. */
900 if (value_id_constant_p (val))
901 return;
903 /* If the value membership changed, add the expression. */
904 if (bitmap_set_bit (set->values, val))
905 bitmap_set_bit (set->expressions, expr->id);
908 /* Print out EXPR to outfile. */
910 static void
911 print_pre_expr (FILE *outfile, const pre_expr expr)
913 switch (expr->kind)
915 case CONSTANT:
916 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
917 break;
918 case NAME:
919 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
920 break;
921 case NARY:
923 unsigned int i;
924 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
925 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
926 for (i = 0; i < nary->length; i++)
928 print_generic_expr (outfile, nary->op[i], 0);
929 if (i != (unsigned) nary->length - 1)
930 fprintf (outfile, ",");
932 fprintf (outfile, "}");
934 break;
936 case REFERENCE:
938 vn_reference_op_t vro;
939 unsigned int i;
940 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
941 fprintf (outfile, "{");
942 for (i = 0;
943 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
944 i++)
946 bool closebrace = false;
947 if (vro->opcode != SSA_NAME
948 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
950 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
951 if (vro->op0)
953 fprintf (outfile, "<");
954 closebrace = true;
957 if (vro->op0)
959 print_generic_expr (outfile, vro->op0, 0);
960 if (vro->op1)
962 fprintf (outfile, ",");
963 print_generic_expr (outfile, vro->op1, 0);
965 if (vro->op2)
967 fprintf (outfile, ",");
968 print_generic_expr (outfile, vro->op2, 0);
971 if (closebrace)
972 fprintf (outfile, ">");
973 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
974 fprintf (outfile, ",");
976 fprintf (outfile, "}");
977 if (ref->vuse)
979 fprintf (outfile, "@");
980 print_generic_expr (outfile, ref->vuse, 0);
983 break;
986 void debug_pre_expr (pre_expr);
988 /* Like print_pre_expr but always prints to stderr. */
989 void
990 debug_pre_expr (pre_expr e)
992 print_pre_expr (stderr, e);
993 fprintf (stderr, "\n");
996 /* Print out SET to OUTFILE. */
998 static void
999 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1000 const char *setname, int blockindex)
1002 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1003 if (set)
1005 bool first = true;
1006 unsigned i;
1007 bitmap_iterator bi;
1009 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1011 const pre_expr expr = expression_for_id (i);
1013 if (!first)
1014 fprintf (outfile, ", ");
1015 first = false;
1016 print_pre_expr (outfile, expr);
1018 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1021 fprintf (outfile, " }\n");
1024 void debug_bitmap_set (bitmap_set_t);
1026 void
1027 debug_bitmap_set (bitmap_set_t set)
1029 print_bitmap_set (stderr, set, "debug", 0);
1032 /* Print out the expressions that have VAL to OUTFILE. */
1034 void
1035 print_value_expressions (FILE *outfile, unsigned int val)
1037 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1038 if (set)
1040 char s[10];
1041 sprintf (s, "%04d", val);
1042 print_bitmap_set (outfile, set, s, 0);
1047 void
1048 debug_value_expressions (unsigned int val)
1050 print_value_expressions (stderr, val);
1053 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1054 represent it. */
1056 static pre_expr
1057 get_or_alloc_expr_for_constant (tree constant)
1059 unsigned int result_id;
1060 unsigned int value_id;
1061 struct pre_expr_d expr;
1062 pre_expr newexpr;
1064 expr.kind = CONSTANT;
1065 PRE_EXPR_CONSTANT (&expr) = constant;
1066 result_id = lookup_expression_id (&expr);
1067 if (result_id != 0)
1068 return expression_for_id (result_id);
1070 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1071 newexpr->kind = CONSTANT;
1072 PRE_EXPR_CONSTANT (newexpr) = constant;
1073 alloc_expression_id (newexpr);
1074 value_id = get_or_alloc_constant_value_id (constant);
1075 add_to_value (value_id, newexpr);
1076 return newexpr;
1079 /* Given a value id V, find the actual tree representing the constant
1080 value if there is one, and return it. Return NULL if we can't find
1081 a constant. */
1083 static tree
1084 get_constant_for_value_id (unsigned int v)
1086 if (value_id_constant_p (v))
1088 unsigned int i;
1089 bitmap_iterator bi;
1090 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1092 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1094 pre_expr expr = expression_for_id (i);
1095 if (expr->kind == CONSTANT)
1096 return PRE_EXPR_CONSTANT (expr);
1099 return NULL;
1102 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1103 Currently only supports constants and SSA_NAMES. */
1104 static pre_expr
1105 get_or_alloc_expr_for (tree t)
1107 if (TREE_CODE (t) == SSA_NAME)
1108 return get_or_alloc_expr_for_name (t);
1109 else if (is_gimple_min_invariant (t))
1110 return get_or_alloc_expr_for_constant (t);
1111 else
1113 /* More complex expressions can result from SCCVN expression
1114 simplification that inserts values for them. As they all
1115 do not have VOPs the get handled by the nary ops struct. */
1116 vn_nary_op_t result;
1117 unsigned int result_id;
1118 vn_nary_op_lookup (t, &result);
1119 if (result != NULL)
1121 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1122 e->kind = NARY;
1123 PRE_EXPR_NARY (e) = result;
1124 result_id = lookup_expression_id (e);
1125 if (result_id != 0)
1127 pool_free (pre_expr_pool, e);
1128 e = expression_for_id (result_id);
1129 return e;
1131 alloc_expression_id (e);
1132 return e;
1135 return NULL;
1138 /* Return the folded version of T if T, when folded, is a gimple
1139 min_invariant. Otherwise, return T. */
1141 static pre_expr
1142 fully_constant_expression (pre_expr e)
1144 switch (e->kind)
1146 case CONSTANT:
1147 return e;
1148 case NARY:
1150 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1151 switch (TREE_CODE_CLASS (nary->opcode))
1153 case tcc_expression:
1154 if (nary->opcode == TRUTH_NOT_EXPR)
1155 goto do_unary;
1156 if (nary->opcode != TRUTH_AND_EXPR
1157 && nary->opcode != TRUTH_OR_EXPR
1158 && nary->opcode != TRUTH_XOR_EXPR)
1159 return e;
1160 /* Fallthrough. */
1161 case tcc_binary:
1162 case tcc_comparison:
1164 /* We have to go from trees to pre exprs to value ids to
1165 constants. */
1166 tree naryop0 = nary->op[0];
1167 tree naryop1 = nary->op[1];
1168 tree result;
1169 if (!is_gimple_min_invariant (naryop0))
1171 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1172 unsigned int vrep0 = get_expr_value_id (rep0);
1173 tree const0 = get_constant_for_value_id (vrep0);
1174 if (const0)
1175 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1177 if (!is_gimple_min_invariant (naryop1))
1179 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1180 unsigned int vrep1 = get_expr_value_id (rep1);
1181 tree const1 = get_constant_for_value_id (vrep1);
1182 if (const1)
1183 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1185 result = fold_binary (nary->opcode, nary->type,
1186 naryop0, naryop1);
1187 if (result && is_gimple_min_invariant (result))
1188 return get_or_alloc_expr_for_constant (result);
1189 /* We might have simplified the expression to a
1190 SSA_NAME for example from x_1 * 1. But we cannot
1191 insert a PHI for x_1 unconditionally as x_1 might
1192 not be available readily. */
1193 return e;
1195 case tcc_reference:
1196 if (nary->opcode != REALPART_EXPR
1197 && nary->opcode != IMAGPART_EXPR
1198 && nary->opcode != VIEW_CONVERT_EXPR)
1199 return e;
1200 /* Fallthrough. */
1201 case tcc_unary:
1202 do_unary:
1204 /* We have to go from trees to pre exprs to value ids to
1205 constants. */
1206 tree naryop0 = nary->op[0];
1207 tree const0, result;
1208 if (is_gimple_min_invariant (naryop0))
1209 const0 = naryop0;
1210 else
1212 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1213 unsigned int vrep0 = get_expr_value_id (rep0);
1214 const0 = get_constant_for_value_id (vrep0);
1216 result = NULL;
1217 if (const0)
1219 tree type1 = TREE_TYPE (nary->op[0]);
1220 const0 = fold_convert (type1, const0);
1221 result = fold_unary (nary->opcode, nary->type, const0);
1223 if (result && is_gimple_min_invariant (result))
1224 return get_or_alloc_expr_for_constant (result);
1225 return e;
1227 default:
1228 return e;
1231 case REFERENCE:
1233 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1234 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1235 vn_reference_op_t op;
1237 /* Try to simplify the translated expression if it is
1238 a call to a builtin function with at most two arguments. */
1239 op = VEC_index (vn_reference_op_s, operands, 0);
1240 if (op->opcode == CALL_EXPR
1241 && TREE_CODE (op->op0) == ADDR_EXPR
1242 && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL
1243 && DECL_BUILT_IN (TREE_OPERAND (op->op0, 0))
1244 && VEC_length (vn_reference_op_s, operands) >= 2
1245 && VEC_length (vn_reference_op_s, operands) <= 3)
1247 vn_reference_op_t arg0, arg1 = NULL;
1248 bool anyconst = false;
1249 arg0 = VEC_index (vn_reference_op_s, operands, 1);
1250 if (VEC_length (vn_reference_op_s, operands) > 2)
1251 arg1 = VEC_index (vn_reference_op_s, operands, 2);
1252 if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant
1253 || (arg0->opcode == ADDR_EXPR
1254 && is_gimple_min_invariant (arg0->op0)))
1255 anyconst = true;
1256 if (arg1
1257 && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant
1258 || (arg1->opcode == ADDR_EXPR
1259 && is_gimple_min_invariant (arg1->op0))))
1260 anyconst = true;
1261 if (anyconst)
1263 tree folded = build_call_expr (TREE_OPERAND (op->op0, 0),
1264 arg1 ? 2 : 1,
1265 arg0->op0,
1266 arg1 ? arg1->op0 : NULL);
1267 if (folded
1268 && TREE_CODE (folded) == NOP_EXPR)
1269 folded = TREE_OPERAND (folded, 0);
1270 if (folded
1271 && is_gimple_min_invariant (folded))
1272 return get_or_alloc_expr_for_constant (folded);
1275 return e;
1277 default:
1278 return e;
1280 return e;
1283 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1284 it has the value it would have in BLOCK. Set *SAME_VALID to true
1285 in case the new vuse doesn't change the value id of the OPERANDS. */
1287 static tree
1288 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1289 alias_set_type set, tree type, tree vuse,
1290 basic_block phiblock,
1291 basic_block block, bool *same_valid)
1293 gimple phi = SSA_NAME_DEF_STMT (vuse);
1294 ao_ref ref;
1295 edge e = NULL;
1296 bool use_oracle;
1298 *same_valid = true;
1300 if (gimple_bb (phi) != phiblock)
1301 return vuse;
1303 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1305 /* Use the alias-oracle to find either the PHI node in this block,
1306 the first VUSE used in this block that is equivalent to vuse or
1307 the first VUSE which definition in this block kills the value. */
1308 if (gimple_code (phi) == GIMPLE_PHI)
1309 e = find_edge (block, phiblock);
1310 else if (use_oracle)
1311 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1313 vuse = gimple_vuse (phi);
1314 phi = SSA_NAME_DEF_STMT (vuse);
1315 if (gimple_bb (phi) != phiblock)
1316 return vuse;
1317 if (gimple_code (phi) == GIMPLE_PHI)
1319 e = find_edge (block, phiblock);
1320 break;
1323 else
1324 return NULL_TREE;
1326 if (e)
1328 if (use_oracle)
1330 bitmap visited = NULL;
1331 /* Try to find a vuse that dominates this phi node by skipping
1332 non-clobbering statements. */
1333 vuse = get_continuation_for_phi (phi, &ref, &visited);
1334 if (visited)
1335 BITMAP_FREE (visited);
1337 else
1338 vuse = NULL_TREE;
1339 if (!vuse)
1341 /* If we didn't find any, the value ID can't stay the same,
1342 but return the translated vuse. */
1343 *same_valid = false;
1344 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1346 /* ??? We would like to return vuse here as this is the canonical
1347 upmost vdef that this reference is associated with. But during
1348 insertion of the references into the hash tables we only ever
1349 directly insert with their direct gimple_vuse, hence returning
1350 something else would make us not find the other expression. */
1351 return PHI_ARG_DEF (phi, e->dest_idx);
1354 return NULL_TREE;
1357 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1358 SET2. This is used to avoid making a set consisting of the union
1359 of PA_IN and ANTIC_IN during insert. */
1361 static inline pre_expr
1362 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1364 pre_expr result;
1366 result = bitmap_find_leader (set1, val, NULL);
1367 if (!result && set2)
1368 result = bitmap_find_leader (set2, val, NULL);
1369 return result;
1372 /* Get the tree type for our PRE expression e. */
1374 static tree
1375 get_expr_type (const pre_expr e)
1377 switch (e->kind)
1379 case NAME:
1380 return TREE_TYPE (PRE_EXPR_NAME (e));
1381 case CONSTANT:
1382 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1383 case REFERENCE:
1384 return PRE_EXPR_REFERENCE (e)->type;
1385 case NARY:
1386 return PRE_EXPR_NARY (e)->type;
1388 gcc_unreachable();
1391 /* Get a representative SSA_NAME for a given expression.
1392 Since all of our sub-expressions are treated as values, we require
1393 them to be SSA_NAME's for simplicity.
