PR tree-optimization/48377
[official-gcc.git] / gcc / tree-ssa-pre.c
blobe59a598348b580bb91e2648230bacc18c16fd9ed
1 /* SSA-PRE for trees.
2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
5 <stevenb@suse.de>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
12 any later version.
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 #include "config.h"
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "tree.h"
28 #include "basic-block.h"
29 #include "tree-pretty-print.h"
30 #include "gimple-pretty-print.h"
31 #include "tree-inline.h"
32 #include "tree-flow.h"
33 #include "gimple.h"
34 #include "tree-dump.h"
35 #include "timevar.h"
36 #include "fibheap.h"
37 #include "hashtab.h"
38 #include "tree-iterator.h"
39 #include "alloc-pool.h"
40 #include "obstack.h"
41 #include "tree-pass.h"
42 #include "flags.h"
43 #include "bitmap.h"
44 #include "langhooks.h"
45 #include "cfgloop.h"
46 #include "tree-ssa-sccvn.h"
47 #include "tree-scalar-evolution.h"
48 #include "params.h"
49 #include "dbgcnt.h"
51 /* TODO:
53 1. Avail sets can be shared by making an avail_find_leader that
54 walks up the dominator tree and looks in those avail sets.
55 This might affect code optimality, it's unclear right now.
56 2. Strength reduction can be performed by anticipating expressions
57 we can repair later on.
58 3. We can do back-substitution or smarter value numbering to catch
59 commutative expressions split up over multiple statements.
62 /* For ease of terminology, "expression node" in the below refers to
63 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
64 represent the actual statement containing the expressions we care about,
65 and we cache the value number by putting it in the expression. */
67 /* Basic algorithm
69 First we walk the statements to generate the AVAIL sets, the
70 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
71 generation of values/expressions by a given block. We use them
72 when computing the ANTIC sets. The AVAIL sets consist of
73 SSA_NAME's that represent values, so we know what values are
74 available in what blocks. AVAIL is a forward dataflow problem. In
75 SSA, values are never killed, so we don't need a kill set, or a
76 fixpoint iteration, in order to calculate the AVAIL sets. In
77 traditional parlance, AVAIL sets tell us the downsafety of the
78 expressions/values.
80 Next, we generate the ANTIC sets. These sets represent the
81 anticipatable expressions. ANTIC is a backwards dataflow
82 problem. An expression is anticipatable in a given block if it could
83 be generated in that block. This means that if we had to perform
84 an insertion in that block, of the value of that expression, we
85 could. Calculating the ANTIC sets requires phi translation of
86 expressions, because the flow goes backwards through phis. We must
87 iterate to a fixpoint of the ANTIC sets, because we have a kill
88 set. Even in SSA form, values are not live over the entire
89 function, only from their definition point onwards. So we have to
90 remove values from the ANTIC set once we go past the definition
91 point of the leaders that make them up.
92 compute_antic/compute_antic_aux performs this computation.
94 Third, we perform insertions to make partially redundant
95 expressions fully redundant.
97 An expression is partially redundant (excluding partial
98 anticipation) if:
100 1. It is AVAIL in some, but not all, of the predecessors of a
101 given block.
102 2. It is ANTIC in all the predecessors.
104 In order to make it fully redundant, we insert the expression into
105 the predecessors where it is not available, but is ANTIC.
107 For the partial anticipation case, we only perform insertion if it
108 is partially anticipated in some block, and fully available in all
109 of the predecessors.
111 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
112 performs these steps.
114 Fourth, we eliminate fully redundant expressions.
115 This is a simple statement walk that replaces redundant
116 calculations with the now available values. */
118 /* Representations of value numbers:
120 Value numbers are represented by a representative SSA_NAME. We
121 will create fake SSA_NAME's in situations where we need a
122 representative but do not have one (because it is a complex
123 expression). In order to facilitate storing the value numbers in
124 bitmaps, and keep the number of wasted SSA_NAME's down, we also
125 associate a value_id with each value number, and create full blown
126 ssa_name's only where we actually need them (IE in operands of
127 existing expressions).
129 Theoretically you could replace all the value_id's with
130 SSA_NAME_VERSION, but this would allocate a large number of
131 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
132 It would also require an additional indirection at each point we
133 use the value id. */
135 /* Representation of expressions on value numbers:
137 Expressions consisting of value numbers are represented the same
138 way as our VN internally represents them, with an additional
139 "pre_expr" wrapping around them in order to facilitate storing all
140 of the expressions in the same sets. */
142 /* Representation of sets:
144 The dataflow sets do not need to be sorted in any particular order
145 for the majority of their lifetime, are simply represented as two
146 bitmaps, one that keeps track of values present in the set, and one
147 that keeps track of expressions present in the set.
149 When we need them in topological order, we produce it on demand by
150 transforming the bitmap into an array and sorting it into topo
151 order. */
153 /* Type of expression, used to know which member of the PRE_EXPR union
154 is valid. */
156 enum pre_expr_kind
158 NAME,
159 NARY,
160 REFERENCE,
161 CONSTANT
164 typedef union pre_expr_union_d
166 tree name;
167 tree constant;
168 vn_nary_op_t nary;
169 vn_reference_t reference;
170 } pre_expr_union;
172 typedef struct pre_expr_d
174 enum pre_expr_kind kind;
175 unsigned int id;
176 pre_expr_union u;
177 } *pre_expr;
179 #define PRE_EXPR_NAME(e) (e)->u.name
180 #define PRE_EXPR_NARY(e) (e)->u.nary
181 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
182 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
184 static int
185 pre_expr_eq (const void *p1, const void *p2)
187 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
188 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
190 if (e1->kind != e2->kind)
191 return false;
193 switch (e1->kind)
195 case CONSTANT:
196 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
197 PRE_EXPR_CONSTANT (e2));
198 case NAME:
199 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
200 case NARY:
201 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
202 case REFERENCE:
203 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
204 PRE_EXPR_REFERENCE (e2));
205 default:
206 gcc_unreachable ();
210 static hashval_t
211 pre_expr_hash (const void *p1)
213 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
214 switch (e->kind)
216 case CONSTANT:
217 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
218 case NAME:
219 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
220 case NARY:
221 return PRE_EXPR_NARY (e)->hashcode;
222 case REFERENCE:
223 return PRE_EXPR_REFERENCE (e)->hashcode;
224 default:
225 gcc_unreachable ();
230 /* Next global expression id number. */
231 static unsigned int next_expression_id;
233 /* Mapping from expression to id number we can use in bitmap sets. */
234 DEF_VEC_P (pre_expr);
235 DEF_VEC_ALLOC_P (pre_expr, heap);
236 static VEC(pre_expr, heap) *expressions;
237 static htab_t expression_to_id;
238 static VEC(unsigned, heap) *name_to_id;
240 /* Allocate an expression id for EXPR. */
242 static inline unsigned int
243 alloc_expression_id (pre_expr expr)
245 void **slot;
246 /* Make sure we won't overflow. */
247 gcc_assert (next_expression_id + 1 > next_expression_id);
248 expr->id = next_expression_id++;
249 VEC_safe_push (pre_expr, heap, expressions, expr);
250 if (expr->kind == NAME)
252 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
253 /* VEC_safe_grow_cleared allocates no headroom. Avoid frequent
254 re-allocations by using VEC_reserve upfront. There is no
255 VEC_quick_grow_cleared unfortunately. */
256 VEC_reserve (unsigned, heap, name_to_id, num_ssa_names);
257 VEC_safe_grow_cleared (unsigned, heap, name_to_id, num_ssa_names);
258 gcc_assert (VEC_index (unsigned, name_to_id, version) == 0);
259 VEC_replace (unsigned, name_to_id, version, expr->id);
261 else
263 slot = htab_find_slot (expression_to_id, expr, INSERT);
264 gcc_assert (!*slot);
265 *slot = expr;
267 return next_expression_id - 1;
270 /* Return the expression id for tree EXPR. */
272 static inline unsigned int
273 get_expression_id (const pre_expr expr)
275 return expr->id;
278 static inline unsigned int
279 lookup_expression_id (const pre_expr expr)
281 void **slot;
283 if (expr->kind == NAME)
285 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
286 if (VEC_length (unsigned, name_to_id) <= version)
287 return 0;
288 return VEC_index (unsigned, name_to_id, version);
290 else
292 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
293 if (!slot)
294 return 0;
295 return ((pre_expr)*slot)->id;
299 /* Return the existing expression id for EXPR, or create one if one
300 does not exist yet. */
302 static inline unsigned int
303 get_or_alloc_expression_id (pre_expr expr)
305 unsigned int id = lookup_expression_id (expr);
306 if (id == 0)
307 return alloc_expression_id (expr);
308 return expr->id = id;
311 /* Return the expression that has expression id ID */
313 static inline pre_expr
314 expression_for_id (unsigned int id)
316 return VEC_index (pre_expr, expressions, id);
319 /* Free the expression id field in all of our expressions,
320 and then destroy the expressions array. */
322 static void
323 clear_expression_ids (void)
325 VEC_free (pre_expr, heap, expressions);
328 static alloc_pool pre_expr_pool;
330 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
332 static pre_expr
333 get_or_alloc_expr_for_name (tree name)
335 struct pre_expr_d expr;
336 pre_expr result;
337 unsigned int result_id;
339 expr.kind = NAME;
340 expr.id = 0;
341 PRE_EXPR_NAME (&expr) = name;
342 result_id = lookup_expression_id (&expr);
343 if (result_id != 0)
344 return expression_for_id (result_id);
346 result = (pre_expr) pool_alloc (pre_expr_pool);
347 result->kind = NAME;
348 PRE_EXPR_NAME (result) = name;
349 alloc_expression_id (result);
350 return result;
353 static bool in_fre = false;
355 /* An unordered bitmap set. One bitmap tracks values, the other,
356 expressions. */
357 typedef struct bitmap_set
359 bitmap_head expressions;
360 bitmap_head values;
361 } *bitmap_set_t;
363 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
364 EXECUTE_IF_SET_IN_BITMAP(&(set)->expressions, 0, (id), (bi))
366 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
367 EXECUTE_IF_SET_IN_BITMAP(&(set)->values, 0, (id), (bi))
369 /* Mapping from value id to expressions with that value_id. */
370 DEF_VEC_P (bitmap_set_t);
371 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
372 static VEC(bitmap_set_t, heap) *value_expressions;
374 /* Sets that we need to keep track of. */
375 typedef struct bb_bitmap_sets
377 /* The EXP_GEN set, which represents expressions/values generated in
378 a basic block. */
379 bitmap_set_t exp_gen;
381 /* The PHI_GEN set, which represents PHI results generated in a
382 basic block. */
383 bitmap_set_t phi_gen;
385 /* The TMP_GEN set, which represents results/temporaries generated
386 in a basic block. IE the LHS of an expression. */
387 bitmap_set_t tmp_gen;
389 /* The AVAIL_OUT set, which represents which values are available in
390 a given basic block. */
391 bitmap_set_t avail_out;
393 /* The ANTIC_IN set, which represents which values are anticipatable
394 in a given basic block. */
395 bitmap_set_t antic_in;
397 /* The PA_IN set, which represents which values are
398 partially anticipatable in a given basic block. */
399 bitmap_set_t pa_in;
401 /* The NEW_SETS set, which is used during insertion to augment the
402 AVAIL_OUT set of blocks with the new insertions performed during
403 the current iteration. */
404 bitmap_set_t new_sets;
406 /* A cache for value_dies_in_block_x. */
407 bitmap expr_dies;
409 /* True if we have visited this block during ANTIC calculation. */
410 unsigned int visited : 1;
412 /* True we have deferred processing this block during ANTIC
413 calculation until its successor is processed. */
414 unsigned int deferred : 1;
416 /* True when the block contains a call that might not return. */
417 unsigned int contains_may_not_return_call : 1;
418 } *bb_value_sets_t;
420 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
421 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
422 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
423 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
424 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
425 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
426 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
427 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
428 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
429 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
430 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
433 /* Basic block list in postorder. */
434 static int *postorder;
436 /* This structure is used to keep track of statistics on what
437 optimization PRE was able to perform. */
438 static struct
440 /* The number of RHS computations eliminated by PRE. */
441 int eliminations;
443 /* The number of new expressions/temporaries generated by PRE. */
444 int insertions;
446 /* The number of inserts found due to partial anticipation */
447 int pa_insert;
449 /* The number of new PHI nodes added by PRE. */
450 int phis;
452 /* The number of values found constant. */
453 int constified;
455 } pre_stats;
457 static bool do_partial_partial;
458 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
459 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
460 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
461 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
462 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
463 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
464 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
465 unsigned int, bool);
466 static bitmap_set_t bitmap_set_new (void);
467 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
468 gimple, tree);
469 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
470 gimple);
471 static unsigned int get_expr_value_id (pre_expr);
473 /* We can add and remove elements and entries to and from sets
474 and hash tables, so we use alloc pools for them. */
476 static alloc_pool bitmap_set_pool;
477 static bitmap_obstack grand_bitmap_obstack;
479 /* To avoid adding 300 temporary variables when we only need one, we
480 only create one temporary variable, on demand, and build ssa names
481 off that. We do have to change the variable if the types don't
482 match the current variable's type. */
483 static tree pretemp;
484 static tree storetemp;
485 static tree prephitemp;
487 /* Set of blocks with statements that have had their EH properties changed. */
488 static bitmap need_eh_cleanup;
490 /* Set of blocks with statements that have had their AB properties changed. */
491 static bitmap need_ab_cleanup;
493 /* The phi_translate_table caches phi translations for a given
494 expression and predecessor. */
496 static htab_t phi_translate_table;
498 /* A three tuple {e, pred, v} used to cache phi translations in the
499 phi_translate_table. */
501 typedef struct expr_pred_trans_d
503 /* The expression. */
504 pre_expr e;
506 /* The predecessor block along which we translated the expression. */
507 basic_block pred;
509 /* The value that resulted from the translation. */
510 pre_expr v;
512 /* The hashcode for the expression, pred pair. This is cached for
513 speed reasons. */
514 hashval_t hashcode;
515 } *expr_pred_trans_t;
516 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
518 /* Return the hash value for a phi translation table entry. */
520 static hashval_t
521 expr_pred_trans_hash (const void *p)
523 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
524 return ve->hashcode;
527 /* Return true if two phi translation table entries are the same.