1394 Prior versions of GVNPRE used to use "value handles" here, so that
1395 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1396 either case, the operands are really values (IE we do not expect
1397 them to be usable without finding leaders). */
1399 static tree
1400 get_representative_for (const pre_expr e)
1402 tree exprtype;
1403 tree name;
1404 unsigned int value_id = get_expr_value_id (e);
1406 switch (e->kind)
1408 case NAME:
1409 return PRE_EXPR_NAME (e);
1410 case CONSTANT:
1411 return PRE_EXPR_CONSTANT (e);
1412 case NARY:
1413 case REFERENCE:
1415 /* Go through all of the expressions representing this value
1416 and pick out an SSA_NAME. */
1417 unsigned int i;
1418 bitmap_iterator bi;
1419 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1420 value_id);
1421 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1423 pre_expr rep = expression_for_id (i);
1424 if (rep->kind == NAME)
1425 return PRE_EXPR_NAME (rep);
1428 break;
1430 /* If we reached here we couldn't find an SSA_NAME. This can
1431 happen when we've discovered a value that has never appeared in
1432 the program as set to an SSA_NAME, most likely as the result of
1433 phi translation. */
1434 if (dump_file)
1436 fprintf (dump_file,
1437 "Could not find SSA_NAME representative for expression:");
1438 print_pre_expr (dump_file, e);
1439 fprintf (dump_file, "\n");
1442 exprtype = get_expr_type (e);
1444 /* Build and insert the assignment of the end result to the temporary
1445 that we will return. */
1446 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1448 pretemp = create_tmp_var (exprtype, "pretmp");
1449 get_var_ann (pretemp);
1452 name = make_ssa_name (pretemp, gimple_build_nop ());
1453 VN_INFO_GET (name)->value_id = value_id;
1454 if (e->kind == CONSTANT)
1455 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1456 else
1457 VN_INFO (name)->valnum = name;
1459 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1460 if (dump_file)
1462 fprintf (dump_file, "Created SSA_NAME representative ");
1463 print_generic_expr (dump_file, name, 0);
1464 fprintf (dump_file, " for expression:");
1465 print_pre_expr (dump_file, e);
1466 fprintf (dump_file, "\n");
1469 return name;
1475 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1476 the phis in PRED. Return NULL if we can't find a leader for each part
1477 of the translated expression. */
1479 static pre_expr
1480 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1481 basic_block pred, basic_block phiblock)
1483 pre_expr oldexpr = expr;
1484 pre_expr phitrans;
1486 if (!expr)
1487 return NULL;
1489 /* Constants contain no values that need translation. */
1490 if (expr->kind == CONSTANT)
1491 return expr;
1493 if (value_id_constant_p (get_expr_value_id (expr)))
1494 return expr;
1496 phitrans = phi_trans_lookup (expr, pred);
1497 if (phitrans)
1498 return phitrans;
1500 switch (expr->kind)
1502 case NARY:
1504 unsigned int i;
1505 bool changed = false;
1506 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1507 struct vn_nary_op_s newnary;
1508 /* The NARY structure is only guaranteed to have been
1509 allocated to the nary->length operands. */
1510 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1511 - sizeof (tree) * (4 - nary->length)));
1513 for (i = 0; i < newnary.length; i++)
1515 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1516 continue;
1517 else
1519 pre_expr leader, result;
1520 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1521 leader = find_leader_in_sets (op_val_id, set1, set2);
1522 result = phi_translate (leader, set1, set2, pred, phiblock);
1523 if (result && result != leader)
1525 tree name = get_representative_for (result);
1526 if (!name)
1527 return NULL;
1528 newnary.op[i] = name;
1530 else if (!result)
1531 return NULL;
1533 changed |= newnary.op[i] != nary->op[i];
1536 if (changed)
1538 pre_expr constant;
1539 unsigned int new_val_id;
1541 tree result = vn_nary_op_lookup_pieces (newnary.length,
1542 newnary.opcode,
1543 newnary.type,
1544 newnary.op[0],
1545 newnary.op[1],
1546 newnary.op[2],
1547 newnary.op[3],
1548 &nary);
1549 if (result && is_gimple_min_invariant (result))
1550 return get_or_alloc_expr_for_constant (result);
1552 expr = (pre_expr) pool_alloc (pre_expr_pool);
1553 expr->kind = NARY;
1554 expr->id = 0;
1555 if (nary)
1557 PRE_EXPR_NARY (expr) = nary;
1558 constant = fully_constant_expression (expr);
1559 if (constant != expr)
1560 return constant;
1562 new_val_id = nary->value_id;
1563 get_or_alloc_expression_id (expr);
1565 else
1567 new_val_id = get_next_value_id ();
1568 VEC_safe_grow_cleared (bitmap_set_t, heap,
1569 value_expressions,
1570 get_max_value_id() + 1);
1571 nary = vn_nary_op_insert_pieces (newnary.length,
1572 newnary.opcode,
1573 newnary.type,
1574 newnary.op[0],
1575 newnary.op[1],
1576 newnary.op[2],
1577 newnary.op[3],
1578 result, new_val_id);
1579 PRE_EXPR_NARY (expr) = nary;
1580 constant = fully_constant_expression (expr);
1581 if (constant != expr)
1582 return constant;
1583 get_or_alloc_expression_id (expr);
1585 add_to_value (new_val_id, expr);
1587 phi_trans_add (oldexpr, expr, pred);
1588 return expr;
1590 break;
1592 case REFERENCE:
1594 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1595 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1596 tree vuse = ref->vuse;
1597 tree newvuse = vuse;
1598 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1599 bool changed = false, same_valid = true;
1600 unsigned int i, j;
1601 vn_reference_op_t operand;
1602 vn_reference_t newref;
1604 for (i = 0, j = 0;
1605 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1607 pre_expr opresult;
1608 pre_expr leader;
1609 tree oldop0 = operand->op0;
1610 tree oldop1 = operand->op1;
1611 tree oldop2 = operand->op2;
1612 tree op0 = oldop0;
1613 tree op1 = oldop1;
1614 tree op2 = oldop2;
1615 tree type = operand->type;
1616 vn_reference_op_s newop = *operand;
1618 if (op0 && TREE_CODE (op0) == SSA_NAME)
1620 unsigned int op_val_id = VN_INFO (op0)->value_id;
1621 leader = find_leader_in_sets (op_val_id, set1, set2);
1622 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1623 if (opresult && opresult != leader)
1625 tree name = get_representative_for (opresult);
1626 if (!name)
1627 break;
1628 op0 = name;
1630 else if (!opresult)
1631 break;
1633 changed |= op0 != oldop0;
1635 if (op1 && TREE_CODE (op1) == SSA_NAME)
1637 unsigned int op_val_id = VN_INFO (op1)->value_id;
1638 leader = find_leader_in_sets (op_val_id, set1, set2);
1639 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1640 if (opresult && opresult != leader)
1642 tree name = get_representative_for (opresult);
1643 if (!name)
1644 break;
1645 op1 = name;
1647 else if (!opresult)
1648 break;
1650 /* We can't possibly insert these. */
1651 else if (op1 && !is_gimple_min_invariant (op1))
1652 break;
1653 changed |= op1 != oldop1;
1654 if (op2 && TREE_CODE (op2) == SSA_NAME)
1656 unsigned int op_val_id = VN_INFO (op2)->value_id;
1657 leader = find_leader_in_sets (op_val_id, set1, set2);
1658 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1659 if (opresult && opresult != leader)
1661 tree name = get_representative_for (opresult);
1662 if (!name)
1663 break;
1664 op2 = name;
1666 else if (!opresult)
1667 break;
1669 /* We can't possibly insert these. */
1670 else if (op2 && !is_gimple_min_invariant (op2))
1671 break;
1672 changed |= op2 != oldop2;
1674 if (!newoperands)
1675 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1676 /* We may have changed from an SSA_NAME to a constant */
1677 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1678 newop.opcode = TREE_CODE (op0);
1679 newop.type = type;
1680 newop.op0 = op0;
1681 newop.op1 = op1;
1682 newop.op2 = op2;
1683 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1684 /* If it transforms from an SSA_NAME to an address, fold with
1685 a preceding indirect reference. */
1686 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1687 && VEC_index (vn_reference_op_s,
1688 newoperands, j - 1)->opcode == INDIRECT_REF)
1689 vn_reference_fold_indirect (&newoperands, &j);
1691 if (i != VEC_length (vn_reference_op_s, operands))
1693 if (newoperands)
1694 VEC_free (vn_reference_op_s, heap, newoperands);
1695 return NULL;
1698 if (vuse)
1700 newvuse = translate_vuse_through_block (newoperands,
1701 ref->set, ref->type,
1702 vuse, phiblock, pred,
1703 &same_valid);
1704 if (newvuse == NULL_TREE)
1706 VEC_free (vn_reference_op_s, heap, newoperands);
1707 return NULL;
1711 if (changed || newvuse != vuse)
1713 unsigned int new_val_id;
1714 pre_expr constant;
1716 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1717 ref->type,
1718 newoperands,
1719 &newref, true);
1720 if (newref)
1721 VEC_free (vn_reference_op_s, heap, newoperands);
1723 if (result && is_gimple_min_invariant (result))
1725 gcc_assert (!newoperands);
1726 return get_or_alloc_expr_for_constant (result);
1729 expr = (pre_expr) pool_alloc (pre_expr_pool);
1730 expr->kind = REFERENCE;
1731 expr->id = 0;
1733 if (newref)
1735 PRE_EXPR_REFERENCE (expr) = newref;
1736 constant = fully_constant_expression (expr);
1737 if (constant != expr)
1738 return constant;
1740 new_val_id = newref->value_id;
1741 get_or_alloc_expression_id (expr);
1743 else
1745 if (changed || !same_valid)
1747 new_val_id = get_next_value_id ();
1748 VEC_safe_grow_cleared (bitmap_set_t, heap,
1749 value_expressions,
1750 get_max_value_id() + 1);
1752 else
1753 new_val_id = ref->value_id;
1754 newref = vn_reference_insert_pieces (newvuse, ref->set,
1755 ref->type,
1756 newoperands,
1757 result, new_val_id);
1758 newoperands = NULL;
1759 PRE_EXPR_REFERENCE (expr) = newref;
1760 constant = fully_constant_expression (expr);
1761 if (constant != expr)
1762 return constant;
1763 get_or_alloc_expression_id (expr);
1765 add_to_value (new_val_id, expr);
1767 VEC_free (vn_reference_op_s, heap, newoperands);
1768 phi_trans_add (oldexpr, expr, pred);
1769 return expr;
1771 break;
1773 case NAME:
1775 gimple phi = NULL;
1776 edge e;
1777 gimple def_stmt;
1778 tree name = PRE_EXPR_NAME (expr);
1780 def_stmt = SSA_NAME_DEF_STMT (name);
1781 if (gimple_code (def_stmt) == GIMPLE_PHI
1782 && gimple_bb (def_stmt) == phiblock)
1783 phi = def_stmt;
1784 else
1785 return expr;
1787 e = find_edge (pred, gimple_bb (phi));
1788 if (e)
1790 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1791 pre_expr newexpr;
1793 if (TREE_CODE (def) == SSA_NAME)
1794 def = VN_INFO (def)->valnum;
1796 /* Handle constant. */
1797 if (is_gimple_min_invariant (def))
1798 return get_or_alloc_expr_for_constant (def);
1800 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1801 return NULL;
1803 newexpr = get_or_alloc_expr_for_name (def);
1804 return newexpr;
1807 return expr;
1809 default:
1810 gcc_unreachable ();
1814 /* For each expression in SET, translate the values through phi nodes
1815 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1816 expressions in DEST. */
1818 static void
1819 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1820 basic_block phiblock)
1822 VEC (pre_expr, heap) *exprs;
1823 pre_expr expr;
1824 int i;
1826 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1828 bitmap_set_copy (dest, set);
1829 return;
1832 exprs = sorted_array_from_bitmap_set (set);
1833 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1835 pre_expr translated;
1836 translated = phi_translate (expr, set, NULL, pred, phiblock);
1838 /* Don't add empty translations to the cache */
1839 if (!translated)
1840 continue;
1842 phi_trans_add (expr, translated, pred);
1844 /* We might end up with multiple expressions from SET being
1845 translated to the same value. In this case we do not want
1846 to retain the NARY or REFERENCE expression but prefer a NAME
1847 which would be the leader. */
1848 if (translated->kind == NAME)
1849 bitmap_value_replace_in_set (dest, translated);
1850 else
1851 bitmap_value_insert_into_set (dest, translated);
1853 VEC_free (pre_expr, heap, exprs);
1856 /* Find the leader for a value (i.e., the name representing that
1857 value) in a given set, and return it. If STMT is non-NULL it
1858 makes sure the defining statement for the leader dominates it.
1859 Return NULL if no leader is found. */
1861 static pre_expr
1862 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1864 if (value_id_constant_p (val))
1866 unsigned int i;
1867 bitmap_iterator bi;
1868 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1870 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1872 pre_expr expr = expression_for_id (i);
1873 if (expr->kind == CONSTANT)
1874 return expr;
1877 if (bitmap_set_contains_value (set, val))
1879 /* Rather than walk the entire bitmap of expressions, and see
1880 whether any of them has the value we are looking for, we look
1881 at the reverse mapping, which tells us the set of expressions
1882 that have a given value (IE value->expressions with that
1883 value) and see if any of those expressions are in our set.
1884 The number of expressions per value is usually significantly
1885 less than the number of expressions in the set. In fact, for
1886 large testcases, doing it this way is roughly 5-10x faster
1887 than walking the bitmap.