528 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
530 static int
531 expr_pred_trans_eq (const void *p1, const void *p2)
533 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
534 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
535 basic_block b1 = ve1->pred;
536 basic_block b2 = ve2->pred;
538 /* If they are not translations for the same basic block, they can't
539 be equal. */
540 if (b1 != b2)
541 return false;
542 return pre_expr_eq (ve1->e, ve2->e);
545 /* Search in the phi translation table for the translation of
546 expression E in basic block PRED.
547 Return the translated value, if found, NULL otherwise. */
549 static inline pre_expr
550 phi_trans_lookup (pre_expr e, basic_block pred)
552 void **slot;
553 struct expr_pred_trans_d ept;
555 ept.e = e;
556 ept.pred = pred;
557 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
558 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
559 NO_INSERT);
560 if (!slot)
561 return NULL;
562 else
563 return ((expr_pred_trans_t) *slot)->v;
567 /* Add the tuple mapping from {expression E, basic block PRED} to
568 value V, to the phi translation table. */
570 static inline void
571 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
573 void **slot;
574 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
575 new_pair->e = e;
576 new_pair->pred = pred;
577 new_pair->v = v;
578 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
579 pred->index);
581 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
582 new_pair->hashcode, INSERT);
583 if (*slot)
584 free (*slot);
585 *slot = (void *) new_pair;
589 /* Add expression E to the expression set of value id V. */
591 void
592 add_to_value (unsigned int v, pre_expr e)
594 bitmap_set_t set;
596 gcc_assert (get_expr_value_id (e) == v);
598 if (v >= VEC_length (bitmap_set_t, value_expressions))
600 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
601 v + 1);
604 set = VEC_index (bitmap_set_t, value_expressions, v);
605 if (!set)
607 set = bitmap_set_new ();
608 VEC_replace (bitmap_set_t, value_expressions, v, set);
611 bitmap_insert_into_set_1 (set, e, v, true);
614 /* Create a new bitmap set and return it. */
616 static bitmap_set_t
617 bitmap_set_new (void)
619 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
620 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
621 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
622 return ret;
625 /* Return the value id for a PRE expression EXPR. */
627 static unsigned int
628 get_expr_value_id (pre_expr expr)
630 switch (expr->kind)
632 case CONSTANT:
634 unsigned int id;
635 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
636 if (id == 0)
638 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
639 add_to_value (id, expr);
641 return id;
643 case NAME:
644 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
645 case NARY:
646 return PRE_EXPR_NARY (expr)->value_id;
647 case REFERENCE:
648 return PRE_EXPR_REFERENCE (expr)->value_id;
649 default:
650 gcc_unreachable ();
654 /* Remove an expression EXPR from a bitmapped set. */
656 static void
657 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
659 unsigned int val = get_expr_value_id (expr);
660 if (!value_id_constant_p (val))
662 bitmap_clear_bit (&set->values, val);
663 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
667 static void
668 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
669 unsigned int val, bool allow_constants)
671 if (allow_constants || !value_id_constant_p (val))
673 /* We specifically expect this and only this function to be able to
674 insert constants into a set. */
675 bitmap_set_bit (&set->values, val);
676 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
680 /* Insert an expression EXPR into a bitmapped set. */
682 static void
683 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
685 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
688 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
690 static void
691 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
693 bitmap_copy (&dest->expressions, &orig->expressions);
694 bitmap_copy (&dest->values, &orig->values);
698 /* Free memory used up by SET. */
699 static void
700 bitmap_set_free (bitmap_set_t set)
702 bitmap_clear (&set->expressions);
703 bitmap_clear (&set->values);
707 /* Generate an topological-ordered array of bitmap set SET. */
709 static VEC(pre_expr, heap) *
710 sorted_array_from_bitmap_set (bitmap_set_t set)
712 unsigned int i, j;
713 bitmap_iterator bi, bj;
714 VEC(pre_expr, heap) *result;
716 /* Pre-allocate roughly enough space for the array. */
717 result = VEC_alloc (pre_expr, heap, bitmap_count_bits (&set->values));
719 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
721 /* The number of expressions having a given value is usually
722 relatively small. Thus, rather than making a vector of all
723 the expressions and sorting it by value-id, we walk the values
724 and check in the reverse mapping that tells us what expressions
725 have a given value, to filter those in our set. As a result,
726 the expressions are inserted in value-id order, which means
727 topological order.
729 If this is somehow a significant lose for some cases, we can
730 choose which set to walk based on the set size. */
731 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
732 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
734 if (bitmap_bit_p (&set->expressions, j))
735 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
739 return result;
742 /* Perform bitmapped set operation DEST &= ORIG. */
744 static void
745 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
747 bitmap_iterator bi;
748 unsigned int i;
750 if (dest != orig)
752 bitmap_head temp;
753 bitmap_initialize (&temp, &grand_bitmap_obstack);
755 bitmap_and_into (&dest->values, &orig->values);
756 bitmap_copy (&temp, &dest->expressions);
757 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
759 pre_expr expr = expression_for_id (i);
760 unsigned int value_id = get_expr_value_id (expr);
761 if (!bitmap_bit_p (&dest->values, value_id))
762 bitmap_clear_bit (&dest->expressions, i);
764 bitmap_clear (&temp);
768 /* Subtract all values and expressions contained in ORIG from DEST. */
770 static bitmap_set_t
771 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
773 bitmap_set_t result = bitmap_set_new ();
774 bitmap_iterator bi;
775 unsigned int i;
777 bitmap_and_compl (&result->expressions, &dest->expressions,
778 &orig->expressions);
780 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
782 pre_expr expr = expression_for_id (i);
783 unsigned int value_id = get_expr_value_id (expr);
784 bitmap_set_bit (&result->values, value_id);
787 return result;
790 /* Subtract all the values in bitmap set B from bitmap set A. */
792 static void
793 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
795 unsigned int i;
796 bitmap_iterator bi;
797 bitmap_head temp;
799 bitmap_initialize (&temp, &grand_bitmap_obstack);
801 bitmap_copy (&temp, &a->expressions);
802 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
804 pre_expr expr = expression_for_id (i);
805 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
806 bitmap_remove_from_set (a, expr);
808 bitmap_clear (&temp);
812 /* Return true if bitmapped set SET contains the value VALUE_ID. */
814 static bool
815 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
817 if (value_id_constant_p (value_id))
818 return true;
820 if (!set || bitmap_empty_p (&set->expressions))
821 return false;
823 return bitmap_bit_p (&set->values, value_id);
826 static inline bool
827 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
829 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
832 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
834 static void
835 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
836 const pre_expr expr)
838 bitmap_set_t exprset;
839 unsigned int i;
840 bitmap_iterator bi;
842 if (value_id_constant_p (lookfor))
843 return;
845 if (!bitmap_set_contains_value (set, lookfor))
846 return;
848 /* The number of expressions having a given value is usually
849 significantly less than the total number of expressions in SET.
850 Thus, rather than check, for each expression in SET, whether it
851 has the value LOOKFOR, we walk the reverse mapping that tells us
852 what expressions have a given value, and see if any of those
853 expressions are in our set. For large testcases, this is about
854 5-10x faster than walking the bitmap. If this is somehow a
855 significant lose for some cases, we can choose which set to walk
856 based on the set size. */
857 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
858 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
860 if (bitmap_clear_bit (&set->expressions, i))
862 bitmap_set_bit (&set->expressions, get_expression_id (expr));
863 return;
868 /* Return true if two bitmap sets are equal. */
870 static bool
871 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
873 return bitmap_equal_p (&a->values, &b->values);
876 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
877 and add it otherwise. */
879 static void
880 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
882 unsigned int val = get_expr_value_id (expr);
884 if (bitmap_set_contains_value (set, val))
885 bitmap_set_replace_value (set, val, expr);
886 else
887 bitmap_insert_into_set (set, expr);
890 /* Insert EXPR into SET if EXPR's value is not already present in
891 SET. */
893 static void
894 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
896 unsigned int val = get_expr_value_id (expr);
898 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
900 /* Constant values are always considered to be part of the set. */
901 if (value_id_constant_p (val))
902 return;
904 /* If the value membership changed, add the expression. */
905 if (bitmap_set_bit (&set->values, val))
906 bitmap_set_bit (&set->expressions, expr->id);
909 /* Print out EXPR to outfile. */
911 static void
912 print_pre_expr (FILE *outfile, const pre_expr expr)
914 switch (expr->kind)
916 case CONSTANT:
917 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
918 break;
919 case NAME:
920 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
921 break;
922 case NARY:
924 unsigned int i;
925 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
926 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
927 for (i = 0; i < nary->length; i++)
929 print_generic_expr (outfile, nary->op[i], 0);
930 if (i != (unsigned) nary->length - 1)
931 fprintf (outfile, ",");
933 fprintf (outfile, "}");
935 break;
937 case REFERENCE:
939 vn_reference_op_t vro;
940 unsigned int i;
941 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
942 fprintf (outfile, "{");