1888 If this is somehow a significant lose for some cases, we can
1889 choose which set to walk based on which set is smaller. */
1890 unsigned int i;
1891 bitmap_iterator bi;
1892 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1894 EXECUTE_IF_AND_IN_BITMAP (exprset->expressions,
1895 set->expressions, 0, i, bi)
1897 pre_expr val = expression_for_id (i);
1898 /* At the point where stmt is not null, there should always
1899 be an SSA_NAME first in the list of expressions. */
1900 if (stmt)
1902 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1903 if (gimple_code (def_stmt) != GIMPLE_PHI
1904 && gimple_bb (def_stmt) == gimple_bb (stmt)
1905 && gimple_uid (def_stmt) >= gimple_uid (stmt))
1906 continue;
1908 return val;
1911 return NULL;
1914 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1915 BLOCK by seeing if it is not killed in the block. Note that we are
1916 only determining whether there is a store that kills it. Because
1917 of the order in which clean iterates over values, we are guaranteed
1918 that altered operands will have caused us to be eliminated from the
1919 ANTIC_IN set already. */
1921 static bool
1922 value_dies_in_block_x (pre_expr expr, basic_block block)
1924 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1925 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1926 gimple def;
1927 gimple_stmt_iterator gsi;
1928 unsigned id = get_expression_id (expr);
1929 bool res = false;
1930 ao_ref ref;
1932 if (!vuse)
1933 return false;
1935 /* Lookup a previously calculated result. */
1936 if (EXPR_DIES (block)
1937 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1938 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1940 /* A memory expression {e, VUSE} dies in the block if there is a
1941 statement that may clobber e. If, starting statement walk from the
1942 top of the basic block, a statement uses VUSE there can be no kill
1943 inbetween that use and the original statement that loaded {e, VUSE},
1944 so we can stop walking. */
1945 ref.base = NULL_TREE;
1946 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1948 tree def_vuse, def_vdef;
1949 def = gsi_stmt (gsi);
1950 def_vuse = gimple_vuse (def);
1951 def_vdef = gimple_vdef (def);
1953 /* Not a memory statement. */
1954 if (!def_vuse)
1955 continue;
1957 /* Not a may-def. */
1958 if (!def_vdef)
1960 /* A load with the same VUSE, we're done. */
1961 if (def_vuse == vuse)
1962 break;
1964 continue;
1967 /* Init ref only if we really need it. */
1968 if (ref.base == NULL_TREE
1969 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1970 refx->operands))
1972 res = true;
1973 break;
1975 /* If the statement may clobber expr, it dies. */
1976 if (stmt_may_clobber_ref_p_1 (def, &ref))
1978 res = true;
1979 break;
1983 /* Remember the result. */
1984 if (!EXPR_DIES (block))
1985 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1986 bitmap_set_bit (EXPR_DIES (block), id * 2);
1987 if (res)
1988 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1990 return res;
1994 #define union_contains_value(SET1, SET2, VAL) \
1995 (bitmap_set_contains_value ((SET1), (VAL)) \
1996 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
1998 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2000 static bool
2001 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2002 vn_reference_op_t vro)
2004 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2006 struct pre_expr_d temp;
2007 temp.kind = NAME;
2008 temp.id = 0;
2009 PRE_EXPR_NAME (&temp) = vro->op0;
2010 temp.id = lookup_expression_id (&temp);
2011 if (temp.id == 0)
2012 return false;
2013 if (!union_contains_value (set1, set2,
2014 get_expr_value_id (&temp)))
2015 return false;
2017 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2019 struct pre_expr_d temp;
2020 temp.kind = NAME;
2021 temp.id = 0;
2022 PRE_EXPR_NAME (&temp) = vro->op1;
2023 temp.id = lookup_expression_id (&temp);
2024 if (temp.id == 0)
2025 return false;
2026 if (!union_contains_value (set1, set2,
2027 get_expr_value_id (&temp)))
2028 return false;
2031 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2033 struct pre_expr_d temp;
2034 temp.kind = NAME;
2035 temp.id = 0;
2036 PRE_EXPR_NAME (&temp) = vro->op2;
2037 temp.id = lookup_expression_id (&temp);
2038 if (temp.id == 0)
2039 return false;
2040 if (!union_contains_value (set1, set2,
2041 get_expr_value_id (&temp)))
2042 return false;
2045 return true;
2048 /* Determine if the expression EXPR is valid in SET1 U SET2.
2049 ONLY SET2 CAN BE NULL.
2050 This means that we have a leader for each part of the expression
2051 (if it consists of values), or the expression is an SSA_NAME.
2052 For loads/calls, we also see if the vuse is killed in this block. */
2054 static bool
2055 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2056 basic_block block)
2058 switch (expr->kind)
2060 case NAME:
2061 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2062 case NARY:
2064 unsigned int i;
2065 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2066 for (i = 0; i < nary->length; i++)
2068 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2070 struct pre_expr_d temp;
2071 temp.kind = NAME;
2072 temp.id = 0;
2073 PRE_EXPR_NAME (&temp) = nary->op[i];
2074 temp.id = lookup_expression_id (&temp);
2075 if (temp.id == 0)
2076 return false;
2077 if (!union_contains_value (set1, set2,
2078 get_expr_value_id (&temp)))
2079 return false;
2082 /* If the NARY may trap make sure the block does not contain
2083 a possible exit point.
2084 ??? This is overly conservative if we translate AVAIL_OUT
2085 as the available expression might be after the exit point. */
2086 if (BB_MAY_NOTRETURN (block)
2087 && vn_nary_may_trap (nary))
2088 return false;
2089 return true;
2091 break;
2092 case REFERENCE:
2094 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2095 vn_reference_op_t vro;
2096 unsigned int i;
2098 for (i = 0; VEC_iterate (vn_reference_op_s, ref->operands, i, vro); i++)
2100 if (!vro_valid_in_sets (set1, set2, vro))
2101 return false;
2103 if (ref->vuse)
2105 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2106 if (!gimple_nop_p (def_stmt)
2107 && gimple_bb (def_stmt) != block
2108 && !dominated_by_p (CDI_DOMINATORS,
2109 block, gimple_bb (def_stmt)))
2110 return false;
2112 return !value_dies_in_block_x (expr, block);
2114 default:
2115 gcc_unreachable ();
2119 /* Clean the set of expressions that are no longer valid in SET1 or
2120 SET2. This means expressions that are made up of values we have no
2121 leaders for in SET1 or SET2. This version is used for partial
2122 anticipation, which means it is not valid in either ANTIC_IN or
2123 PA_IN. */
2125 static void
2126 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2128 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2129 pre_expr expr;
2130 int i;
2132 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2134 if (!valid_in_sets (set1, set2, expr, block))
2135 bitmap_remove_from_set (set1, expr);
2137 VEC_free (pre_expr, heap, exprs);
2140 /* Clean the set of expressions that are no longer valid in SET. This
2141 means expressions that are made up of values we have no leaders for
2142 in SET. */
2144 static void
2145 clean (bitmap_set_t set, basic_block block)
2147 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2148 pre_expr expr;
2149 int i;
2151 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2153 if (!valid_in_sets (set, NULL, expr, block))
2154 bitmap_remove_from_set (set, expr);
2156 VEC_free (pre_expr, heap, exprs);
2159 static sbitmap has_abnormal_preds;
2161 /* List of blocks that may have changed during ANTIC computation and
2162 thus need to be iterated over. */
2164 static sbitmap changed_blocks;
2166 /* Decide whether to defer a block for a later iteration, or PHI
2167 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2168 should defer the block, and true if we processed it. */
2170 static bool
2171 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2172 basic_block block, basic_block phiblock)
2174 if (!BB_VISITED (phiblock))
2176 SET_BIT (changed_blocks, block->index);
2177 BB_VISITED (block) = 0;
2178 BB_DEFERRED (block) = 1;
2179 return false;
2181 else
2182 phi_translate_set (dest, source, block, phiblock);
2183 return true;
2186 /* Compute the ANTIC set for BLOCK.
2188 If succs(BLOCK) > 1 then
2189 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2190 else if succs(BLOCK) == 1 then
2191 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2193 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2196 static bool
2197 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2199 bool changed = false;
2200 bitmap_set_t S, old, ANTIC_OUT;
2201 bitmap_iterator bi;
2202 unsigned int bii;
2203 edge e;
2204 edge_iterator ei;
2206 old = ANTIC_OUT = S = NULL;
2207 BB_VISITED (block) = 1;
2209 /* If any edges from predecessors are abnormal, antic_in is empty,
2210 so do nothing. */
2211 if (block_has_abnormal_pred_edge)
2212 goto maybe_dump_sets;
2214 old = ANTIC_IN (block);
2215 ANTIC_OUT = bitmap_set_new ();
2217 /* If the block has no successors, ANTIC_OUT is empty. */
2218 if (EDGE_COUNT (block->succs) == 0)
2220 /* If we have one successor, we could have some phi nodes to
2221 translate through. */
2222 else if (single_succ_p (block))
2224 basic_block succ_bb = single_succ (block);
2226 /* We trade iterations of the dataflow equations for having to
2227 phi translate the maximal set, which is incredibly slow
2228 (since the maximal set often has 300+ members, even when you
2229 have a small number of blocks).
2230 Basically, we defer the computation of ANTIC for this block
2231 until we have processed it's successor, which will inevitably
2232 have a *much* smaller set of values to phi translate once
2233 clean has been run on it.
2234 The cost of doing this is that we technically perform more
2235 iterations, however, they are lower cost iterations.
2237 Timings for PRE on tramp3d-v4:
2238 without maximal set fix: 11 seconds
2239 with maximal set fix/without deferring: 26 seconds
2240 with maximal set fix/with deferring: 11 seconds
2243 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2244 block, succ_bb))
2246 changed = true;
2247 goto maybe_dump_sets;
2250 /* If we have multiple successors, we take the intersection of all of
2251 them. Note that in the case of loop exit phi nodes, we may have
2252 phis to translate through. */
2253 else
2255 VEC(basic_block, heap) * worklist;
2256 size_t i;
2257 basic_block bprime, first = NULL;
2259 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2260 FOR_EACH_EDGE (e, ei, block->succs)
2262 if (!first
2263 && BB_VISITED (e->dest))
2264 first = e->dest;
2265 else if (BB_VISITED (e->dest))
2266 VEC_quick_push (basic_block, worklist, e->dest);
2269 /* Of multiple successors we have to have visited one already. */
2270 if (!first)
2272 SET_BIT (changed_blocks, block->index);
2273 BB_VISITED (block) = 0;
2274 BB_DEFERRED (block) = 1;
2275 changed = true;
2276 VEC_free (basic_block, heap, worklist);
2277 goto maybe_dump_sets;
2280 if (!gimple_seq_empty_p (phi_nodes (first)))
2281 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2282 else
2283 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2285 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2287 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2289 bitmap_set_t tmp = bitmap_set_new ();
2290 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2291 bitmap_set_and (ANTIC_OUT, tmp);
2292 bitmap_set_free (tmp);
2294 else
2295 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2297 VEC_free (basic_block, heap, worklist);
2300 /* Generate ANTIC_OUT - TMP_GEN. */
2301 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2303 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2304 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2305 TMP_GEN (block));
2307 /* Then union in the ANTIC_OUT - TMP_GEN values,
2308 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2309 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2310 bitmap_value_insert_into_set (ANTIC_IN (block),
2311 expression_for_id (bii));
2313 clean (ANTIC_IN (block), block);
2315 /* !old->expressions can happen when we deferred a block. */
2316 if (!old->expressions || !bitmap_set_equal (old, ANTIC_IN (block)))
2318 changed = true;
2319 SET_BIT (changed_blocks, block->index);
2320 FOR_EACH_EDGE (e, ei, block->preds)
2321 SET_BIT (changed_blocks, e->src->index);
2323 else
2324 RESET_BIT (changed_blocks, block->index);
2326 maybe_dump_sets:
2327 if (dump_file && (dump_flags & TDF_DETAILS))
2329 if (!BB_DEFERRED (block) || BB_VISITED (block))
2331 if (ANTIC_OUT)
2332 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2334 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2335 block->index);
2337 if (S)
2338 print_bitmap_set (dump_file, S, "S", block->index);
2340 else
2342 fprintf (dump_file,
2343 "Block %d was deferred for a future iteration.\n",
2344 block->index);
2347 if (old)
2348 bitmap_set_free (old);
2349 if (S)
2350 bitmap_set_free (S);
2351 if (ANTIC_OUT)
2352 bitmap_set_free (ANTIC_OUT);
2353 return changed;
2356 /* Compute PARTIAL_ANTIC for BLOCK.