943 for (i = 0;
944 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
945 i++)
947 bool closebrace = false;
948 if (vro->opcode != SSA_NAME
949 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
951 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
952 if (vro->op0)
954 fprintf (outfile, "<");
955 closebrace = true;
958 if (vro->op0)
960 print_generic_expr (outfile, vro->op0, 0);
961 if (vro->op1)
963 fprintf (outfile, ",");
964 print_generic_expr (outfile, vro->op1, 0);
966 if (vro->op2)
968 fprintf (outfile, ",");
969 print_generic_expr (outfile, vro->op2, 0);
972 if (closebrace)
973 fprintf (outfile, ">");
974 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
975 fprintf (outfile, ",");
977 fprintf (outfile, "}");
978 if (ref->vuse)
980 fprintf (outfile, "@");
981 print_generic_expr (outfile, ref->vuse, 0);
984 break;
987 void debug_pre_expr (pre_expr);
989 /* Like print_pre_expr but always prints to stderr. */
990 DEBUG_FUNCTION void
991 debug_pre_expr (pre_expr e)
993 print_pre_expr (stderr, e);
994 fprintf (stderr, "\n");
997 /* Print out SET to OUTFILE. */
999 static void
1000 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1001 const char *setname, int blockindex)
1003 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1004 if (set)
1006 bool first = true;
1007 unsigned i;
1008 bitmap_iterator bi;
1010 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1012 const pre_expr expr = expression_for_id (i);
1014 if (!first)
1015 fprintf (outfile, ", ");
1016 first = false;
1017 print_pre_expr (outfile, expr);
1019 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1022 fprintf (outfile, " }\n");
1025 void debug_bitmap_set (bitmap_set_t);
1027 DEBUG_FUNCTION void
1028 debug_bitmap_set (bitmap_set_t set)
1030 print_bitmap_set (stderr, set, "debug", 0);
1033 /* Print out the expressions that have VAL to OUTFILE. */
1035 void
1036 print_value_expressions (FILE *outfile, unsigned int val)
1038 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1039 if (set)
1041 char s[10];
1042 sprintf (s, "%04d", val);
1043 print_bitmap_set (outfile, set, s, 0);
1048 DEBUG_FUNCTION void
1049 debug_value_expressions (unsigned int val)
1051 print_value_expressions (stderr, val);
1054 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1055 represent it. */
1057 static pre_expr
1058 get_or_alloc_expr_for_constant (tree constant)
1060 unsigned int result_id;
1061 unsigned int value_id;
1062 struct pre_expr_d expr;
1063 pre_expr newexpr;
1065 expr.kind = CONSTANT;
1066 PRE_EXPR_CONSTANT (&expr) = constant;
1067 result_id = lookup_expression_id (&expr);
1068 if (result_id != 0)
1069 return expression_for_id (result_id);
1071 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1072 newexpr->kind = CONSTANT;
1073 PRE_EXPR_CONSTANT (newexpr) = constant;
1074 alloc_expression_id (newexpr);
1075 value_id = get_or_alloc_constant_value_id (constant);
1076 add_to_value (value_id, newexpr);
1077 return newexpr;
1080 /* Given a value id V, find the actual tree representing the constant
1081 value if there is one, and return it. Return NULL if we can't find
1082 a constant. */
1084 static tree
1085 get_constant_for_value_id (unsigned int v)
1087 if (value_id_constant_p (v))
1089 unsigned int i;
1090 bitmap_iterator bi;
1091 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1093 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1095 pre_expr expr = expression_for_id (i);
1096 if (expr->kind == CONSTANT)
1097 return PRE_EXPR_CONSTANT (expr);
1100 return NULL;
1103 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1104 Currently only supports constants and SSA_NAMES. */
1105 static pre_expr
1106 get_or_alloc_expr_for (tree t)
1108 if (TREE_CODE (t) == SSA_NAME)
1109 return get_or_alloc_expr_for_name (t);
1110 else if (is_gimple_min_invariant (t))
1111 return get_or_alloc_expr_for_constant (t);
1112 else
1114 /* More complex expressions can result from SCCVN expression
1115 simplification that inserts values for them. As they all
1116 do not have VOPs the get handled by the nary ops struct. */
1117 vn_nary_op_t result;
1118 unsigned int result_id;
1119 vn_nary_op_lookup (t, &result);
1120 if (result != NULL)
1122 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1123 e->kind = NARY;
1124 PRE_EXPR_NARY (e) = result;
1125 result_id = lookup_expression_id (e);
1126 if (result_id != 0)
1128 pool_free (pre_expr_pool, e);
1129 e = expression_for_id (result_id);
1130 return e;
1132 alloc_expression_id (e);
1133 return e;
1136 return NULL;
1139 /* Return the folded version of T if T, when folded, is a gimple
1140 min_invariant. Otherwise, return T. */
1142 static pre_expr
1143 fully_constant_expression (pre_expr e)
1145 switch (e->kind)
1147 case CONSTANT:
1148 return e;
1149 case NARY:
1151 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1152 switch (TREE_CODE_CLASS (nary->opcode))
1154 case tcc_expression:
1155 if (nary->opcode == TRUTH_NOT_EXPR)
1156 goto do_unary;
1157 if (nary->opcode != TRUTH_AND_EXPR
1158 && nary->opcode != TRUTH_OR_EXPR
1159 && nary->opcode != TRUTH_XOR_EXPR)
1160 return e;
1161 /* Fallthrough. */
1162 case tcc_binary:
1163 case tcc_comparison:
1165 /* We have to go from trees to pre exprs to value ids to
1166 constants. */
1167 tree naryop0 = nary->op[0];
1168 tree naryop1 = nary->op[1];
1169 tree result;
1170 if (!is_gimple_min_invariant (naryop0))
1172 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1173 unsigned int vrep0 = get_expr_value_id (rep0);
1174 tree const0 = get_constant_for_value_id (vrep0);
1175 if (const0)
1176 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1178 if (!is_gimple_min_invariant (naryop1))
1180 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1181 unsigned int vrep1 = get_expr_value_id (rep1);
1182 tree const1 = get_constant_for_value_id (vrep1);
1183 if (const1)
1184 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1186 result = fold_binary (nary->opcode, nary->type,
1187 naryop0, naryop1);
1188 if (result && is_gimple_min_invariant (result))
1189 return get_or_alloc_expr_for_constant (result);
1190 /* We might have simplified the expression to a
1191 SSA_NAME for example from x_1 * 1. But we cannot
1192 insert a PHI for x_1 unconditionally as x_1 might
1193 not be available readily. */
1194 return e;
1196 case tcc_reference:
1197 if (nary->opcode != REALPART_EXPR
1198 && nary->opcode != IMAGPART_EXPR
1199 && nary->opcode != VIEW_CONVERT_EXPR)
1200 return e;
1201 /* Fallthrough. */
1202 case tcc_unary:
1203 do_unary:
1205 /* We have to go from trees to pre exprs to value ids to
1206 constants. */
1207 tree naryop0 = nary->op[0];
1208 tree const0, result;
1209 if (is_gimple_min_invariant (naryop0))
1210 const0 = naryop0;
1211 else
1213 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1214 unsigned int vrep0 = get_expr_value_id (rep0);
1215 const0 = get_constant_for_value_id (vrep0);
1217 result = NULL;
1218 if (const0)
1220 tree type1 = TREE_TYPE (nary->op[0]);
1221 const0 = fold_convert (type1, const0);
1222 result = fold_unary (nary->opcode, nary->type, const0);
1224 if (result && is_gimple_min_invariant (result))
1225 return get_or_alloc_expr_for_constant (result);
1226 return e;
1228 default:
1229 return e;
1232 case REFERENCE:
1234 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1235 tree folded;
1236 if ((folded = fully_constant_vn_reference_p (ref)))
1237 return get_or_alloc_expr_for_constant (folded);
1238 return e;
1240 default:
1241 return e;
1243 return e;
1246 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1247 it has the value it would have in BLOCK. Set *SAME_VALID to true
1248 in case the new vuse doesn't change the value id of the OPERANDS. */
1250 static tree
1251 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1252 alias_set_type set, tree type, tree vuse,
1253 basic_block phiblock,
1254 basic_block block, bool *same_valid)
1256 gimple phi = SSA_NAME_DEF_STMT (vuse);
1257 ao_ref ref;
1258 edge e = NULL;
1259 bool use_oracle;
1261 *same_valid = true;
1263 if (gimple_bb (phi) != phiblock)
1264 return vuse;
1266 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1268 /* Use the alias-oracle to find either the PHI node in this block,
1269 the first VUSE used in this block that is equivalent to vuse or
1270 the first VUSE which definition in this block kills the value. */
1271 if (gimple_code (phi) == GIMPLE_PHI)
1272 e = find_edge (block, phiblock);
1273 else if (use_oracle)
1274 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1276 vuse = gimple_vuse (phi);
1277 phi = SSA_NAME_DEF_STMT (vuse);
1278 if (gimple_bb (phi) != phiblock)
1279 return vuse;
1280 if (gimple_code (phi) == GIMPLE_PHI)
1282 e = find_edge (block, phiblock);
1283 break;
1286 else
1287 return NULL_TREE;
1289 if (e)
1291 if (use_oracle)
1293 bitmap visited = NULL;
1294 /* Try to find a vuse that dominates this phi node by skipping
1295 non-clobbering statements. */
1296 vuse = get_continuation_for_phi (phi, &ref, &visited);
1297 if (visited)
1298 BITMAP_FREE (visited);
1300 else
1301 vuse = NULL_TREE;
1302 if (!vuse)
1304 /* If we didn't find any, the value ID can't stay the same,
1305 but return the translated vuse. */
1306 *same_valid = false;
1307 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1309 /* ??? We would like to return vuse here as this is the canonical
1310 upmost vdef that this reference is associated with. But during
1311 insertion of the references into the hash tables we only ever
1312 directly insert with their direct gimple_vuse, hence returning
1313 something else would make us not find the other expression. */
1314 return PHI_ARG_DEF (phi, e->dest_idx);
1317 return NULL_TREE;
1320 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1321 SET2. This is used to avoid making a set consisting of the union
1322 of PA_IN and ANTIC_IN during insert. */
1324 static inline pre_expr
1325 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1327 pre_expr result;
1329 result = bitmap_find_leader (set1, val, NULL);
1330 if (!result && set2)
1331 result = bitmap_find_leader (set2, val, NULL);
1332 return result;
1335 /* Get the tree type for our PRE expression e. */
1337 static tree
1338 get_expr_type (const pre_expr e)
1340 switch (e->kind)
1342 case NAME:
1343 return TREE_TYPE (PRE_EXPR_NAME (e));
1344 case CONSTANT:
1345 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1346 case REFERENCE:
1347 return PRE_EXPR_REFERENCE (e)->type;
1348 case NARY:
1349 return PRE_EXPR_NARY (e)->type;
1351 gcc_unreachable();
1354 /* Get a representative SSA_NAME for a given expression.
1355 Since all of our sub-expressions are treated as values, we require
1356 them to be SSA_NAME's for simplicity.
1357 Prior versions of GVNPRE used to use "value handles" here, so that
1358 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1359 either case, the operands are really values (IE we do not expect
1360 them to be usable without finding leaders). */
1362 static tree
1363 get_representative_for (const pre_expr e)
1365 tree exprtype;
1366 tree name;
1367 unsigned int value_id = get_expr_value_id (e);
1369 switch (e->kind)
1371 case NAME:
1372 return PRE_EXPR_NAME (e);
1373 case CONSTANT:
1374 return PRE_EXPR_CONSTANT (e);
1375 case NARY:
1376 case REFERENCE:
1378 /* Go through all of the expressions representing this value
1379 and pick out an SSA_NAME. */
1380 unsigned int i;
1381 bitmap_iterator bi;
1382 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1383 value_id);
1384 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1386 pre_expr rep = expression_for_id (i);
1387 if (rep->kind == NAME)
1388 return PRE_EXPR_NAME (rep);
1391 break;
1393 /* If we reached here we couldn't find an SSA_NAME. This can
1394 happen when we've discovered a value that has never appeared in
1395 the program as set to an SSA_NAME, most likely as the result of
1396 phi translation. */
1397 if (dump_file)
1399 fprintf (dump_file,
1400 "Could not find SSA_NAME representative for expression:");
1401 print_pre_expr (dump_file, e);
1402 fprintf (dump_file, "\n");
1405 exprtype = get_expr_type (e);
1407 /* Build and insert the assignment of the end result to the temporary
1408 that we will return. */
1409 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1411 pretemp = create_tmp_reg (exprtype, "pretmp");
1412 get_var_ann (pretemp);
1415 name = make_ssa_name (pretemp, gimple_build_nop ());
1416 VN_INFO_GET (name)->value_id = value_id;
1417 if (e->kind == CONSTANT)
1418 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1419 else
1420 VN_INFO (name)->valnum = name;
1422 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1423 if (dump_file)
1425 fprintf (dump_file, "Created SSA_NAME representative ");
1426 print_generic_expr (dump_file, name, 0);
1427 fprintf (dump_file, " for expression:");
1428 print_pre_expr (dump_file, e);
1429 fprintf (dump_file, "\n");
1432 return name;
1437 static pre_expr
1438 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1439 basic_block pred, basic_block phiblock);
1441 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1442 the phis in PRED. Return NULL if we can't find a leader for each part
1443 of the translated expression. */
1445 static pre_expr
1446 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1447 basic_block pred, basic_block phiblock)
1449 switch (expr->kind)
1451 case NARY:
1453 unsigned int i;
1454 bool changed = false;
1455 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1456 struct vn_nary_op_s newnary;
1457 /* The NARY structure is only guaranteed to have been
1458 allocated to the nary->length operands. */
1459 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1460 - sizeof (tree) * (4 - nary->length)));
1462 for (i = 0; i < newnary.length; i++)
1464 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1465 continue;
1466 else
1468 pre_expr leader, result;
1469 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1470 leader = find_leader_in_sets (op_val_id, set1, set2);
1471 result = phi_translate (leader, set1, set2, pred, phiblock);
1472 if (result && result != leader)
1474 tree name = get_representative_for (result);
1475 if (!name)
1476 return NULL;
1477 newnary.op[i] = name;
1479 else if (!result)
1480 return NULL;
1482 changed |= newnary.op[i] != nary->op[i];
1485 if (changed)
1487 pre_expr constant;
1488 unsigned int new_val_id;
1490 tree result = vn_nary_op_lookup_pieces (newnary.length,
1491 newnary.opcode,
1492 newnary.type,
1493 newnary.op[0],
1494 newnary.op[1],
1495 newnary.op[2],
1496 newnary.op[3],
1497 &nary);
1498 if (result && is_gimple_min_invariant (result))
1499 return get_or_alloc_expr_for_constant (result);
1501 expr = (pre_expr) pool_alloc (pre_expr_pool);
1502 expr->kind = NARY;
1503 expr->id = 0;
1504 if (nary)
1506 PRE_EXPR_NARY (expr) = nary;
1507 constant = fully_constant_expression (expr);
1508 if (constant != expr)
1509 return constant;
1511 new_val_id = nary->value_id;
1512 get_or_alloc_expression_id (expr);
1514 else
1516 new_val_id = get_next_value_id ();
1517 VEC_safe_grow_cleared (bitmap_set_t, heap,
1518 value_expressions,
1519 get_max_value_id() + 1);
1520 nary = vn_nary_op_insert_pieces (newnary.length,
1521 newnary.opcode,
1522 newnary.type,
1523 newnary.op[0],
1524 newnary.op[1],
1525 newnary.op[2],
1526 newnary.op[3],
1527 result, new_val_id);
1528 PRE_EXPR_NARY (expr) = nary;
1529 constant = fully_constant_expression (expr);
1530 if (constant != expr)
1531 return constant;
1532 get_or_alloc_expression_id (expr);
1534 add_to_value (new_val_id, expr);
1536 return expr;
1538 break;
1540 case REFERENCE:
1542 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1543 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1544 tree vuse = ref->vuse;
1545 tree newvuse = vuse;
1546 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1547 bool changed = false, same_valid = true;
1548 unsigned int i, j;
1549 vn_reference_op_t operand;
1550 vn_reference_t newref;
1552 for (i = 0, j = 0;
1553 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1555 pre_expr opresult;
1556 pre_expr leader;
1557 tree oldop0 = operand->op0;
1558 tree oldop1 = operand->op1;
1559 tree oldop2 = operand->op2;
1560 tree op0 = oldop0;
1561 tree op1 = oldop1;
1562 tree op2 = oldop2;
1563 tree type = operand->type;
1564 vn_reference_op_s newop = *operand;
1566 if (op0 && TREE_CODE (op0) == SSA_NAME)
1568 unsigned int op_val_id = VN_INFO (op0)->value_id;
1569 leader = find_leader_in_sets (op_val_id, set1, set2);
1570 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1571 if (opresult && opresult != leader)
1573 tree name = get_representative_for (opresult);
1574 if (!name)
1575 break;
1576 op0 = name;
1578 else if (!opresult)
1579 break;
1581 changed |= op0 != oldop0;
1583 if (op1 && TREE_CODE (op1) == SSA_NAME)
1585 unsigned int op_val_id = VN_INFO (op1)->value_id;
1586 leader = find_leader_in_sets (op_val_id, set1, set2);
1587 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1588 if (opresult && opresult != leader)
1590 tree name = get_representative_for (opresult);
1591 if (!name)
1592 break;
1593 op1 = name;
1595 else if (!opresult)
1596 break;
1598 /* We can't possibly insert these. */
1599 else if (op1 && !is_gimple_min_invariant (op1))
1600 break;
1601 changed |= op1 != oldop1;
1602 if (op2 && TREE_CODE (op2) == SSA_NAME)
1604 unsigned int op_val_id = VN_INFO (op2)->value_id;
1605 leader = find_leader_in_sets (op_val_id, set1, set2);
1606 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1607 if (opresult && opresult != leader)
1609 tree name = get_representative_for (opresult);
1610 if (!name)
1611 break;
1612 op2 = name;
1614 else if (!opresult)
1615 break;
1617 /* We can't possibly insert these. */
1618 else if (op2 && !is_gimple_min_invariant (op2))
1619 break;
1620 changed |= op2 != oldop2;
1622 if (!newoperands)
1623 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1624 /* We may have changed from an SSA_NAME to a constant */
1625 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1626 newop.opcode = TREE_CODE (op0);
1627 newop.type = type;
1628 newop.op0 = op0;
1629 newop.op1 = op1;
1630 newop.op2 = op2;
1631 /* If it transforms a non-constant ARRAY_REF into a constant
1632 one, adjust the constant offset. */
1633 if (newop.opcode == ARRAY_REF
1634 && newop.off == -1
1635 && TREE_CODE (op0) == INTEGER_CST
1636 && TREE_CODE (op1) == INTEGER_CST
1637 && TREE_CODE (op2) == INTEGER_CST)
1639 double_int off = tree_to_double_int (op0);
1640 off = double_int_add (off,
1641 double_int_neg
1642 (tree_to_double_int (op1)));
1643 off = double_int_mul (off, tree_to_double_int (op2));
1644 if (double_int_fits_in_shwi_p (off))
1645 newop.off = off.low;
1647 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1648 /* If it transforms from an SSA_NAME to an address, fold with
1649 a preceding indirect reference. */
1650 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1651 && VEC_index (vn_reference_op_s,
1652 newoperands, j - 1)->opcode == MEM_REF)
1653 vn_reference_fold_indirect (&newoperands, &j);
1655 if (i != VEC_length (vn_reference_op_s, operands))
1657 if (newoperands)
1658 VEC_free (vn_reference_op_s, heap, newoperands);
1659 return NULL;
1662 if (vuse)
1664 newvuse = translate_vuse_through_block (newoperands,
1665 ref->set, ref->type,
1666 vuse, phiblock, pred,
1667 &same_valid);
1668 if (newvuse == NULL_TREE)
1670 VEC_free (vn_reference_op_s, heap, newoperands);
1671 return NULL;
1675 if (changed || newvuse != vuse)
1677 unsigned int new_val_id;
1678 pre_expr constant;
1679 bool converted = false;
1681 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1682 ref->type,
1683 newoperands,
1684 &newref, VN_WALK);
1685 if (result)
1686 VEC_free (vn_reference_op_s, heap, newoperands);
1688 if (result
1689 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1691 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1692 converted = true;
1694 else if (!result && newref
1695 && !useless_type_conversion_p (ref->type, newref->type))
1697 VEC_free (vn_reference_op_s, heap, newoperands);
1698 return NULL;
1701 if (result && is_gimple_min_invariant (result))
1703 gcc_assert (!newoperands);
1704 return get_or_alloc_expr_for_constant (result);
1707 expr = (pre_expr) pool_alloc (pre_expr_pool);
1708 expr->kind = REFERENCE;
1709 expr->id = 0;
1711 if (converted)
1713 vn_nary_op_t nary;
1714 tree nresult;
1716 gcc_assert (CONVERT_EXPR_P (result)
1717 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1719 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1720 TREE_TYPE (result),
1721 TREE_OPERAND (result, 0),
1722 NULL_TREE, NULL_TREE,
1723 NULL_TREE,
1724 &nary);
1725 if (nresult && is_gimple_min_invariant (nresult))
1726 return get_or_alloc_expr_for_constant (nresult);
1728 expr->kind = NARY;
1729 if (nary)
1731 PRE_EXPR_NARY (expr) = nary;
1732 constant = fully_constant_expression (expr);
1733 if (constant != expr)
1734 return constant;
1736 new_val_id = nary->value_id;
1737 get_or_alloc_expression_id (expr);
1739 else
1741 new_val_id = get_next_value_id ();
1742 VEC_safe_grow_cleared (bitmap_set_t, heap,
1743 value_expressions,
1744 get_max_value_id() + 1);
1745 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1746 TREE_TYPE (result),
1747 TREE_OPERAND (result, 0),
1748 NULL_TREE, NULL_TREE,
1749 NULL_TREE, NULL_TREE,
1750 new_val_id);
1751 PRE_EXPR_NARY (expr) = nary;
1752 constant = fully_constant_expression (expr);
1753 if (constant != expr)
1754 return constant;
1755 get_or_alloc_expression_id (expr);
1758 else if (newref)
1760 PRE_EXPR_REFERENCE (expr) = newref;
1761 constant = fully_constant_expression (expr);
1762 if (constant != expr)
1763 return constant;
1765 new_val_id = newref->value_id;
1766 get_or_alloc_expression_id (expr);
1768 else
1770 if (changed || !same_valid)
1772 new_val_id = get_next_value_id ();
1773 VEC_safe_grow_cleared (bitmap_set_t, heap,
1774 value_expressions,
1775 get_max_value_id() + 1);
1777 else
1778 new_val_id = ref->value_id;
1779 newref = vn_reference_insert_pieces (newvuse, ref->set,
1780 ref->type,
1781 newoperands,
1782 result, new_val_id);
1783 newoperands = NULL;
1784 PRE_EXPR_REFERENCE (expr) = newref;
1785 constant = fully_constant_expression (expr);
1786 if (constant != expr)
1787 return constant;
1788 get_or_alloc_expression_id (expr);
1790 add_to_value (new_val_id, expr);
1792 VEC_free (vn_reference_op_s, heap, newoperands);
1793 return expr;
1795 break;
1797 case NAME:
1799 gimple phi = NULL;
1800 edge e;
1801 gimple def_stmt;
1802 tree name = PRE_EXPR_NAME (expr);
1804 def_stmt = SSA_NAME_DEF_STMT (name);
1805 if (gimple_code (def_stmt) == GIMPLE_PHI
1806 && gimple_bb (def_stmt) == phiblock)
1807 phi = def_stmt;
1808 else
1809 return expr;
1811 e = find_edge (pred, gimple_bb (phi));
1812 if (e)
1814 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1815 pre_expr newexpr;
1817 if (TREE_CODE (def) == SSA_NAME)
1818 def = VN_INFO (def)->valnum;
1820 /* Handle constant. */
1821 if (is_gimple_min_invariant (def))
1822 return get_or_alloc_expr_for_constant (def);
1824 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1825 return NULL;
1827 newexpr = get_or_alloc_expr_for_name (def);
1828 return newexpr;
1831 return expr;
1833 default:
1834 gcc_unreachable ();
1838 /* Wrapper around phi_translate_1 providing caching functionality. */
1840 static pre_expr
1841 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1842 basic_block pred, basic_block phiblock)
1844 pre_expr phitrans;
1846 if (!expr)
1847 return NULL;
1849 /* Constants contain no values that need translation. */
1850 if (expr->kind == CONSTANT)
1851 return expr;
1853 if (value_id_constant_p (get_expr_value_id (expr)))
1854 return expr;
1856 if (expr->kind != NAME)
1858 phitrans = phi_trans_lookup (expr, pred);
1859 if (phitrans)
1860 return phitrans;
1863 /* Translate. */
1864 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1866 /* Don't add empty translations to the cache. Neither add
1867 translations of NAMEs as those are cheap to translate. */
1868 if (phitrans
1869 && expr->kind != NAME)
1870 phi_trans_add (expr, phitrans, pred);
1872 return phitrans;
1876 /* For each expression in SET, translate the values through phi nodes
1877 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1878 expressions in DEST. */
1880 static void
1881 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1882 basic_block phiblock)
1884 VEC (pre_expr, heap) *exprs;
1885 pre_expr expr;
1886 int i;
1888 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1890 bitmap_set_copy (dest, set);
1891 return;
1894 exprs = sorted_array_from_bitmap_set (set);
1895 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1897 pre_expr translated;
1898 translated = phi_translate (expr, set, NULL, pred, phiblock);
1899 if (!translated)
1900 continue;
1902 /* We might end up with multiple expressions from SET being
1903 translated to the same value. In this case we do not want
1904 to retain the NARY or REFERENCE expression but prefer a NAME
1905 which would be the leader. */
1906 if (translated->kind == NAME)
1907 bitmap_value_replace_in_set (dest, translated);
1908 else
1909 bitmap_value_insert_into_set (dest, translated);
1911 VEC_free (pre_expr, heap, exprs);
1914 /* Find the leader for a value (i.e., the name representing that
1915 value) in a given set, and return it. If STMT is non-NULL it
1916 makes sure the defining statement for the leader dominates it.
1917 Return NULL if no leader is found. */
1919 static pre_expr
1920 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1922 if (value_id_constant_p (val))
1924 unsigned int i;
1925 bitmap_iterator bi;
1926 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1928 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1930 pre_expr expr = expression_for_id (i);
1931 if (expr->kind == CONSTANT)
1932 return expr;
1935 if (bitmap_set_contains_value (set, val))
1937 /* Rather than walk the entire bitmap of expressions, and see
1938 whether any of them has the value we are looking for, we look
1939 at the reverse mapping, which tells us the set of expressions
1940 that have a given value (IE value->expressions with that
1941 value) and see if any of those expressions are in our set.