2358 If succs(BLOCK) > 1 then
2359 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2360 in ANTIC_OUT for all succ(BLOCK)
2361 else if succs(BLOCK) == 1 then
2362 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2364 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2365 - ANTIC_IN[BLOCK])
2368 static bool
2369 compute_partial_antic_aux (basic_block block,
2370 bool block_has_abnormal_pred_edge)
2372 bool changed = false;
2373 bitmap_set_t old_PA_IN;
2374 bitmap_set_t PA_OUT;
2375 edge e;
2376 edge_iterator ei;
2377 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2379 old_PA_IN = PA_OUT = NULL;
2381 /* If any edges from predecessors are abnormal, antic_in is empty,
2382 so do nothing. */
2383 if (block_has_abnormal_pred_edge)
2384 goto maybe_dump_sets;
2386 /* If there are too many partially anticipatable values in the
2387 block, phi_translate_set can take an exponential time: stop
2388 before the translation starts. */
2389 if (max_pa
2390 && single_succ_p (block)
2391 && bitmap_count_bits (PA_IN (single_succ (block))->values) > max_pa)
2392 goto maybe_dump_sets;
2394 old_PA_IN = PA_IN (block);
2395 PA_OUT = bitmap_set_new ();
2397 /* If the block has no successors, ANTIC_OUT is empty. */
2398 if (EDGE_COUNT (block->succs) == 0)
2400 /* If we have one successor, we could have some phi nodes to
2401 translate through. Note that we can't phi translate across DFS
2402 back edges in partial antic, because it uses a union operation on
2403 the successors. For recurrences like IV's, we will end up
2404 generating a new value in the set on each go around (i + 3 (VH.1)
2405 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2406 else if (single_succ_p (block))
2408 basic_block succ = single_succ (block);
2409 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2410 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2412 /* If we have multiple successors, we take the union of all of
2413 them. */
2414 else
2416 VEC(basic_block, heap) * worklist;
2417 size_t i;
2418 basic_block bprime;
2420 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2421 FOR_EACH_EDGE (e, ei, block->succs)
2423 if (e->flags & EDGE_DFS_BACK)
2424 continue;
2425 VEC_quick_push (basic_block, worklist, e->dest);
2427 if (VEC_length (basic_block, worklist) > 0)
2429 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2431 unsigned int i;
2432 bitmap_iterator bi;
2434 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2435 bitmap_value_insert_into_set (PA_OUT,
2436 expression_for_id (i));
2437 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2439 bitmap_set_t pa_in = bitmap_set_new ();
2440 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2441 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2442 bitmap_value_insert_into_set (PA_OUT,
2443 expression_for_id (i));
2444 bitmap_set_free (pa_in);
2446 else
2447 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2448 bitmap_value_insert_into_set (PA_OUT,
2449 expression_for_id (i));
2452 VEC_free (basic_block, heap, worklist);
2455 /* PA_IN starts with PA_OUT - TMP_GEN.
2456 Then we subtract things from ANTIC_IN. */
2457 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2459 /* For partial antic, we want to put back in the phi results, since
2460 we will properly avoid making them partially antic over backedges. */
2461 bitmap_ior_into (PA_IN (block)->values, PHI_GEN (block)->values);
2462 bitmap_ior_into (PA_IN (block)->expressions, PHI_GEN (block)->expressions);
2464 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2465 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2467 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2469 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2471 changed = true;
2472 SET_BIT (changed_blocks, block->index);
2473 FOR_EACH_EDGE (e, ei, block->preds)
2474 SET_BIT (changed_blocks, e->src->index);
2476 else
2477 RESET_BIT (changed_blocks, block->index);
2479 maybe_dump_sets:
2480 if (dump_file && (dump_flags & TDF_DETAILS))
2482 if (PA_OUT)
2483 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2485 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2487 if (old_PA_IN)
2488 bitmap_set_free (old_PA_IN);
2489 if (PA_OUT)
2490 bitmap_set_free (PA_OUT);
2491 return changed;
2494 /* Compute ANTIC and partial ANTIC sets. */
2496 static void
2497 compute_antic (void)
2499 bool changed = true;
2500 int num_iterations = 0;
2501 basic_block block;
2502 int i;
2504 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2505 We pre-build the map of blocks with incoming abnormal edges here. */
2506 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2507 sbitmap_zero (has_abnormal_preds);
2509 FOR_EACH_BB (block)
2511 edge_iterator ei;
2512 edge e;
2514 FOR_EACH_EDGE (e, ei, block->preds)
2516 e->flags &= ~EDGE_DFS_BACK;
2517 if (e->flags & EDGE_ABNORMAL)
2519 SET_BIT (has_abnormal_preds, block->index);
2520 break;
2524 BB_VISITED (block) = 0;
2525 BB_DEFERRED (block) = 0;
2527 /* While we are here, give empty ANTIC_IN sets to each block. */
2528 ANTIC_IN (block) = bitmap_set_new ();
2529 PA_IN (block) = bitmap_set_new ();
2532 /* At the exit block we anticipate nothing. */
2533 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2534 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2535 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2537 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2538 sbitmap_ones (changed_blocks);
2539 while (changed)
2541 if (dump_file && (dump_flags & TDF_DETAILS))
2542 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2543 num_iterations++;
2544 changed = false;
2545 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2547 if (TEST_BIT (changed_blocks, postorder[i]))
2549 basic_block block = BASIC_BLOCK (postorder[i]);
2550 changed |= compute_antic_aux (block,
2551 TEST_BIT (has_abnormal_preds,
2552 block->index));
2555 #ifdef ENABLE_CHECKING
2556 /* Theoretically possible, but *highly* unlikely. */
2557 gcc_assert (num_iterations < 500);
2558 #endif
2561 statistics_histogram_event (cfun, "compute_antic iterations",
2562 num_iterations);
2564 if (do_partial_partial)
2566 sbitmap_ones (changed_blocks);
2567 mark_dfs_back_edges ();
2568 num_iterations = 0;
2569 changed = true;
2570 while (changed)
2572 if (dump_file && (dump_flags & TDF_DETAILS))
2573 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2574 num_iterations++;
2575 changed = false;
2576 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2578 if (TEST_BIT (changed_blocks, postorder[i]))
2580 basic_block block = BASIC_BLOCK (postorder[i]);
2581 changed
2582 |= compute_partial_antic_aux (block,
2583 TEST_BIT (has_abnormal_preds,
2584 block->index));
2587 #ifdef ENABLE_CHECKING
2588 /* Theoretically possible, but *highly* unlikely. */
2589 gcc_assert (num_iterations < 500);
2590 #endif
2592 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2593 num_iterations);
2595 sbitmap_free (has_abnormal_preds);
2596 sbitmap_free (changed_blocks);
2599 /* Return true if we can value number the call in STMT. This is true
2600 if we have a pure or constant call. */
2602 static bool
2603 can_value_number_call (gimple stmt)
2605 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2606 return true;
2607 return false;
2610 /* Return true if OP is a tree which we can perform PRE on.
2611 This may not match the operations we can value number, but in
2612 a perfect world would. */
2614 static bool
2615 can_PRE_operation (tree op)
2617 return UNARY_CLASS_P (op)
2618 || BINARY_CLASS_P (op)
2619 || COMPARISON_CLASS_P (op)
2620 || TREE_CODE (op) == INDIRECT_REF
2621 || TREE_CODE (op) == COMPONENT_REF
2622 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2623 || TREE_CODE (op) == CALL_EXPR
2624 || TREE_CODE (op) == ARRAY_REF;
2628 /* Inserted expressions are placed onto this worklist, which is used
2629 for performing quick dead code elimination of insertions we made
2630 that didn't turn out to be necessary. */
2631 static VEC(gimple,heap) *inserted_exprs;
2632 static bitmap inserted_phi_names;
2634 /* Pool allocated fake store expressions are placed onto this
2635 worklist, which, after performing dead code elimination, is walked
2636 to see which expressions need to be put into GC'able memory */
2637 static VEC(gimple, heap) *need_creation;
2639 /* The actual worker for create_component_ref_by_pieces. */
2641 static tree
2642 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2643 unsigned int *operand, gimple_seq *stmts,
2644 gimple domstmt)
2646 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2647 *operand);
2648 tree genop;
2649 ++*operand;
2650 switch (currop->opcode)
2652 case CALL_EXPR:
2654 tree folded, sc = currop->op1;
2655 unsigned int nargs = 0;
2656 tree *args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2657 ref->operands) - 1);
2658 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2660 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2661 operand, stmts,
2662 domstmt);
2663 nargs++;
2665 folded = build_call_array (currop->type,
2666 TREE_CODE (currop->op0) == FUNCTION_DECL
2667 ? build_fold_addr_expr (currop->op0)
2668 : currop->op0,
2669 nargs, args);
2670 free (args);
2671 if (sc)
2673 pre_expr scexpr = get_or_alloc_expr_for (sc);
2674 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2675 if (!sc)
2676 return NULL_TREE;
2677 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2679 return folded;
2681 break;
2682 case TARGET_MEM_REF:
2684 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2685 *operand);
2686 pre_expr op0expr;
2687 tree genop0 = NULL_TREE;
2688 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2689 stmts, domstmt);
2690 if (!baseop)
2691 return NULL_TREE;
2692 if (currop->op0)
2694 op0expr = get_or_alloc_expr_for (currop->op0);
2695 genop0 = find_or_generate_expression (block, op0expr,
2696 stmts, domstmt);
2697 if (!genop0)
2698 return NULL_TREE;
2700 if (DECL_P (baseop))
2701 return build6 (TARGET_MEM_REF, currop->type,
2702 baseop, NULL_TREE,
2703 genop0, currop->op1, currop->op2,
2704 unshare_expr (nextop->op1));
2705 else
2706 return build6 (TARGET_MEM_REF, currop->type,
2707 NULL_TREE, baseop,
2708 genop0, currop->op1, currop->op2,
2709 unshare_expr (nextop->op1));
2711 break;
2712 case ADDR_EXPR:
2713 if (currop->op0)
2715 gcc_assert (is_gimple_min_invariant (currop->op0));
2716 return currop->op0;
2718 /* Fallthrough. */
2719 case REALPART_EXPR:
2720 case IMAGPART_EXPR:
2721 case VIEW_CONVERT_EXPR:
2723 tree folded;
2724 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2725 operand,
2726 stmts, domstmt);
2727 if (!genop0)
2728 return NULL_TREE;
2729 folded = fold_build1 (currop->opcode, currop->type,
2730 genop0);
2731 return folded;
2733 break;
2734 case ALIGN_INDIRECT_REF:
2735 case MISALIGNED_INDIRECT_REF:
2736 case INDIRECT_REF:
2738 tree folded;
2739 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2740 operand,
2741 stmts, domstmt);
2742 if (!genop1)
2743 return NULL_TREE;
2744 genop1 = fold_convert (build_pointer_type (currop->type),
2745 genop1);
2747 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2748 folded = fold_build2 (currop->opcode, currop->type,
2749 genop1, currop->op1);
2750 else
2751 folded = fold_build1 (currop->opcode, currop->type,
2752 genop1);
2753 return folded;
2755 break;
2756 case BIT_FIELD_REF:
2758 tree folded;
2759 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2760 stmts, domstmt);
2761 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2762 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2763 tree genop1;
2764 tree genop2;
2766 if (!genop0)
2767 return NULL_TREE;
2768 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2769 if (!genop1)
2770 return NULL_TREE;
2771 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2772 if (!genop2)
2773 return NULL_TREE;
2774 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2775 genop2);
2776 return folded;
2779 /* For array ref vn_reference_op's, operand 1 of the array ref
2780 is op0 of the reference op and operand 3 of the array ref is
2781 op1. */
2782 case ARRAY_RANGE_REF:
2783 case ARRAY_REF:
2785 tree genop0;
2786 tree genop1 = currop->op0;
2787 pre_expr op1expr;
2788 tree genop2 = currop->op1;
2789 pre_expr op2expr;
2790 tree genop3 = currop->op2;
2791 pre_expr op3expr;
2792 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2793 stmts, domstmt);
2794 if (!genop0)
2795 return NULL_TREE;
2796 op1expr = get_or_alloc_expr_for (genop1);
2797 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2798 if (!genop1)
2799 return NULL_TREE;
2800 if (genop2)
2802 op2expr = get_or_alloc_expr_for (genop2);
2803 genop2 = find_or_generate_expression (block, op2expr, stmts,
2804 domstmt);
2805 if (!genop2)
2806 return NULL_TREE;
2808 if (genop3)
2810 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2811 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2812 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2813 op3expr = get_or_alloc_expr_for (genop3);
2814 genop3 = find_or_generate_expression (block, op3expr, stmts,
2815 domstmt);
2816 if (!genop3)
2817 return NULL_TREE;
2819 return build4 (currop->opcode, currop->type, genop0, genop1,
2820 genop2, genop3);
2822 case COMPONENT_REF:
2824 tree op0;
2825 tree op1;
2826 tree genop2 = currop->op1;
2827 pre_expr op2expr;
2828 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2829 stmts, domstmt);
2830 if (!op0)
2831 return NULL_TREE;
2832 /* op1 should be a FIELD_DECL, which are represented by
2833 themselves. */
2834 op1 = currop->op0;
2835 if (genop2)
2837 op2expr = get_or_alloc_expr_for (genop2);
2838 genop2 = find_or_generate_expression (block, op2expr, stmts,
2839 domstmt);
2840 if (!genop2)
2841 return NULL_TREE;
2844 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2845 genop2);
2847 break;
2848 case SSA_NAME:
2850 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2851 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2852 return genop;
2854 case STRING_CST:
2855 case INTEGER_CST:
2856 case COMPLEX_CST:
2857 case VECTOR_CST:
2858 case REAL_CST:
2859 case CONSTRUCTOR:
2860 case VAR_DECL:
2861 case PARM_DECL:
2862 case CONST_DECL:
2863 case RESULT_DECL:
2864 case FUNCTION_DECL:
2865 return currop->op0;
2867 default:
2868 gcc_unreachable ();
2872 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2873 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2874 trying to rename aggregates into ssa form directly, which is a no no.
2876 Thus, this routine doesn't create temporaries, it just builds a
2877 single access expression for the array, calling
2878 find_or_generate_expression to build the innermost pieces.