1942 The number of expressions per value is usually significantly
1943 less than the number of expressions in the set. In fact, for
1944 large testcases, doing it this way is roughly 5-10x faster
1945 than walking the bitmap.
1946 If this is somehow a significant lose for some cases, we can
1947 choose which set to walk based on which set is smaller. */
1948 unsigned int i;
1949 bitmap_iterator bi;
1950 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1952 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1953 &set->expressions, 0, i, bi)
1955 pre_expr val = expression_for_id (i);
1956 /* At the point where stmt is not null, there should always
1957 be an SSA_NAME first in the list of expressions. */
1958 if (stmt)
1960 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1961 if (gimple_code (def_stmt) != GIMPLE_PHI
1962 && gimple_bb (def_stmt) == gimple_bb (stmt)
1963 /* PRE insertions are at the end of the basic-block
1964 and have UID 0. */
1965 && (gimple_uid (def_stmt) == 0
1966 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1967 continue;
1969 return val;
1972 return NULL;
1975 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1976 BLOCK by seeing if it is not killed in the block. Note that we are
1977 only determining whether there is a store that kills it. Because
1978 of the order in which clean iterates over values, we are guaranteed
1979 that altered operands will have caused us to be eliminated from the
1980 ANTIC_IN set already. */
1982 static bool
1983 value_dies_in_block_x (pre_expr expr, basic_block block)
1985 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1986 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1987 gimple def;
1988 gimple_stmt_iterator gsi;
1989 unsigned id = get_expression_id (expr);
1990 bool res = false;
1991 ao_ref ref;
1993 if (!vuse)
1994 return false;
1996 /* Lookup a previously calculated result. */
1997 if (EXPR_DIES (block)
1998 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1999 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
2001 /* A memory expression {e, VUSE} dies in the block if there is a
2002 statement that may clobber e. If, starting statement walk from the
2003 top of the basic block, a statement uses VUSE there can be no kill
2004 inbetween that use and the original statement that loaded {e, VUSE},
2005 so we can stop walking. */
2006 ref.base = NULL_TREE;
2007 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
2009 tree def_vuse, def_vdef;
2010 def = gsi_stmt (gsi);
2011 def_vuse = gimple_vuse (def);
2012 def_vdef = gimple_vdef (def);
2014 /* Not a memory statement. */
2015 if (!def_vuse)
2016 continue;
2018 /* Not a may-def. */
2019 if (!def_vdef)
2021 /* A load with the same VUSE, we're done. */
2022 if (def_vuse == vuse)
2023 break;
2025 continue;
2028 /* Init ref only if we really need it. */
2029 if (ref.base == NULL_TREE
2030 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
2031 refx->operands))
2033 res = true;
2034 break;
2036 /* If the statement may clobber expr, it dies. */
2037 if (stmt_may_clobber_ref_p_1 (def, &ref))
2039 res = true;
2040 break;
2044 /* Remember the result. */
2045 if (!EXPR_DIES (block))
2046 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2047 bitmap_set_bit (EXPR_DIES (block), id * 2);
2048 if (res)
2049 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2051 return res;
2055 #define union_contains_value(SET1, SET2, VAL) \
2056 (bitmap_set_contains_value ((SET1), (VAL)) \
2057 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
2059 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2061 static bool
2062 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2063 vn_reference_op_t vro)
2065 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2067 struct pre_expr_d temp;
2068 temp.kind = NAME;
2069 temp.id = 0;
2070 PRE_EXPR_NAME (&temp) = vro->op0;
2071 temp.id = lookup_expression_id (&temp);
2072 if (temp.id == 0)
2073 return false;
2074 if (!union_contains_value (set1, set2,
2075 get_expr_value_id (&temp)))
2076 return false;
2078 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2080 struct pre_expr_d temp;
2081 temp.kind = NAME;
2082 temp.id = 0;
2083 PRE_EXPR_NAME (&temp) = vro->op1;
2084 temp.id = lookup_expression_id (&temp);
2085 if (temp.id == 0)
2086 return false;
2087 if (!union_contains_value (set1, set2,
2088 get_expr_value_id (&temp)))
2089 return false;
2092 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2094 struct pre_expr_d temp;
2095 temp.kind = NAME;
2096 temp.id = 0;
2097 PRE_EXPR_NAME (&temp) = vro->op2;
2098 temp.id = lookup_expression_id (&temp);
2099 if (temp.id == 0)
2100 return false;
2101 if (!union_contains_value (set1, set2,
2102 get_expr_value_id (&temp)))
2103 return false;
2106 return true;
2109 /* Determine if the expression EXPR is valid in SET1 U SET2.
2110 ONLY SET2 CAN BE NULL.
2111 This means that we have a leader for each part of the expression
2112 (if it consists of values), or the expression is an SSA_NAME.
2113 For loads/calls, we also see if the vuse is killed in this block. */
2115 static bool
2116 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2117 basic_block block)
2119 switch (expr->kind)
2121 case NAME:
2122 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2123 case NARY:
2125 unsigned int i;
2126 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2127 for (i = 0; i < nary->length; i++)
2129 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2131 struct pre_expr_d temp;
2132 temp.kind = NAME;
2133 temp.id = 0;
2134 PRE_EXPR_NAME (&temp) = nary->op[i];
2135 temp.id = lookup_expression_id (&temp);
2136 if (temp.id == 0)
2137 return false;
2138 if (!union_contains_value (set1, set2,
2139 get_expr_value_id (&temp)))
2140 return false;
2143 /* If the NARY may trap make sure the block does not contain
2144 a possible exit point.
2145 ??? This is overly conservative if we translate AVAIL_OUT
2146 as the available expression might be after the exit point. */
2147 if (BB_MAY_NOTRETURN (block)
2148 && vn_nary_may_trap (nary))
2149 return false;
2150 return true;
2152 break;
2153 case REFERENCE:
2155 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2156 vn_reference_op_t vro;
2157 unsigned int i;
2159 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2161 if (!vro_valid_in_sets (set1, set2, vro))
2162 return false;
2164 if (ref->vuse)
2166 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2167 if (!gimple_nop_p (def_stmt)
2168 && gimple_bb (def_stmt) != block
2169 && !dominated_by_p (CDI_DOMINATORS,
2170 block, gimple_bb (def_stmt)))
2171 return false;
2173 return !value_dies_in_block_x (expr, block);
2175 default:
2176 gcc_unreachable ();
2180 /* Clean the set of expressions that are no longer valid in SET1 or
2181 SET2. This means expressions that are made up of values we have no
2182 leaders for in SET1 or SET2. This version is used for partial
2183 anticipation, which means it is not valid in either ANTIC_IN or
2184 PA_IN. */
2186 static void
2187 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2189 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2190 pre_expr expr;
2191 int i;
2193 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2195 if (!valid_in_sets (set1, set2, expr, block))
2196 bitmap_remove_from_set (set1, expr);
2198 VEC_free (pre_expr, heap, exprs);
2201 /* Clean the set of expressions that are no longer valid in SET. This
2202 means expressions that are made up of values we have no leaders for
2203 in SET. */
2205 static void
2206 clean (bitmap_set_t set, basic_block block)
2208 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2209 pre_expr expr;
2210 int i;
2212 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2214 if (!valid_in_sets (set, NULL, expr, block))
2215 bitmap_remove_from_set (set, expr);
2217 VEC_free (pre_expr, heap, exprs);
2220 static sbitmap has_abnormal_preds;
2222 /* List of blocks that may have changed during ANTIC computation and
2223 thus need to be iterated over. */
2225 static sbitmap changed_blocks;
2227 /* Decide whether to defer a block for a later iteration, or PHI
2228 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2229 should defer the block, and true if we processed it. */
2231 static bool
2232 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2233 basic_block block, basic_block phiblock)
2235 if (!BB_VISITED (phiblock))
2237 SET_BIT (changed_blocks, block->index);
2238 BB_VISITED (block) = 0;
2239 BB_DEFERRED (block) = 1;
2240 return false;
2242 else
2243 phi_translate_set (dest, source, block, phiblock);
2244 return true;
2247 /* Compute the ANTIC set for BLOCK.
2249 If succs(BLOCK) > 1 then
2250 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2251 else if succs(BLOCK) == 1 then
2252 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2254 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2257 static bool
2258 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2260 bool changed = false;
2261 bitmap_set_t S, old, ANTIC_OUT;
2262 bitmap_iterator bi;
2263 unsigned int bii;
2264 edge e;
2265 edge_iterator ei;
2267 old = ANTIC_OUT = S = NULL;
2268 BB_VISITED (block) = 1;
2270 /* If any edges from predecessors are abnormal, antic_in is empty,
2271 so do nothing. */
2272 if (block_has_abnormal_pred_edge)
2273 goto maybe_dump_sets;
2275 old = ANTIC_IN (block);
2276 ANTIC_OUT = bitmap_set_new ();
2278 /* If the block has no successors, ANTIC_OUT is empty. */
2279 if (EDGE_COUNT (block->succs) == 0)
2281 /* If we have one successor, we could have some phi nodes to
2282 translate through. */
2283 else if (single_succ_p (block))
2285 basic_block succ_bb = single_succ (block);
2287 /* We trade iterations of the dataflow equations for having to
2288 phi translate the maximal set, which is incredibly slow
2289 (since the maximal set often has 300+ members, even when you
2290 have a small number of blocks).
2291 Basically, we defer the computation of ANTIC for this block
2292 until we have processed it's successor, which will inevitably
2293 have a *much* smaller set of values to phi translate once
2294 clean has been run on it.
2295 The cost of doing this is that we technically perform more
2296 iterations, however, they are lower cost iterations.
2298 Timings for PRE on tramp3d-v4:
2299 without maximal set fix: 11 seconds
2300 with maximal set fix/without deferring: 26 seconds
2301 with maximal set fix/with deferring: 11 seconds
2304 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2305 block, succ_bb))
2307 changed = true;
2308 goto maybe_dump_sets;
2311 /* If we have multiple successors, we take the intersection of all of
2312 them. Note that in the case of loop exit phi nodes, we may have
2313 phis to translate through. */
2314 else
2316 VEC(basic_block, heap) * worklist;
2317 size_t i;
2318 basic_block bprime, first = NULL;
2320 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2321 FOR_EACH_EDGE (e, ei, block->succs)
2323 if (!first
2324 && BB_VISITED (e->dest))
2325 first = e->dest;
2326 else if (BB_VISITED (e->dest))
2327 VEC_quick_push (basic_block, worklist, e->dest);
2330 /* Of multiple successors we have to have visited one already. */
2331 if (!first)
2333 SET_BIT (changed_blocks, block->index);
2334 BB_VISITED (block) = 0;
2335 BB_DEFERRED (block) = 1;
2336 changed = true;
2337 VEC_free (basic_block, heap, worklist);
2338 goto maybe_dump_sets;
2341 if (!gimple_seq_empty_p (phi_nodes (first)))
2342 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2343 else
2344 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2346 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2348 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2350 bitmap_set_t tmp = bitmap_set_new ();
2351 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2352 bitmap_set_and (ANTIC_OUT, tmp);
2353 bitmap_set_free (tmp);
2355 else
2356 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2358 VEC_free (basic_block, heap, worklist);
2361 /* Generate ANTIC_OUT - TMP_GEN. */
2362 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2364 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2365 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2366 TMP_GEN (block));
2368 /* Then union in the ANTIC_OUT - TMP_GEN values,
2369 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2370 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2371 bitmap_value_insert_into_set (ANTIC_IN (block),
2372 expression_for_id (bii));
2374 clean (ANTIC_IN (block), block);
2376 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2378 changed = true;
2379 SET_BIT (changed_blocks, block->index);
2380 FOR_EACH_EDGE (e, ei, block->preds)
2381 SET_BIT (changed_blocks, e->src->index);
2383 else
2384 RESET_BIT (changed_blocks, block->index);
2386 maybe_dump_sets:
2387 if (dump_file && (dump_flags & TDF_DETAILS))
2389 if (!BB_DEFERRED (block) || BB_VISITED (block))
2391 if (ANTIC_OUT)
2392 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2394 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2395 block->index);
2397 if (S)
2398 print_bitmap_set (dump_file, S, "S", block->index);
2400 else
2402 fprintf (dump_file,
2403 "Block %d was deferred for a future iteration.\n",
2404 block->index);
2407 if (old)
2408 bitmap_set_free (old);
2409 if (S)
2410 bitmap_set_free (S);
2411 if (ANTIC_OUT)
2412 bitmap_set_free (ANTIC_OUT);
2413 return changed;
2416 /* Compute PARTIAL_ANTIC for BLOCK.
2418 If succs(BLOCK) > 1 then
2419 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2420 in ANTIC_OUT for all succ(BLOCK)
2421 else if succs(BLOCK) == 1 then
2422 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2424 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2425 - ANTIC_IN[BLOCK])
2428 static bool
2429 compute_partial_antic_aux (basic_block block,
2430 bool block_has_abnormal_pred_edge)
2432 bool changed = false;
2433 bitmap_set_t old_PA_IN;
2434 bitmap_set_t PA_OUT;
2435 edge e;
2436 edge_iterator ei;
2437 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2439 old_PA_IN = PA_OUT = NULL;
2441 /* If any edges from predecessors are abnormal, antic_in is empty,
2442 so do nothing. */
2443 if (block_has_abnormal_pred_edge)
2444 goto maybe_dump_sets;
2446 /* If there are too many partially anticipatable values in the
2447 block, phi_translate_set can take an exponential time: stop
2448 before the translation starts. */
2449 if (max_pa
2450 && single_succ_p (block)
2451 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2452 goto maybe_dump_sets;
2454 old_PA_IN = PA_IN (block);
2455 PA_OUT = bitmap_set_new ();
2457 /* If the block has no successors, ANTIC_OUT is empty. */
2458 if (EDGE_COUNT (block->succs) == 0)
2460 /* If we have one successor, we could have some phi nodes to
2461 translate through. Note that we can't phi translate across DFS
2462 back edges in partial antic, because it uses a union operation on
2463 the successors. For recurrences like IV's, we will end up
2464 generating a new value in the set on each go around (i + 3 (VH.1)
2465 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2466 else if (single_succ_p (block))
2468 basic_block succ = single_succ (block);
2469 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2470 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2472 /* If we have multiple successors, we take the union of all of
2473 them. */
2474 else
2476 VEC(basic_block, heap) * worklist;
2477 size_t i;
2478 basic_block bprime;
2480 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2481 FOR_EACH_EDGE (e, ei, block->succs)
2483 if (e->flags & EDGE_DFS_BACK)
2484 continue;
2485 VEC_quick_push (basic_block, worklist, e->dest);
2487 if (VEC_length (basic_block, worklist) > 0)
2489 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2491 unsigned int i;
2492 bitmap_iterator bi;
2494 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2495 bitmap_value_insert_into_set (PA_OUT,
2496 expression_for_id (i));
2497 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2499 bitmap_set_t pa_in = bitmap_set_new ();
2500 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2501 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2502 bitmap_value_insert_into_set (PA_OUT,
2503 expression_for_id (i));
2504 bitmap_set_free (pa_in);
2506 else
2507 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2508 bitmap_value_insert_into_set (PA_OUT,
2509 expression_for_id (i));
2512 VEC_free (basic_block, heap, worklist);
2515 /* PA_IN starts with PA_OUT - TMP_GEN.
2516 Then we subtract things from ANTIC_IN. */
2517 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2519 /* For partial antic, we want to put back in the phi results, since
2520 we will properly avoid making them partially antic over backedges. */
2521 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2522 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2524 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2525 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2527 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2529 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2531 changed = true;
2532 SET_BIT (changed_blocks, block->index);
2533 FOR_EACH_EDGE (e, ei, block->preds)
2534 SET_BIT (changed_blocks, e->src->index);
2536 else
2537 RESET_BIT (changed_blocks, block->index);
2539 maybe_dump_sets:
2540 if (dump_file && (dump_flags & TDF_DETAILS))
2542 if (PA_OUT)
2543 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2545 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2547 if (old_PA_IN)
2548 bitmap_set_free (old_PA_IN);
2549 if (PA_OUT)
2550 bitmap_set_free (PA_OUT);
2551 return changed;
2554 /* Compute ANTIC and partial ANTIC sets. */
2556 static void
2557 compute_antic (void)
2559 bool changed = true;
2560 int num_iterations = 0;
2561 basic_block block;
2562 int i;
2564 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2565 We pre-build the map of blocks with incoming abnormal edges here. */
2566 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2567 sbitmap_zero (has_abnormal_preds);
2569 FOR_EACH_BB (block)
2571 edge_iterator ei;
2572 edge e;
2574 FOR_EACH_EDGE (e, ei, block->preds)
2576 e->flags &= ~EDGE_DFS_BACK;
2577 if (e->flags & EDGE_ABNORMAL)
2579 SET_BIT (has_abnormal_preds, block->index);
2580 break;
2584 BB_VISITED (block) = 0;
2585 BB_DEFERRED (block) = 0;
2587 /* While we are here, give empty ANTIC_IN sets to each block. */
2588 ANTIC_IN (block) = bitmap_set_new ();
2589 PA_IN (block) = bitmap_set_new ();
2592 /* At the exit block we anticipate nothing. */
2593 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2594 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2595 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2597 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2598 sbitmap_ones (changed_blocks);
2599 while (changed)
2601 if (dump_file && (dump_flags & TDF_DETAILS))
2602 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2603 /* ??? We need to clear our PHI translation cache here as the
2604 ANTIC sets shrink and we restrict valid translations to
2605 those having operands with leaders in ANTIC. Same below
2606 for PA ANTIC computation. */
2607 num_iterations++;
2608 changed = false;
2609 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2611 if (TEST_BIT (changed_blocks, postorder[i]))
2613 basic_block block = BASIC_BLOCK (postorder[i]);
2614 changed |= compute_antic_aux (block,
2615 TEST_BIT (has_abnormal_preds,
2616 block->index));
2619 /* Theoretically possible, but *highly* unlikely. */
2620 gcc_checking_assert (num_iterations < 500);
2623 statistics_histogram_event (cfun, "compute_antic iterations",
2624 num_iterations);
2626 if (do_partial_partial)
2628 sbitmap_ones (changed_blocks);
2629 mark_dfs_back_edges ();
2630 num_iterations = 0;
2631 changed = true;
2632 while (changed)
2634 if (dump_file && (dump_flags & TDF_DETAILS))
2635 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2636 num_iterations++;
2637 changed = false;
2638 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2640 if (TEST_BIT (changed_blocks, postorder[i]))
2642 basic_block block = BASIC_BLOCK (postorder[i]);
2643 changed
2644 |= compute_partial_antic_aux (block,
2645 TEST_BIT (has_abnormal_preds,
2646 block->index));
2649 /* Theoretically possible, but *highly* unlikely. */
2650 gcc_checking_assert (num_iterations < 500);
2652 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2653 num_iterations);
2655 sbitmap_free (has_abnormal_preds);
2656 sbitmap_free (changed_blocks);
2659 /* Return true if we can value number the call in STMT. This is true
2660 if we have a pure or constant call. */
2662 static bool
2663 can_value_number_call (gimple stmt)
2665 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2666 return true;
2667 return false;
2670 /* Return true if OP is a tree which we can perform PRE on.
2671 This may not match the operations we can value number, but in
2672 a perfect world would. */
2674 static bool
2675 can_PRE_operation (tree op)
2677 return UNARY_CLASS_P (op)
2678 || BINARY_CLASS_P (op)
2679 || COMPARISON_CLASS_P (op)
2680 || TREE_CODE (op) == MEM_REF
2681 || TREE_CODE (op) == COMPONENT_REF
2682 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2683 || TREE_CODE (op) == CALL_EXPR
2684 || TREE_CODE (op) == ARRAY_REF;
2688 /* Inserted expressions are placed onto this worklist, which is used
2689 for performing quick dead code elimination of insertions we made
2690 that didn't turn out to be necessary. */
2691 static bitmap inserted_exprs;
2693 /* Pool allocated fake store expressions are placed onto this
2694 worklist, which, after performing dead code elimination, is walked
2695 to see which expressions need to be put into GC'able memory */
2696 static VEC(gimple, heap) *need_creation;
2698 /* The actual worker for create_component_ref_by_pieces. */
2700 static tree
2701 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2702 unsigned int *operand, gimple_seq *stmts,
2703 gimple domstmt)
2705 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2706 *operand);
2707 tree genop;
2708 ++*operand;
2709 switch (currop->opcode)
2711 case CALL_EXPR:
2713 tree folded, sc = NULL_TREE;
2714 unsigned int nargs = 0;
2715 tree fn, *args;
2716 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2717 fn = currop->op0;
2718 else
2720 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2721 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2722 if (!fn)
2723 return NULL_TREE;
2725 if (currop->op1)
2727 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2728 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2729 if (!sc)
2730 return NULL_TREE;
2732 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2733 ref->operands) - 1);
2734 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2736 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2737 operand, stmts,
2738 domstmt);
2739 if (!args[nargs])
2741 free (args);
2742 return NULL_TREE;
2744 nargs++;
2746 folded = build_call_array (currop->type,
2747 (TREE_CODE (fn) == FUNCTION_DECL
2748 ? build_fold_addr_expr (fn) : fn),
2749 nargs, args);
2750 free (args);
2751 if (sc)
2752 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2753 return folded;
2755 break;
2756 case MEM_REF:
2758 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2759 stmts, domstmt);
2760 tree offset = currop->op0;
2761 if (!baseop)
2762 return NULL_TREE;
2763 if (TREE_CODE (baseop) == ADDR_EXPR
2764 && handled_component_p (TREE_OPERAND (baseop, 0)))
2766 HOST_WIDE_INT off;
2767 tree base;
2768 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2769 &off);
2770 gcc_assert (base);
2771 offset = int_const_binop (PLUS_EXPR, offset,
2772 build_int_cst (TREE_TYPE (offset),
2773 off), 0);
2774 baseop = build_fold_addr_expr (base);
2776 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2778 break;
2779 case TARGET_MEM_REF:
2781 pre_expr op0expr, op1expr;
2782 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2783 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2784 ++*operand);
2785 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2786 stmts, domstmt);
2787 if (!baseop)
2788 return NULL_TREE;
2789 if (currop->op0)
2791 op0expr = get_or_alloc_expr_for (currop->op0);
2792 genop0 = find_or_generate_expression (block, op0expr,
2793 stmts, domstmt);
2794 if (!genop0)
2795 return NULL_TREE;
2797 if (nextop->op0)
2799 op1expr = get_or_alloc_expr_for (nextop->op0);
2800 genop1 = find_or_generate_expression (block, op1expr,
2801 stmts, domstmt);
2802 if (!genop1)
2803 return NULL_TREE;
2805 return build5 (TARGET_MEM_REF, currop->type,
2806 baseop, currop->op2, genop0, currop->op1, genop1);
2808 break;
2809 case ADDR_EXPR:
2810 if (currop->op0)
2812 gcc_assert (is_gimple_min_invariant (currop->op0));
2813 return currop->op0;
2815 /* Fallthrough. */
2816 case REALPART_EXPR:
2817 case IMAGPART_EXPR:
2818 case VIEW_CONVERT_EXPR:
2820 tree folded;
2821 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2822 operand,
2823 stmts, domstmt);
2824 if (!genop0)
2825 return NULL_TREE;
2826 folded = fold_build1 (currop->opcode, currop->type,
2827 genop0);
2828 return folded;
2830 break;
2831 case BIT_FIELD_REF:
2833 tree folded;
2834 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2835 stmts, domstmt);
2836 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2837 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2838 tree genop1;
2839 tree genop2;
2841 if (!genop0)
2842 return NULL_TREE;
2843 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2844 if (!genop1)
2845 return NULL_TREE;
2846 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2847 if (!genop2)
2848 return NULL_TREE;
2849 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2850 genop2);
2851 return folded;
2854 /* For array ref vn_reference_op's, operand 1 of the array ref
2855 is op0 of the reference op and operand 3 of the array ref is
2856 op1. */
2857 case ARRAY_RANGE_REF:
2858 case ARRAY_REF:
2860 tree genop0;
2861 tree genop1 = currop->op0;
2862 pre_expr op1expr;
2863 tree genop2 = currop->op1;
2864 pre_expr op2expr;
2865 tree genop3 = currop->op2;
2866 pre_expr op3expr;
2867 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2868 stmts, domstmt);
2869 if (!genop0)
2870 return NULL_TREE;
2871 op1expr = get_or_alloc_expr_for (genop1);
2872 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2873 if (!genop1)
2874 return NULL_TREE;
2875 if (genop2)
2877 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2878 /* Drop zero minimum index if redundant. */
2879 if (integer_zerop (genop2)
2880 && (!domain_type
2881 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2882 genop2 = NULL_TREE;
2883 else
2885 op2expr = get_or_alloc_expr_for (genop2);
2886 genop2 = find_or_generate_expression (block, op2expr, stmts,
2887 domstmt);
2888 if (!genop2)
2889 return NULL_TREE;
2892 if (genop3)
2894 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2895 /* We can't always put a size in units of the element alignment
2896 here as the element alignment may be not visible. See
2897 PR43783. Simply drop the element size for constant
2898 sizes. */
2899 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2900 genop3 = NULL_TREE;
2901 else
2903 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2904 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2905 op3expr = get_or_alloc_expr_for (genop3);
2906 genop3 = find_or_generate_expression (block, op3expr, stmts,
2907 domstmt);
2908 if (!genop3)
2909 return NULL_TREE;
2912 return build4 (currop->opcode, currop->type, genop0, genop1,
2913 genop2, genop3);
2915 case COMPONENT_REF:
2917 tree op0;
2918 tree op1;
2919 tree genop2 = currop->op1;
2920 pre_expr op2expr;
2921 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2922 stmts, domstmt);
2923 if (!op0)
2924 return NULL_TREE;
2925 /* op1 should be a FIELD_DECL, which are represented by
2926 themselves. */
2927 op1 = currop->op0;
2928 if (genop2)
2930 op2expr = get_or_alloc_expr_for (genop2);
2931 genop2 = find_or_generate_expression (block, op2expr, stmts,
2932 domstmt);
2933 if (!genop2)
2934 return NULL_TREE;
2937 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2938 genop2);
2940 break;
2941 case SSA_NAME:
2943 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2944 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2945 return genop;
2947 case STRING_CST:
2948 case INTEGER_CST:
2949 case COMPLEX_CST:
2950 case VECTOR_CST:
2951 case REAL_CST:
2952 case CONSTRUCTOR:
2953 case VAR_DECL:
2954 case PARM_DECL:
2955 case CONST_DECL:
2956 case RESULT_DECL:
2957 case FUNCTION_DECL:
2958 return currop->op0;
2960 default:
2961 gcc_unreachable ();
2965 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2966 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2967 trying to rename aggregates into ssa form directly, which is a no no.