2880 This function is a subroutine of create_expression_by_pieces, and
2881 should not be called on it's own unless you really know what you
2882 are doing. */
2884 static tree
2885 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2886 gimple_seq *stmts, gimple domstmt)
2888 unsigned int op = 0;
2889 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2892 /* Find a leader for an expression, or generate one using
2893 create_expression_by_pieces if it's ANTIC but
2894 complex.
2895 BLOCK is the basic_block we are looking for leaders in.
2896 EXPR is the expression to find a leader or generate for.
2897 STMTS is the statement list to put the inserted expressions on.
2898 Returns the SSA_NAME of the LHS of the generated expression or the
2899 leader.
2900 DOMSTMT if non-NULL is a statement that should be dominated by
2901 all uses in the generated expression. If DOMSTMT is non-NULL this
2902 routine can fail and return NULL_TREE. Otherwise it will assert
2903 on failure. */
2905 static tree
2906 find_or_generate_expression (basic_block block, pre_expr expr,
2907 gimple_seq *stmts, gimple domstmt)
2909 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2910 get_expr_value_id (expr), domstmt);
2911 tree genop = NULL;
2912 if (leader)
2914 if (leader->kind == NAME)
2915 genop = PRE_EXPR_NAME (leader);
2916 else if (leader->kind == CONSTANT)
2917 genop = PRE_EXPR_CONSTANT (leader);
2920 /* If it's still NULL, it must be a complex expression, so generate
2921 it recursively. Not so for FRE though. */
2922 if (genop == NULL
2923 && !in_fre)
2925 bitmap_set_t exprset;
2926 unsigned int lookfor = get_expr_value_id (expr);
2927 bool handled = false;
2928 bitmap_iterator bi;
2929 unsigned int i;
2931 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2932 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2934 pre_expr temp = expression_for_id (i);
2935 if (temp->kind != NAME)
2937 handled = true;
2938 genop = create_expression_by_pieces (block, temp, stmts,
2939 domstmt,
2940 get_expr_type (expr));
2941 break;
2944 if (!handled && domstmt)
2945 return NULL_TREE;
2947 gcc_assert (handled);
2949 return genop;
2952 #define NECESSARY GF_PLF_1
2954 /* Create an expression in pieces, so that we can handle very complex
2955 expressions that may be ANTIC, but not necessary GIMPLE.
2956 BLOCK is the basic block the expression will be inserted into,
2957 EXPR is the expression to insert (in value form)
2958 STMTS is a statement list to append the necessary insertions into.
2960 This function will die if we hit some value that shouldn't be
2961 ANTIC but is (IE there is no leader for it, or its components).
2962 This function may also generate expressions that are themselves
2963 partially or fully redundant. Those that are will be either made
2964 fully redundant during the next iteration of insert (for partially
2965 redundant ones), or eliminated by eliminate (for fully redundant
2966 ones).
2968 If DOMSTMT is non-NULL then we make sure that all uses in the
2969 expressions dominate that statement. In this case the function
2970 can return NULL_TREE to signal failure. */
2972 static tree
2973 create_expression_by_pieces (basic_block block, pre_expr expr,
2974 gimple_seq *stmts, gimple domstmt, tree type)
2976 tree temp, name;
2977 tree folded;
2978 gimple_seq forced_stmts = NULL;
2979 unsigned int value_id;
2980 gimple_stmt_iterator gsi;
2981 tree exprtype = type ? type : get_expr_type (expr);
2982 pre_expr nameexpr;
2983 gimple newstmt;
2985 switch (expr->kind)
2987 /* We may hit the NAME/CONSTANT case if we have to convert types
2988 that value numbering saw through. */
2989 case NAME:
2990 folded = PRE_EXPR_NAME (expr);
2991 break;
2992 case CONSTANT:
2993 folded = PRE_EXPR_CONSTANT (expr);
2994 break;
2995 case REFERENCE:
2997 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2998 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3000 break;
3001 case NARY:
3003 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3004 switch (nary->length)
3006 case 2:
3008 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3009 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
3010 tree genop1 = find_or_generate_expression (block, op1,
3011 stmts, domstmt);
3012 tree genop2 = find_or_generate_expression (block, op2,
3013 stmts, domstmt);
3014 if (!genop1 || !genop2)
3015 return NULL_TREE;
3016 genop1 = fold_convert (TREE_TYPE (nary->op[0]),
3017 genop1);
3018 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
3019 may be a constant with the wrong type. */
3020 if (nary->opcode == POINTER_PLUS_EXPR)
3021 genop2 = fold_convert (sizetype, genop2);
3022 else
3023 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
3025 folded = fold_build2 (nary->opcode, nary->type,
3026 genop1, genop2);
3028 break;
3029 case 1:
3031 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3032 tree genop1 = find_or_generate_expression (block, op1,
3033 stmts, domstmt);
3034 if (!genop1)
3035 return NULL_TREE;
3036 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3038 folded = fold_build1 (nary->opcode, nary->type,
3039 genop1);
3041 break;
3042 default:
3043 return NULL_TREE;
3046 break;
3047 default:
3048 return NULL_TREE;
3051 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3052 folded = fold_convert (exprtype, folded);
3054 /* Force the generated expression to be a sequence of GIMPLE
3055 statements.
3056 We have to call unshare_expr because force_gimple_operand may
3057 modify the tree we pass to it. */
3058 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3059 false, NULL);
3061 /* If we have any intermediate expressions to the value sets, add them
3062 to the value sets and chain them in the instruction stream. */
3063 if (forced_stmts)
3065 gsi = gsi_start (forced_stmts);
3066 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3068 gimple stmt = gsi_stmt (gsi);
3069 tree forcedname = gimple_get_lhs (stmt);
3070 pre_expr nameexpr;
3072 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3073 if (TREE_CODE (forcedname) == SSA_NAME)
3075 VN_INFO_GET (forcedname)->valnum = forcedname;
3076 VN_INFO (forcedname)->value_id = get_next_value_id ();
3077 nameexpr = get_or_alloc_expr_for_name (forcedname);
3078 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3079 if (!in_fre)
3080 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3081 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3083 mark_symbols_for_renaming (stmt);
3085 gimple_seq_add_seq (stmts, forced_stmts);
3088 /* Build and insert the assignment of the end result to the temporary
3089 that we will return. */
3090 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3092 pretemp = create_tmp_var (exprtype, "pretmp");
3093 get_var_ann (pretemp);
3096 temp = pretemp;
3097 add_referenced_var (temp);
3099 if (TREE_CODE (exprtype) == COMPLEX_TYPE
3100 || TREE_CODE (exprtype) == VECTOR_TYPE)
3101 DECL_GIMPLE_REG_P (temp) = 1;
3103 newstmt = gimple_build_assign (temp, folded);
3104 name = make_ssa_name (temp, newstmt);
3105 gimple_assign_set_lhs (newstmt, name);
3106 gimple_set_plf (newstmt, NECESSARY, false);
3108 gimple_seq_add_stmt (stmts, newstmt);
3109 VEC_safe_push (gimple, heap, inserted_exprs, newstmt);
3111 /* All the symbols in NEWEXPR should be put into SSA form. */
3112 mark_symbols_for_renaming (newstmt);
3114 /* Add a value number to the temporary.
3115 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3116 we are creating the expression by pieces, and this particular piece of
3117 the expression may have been represented. There is no harm in replacing
3118 here. */
3119 VN_INFO_GET (name)->valnum = name;
3120 value_id = get_expr_value_id (expr);
3121 VN_INFO (name)->value_id = value_id;
3122 nameexpr = get_or_alloc_expr_for_name (name);
3123 add_to_value (value_id, nameexpr);
3124 if (!in_fre)
3125 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3126 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3128 pre_stats.insertions++;
3129 if (dump_file && (dump_flags & TDF_DETAILS))
3131 fprintf (dump_file, "Inserted ");
3132 print_gimple_stmt (dump_file, newstmt, 0, 0);
3133 fprintf (dump_file, " in predecessor %d\n", block->index);
3136 return name;
3140 /* Returns true if we want to inhibit the insertions of PHI nodes
3141 for the given EXPR for basic block BB (a member of a loop).
3142 We want to do this, when we fear that the induction variable we
3143 create might inhibit vectorization. */
3145 static bool
3146 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3148 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3149 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3150 vn_reference_op_t op;
3151 unsigned i;
3153 /* If we aren't going to vectorize we don't inhibit anything. */
3154 if (!flag_tree_vectorize)
3155 return false;
3157 /* Otherwise we inhibit the insertion when the address of the
3158 memory reference is a simple induction variable. In other
3159 cases the vectorizer won't do anything anyway (either it's
3160 loop invariant or a complicated expression). */
3161 for (i = 0; VEC_iterate (vn_reference_op_s, ops, i, op); ++i)
3163 switch (op->opcode)
3165 case ARRAY_REF:
3166 case ARRAY_RANGE_REF:
3167 if (TREE_CODE (op->op0) != SSA_NAME)
3168 break;
3169 /* Fallthru. */
3170 case SSA_NAME:
3172 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3173 affine_iv iv;
3174 /* Default defs are loop invariant. */
3175 if (!defbb)
3176 break;
3177 /* Defined outside this loop, also loop invariant. */
3178 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3179 break;
3180 /* If it's a simple induction variable inhibit insertion,
3181 the vectorizer might be interested in this one. */
3182 if (simple_iv (bb->loop_father, bb->loop_father,
3183 op->op0, &iv, true))
3184 return true;
3185 /* No simple IV, vectorizer can't do anything, hence no
3186 reason to inhibit the transformation for this operand. */
3187 break;
3189 default:
3190 break;
3193 return false;
3196 /* Insert the to-be-made-available values of expression EXPRNUM for each
3197 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3198 merge the result with a phi node, given the same value number as
3199 NODE. Return true if we have inserted new stuff. */
3201 static bool
3202 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3203 pre_expr *avail)
3205 pre_expr expr = expression_for_id (exprnum);
3206 pre_expr newphi;
3207 unsigned int val = get_expr_value_id (expr);
3208 edge pred;
3209 bool insertions = false;
3210 bool nophi = false;
3211 basic_block bprime;
3212 pre_expr eprime;
3213 edge_iterator ei;
3214 tree type = get_expr_type (expr);
3215 tree temp;
3216 gimple phi;
3218 if (dump_file && (dump_flags & TDF_DETAILS))
3220 fprintf (dump_file, "Found partial redundancy for expression ");
3221 print_pre_expr (dump_file, expr);
3222 fprintf (dump_file, " (%04d)\n", val);
3225 /* Make sure we aren't creating an induction variable. */
3226 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3228 bool firstinsideloop = false;
3229 bool secondinsideloop = false;
3230 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3231 EDGE_PRED (block, 0)->src);
3232 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3233 EDGE_PRED (block, 1)->src);
3234 /* Induction variables only have one edge inside the loop. */
3235 if ((firstinsideloop ^ secondinsideloop)
3236 && (expr->kind != REFERENCE
3237 || inhibit_phi_insertion (block, expr)))
3239 if (dump_file && (dump_flags & TDF_DETAILS))
3240 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3241 nophi = true;
3245 /* Make the necessary insertions. */
3246 FOR_EACH_EDGE (pred, ei, block->preds)
3248 gimple_seq stmts = NULL;
3249 tree builtexpr;
3250 bprime = pred->src;
3251 eprime = avail[bprime->index];
3253 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3255 builtexpr = create_expression_by_pieces (bprime,
3256 eprime,
3257 &stmts, NULL,
3258 type);
3259 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3260 gsi_insert_seq_on_edge (pred, stmts);
3261 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3262 insertions = true;
3264 else if (eprime->kind == CONSTANT)
3266 /* Constants may not have the right type, fold_convert
3267 should give us back a constant with the right type.
3269 tree constant = PRE_EXPR_CONSTANT (eprime);
3270 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3272 tree builtexpr = fold_convert (type, constant);
3273 if (!is_gimple_min_invariant (builtexpr))
3275 tree forcedexpr = force_gimple_operand (builtexpr,
3276 &stmts, true,
3277 NULL);
3278 if (!is_gimple_min_invariant (forcedexpr))
3280 if (forcedexpr != builtexpr)
3282 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3283 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3285 if (stmts)
3287 gimple_stmt_iterator gsi;
3288 gsi = gsi_start (stmts);
3289 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3291 gimple stmt = gsi_stmt (gsi);
3292 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3293 gimple_set_plf (stmt, NECESSARY, false);
3295 gsi_insert_seq_on_edge (pred, stmts);
3297 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3302 else if (eprime->kind == NAME)
3304 /* We may have to do a conversion because our value
3305 numbering can look through types in certain cases, but
3306 our IL requires all operands of a phi node have the same
3307 type. */
3308 tree name = PRE_EXPR_NAME (eprime);
3309 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3311 tree builtexpr;
3312 tree forcedexpr;
3313 builtexpr = fold_convert (type, name);
3314 forcedexpr = force_gimple_operand (builtexpr,
3315 &stmts, true,
3316 NULL);
3318 if (forcedexpr != name)
3320 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3321 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3324 if (stmts)
3326 gimple_stmt_iterator gsi;
3327 gsi = gsi_start (stmts);
3328 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3330 gimple stmt = gsi_stmt (gsi);
3331 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3332 gimple_set_plf (stmt, NECESSARY, false);
3334 gsi_insert_seq_on_edge (pred, stmts);
3336 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3340 /* If we didn't want a phi node, and we made insertions, we still have
3341 inserted new stuff, and thus return true. If we didn't want a phi node,
3342 and didn't make insertions, we haven't added anything new, so return
3343 false. */
3344 if (nophi && insertions)
3345 return true;
3346 else if (nophi && !insertions)
3347 return false;
3349 /* Now build a phi for the new variable. */
3350 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3352 prephitemp = create_tmp_var (type, "prephitmp");
3353 get_var_ann (prephitemp);
3356 temp = prephitemp;
3357 add_referenced_var (temp);
3359 if (TREE_CODE (type) == COMPLEX_TYPE
3360 || TREE_CODE (type) == VECTOR_TYPE)
3361 DECL_GIMPLE_REG_P (temp) = 1;
3362 phi = create_phi_node (temp, block);
3364 gimple_set_plf (phi, NECESSARY, false);
3365 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3366 VN_INFO (gimple_phi_result (phi))->value_id = val;
3367 VEC_safe_push (gimple, heap, inserted_exprs, phi);
3368 bitmap_set_bit (inserted_phi_names,
3369 SSA_NAME_VERSION (gimple_phi_result (phi)));
3370 FOR_EACH_EDGE (pred, ei, block->preds)
3372 pre_expr ae = avail[pred->src->index];
3373 gcc_assert (get_expr_type (ae) == type
3374 || useless_type_conversion_p (type, get_expr_type (ae)));
3375 if (ae->kind == CONSTANT)
3376 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3377 else
3378 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3379 UNKNOWN_LOCATION);
3382 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3383 add_to_value (val, newphi);
3385 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3386 this insertion, since we test for the existence of this value in PHI_GEN
3387 before proceeding with the partial redundancy checks in insert_aux.
3389 The value may exist in AVAIL_OUT, in particular, it could be represented
3390 by the expression we are trying to eliminate, in which case we want the
3391 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3392 inserted there.
3394 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3395 this block, because if it did, it would have existed in our dominator's
3396 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3399 bitmap_insert_into_set (PHI_GEN (block), newphi);
3400 bitmap_value_replace_in_set (AVAIL_OUT (block),
3401 newphi);
3402 bitmap_insert_into_set (NEW_SETS (block),
3403 newphi);
3405 if (dump_file && (dump_flags & TDF_DETAILS))
3407 fprintf (dump_file, "Created phi ");
3408 print_gimple_stmt (dump_file, phi, 0, 0);
3409 fprintf (dump_file, " in block %d\n", block->index);
3411 pre_stats.phis++;
3412 return true;
3417 /* Perform insertion of partially redundant values.
3418 For BLOCK, do the following:
3419 1. Propagate the NEW_SETS of the dominator into the current block.
3420 If the block has multiple predecessors,
3421 2a. Iterate over the ANTIC expressions for the block to see if
3422 any of them are partially redundant.
3423 2b. If so, insert them into the necessary predecessors to make
3424 the expression fully redundant.
3425 2c. Insert a new PHI merging the values of the predecessors.
3426 2d. Insert the new PHI, and the new expressions, into the
3427 NEW_SETS set.
3428 3. Recursively call ourselves on the dominator children of BLOCK.
3430 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3431 do_regular_insertion and do_partial_insertion.
3435 static bool
3436 do_regular_insertion (basic_block block, basic_block dom)
3438 bool new_stuff = false;
3439 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3440 pre_expr expr;
3441 int i;
3443 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3445 if (expr->kind != NAME)
3447 pre_expr *avail;
3448 unsigned int val;
3449 bool by_some = false;
3450 bool cant_insert = false;
3451 bool all_same = true;
3452 pre_expr first_s = NULL;
3453 edge pred;
3454 basic_block bprime;
3455 pre_expr eprime = NULL;
3456 edge_iterator ei;
3457 pre_expr edoubleprime = NULL;
3458 bool do_insertion = false;
3460 val = get_expr_value_id (expr);
3461 if (bitmap_set_contains_value (PHI_GEN (block), val))
3462 continue;
3463 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3465 if (dump_file && (dump_flags & TDF_DETAILS))
3466 fprintf (dump_file, "Found fully redundant value\n");
3467 continue;
3470 avail = XCNEWVEC (pre_expr, last_basic_block);
3471 FOR_EACH_EDGE (pred, ei, block->preds)
3473 unsigned int vprime;
3475 /* We should never run insertion for the exit block
3476 and so not come across fake pred edges. */
3477 gcc_assert (!(pred->flags & EDGE_FAKE));
3478 bprime = pred->src;
3479 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3480 bprime, block);
3482 /* eprime will generally only be NULL if the
3483 value of the expression, translated
3484 through the PHI for this predecessor, is
3485 undefined. If that is the case, we can't
3486 make the expression fully redundant,
3487 because its value is undefined along a
3488 predecessor path. We can thus break out
3489 early because it doesn't matter what the
3490 rest of the results are. */
3491 if (eprime == NULL)
3493 cant_insert = true;
3494 break;
3497 eprime = fully_constant_expression (eprime);
3498 vprime = get_expr_value_id (eprime);
3499 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3500 vprime, NULL);
3501 if (edoubleprime == NULL)
3503 avail[bprime->index] = eprime;
3504 all_same = false;
3506 else
3508 avail[bprime->index] = edoubleprime;
3509 by_some = true;
3510 /* We want to perform insertions to remove a redundancy on
3511 a path in the CFG we want to optimize for speed. */
3512 if (optimize_edge_for_speed_p (pred))
3513 do_insertion = true;
3514 if (first_s == NULL)
3515 first_s = edoubleprime;
3516 else if (!pre_expr_eq (first_s, edoubleprime))
3517 all_same = false;
3520 /* If we can insert it, it's not the same value
3521 already existing along every predecessor, and
3522 it's defined by some predecessor, it is
3523 partially redundant. */
3524 if (!cant_insert && !all_same && by_some && do_insertion
3525 && dbg_cnt (treepre_insert))
3527 if (insert_into_preds_of_block (block, get_expression_id (expr),
3528 avail))
3529 new_stuff = true;
3531 /* If all edges produce the same value and that value is
3532 an invariant, then the PHI has the same value on all
3533 edges. Note this. */
3534 else if (!cant_insert && all_same && eprime
3535 && (edoubleprime->kind == CONSTANT
3536 || edoubleprime->kind == NAME)
3537 && !value_id_constant_p (val))
3539 unsigned int j;
3540 bitmap_iterator bi;
3541 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3542 value_expressions, val);
3544 unsigned int new_val = get_expr_value_id (edoubleprime);
3545 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3547 pre_expr expr = expression_for_id (j);
3549 if (expr->kind == NAME)
3551 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3552 /* Just reset the value id and valnum so it is
3553 the same as the constant we have discovered. */
3554 if (edoubleprime->kind == CONSTANT)
3556 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3557 pre_stats.constified++;
3559 else
3560 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3561 info->value_id = new_val;
3565 free (avail);
3569 VEC_free (pre_expr, heap, exprs);
3570 return new_stuff;
3574 /* Perform insertion for partially anticipatable expressions. There
3575 is only one case we will perform insertion for these. This case is
3576 if the expression is partially anticipatable, and fully available.
3577 In this case, we know that putting it earlier will enable us to
3578 remove the later computation. */
3581 static bool
3582 do_partial_partial_insertion (basic_block block, basic_block dom)
3584 bool new_stuff = false;
3585 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3586 pre_expr expr;
3587 int i;
3589 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3591 if (expr->kind != NAME)
3593 pre_expr *avail;
3594 unsigned int val;
3595 bool by_all = true;
3596 bool cant_insert = false;
3597 edge pred;
3598 basic_block bprime;
3599 pre_expr eprime = NULL;
3600 edge_iterator ei;
3602 val = get_expr_value_id (expr);
3603 if (bitmap_set_contains_value (PHI_GEN (block), val))
3604 continue;
3605 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3606 continue;
3608 avail = XCNEWVEC (pre_expr, last_basic_block);
3609 FOR_EACH_EDGE (pred, ei, block->preds)
3611 unsigned int vprime;
3612 pre_expr edoubleprime;
3614 /* We should never run insertion for the exit block
3615 and so not come across fake pred edges. */
3616 gcc_assert (!(pred->flags & EDGE_FAKE));
3617 bprime = pred->src;
3618 eprime = phi_translate (expr, ANTIC_IN (block),
3619 PA_IN (block),
3620 bprime, block);
3622 /* eprime will generally only be NULL if the
3623 value of the expression, translated
3624 through the PHI for this predecessor, is
3625 undefined. If that is the case, we can't
3626 make the expression fully redundant,
3627 because its value is undefined along a
3628 predecessor path. We can thus break out
3629 early because it doesn't matter what the
3630 rest of the results are. */
3631 if (eprime == NULL)
3633 cant_insert = true;
3634 break;
3637 eprime = fully_constant_expression (eprime);
3638 vprime = get_expr_value_id (eprime);
3639 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3640 vprime, NULL);
3641 if (edoubleprime == NULL)
3643 by_all = false;
3644 break;
3646 else
3647 avail[bprime->index] = edoubleprime;
3651 /* If we can insert it, it's not the same value
3652 already existing along every predecessor, and
3653 it's defined by some predecessor, it is
3654 partially redundant. */
3655 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3657 pre_stats.pa_insert++;
3658 if (insert_into_preds_of_block (block, get_expression_id (expr),
3659 avail))
3660 new_stuff = true;
3662 free (avail);
3666 VEC_free (pre_expr, heap, exprs);
3667 return new_stuff;
3670 static bool
3671 insert_aux (basic_block block)
3673 basic_block son;
3674 bool new_stuff = false;
3676 if (block)
3678 basic_block dom;
3679 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3680 if (dom)
3682 unsigned i;
3683 bitmap_iterator bi;
3684 bitmap_set_t newset = NEW_SETS (dom);
3685 if (newset)
3687 /* Note that we need to value_replace both NEW_SETS, and
3688 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3689 represented by some non-simple expression here that we want
3690 to replace it with. */
3691 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3693 pre_expr expr = expression_for_id (i);
3694 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3695 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3698 if (!single_pred_p (block))
3700 new_stuff |= do_regular_insertion (block, dom);
3701 if (do_partial_partial)
3702 new_stuff |= do_partial_partial_insertion (block, dom);
3706 for (son = first_dom_son (CDI_DOMINATORS, block);
3707 son;
3708 son = next_dom_son (CDI_DOMINATORS, son))
3710 new_stuff |= insert_aux (son);
3713 return new_stuff;
3716 /* Perform insertion of partially redundant values. */
3718 static void
3719 insert (void)
3721 bool new_stuff = true;
3722 basic_block bb;
3723 int num_iterations = 0;
3725 FOR_ALL_BB (bb)
3726 NEW_SETS (bb) = bitmap_set_new ();
3728 while (new_stuff)
3730 num_iterations++;
3731 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3733 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3737 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3739 static void
3740 add_to_exp_gen (basic_block block, tree op)
3742 if (!in_fre)
3744 pre_expr result;
3745 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3746 return;
3747 result = get_or_alloc_expr_for_name (op);
3748 bitmap_value_insert_into_set (EXP_GEN (block), result);
3752 /* Create value ids for PHI in BLOCK. */
3754 static void
3755 make_values_for_phi (gimple phi, basic_block block)
3757 tree result = gimple_phi_result (phi);
3759 /* We have no need for virtual phis, as they don't represent
3760 actual computations. */
3761 if (is_gimple_reg (result))
3763 pre_expr e = get_or_alloc_expr_for_name (result);
3764 add_to_value (get_expr_value_id (e), e);
3765 bitmap_insert_into_set (PHI_GEN (block), e);
3766 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3767 if (!in_fre)
3769 unsigned i;
3770 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3772 tree arg = gimple_phi_arg_def (phi, i);
3773 if (TREE_CODE (arg) == SSA_NAME)
3775 e = get_or_alloc_expr_for_name (arg);
3776 add_to_value (get_expr_value_id (e), e);
3783 /* Compute the AVAIL set for all basic blocks.
3785 This function performs value numbering of the statements in each basic
3786 block. The AVAIL sets are built from information we glean while doing
3787 this value numbering, since the AVAIL sets contain only one entry per
3788 value.
3790 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3791 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3793 static void
3794 compute_avail (void)
3797 basic_block block, son;
3798 basic_block *worklist;
3799 size_t sp = 0;
3800 unsigned i;
3802 /* We pretend that default definitions are defined in the entry block.
3803 This includes function arguments and the static chain decl. */
3804 for (i = 1; i < num_ssa_names; ++i)
3806 tree name = ssa_name (i);
3807 pre_expr e;
3808 if (!name
3809 || !SSA_NAME_IS_DEFAULT_DEF (name)
3810 || has_zero_uses (name)
3811 || !is_gimple_reg (name))
3812 continue;
3814 e = get_or_alloc_expr_for_name (name);
3815 add_to_value (get_expr_value_id (e), e);
3816 if (!in_fre)
3817 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3818 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3821 /* Allocate the worklist. */
3822 worklist = XNEWVEC (basic_block, n_basic_blocks);
3824 /* Seed the algorithm by putting the dominator children of the entry
3825 block on the worklist. */
3826 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3827 son;
3828 son = next_dom_son (CDI_DOMINATORS, son))
3829 worklist[sp++] = son;
3831 /* Loop until the worklist is empty. */
3832 while (sp)
3834 gimple_stmt_iterator gsi;
3835 gimple stmt;
3836 basic_block dom;
3837 unsigned int stmt_uid = 1;
3839 /* Pick a block from the worklist. */
3840 block = worklist[--sp];
3842 /* Initially, the set of available values in BLOCK is that of
3843 its immediate dominator. */
3844 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3845 if (dom)
3846 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3848 /* Generate values for PHI nodes. */
3849 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3850 make_values_for_phi (gsi_stmt (gsi), block);
3852 BB_MAY_NOTRETURN (block) = 0;
3854 /* Now compute value numbers and populate value sets with all
3855 the expressions computed in BLOCK. */
3856 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3858 ssa_op_iter iter;
3859 tree op;
3861 stmt = gsi_stmt (gsi);
3862 gimple_set_uid (stmt, stmt_uid++);
3864 /* Cache whether the basic-block has any non-visible side-effect
3865 or control flow.