2969 Thus, this routine doesn't create temporaries, it just builds a
2970 single access expression for the array, calling
2971 find_or_generate_expression to build the innermost pieces.
2973 This function is a subroutine of create_expression_by_pieces, and
2974 should not be called on it's own unless you really know what you
2975 are doing. */
2977 static tree
2978 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2979 gimple_seq *stmts, gimple domstmt)
2981 unsigned int op = 0;
2982 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2985 /* Find a leader for an expression, or generate one using
2986 create_expression_by_pieces if it's ANTIC but
2987 complex.
2988 BLOCK is the basic_block we are looking for leaders in.
2989 EXPR is the expression to find a leader or generate for.
2990 STMTS is the statement list to put the inserted expressions on.
2991 Returns the SSA_NAME of the LHS of the generated expression or the
2992 leader.
2993 DOMSTMT if non-NULL is a statement that should be dominated by
2994 all uses in the generated expression. If DOMSTMT is non-NULL this
2995 routine can fail and return NULL_TREE. Otherwise it will assert
2996 on failure. */
2998 static tree
2999 find_or_generate_expression (basic_block block, pre_expr expr,
3000 gimple_seq *stmts, gimple domstmt)
3002 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
3003 get_expr_value_id (expr), domstmt);
3004 tree genop = NULL;
3005 if (leader)
3007 if (leader->kind == NAME)
3008 genop = PRE_EXPR_NAME (leader);
3009 else if (leader->kind == CONSTANT)
3010 genop = PRE_EXPR_CONSTANT (leader);
3013 /* If it's still NULL, it must be a complex expression, so generate
3014 it recursively. Not so if inserting expressions for values generated
3015 by SCCVN. */
3016 if (genop == NULL
3017 && !domstmt)
3019 bitmap_set_t exprset;
3020 unsigned int lookfor = get_expr_value_id (expr);
3021 bool handled = false;
3022 bitmap_iterator bi;
3023 unsigned int i;
3025 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3026 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3028 pre_expr temp = expression_for_id (i);
3029 if (temp->kind != NAME)
3031 handled = true;
3032 genop = create_expression_by_pieces (block, temp, stmts,
3033 domstmt,
3034 get_expr_type (expr));
3035 break;
3038 if (!handled && domstmt)
3039 return NULL_TREE;
3041 gcc_assert (handled);
3043 return genop;
3046 #define NECESSARY GF_PLF_1
3048 /* Create an expression in pieces, so that we can handle very complex
3049 expressions that may be ANTIC, but not necessary GIMPLE.
3050 BLOCK is the basic block the expression will be inserted into,
3051 EXPR is the expression to insert (in value form)
3052 STMTS is a statement list to append the necessary insertions into.
3054 This function will die if we hit some value that shouldn't be
3055 ANTIC but is (IE there is no leader for it, or its components).
3056 This function may also generate expressions that are themselves
3057 partially or fully redundant. Those that are will be either made
3058 fully redundant during the next iteration of insert (for partially
3059 redundant ones), or eliminated by eliminate (for fully redundant
3060 ones).
3062 If DOMSTMT is non-NULL then we make sure that all uses in the
3063 expressions dominate that statement. In this case the function
3064 can return NULL_TREE to signal failure. */
3066 static tree
3067 create_expression_by_pieces (basic_block block, pre_expr expr,
3068 gimple_seq *stmts, gimple domstmt, tree type)
3070 tree temp, name;
3071 tree folded;
3072 gimple_seq forced_stmts = NULL;
3073 unsigned int value_id;
3074 gimple_stmt_iterator gsi;
3075 tree exprtype = type ? type : get_expr_type (expr);
3076 pre_expr nameexpr;
3077 gimple newstmt;
3079 switch (expr->kind)
3081 /* We may hit the NAME/CONSTANT case if we have to convert types
3082 that value numbering saw through. */
3083 case NAME:
3084 folded = PRE_EXPR_NAME (expr);
3085 break;
3086 case CONSTANT:
3087 folded = PRE_EXPR_CONSTANT (expr);
3088 break;
3089 case REFERENCE:
3091 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3092 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3094 break;
3095 case NARY:
3097 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3098 switch (nary->length)
3100 case 2:
3102 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3103 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
3104 tree genop1 = find_or_generate_expression (block, op1,
3105 stmts, domstmt);
3106 tree genop2 = find_or_generate_expression (block, op2,
3107 stmts, domstmt);
3108 if (!genop1 || !genop2)
3109 return NULL_TREE;
3110 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
3111 may be a constant with the wrong type. */
3112 if (nary->opcode == POINTER_PLUS_EXPR)
3114 genop1 = fold_convert (nary->type, genop1);
3115 genop2 = fold_convert (sizetype, genop2);
3117 else
3119 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3120 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
3123 folded = fold_build2 (nary->opcode, nary->type,
3124 genop1, genop2);
3126 break;
3127 case 1:
3129 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3130 tree genop1 = find_or_generate_expression (block, op1,
3131 stmts, domstmt);
3132 if (!genop1)
3133 return NULL_TREE;
3134 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3136 folded = fold_build1 (nary->opcode, nary->type,
3137 genop1);
3139 break;
3140 default:
3141 return NULL_TREE;
3144 break;
3145 default:
3146 return NULL_TREE;
3149 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3150 folded = fold_convert (exprtype, folded);
3152 /* Force the generated expression to be a sequence of GIMPLE
3153 statements.
3154 We have to call unshare_expr because force_gimple_operand may
3155 modify the tree we pass to it. */
3156 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3157 false, NULL);
3159 /* If we have any intermediate expressions to the value sets, add them
3160 to the value sets and chain them in the instruction stream. */
3161 if (forced_stmts)
3163 gsi = gsi_start (forced_stmts);
3164 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3166 gimple stmt = gsi_stmt (gsi);
3167 tree forcedname = gimple_get_lhs (stmt);
3168 pre_expr nameexpr;
3170 if (TREE_CODE (forcedname) == SSA_NAME)
3172 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3173 VN_INFO_GET (forcedname)->valnum = forcedname;
3174 VN_INFO (forcedname)->value_id = get_next_value_id ();
3175 nameexpr = get_or_alloc_expr_for_name (forcedname);
3176 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3177 if (!in_fre)
3178 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3179 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3181 mark_symbols_for_renaming (stmt);
3183 gimple_seq_add_seq (stmts, forced_stmts);
3186 /* Build and insert the assignment of the end result to the temporary
3187 that we will return. */
3188 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3190 pretemp = create_tmp_reg (exprtype, "pretmp");
3191 get_var_ann (pretemp);
3194 temp = pretemp;
3195 add_referenced_var (temp);
3197 newstmt = gimple_build_assign (temp, folded);
3198 name = make_ssa_name (temp, newstmt);
3199 gimple_assign_set_lhs (newstmt, name);
3200 gimple_set_plf (newstmt, NECESSARY, false);
3202 gimple_seq_add_stmt (stmts, newstmt);
3203 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3205 /* All the symbols in NEWEXPR should be put into SSA form. */
3206 mark_symbols_for_renaming (newstmt);
3208 /* Add a value number to the temporary.
3209 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3210 we are creating the expression by pieces, and this particular piece of
3211 the expression may have been represented. There is no harm in replacing
3212 here. */
3213 VN_INFO_GET (name)->valnum = name;
3214 value_id = get_expr_value_id (expr);
3215 VN_INFO (name)->value_id = value_id;
3216 nameexpr = get_or_alloc_expr_for_name (name);
3217 add_to_value (value_id, nameexpr);
3218 if (NEW_SETS (block))
3219 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3220 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3222 pre_stats.insertions++;
3223 if (dump_file && (dump_flags & TDF_DETAILS))
3225 fprintf (dump_file, "Inserted ");
3226 print_gimple_stmt (dump_file, newstmt, 0, 0);
3227 fprintf (dump_file, " in predecessor %d\n", block->index);
3230 return name;
3234 /* Returns true if we want to inhibit the insertions of PHI nodes
3235 for the given EXPR for basic block BB (a member of a loop).
3236 We want to do this, when we fear that the induction variable we
3237 create might inhibit vectorization. */
3239 static bool
3240 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3242 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3243 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3244 vn_reference_op_t op;
3245 unsigned i;
3247 /* If we aren't going to vectorize we don't inhibit anything. */
3248 if (!flag_tree_vectorize)
3249 return false;
3251 /* Otherwise we inhibit the insertion when the address of the
3252 memory reference is a simple induction variable. In other
3253 cases the vectorizer won't do anything anyway (either it's
3254 loop invariant or a complicated expression). */
3255 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3257 switch (op->opcode)
3259 case ARRAY_REF:
3260 case ARRAY_RANGE_REF:
3261 if (TREE_CODE (op->op0) != SSA_NAME)
3262 break;
3263 /* Fallthru. */
3264 case SSA_NAME:
3266 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3267 affine_iv iv;
3268 /* Default defs are loop invariant. */
3269 if (!defbb)
3270 break;
3271 /* Defined outside this loop, also loop invariant. */
3272 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3273 break;
3274 /* If it's a simple induction variable inhibit insertion,
3275 the vectorizer might be interested in this one. */
3276 if (simple_iv (bb->loop_father, bb->loop_father,
3277 op->op0, &iv, true))
3278 return true;
3279 /* No simple IV, vectorizer can't do anything, hence no
3280 reason to inhibit the transformation for this operand. */
3281 break;
3283 default:
3284 break;
3287 return false;
3290 /* Insert the to-be-made-available values of expression EXPRNUM for each
3291 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3292 merge the result with a phi node, given the same value number as
3293 NODE. Return true if we have inserted new stuff. */
3295 static bool
3296 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3297 pre_expr *avail)
3299 pre_expr expr = expression_for_id (exprnum);
3300 pre_expr newphi;
3301 unsigned int val = get_expr_value_id (expr);
3302 edge pred;
3303 bool insertions = false;
3304 bool nophi = false;
3305 basic_block bprime;
3306 pre_expr eprime;
3307 edge_iterator ei;
3308 tree type = get_expr_type (expr);
3309 tree temp;
3310 gimple phi;
3312 if (dump_file && (dump_flags & TDF_DETAILS))
3314 fprintf (dump_file, "Found partial redundancy for expression ");
3315 print_pre_expr (dump_file, expr);
3316 fprintf (dump_file, " (%04d)\n", val);
3319 /* Make sure we aren't creating an induction variable. */
3320 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3322 bool firstinsideloop = false;
3323 bool secondinsideloop = false;
3324 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3325 EDGE_PRED (block, 0)->src);
3326 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3327 EDGE_PRED (block, 1)->src);
3328 /* Induction variables only have one edge inside the loop. */
3329 if ((firstinsideloop ^ secondinsideloop)
3330 && (expr->kind != REFERENCE
3331 || inhibit_phi_insertion (block, expr)))
3333 if (dump_file && (dump_flags & TDF_DETAILS))
3334 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3335 nophi = true;
3339 /* Make the necessary insertions. */
3340 FOR_EACH_EDGE (pred, ei, block->preds)
3342 gimple_seq stmts = NULL;
3343 tree builtexpr;
3344 bprime = pred->src;
3345 eprime = avail[bprime->index];
3347 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3349 builtexpr = create_expression_by_pieces (bprime,
3350 eprime,
3351 &stmts, NULL,
3352 type);
3353 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3354 gsi_insert_seq_on_edge (pred, stmts);
3355 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3356 insertions = true;
3358 else if (eprime->kind == CONSTANT)
3360 /* Constants may not have the right type, fold_convert
3361 should give us back a constant with the right type.