3866 If this isn't a call or it is the last stmt in the
3867 basic-block then the CFG represents things correctly. */
3868 if (is_gimple_call (stmt)
3869 && !stmt_ends_bb_p (stmt))
3871 /* Non-looping const functions always return normally.
3872 Otherwise the call might not return or have side-effects
3873 that forbids hoisting possibly trapping expressions
3874 before it. */
3875 int flags = gimple_call_flags (stmt);
3876 if (!(flags & ECF_CONST)
3877 || (flags & ECF_LOOPING_CONST_OR_PURE))
3878 BB_MAY_NOTRETURN (block) = 1;
3881 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3883 pre_expr e = get_or_alloc_expr_for_name (op);
3885 add_to_value (get_expr_value_id (e), e);
3886 if (!in_fre)
3887 bitmap_insert_into_set (TMP_GEN (block), e);
3888 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3891 if (gimple_has_volatile_ops (stmt)
3892 || stmt_could_throw_p (stmt))
3893 continue;
3895 switch (gimple_code (stmt))
3897 case GIMPLE_RETURN:
3898 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3899 add_to_exp_gen (block, op);
3900 continue;
3902 case GIMPLE_CALL:
3904 vn_reference_t ref;
3905 unsigned int i;
3906 vn_reference_op_t vro;
3907 pre_expr result = NULL;
3908 VEC(vn_reference_op_s, heap) *ops = NULL;
3910 if (!can_value_number_call (stmt))
3911 continue;
3913 copy_reference_ops_from_call (stmt, &ops);
3914 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3915 gimple_expr_type (stmt),
3916 ops, &ref, false);
3917 VEC_free (vn_reference_op_s, heap, ops);
3918 if (!ref)
3919 continue;
3921 for (i = 0; VEC_iterate (vn_reference_op_s,
3922 ref->operands, i,
3923 vro); i++)
3925 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3926 add_to_exp_gen (block, vro->op0);
3927 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3928 add_to_exp_gen (block, vro->op1);
3929 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3930 add_to_exp_gen (block, vro->op2);
3932 result = (pre_expr) pool_alloc (pre_expr_pool);
3933 result->kind = REFERENCE;
3934 result->id = 0;
3935 PRE_EXPR_REFERENCE (result) = ref;
3937 get_or_alloc_expression_id (result);
3938 add_to_value (get_expr_value_id (result), result);
3939 if (!in_fre)
3940 bitmap_value_insert_into_set (EXP_GEN (block), result);
3941 continue;
3944 case GIMPLE_ASSIGN:
3946 pre_expr result = NULL;
3947 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
3949 case tcc_unary:
3950 case tcc_binary:
3951 case tcc_comparison:
3953 vn_nary_op_t nary;
3954 unsigned int i;
3956 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
3957 gimple_assign_rhs_code (stmt),
3958 gimple_expr_type (stmt),
3959 gimple_assign_rhs1 (stmt),
3960 gimple_assign_rhs2 (stmt),
3961 NULL_TREE, NULL_TREE, &nary);
3963 if (!nary)
3964 continue;
3966 for (i = 0; i < nary->length; i++)
3967 if (TREE_CODE (nary->op[i]) == SSA_NAME)
3968 add_to_exp_gen (block, nary->op[i]);
3970 result = (pre_expr) pool_alloc (pre_expr_pool);
3971 result->kind = NARY;
3972 result->id = 0;
3973 PRE_EXPR_NARY (result) = nary;
3974 break;
3977 case tcc_declaration:
3978 case tcc_reference:
3980 vn_reference_t ref;
3981 unsigned int i;
3982 vn_reference_op_t vro;
3984 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3985 gimple_vuse (stmt),
3986 true, &ref);
3987 if (!ref)
3988 continue;
3990 for (i = 0; VEC_iterate (vn_reference_op_s,
3991 ref->operands, i,
3992 vro); i++)
3994 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3995 add_to_exp_gen (block, vro->op0);
3996 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3997 add_to_exp_gen (block, vro->op1);
3998 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3999 add_to_exp_gen (block, vro->op2);
4001 result = (pre_expr) pool_alloc (pre_expr_pool);
4002 result->kind = REFERENCE;
4003 result->id = 0;
4004 PRE_EXPR_REFERENCE (result) = ref;
4005 break;
4008 default:
4009 /* For any other statement that we don't
4010 recognize, simply add all referenced
4011 SSA_NAMEs to EXP_GEN. */
4012 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4013 add_to_exp_gen (block, op);
4014 continue;
4017 get_or_alloc_expression_id (result);
4018 add_to_value (get_expr_value_id (result), result);
4019 if (!in_fre)
4020 bitmap_value_insert_into_set (EXP_GEN (block), result);
4022 continue;
4024 default:
4025 break;
4029 /* Put the dominator children of BLOCK on the worklist of blocks
4030 to compute available sets for. */
4031 for (son = first_dom_son (CDI_DOMINATORS, block);
4032 son;
4033 son = next_dom_son (CDI_DOMINATORS, son))
4034 worklist[sp++] = son;
4037 free (worklist);
4040 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4041 than the available expressions for it. The insertion point is
4042 right before the first use in STMT. Returns the SSA_NAME that should
4043 be used for replacement. */
4045 static tree
4046 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4048 basic_block bb = gimple_bb (stmt);
4049 gimple_stmt_iterator gsi;
4050 gimple_seq stmts = NULL;
4051 tree expr;
4052 pre_expr e;
4054 /* First create a value expression from the expression we want
4055 to insert and associate it with the value handle for SSA_VN. */
4056 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4057 if (e == NULL)
4058 return NULL_TREE;
4060 /* Then use create_expression_by_pieces to generate a valid
4061 expression to insert at this point of the IL stream. */
4062 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4063 if (expr == NULL_TREE)
4064 return NULL_TREE;
4065 gsi = gsi_for_stmt (stmt);
4066 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4068 return expr;
4071 /* Eliminate fully redundant computations. */
4073 static unsigned int
4074 eliminate (void)
4076 VEC (gimple, heap) *to_remove = NULL;
4077 basic_block b;
4078 unsigned int todo = 0;
4079 gimple_stmt_iterator gsi;
4080 gimple stmt;
4081 unsigned i;
4083 FOR_EACH_BB (b)
4085 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4087 stmt = gsi_stmt (gsi);
4089 /* Lookup the RHS of the expression, see if we have an
4090 available computation for it. If so, replace the RHS with
4091 the available computation. */
4092 if (gimple_has_lhs (stmt)
4093 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4094 && !gimple_assign_ssa_name_copy_p (stmt)
4095 && (!gimple_assign_single_p (stmt)
4096 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4097 && !gimple_has_volatile_ops (stmt)
4098 && !has_zero_uses (gimple_get_lhs (stmt)))
4100 tree lhs = gimple_get_lhs (stmt);
4101 tree rhs = NULL_TREE;
4102 tree sprime = NULL;
4103 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4104 pre_expr sprimeexpr;
4106 if (gimple_assign_single_p (stmt))
4107 rhs = gimple_assign_rhs1 (stmt);
4109 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4110 get_expr_value_id (lhsexpr),
4111 NULL);
4113 if (sprimeexpr)
4115 if (sprimeexpr->kind == CONSTANT)
4116 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4117 else if (sprimeexpr->kind == NAME)
4118 sprime = PRE_EXPR_NAME (sprimeexpr);
4119 else
4120 gcc_unreachable ();
4123 /* If there is no existing leader but SCCVN knows this
4124 value is constant, use that constant. */
4125 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4127 sprime = VN_INFO (lhs)->valnum;
4128 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4129 TREE_TYPE (sprime)))
4130 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4132 if (dump_file && (dump_flags & TDF_DETAILS))
4134 fprintf (dump_file, "Replaced ");
4135 print_gimple_expr (dump_file, stmt, 0, 0);
4136 fprintf (dump_file, " with ");
4137 print_generic_expr (dump_file, sprime, 0);
4138 fprintf (dump_file, " in ");
4139 print_gimple_stmt (dump_file, stmt, 0, 0);
4141 pre_stats.eliminations++;
4142 propagate_tree_value_into_stmt (&gsi, sprime);
4143 stmt = gsi_stmt (gsi);
4144 update_stmt (stmt);
4145 continue;
4148 /* If there is no existing usable leader but SCCVN thinks
4149 it has an expression it wants to use as replacement,
4150 insert that. */
4151 if (!sprime || sprime == lhs)
4153 tree val = VN_INFO (lhs)->valnum;
4154 if (val != VN_TOP
4155 && TREE_CODE (val) == SSA_NAME
4156 && VN_INFO (val)->needs_insertion
4157 && can_PRE_operation (vn_get_expr_for (val)))
4158 sprime = do_SCCVN_insertion (stmt, val);
4160 if (sprime
4161 && sprime != lhs
4162 && (rhs == NULL_TREE
4163 || TREE_CODE (rhs) != SSA_NAME
4164 || may_propagate_copy (rhs, sprime)))
4166 gcc_assert (sprime != rhs);
4168 if (dump_file && (dump_flags & TDF_DETAILS))
4170 fprintf (dump_file, "Replaced ");
4171 print_gimple_expr (dump_file, stmt, 0, 0);
4172 fprintf (dump_file, " with ");
4173 print_generic_expr (dump_file, sprime, 0);
4174 fprintf (dump_file, " in ");
4175 print_gimple_stmt (dump_file, stmt, 0, 0);
4178 if (TREE_CODE (sprime) == SSA_NAME)
4179 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4180 NECESSARY, true);
4181 /* We need to make sure the new and old types actually match,
4182 which may require adding a simple cast, which fold_convert
4183 will do for us. */
4184 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4185 && !useless_type_conversion_p (gimple_expr_type (stmt),
4186 TREE_TYPE (sprime)))
4187 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4189 pre_stats.eliminations++;
4190 propagate_tree_value_into_stmt (&gsi, sprime);
4191 stmt = gsi_stmt (gsi);
4192 update_stmt (stmt);
4194 /* If we removed EH side effects from the statement, clean
4195 its EH information. */
4196 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4198 bitmap_set_bit (need_eh_cleanup,
4199 gimple_bb (stmt)->index);
4200 if (dump_file && (dump_flags & TDF_DETAILS))
4201 fprintf (dump_file, " Removed EH side effects.\n");
4205 /* If the statement is a scalar store, see if the expression
4206 has the same value number as its rhs. If so, the store is
4207 dead. */
4208 else if (gimple_assign_single_p (stmt)
4209 && !is_gimple_reg (gimple_assign_lhs (stmt))
4210 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4211 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4213 tree rhs = gimple_assign_rhs1 (stmt);
4214 tree val;
4215 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4216 gimple_vuse (stmt), true, NULL);
4217 if (TREE_CODE (rhs) == SSA_NAME)
4218 rhs = VN_INFO (rhs)->valnum;
4219 if (val
4220 && operand_equal_p (val, rhs, 0))
4222 if (dump_file && (dump_flags & TDF_DETAILS))
4224 fprintf (dump_file, "Deleted redundant store ");
4225 print_gimple_stmt (dump_file, stmt, 0, 0);
4228 /* Queue stmt for removal. */
4229 VEC_safe_push (gimple, heap, to_remove, stmt);
4232 /* Visit COND_EXPRs and fold the comparison with the
4233 available value-numbers. */
4234 else if (gimple_code (stmt) == GIMPLE_COND)
4236 tree op0 = gimple_cond_lhs (stmt);
4237 tree op1 = gimple_cond_rhs (stmt);
4238 tree result;
4240 if (TREE_CODE (op0) == SSA_NAME)
4241 op0 = VN_INFO (op0)->valnum;
4242 if (TREE_CODE (op1) == SSA_NAME)
4243 op1 = VN_INFO (op1)->valnum;
4244 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4245 op0, op1);
4246 if (result && TREE_CODE (result) == INTEGER_CST)
4248 if (integer_zerop (result))
4249 gimple_cond_make_false (stmt);
4250 else
4251 gimple_cond_make_true (stmt);
4252 update_stmt (stmt);
4253 todo = TODO_cleanup_cfg;
4256 /* Visit indirect calls and turn them into direct calls if
4257 possible. */
4258 if (gimple_code (stmt) == GIMPLE_CALL
4259 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4261 tree fn = VN_INFO (gimple_call_fn (stmt))->valnum;
4262 if (TREE_CODE (fn) == ADDR_EXPR
4263 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
4265 if (dump_file && (dump_flags & TDF_DETAILS))
4267 fprintf (dump_file, "Replacing call target with ");
4268 print_generic_expr (dump_file, fn, 0);
4269 fprintf (dump_file, " in ");
4270 print_gimple_stmt (dump_file, stmt, 0, 0);
4273 gimple_call_set_fn (stmt, fn);
4274 update_stmt (stmt);
4275 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4277 bitmap_set_bit (need_eh_cleanup,
4278 gimple_bb (stmt)->index);
4279 if (dump_file && (dump_flags & TDF_DETAILS))
4280 fprintf (dump_file, " Removed EH side effects.\n");
4283 /* Changing an indirect call to a direct call may
4284 have exposed different semantics. This may
4285 require an SSA update. */
4286 todo |= TODO_update_ssa_only_virtuals;
4291 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4293 gimple stmt, phi = gsi_stmt (gsi);
4294 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4295 pre_expr sprimeexpr, resexpr;
4296 gimple_stmt_iterator gsi2;
4298 /* We want to perform redundant PHI elimination. Do so by
4299 replacing the PHI with a single copy if possible.