3363 tree constant = PRE_EXPR_CONSTANT (eprime);
3364 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3366 tree builtexpr = fold_convert (type, constant);
3367 if (!is_gimple_min_invariant (builtexpr))
3369 tree forcedexpr = force_gimple_operand (builtexpr,
3370 &stmts, true,
3371 NULL);
3372 if (!is_gimple_min_invariant (forcedexpr))
3374 if (forcedexpr != builtexpr)
3376 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3377 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3379 if (stmts)
3381 gimple_stmt_iterator gsi;
3382 gsi = gsi_start (stmts);
3383 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3385 gimple stmt = gsi_stmt (gsi);
3386 tree lhs = gimple_get_lhs (stmt);
3387 if (TREE_CODE (lhs) == SSA_NAME)
3388 bitmap_set_bit (inserted_exprs,
3389 SSA_NAME_VERSION (lhs));
3390 gimple_set_plf (stmt, NECESSARY, false);
3392 gsi_insert_seq_on_edge (pred, stmts);
3394 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3397 else
3398 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3401 else if (eprime->kind == NAME)
3403 /* We may have to do a conversion because our value
3404 numbering can look through types in certain cases, but
3405 our IL requires all operands of a phi node have the same
3406 type. */
3407 tree name = PRE_EXPR_NAME (eprime);
3408 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3410 tree builtexpr;
3411 tree forcedexpr;
3412 builtexpr = fold_convert (type, name);
3413 forcedexpr = force_gimple_operand (builtexpr,
3414 &stmts, true,
3415 NULL);
3417 if (forcedexpr != name)
3419 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3420 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3423 if (stmts)
3425 gimple_stmt_iterator gsi;
3426 gsi = gsi_start (stmts);
3427 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3429 gimple stmt = gsi_stmt (gsi);
3430 tree lhs = gimple_get_lhs (stmt);
3431 if (TREE_CODE (lhs) == SSA_NAME)
3432 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3433 gimple_set_plf (stmt, NECESSARY, false);
3435 gsi_insert_seq_on_edge (pred, stmts);
3437 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3441 /* If we didn't want a phi node, and we made insertions, we still have
3442 inserted new stuff, and thus return true. If we didn't want a phi node,
3443 and didn't make insertions, we haven't added anything new, so return
3444 false. */
3445 if (nophi && insertions)
3446 return true;
3447 else if (nophi && !insertions)
3448 return false;
3450 /* Now build a phi for the new variable. */
3451 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3453 prephitemp = create_tmp_var (type, "prephitmp");
3454 get_var_ann (prephitemp);
3457 temp = prephitemp;
3458 add_referenced_var (temp);
3460 if (TREE_CODE (type) == COMPLEX_TYPE
3461 || TREE_CODE (type) == VECTOR_TYPE)
3462 DECL_GIMPLE_REG_P (temp) = 1;
3463 phi = create_phi_node (temp, block);
3465 gimple_set_plf (phi, NECESSARY, false);
3466 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3467 VN_INFO (gimple_phi_result (phi))->value_id = val;
3468 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3469 FOR_EACH_EDGE (pred, ei, block->preds)
3471 pre_expr ae = avail[pred->src->index];
3472 gcc_assert (get_expr_type (ae) == type
3473 || useless_type_conversion_p (type, get_expr_type (ae)));
3474 if (ae->kind == CONSTANT)
3475 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3476 else
3477 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3478 UNKNOWN_LOCATION);
3481 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3482 add_to_value (val, newphi);
3484 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3485 this insertion, since we test for the existence of this value in PHI_GEN
3486 before proceeding with the partial redundancy checks in insert_aux.
3488 The value may exist in AVAIL_OUT, in particular, it could be represented
3489 by the expression we are trying to eliminate, in which case we want the
3490 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3491 inserted there.
3493 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3494 this block, because if it did, it would have existed in our dominator's
3495 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3498 bitmap_insert_into_set (PHI_GEN (block), newphi);
3499 bitmap_value_replace_in_set (AVAIL_OUT (block),
3500 newphi);
3501 bitmap_insert_into_set (NEW_SETS (block),
3502 newphi);
3504 if (dump_file && (dump_flags & TDF_DETAILS))
3506 fprintf (dump_file, "Created phi ");
3507 print_gimple_stmt (dump_file, phi, 0, 0);
3508 fprintf (dump_file, " in block %d\n", block->index);
3510 pre_stats.phis++;
3511 return true;
3516 /* Perform insertion of partially redundant values.
3517 For BLOCK, do the following:
3518 1. Propagate the NEW_SETS of the dominator into the current block.
3519 If the block has multiple predecessors,
3520 2a. Iterate over the ANTIC expressions for the block to see if
3521 any of them are partially redundant.
3522 2b. If so, insert them into the necessary predecessors to make
3523 the expression fully redundant.
3524 2c. Insert a new PHI merging the values of the predecessors.
3525 2d. Insert the new PHI, and the new expressions, into the
3526 NEW_SETS set.
3527 3. Recursively call ourselves on the dominator children of BLOCK.
3529 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3530 do_regular_insertion and do_partial_insertion.
3534 static bool
3535 do_regular_insertion (basic_block block, basic_block dom)
3537 bool new_stuff = false;
3538 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3539 pre_expr expr;
3540 int i;
3542 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3544 if (expr->kind != NAME)
3546 pre_expr *avail;
3547 unsigned int val;
3548 bool by_some = false;
3549 bool cant_insert = false;
3550 bool all_same = true;
3551 pre_expr first_s = NULL;
3552 edge pred;
3553 basic_block bprime;
3554 pre_expr eprime = NULL;
3555 edge_iterator ei;
3556 pre_expr edoubleprime = NULL;
3557 bool do_insertion = false;
3559 val = get_expr_value_id (expr);
3560 if (bitmap_set_contains_value (PHI_GEN (block), val))
3561 continue;
3562 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3564 if (dump_file && (dump_flags & TDF_DETAILS))
3565 fprintf (dump_file, "Found fully redundant value\n");
3566 continue;
3569 avail = XCNEWVEC (pre_expr, last_basic_block);
3570 FOR_EACH_EDGE (pred, ei, block->preds)
3572 unsigned int vprime;
3574 /* We should never run insertion for the exit block
3575 and so not come across fake pred edges. */
3576 gcc_assert (!(pred->flags & EDGE_FAKE));
3577 bprime = pred->src;
3578 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3579 bprime, block);
3581 /* eprime will generally only be NULL if the
3582 value of the expression, translated
3583 through the PHI for this predecessor, is
3584 undefined. If that is the case, we can't
3585 make the expression fully redundant,
3586 because its value is undefined along a
3587 predecessor path. We can thus break out
3588 early because it doesn't matter what the
3589 rest of the results are. */
3590 if (eprime == NULL)
3592 cant_insert = true;
3593 break;
3596 eprime = fully_constant_expression (eprime);
3597 vprime = get_expr_value_id (eprime);
3598 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3599 vprime, NULL);
3600 if (edoubleprime == NULL)
3602 avail[bprime->index] = eprime;
3603 all_same = false;
3605 else
3607 avail[bprime->index] = edoubleprime;
3608 by_some = true;
3609 /* We want to perform insertions to remove a redundancy on
3610 a path in the CFG we want to optimize for speed. */
3611 if (optimize_edge_for_speed_p (pred))
3612 do_insertion = true;
3613 if (first_s == NULL)
3614 first_s = edoubleprime;
3615 else if (!pre_expr_eq (first_s, edoubleprime))
3616 all_same = false;
3619 /* If we can insert it, it's not the same value
3620 already existing along every predecessor, and
3621 it's defined by some predecessor, it is
3622 partially redundant. */
3623 if (!cant_insert && !all_same && by_some)
3625 if (!do_insertion)
3627 if (dump_file && (dump_flags & TDF_DETAILS))
3629 fprintf (dump_file, "Skipping partial redundancy for "
3630 "expression ");
3631 print_pre_expr (dump_file, expr);
3632 fprintf (dump_file, " (%04d), no redundancy on to be "
3633 "optimized for speed edge\n", val);
3636 else if (dbg_cnt (treepre_insert)
3637 && insert_into_preds_of_block (block,
3638 get_expression_id (expr),
3639 avail))
3640 new_stuff = true;
3642 /* If all edges produce the same value and that value is
3643 an invariant, then the PHI has the same value on all
3644 edges. Note this. */
3645 else if (!cant_insert && all_same && eprime
3646 && (edoubleprime->kind == CONSTANT
3647 || edoubleprime->kind == NAME)
3648 && !value_id_constant_p (val))
3650 unsigned int j;
3651 bitmap_iterator bi;
3652 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3653 value_expressions, val);
3655 unsigned int new_val = get_expr_value_id (edoubleprime);
3656 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3658 pre_expr expr = expression_for_id (j);
3660 if (expr->kind == NAME)
3662 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3663 /* Just reset the value id and valnum so it is
3664 the same as the constant we have discovered. */
3665 if (edoubleprime->kind == CONSTANT)
3667 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3668 pre_stats.constified++;
3670 else
3671 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3672 info->value_id = new_val;
3676 free (avail);
3680 VEC_free (pre_expr, heap, exprs);
3681 return new_stuff;
3685 /* Perform insertion for partially anticipatable expressions. There
3686 is only one case we will perform insertion for these. This case is
3687 if the expression is partially anticipatable, and fully available.
3688 In this case, we know that putting it earlier will enable us to
3689 remove the later computation. */
3692 static bool
3693 do_partial_partial_insertion (basic_block block, basic_block dom)
3695 bool new_stuff = false;
3696 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3697 pre_expr expr;
3698 int i;
3700 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3702 if (expr->kind != NAME)
3704 pre_expr *avail;
3705 unsigned int val;
3706 bool by_all = true;
3707 bool cant_insert = false;
3708 edge pred;
3709 basic_block bprime;
3710 pre_expr eprime = NULL;
3711 edge_iterator ei;
3713 val = get_expr_value_id (expr);
3714 if (bitmap_set_contains_value (PHI_GEN (block), val))
3715 continue;
3716 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3717 continue;
3719 avail = XCNEWVEC (pre_expr, last_basic_block);
3720 FOR_EACH_EDGE (pred, ei, block->preds)
3722 unsigned int vprime;
3723 pre_expr edoubleprime;
3725 /* We should never run insertion for the exit block
3726 and so not come across fake pred edges. */
3727 gcc_assert (!(pred->flags & EDGE_FAKE));
3728 bprime = pred->src;
3729 eprime = phi_translate (expr, ANTIC_IN (block),
3730 PA_IN (block),
3731 bprime, block);
3733 /* eprime will generally only be NULL if the
3734 value of the expression, translated
3735 through the PHI for this predecessor, is
3736 undefined. If that is the case, we can't
3737 make the expression fully redundant,
3738 because its value is undefined along a
3739 predecessor path. We can thus break out
3740 early because it doesn't matter what the
3741 rest of the results are. */
3742 if (eprime == NULL)
3744 cant_insert = true;
3745 break;
3748 eprime = fully_constant_expression (eprime);
3749 vprime = get_expr_value_id (eprime);
3750 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3751 vprime, NULL);
3752 if (edoubleprime == NULL)
3754 by_all = false;
3755 break;
3757 else
3758 avail[bprime->index] = edoubleprime;
3762 /* If we can insert it, it's not the same value
3763 already existing along every predecessor, and
3764 it's defined by some predecessor, it is
3765 partially redundant. */
3766 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3768 pre_stats.pa_insert++;