4300 Do not touch inserted, single-argument or virtual PHIs. */
4301 if (gimple_phi_num_args (phi) == 1
4302 || !is_gimple_reg (res)
4303 || bitmap_bit_p (inserted_phi_names, SSA_NAME_VERSION (res)))
4305 gsi_next (&gsi);
4306 continue;
4309 resexpr = get_or_alloc_expr_for_name (res);
4310 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4311 get_expr_value_id (resexpr), NULL);
4312 if (sprimeexpr)
4314 if (sprimeexpr->kind == CONSTANT)
4315 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4316 else if (sprimeexpr->kind == NAME)
4317 sprime = PRE_EXPR_NAME (sprimeexpr);
4318 else
4319 gcc_unreachable ();
4321 if (!sprimeexpr
4322 || sprime == res)
4324 gsi_next (&gsi);
4325 continue;
4328 if (dump_file && (dump_flags & TDF_DETAILS))
4330 fprintf (dump_file, "Replaced redundant PHI node defining ");
4331 print_generic_expr (dump_file, res, 0);
4332 fprintf (dump_file, " with ");
4333 print_generic_expr (dump_file, sprime, 0);
4334 fprintf (dump_file, "\n");
4337 remove_phi_node (&gsi, false);
4339 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4340 sprime = fold_convert (TREE_TYPE (res), sprime);
4341 stmt = gimple_build_assign (res, sprime);
4342 SSA_NAME_DEF_STMT (res) = stmt;
4343 if (TREE_CODE (sprime) == SSA_NAME)
4344 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4345 NECESSARY, true);
4346 gsi2 = gsi_after_labels (b);
4347 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4348 /* Queue the copy for eventual removal. */
4349 VEC_safe_push (gimple, heap, to_remove, stmt);
4350 pre_stats.eliminations++;
4354 /* We cannot remove stmts during BB walk, especially not release SSA
4355 names there as this confuses the VN machinery. The stmts ending
4356 up in to_remove are either stores or simple copies. */
4357 for (i = 0; VEC_iterate (gimple, to_remove, i, stmt); ++i)
4359 tree lhs = gimple_assign_lhs (stmt);
4360 use_operand_p use_p;
4361 gimple use_stmt;
4363 /* If there is a single use only, propagate the equivalency
4364 instead of keeping the copy. */
4365 if (TREE_CODE (lhs) == SSA_NAME
4366 && single_imm_use (lhs, &use_p, &use_stmt)
4367 && may_propagate_copy (USE_FROM_PTR (use_p),
4368 gimple_assign_rhs1 (stmt)))
4370 SET_USE (use_p, gimple_assign_rhs1 (stmt));
4371 update_stmt (use_stmt);
4374 /* If this is a store or a now unused copy, remove it. */
4375 if (TREE_CODE (lhs) != SSA_NAME
4376 || has_zero_uses (lhs))
4378 gsi = gsi_for_stmt (stmt);
4379 unlink_stmt_vdef (stmt);
4380 gsi_remove (&gsi, true);
4381 release_defs (stmt);
4384 VEC_free (gimple, heap, to_remove);
4386 return todo;
4389 /* Borrow a bit of tree-ssa-dce.c for the moment.
4390 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4391 this may be a bit faster, and we may want critical edges kept split. */
4393 /* If OP's defining statement has not already been determined to be necessary,
4394 mark that statement necessary. Return the stmt, if it is newly
4395 necessary. */
4397 static inline gimple
4398 mark_operand_necessary (tree op)
4400 gimple stmt;
4402 gcc_assert (op);
4404 if (TREE_CODE (op) != SSA_NAME)
4405 return NULL;
4407 stmt = SSA_NAME_DEF_STMT (op);
4408 gcc_assert (stmt);
4410 if (gimple_plf (stmt, NECESSARY)
4411 || gimple_nop_p (stmt))
4412 return NULL;
4414 gimple_set_plf (stmt, NECESSARY, true);
4415 return stmt;
4418 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4419 to insert PHI nodes sometimes, and because value numbering of casts isn't
4420 perfect, we sometimes end up inserting dead code. This simple DCE-like
4421 pass removes any insertions we made that weren't actually used. */
4423 static void
4424 remove_dead_inserted_code (void)
4426 VEC(gimple,heap) *worklist = NULL;
4427 int i;
4428 gimple t;
4430 worklist = VEC_alloc (gimple, heap, VEC_length (gimple, inserted_exprs));
4431 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4433 if (gimple_plf (t, NECESSARY))
4434 VEC_quick_push (gimple, worklist, t);
4436 while (VEC_length (gimple, worklist) > 0)
4438 t = VEC_pop (gimple, worklist);
4440 /* PHI nodes are somewhat special in that each PHI alternative has
4441 data and control dependencies. All the statements feeding the
4442 PHI node's arguments are always necessary. */
4443 if (gimple_code (t) == GIMPLE_PHI)
4445 unsigned k;
4447 VEC_reserve (gimple, heap, worklist, gimple_phi_num_args (t));
4448 for (k = 0; k < gimple_phi_num_args (t); k++)
4450 tree arg = PHI_ARG_DEF (t, k);
4451 if (TREE_CODE (arg) == SSA_NAME)
4453 gimple n = mark_operand_necessary (arg);
4454 if (n)
4455 VEC_quick_push (gimple, worklist, n);
4459 else
4461 /* Propagate through the operands. Examine all the USE, VUSE and
4462 VDEF operands in this statement. Mark all the statements
4463 which feed this statement's uses as necessary. */
4464 ssa_op_iter iter;
4465 tree use;
4467 /* The operands of VDEF expressions are also needed as they
4468 represent potential definitions that may reach this
4469 statement (VDEF operands allow us to follow def-def
4470 links). */
4472 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4474 gimple n = mark_operand_necessary (use);
4475 if (n)
4476 VEC_safe_push (gimple, heap, worklist, n);
4481 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4483 if (!gimple_plf (t, NECESSARY))
4485 gimple_stmt_iterator gsi;
4487 if (dump_file && (dump_flags & TDF_DETAILS))
4489 fprintf (dump_file, "Removing unnecessary insertion:");
4490 print_gimple_stmt (dump_file, t, 0, 0);
4493 gsi = gsi_for_stmt (t);
4494 if (gimple_code (t) == GIMPLE_PHI)
4495 remove_phi_node (&gsi, true);
4496 else
4498 gsi_remove (&gsi, true);
4499 release_defs (t);
4503 VEC_free (gimple, heap, worklist);
4506 /* Initialize data structures used by PRE. */
4508 static void
4509 init_pre (bool do_fre)
4511 basic_block bb;
4513 next_expression_id = 1;
4514 expressions = NULL;
4515 VEC_safe_push (pre_expr, heap, expressions, NULL);
4516 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4517 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4518 get_max_value_id() + 1);
4519 name_to_id = NULL;
4521 in_fre = do_fre;
4523 inserted_exprs = NULL;
4524 need_creation = NULL;
4525 pretemp = NULL_TREE;
4526 storetemp = NULL_TREE;
4527 prephitemp = NULL_TREE;
4529 connect_infinite_loops_to_exit ();
4530 memset (&pre_stats, 0, sizeof (pre_stats));
4533 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4534 post_order_compute (postorder, false, false);
4536 FOR_ALL_BB (bb)
4537 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4539 calculate_dominance_info (CDI_POST_DOMINATORS);
4540 calculate_dominance_info (CDI_DOMINATORS);
4542 bitmap_obstack_initialize (&grand_bitmap_obstack);
4543 inserted_phi_names = BITMAP_ALLOC (&grand_bitmap_obstack);
4544 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4545 expr_pred_trans_eq, free);
4546 expression_to_id = htab_create (num_ssa_names * 3,
4547 pre_expr_hash,
4548 pre_expr_eq, NULL);
4549 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4550 sizeof (struct bitmap_set), 30);
4551 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4552 sizeof (struct pre_expr_d), 30);
4553 FOR_ALL_BB (bb)
4555 EXP_GEN (bb) = bitmap_set_new ();
4556 PHI_GEN (bb) = bitmap_set_new ();
4557 TMP_GEN (bb) = bitmap_set_new ();
4558 AVAIL_OUT (bb) = bitmap_set_new ();
4561 need_eh_cleanup = BITMAP_ALLOC (NULL);
4565 /* Deallocate data structures used by PRE. */
4567 static void
4568 fini_pre (bool do_fre)
4570 basic_block bb;
4572 free (postorder);
4573 VEC_free (bitmap_set_t, heap, value_expressions);
4574 VEC_free (gimple, heap, inserted_exprs);
4575 VEC_free (gimple, heap, need_creation);
4576 bitmap_obstack_release (&grand_bitmap_obstack);
4577 free_alloc_pool (bitmap_set_pool);
4578 free_alloc_pool (pre_expr_pool);
4579 htab_delete (phi_translate_table);
4580 htab_delete (expression_to_id);
4581 VEC_free (unsigned, heap, name_to_id);
4583 FOR_ALL_BB (bb)
4585 free (bb->aux);
4586 bb->aux = NULL;
4589 free_dominance_info (CDI_POST_DOMINATORS);
4591 if (!bitmap_empty_p (need_eh_cleanup))
4593 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4594 cleanup_tree_cfg ();
4597 BITMAP_FREE (need_eh_cleanup);
4599 if (!do_fre)
4600 loop_optimizer_finalize ();
4603 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4604 only wants to do full redundancy elimination. */
4606 static unsigned int
4607 execute_pre (bool do_fre)
4609 unsigned int todo = 0;
4611 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4613 /* This has to happen before SCCVN runs because
4614 loop_optimizer_init may create new phis, etc. */
4615 if (!do_fre)
4616 loop_optimizer_init (LOOPS_NORMAL);
4618 if (!run_scc_vn (do_fre))
4620 if (!do_fre)
4622 remove_dead_inserted_code ();
4623 loop_optimizer_finalize ();
4626 return 0;
4628 init_pre (do_fre);
4629 scev_initialize ();
4632 /* Collect and value number expressions computed in each basic block. */
4633 compute_avail ();
4635 if (dump_file && (dump_flags & TDF_DETAILS))
4637 basic_block bb;
4639 FOR_ALL_BB (bb)
4641 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4642 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4643 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4644 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4648 /* Insert can get quite slow on an incredibly large number of basic
4649 blocks due to some quadratic behavior. Until this behavior is
4650 fixed, don't run it when he have an incredibly large number of
4651 bb's. If we aren't going to run insert, there is no point in
4652 computing ANTIC, either, even though it's plenty fast. */
4653 if (!do_fre && n_basic_blocks < 4000)
4655 compute_antic ();
4656 insert ();
4659 /* Remove all the redundant expressions. */
4660 todo |= eliminate ();
4662 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4663 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4664 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4665 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4666 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4668 /* Make sure to remove fake edges before committing our inserts.
4669 This makes sure we don't end up with extra critical edges that
4670 we would need to split. */
4671 remove_fake_exit_edges ();
4672 gsi_commit_edge_inserts ();
4674 clear_expression_ids ();
4675 free_scc_vn ();
4676 if (!do_fre)
4677 remove_dead_inserted_code ();
4679 scev_finalize ();
4680 fini_pre (do_fre);
4682 return todo;
4685 /* Gate and execute functions for PRE. */
4687 static unsigned int
4688 do_pre (void)
4690 return execute_pre (false);
4693 static bool
4694 gate_pre (void)
4696 return flag_tree_pre != 0;
4699 struct gimple_opt_pass pass_pre =
4702 GIMPLE_PASS,
4703 "pre", /* name */
4704 gate_pre, /* gate */
4705 do_pre, /* execute */
4706 NULL, /* sub */
4707 NULL, /* next */
4708 0, /* static_pass_number */
4709 TV_TREE_PRE, /* tv_id */
4710 PROP_no_crit_edges | PROP_cfg
4711 | PROP_ssa, /* properties_required */
4712 0, /* properties_provided */
4713 0, /* properties_destroyed */
4714 TODO_rebuild_alias, /* todo_flags_start */
4715 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4716 | TODO_verify_ssa /* todo_flags_finish */
4721 /* Gate and execute functions for FRE. */
4723 static unsigned int
4724 execute_fre (void)
4726 return execute_pre (true);
4729 static bool
4730 gate_fre (void)
4732 return flag_tree_fre != 0;
4735 struct gimple_opt_pass pass_fre =
4738 GIMPLE_PASS,
4739 "fre", /* name */
4740 gate_fre, /* gate */
4741 execute_fre, /* execute */
4742 NULL, /* sub */
4743 NULL, /* next */
4744 0, /* static_pass_number */
4745 TV_TREE_FRE, /* tv_id */
4746 PROP_cfg | PROP_ssa, /* properties_required */
4747 0, /* properties_provided */
4748 0, /* properties_destroyed */
4749 0, /* todo_flags_start */
4750 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */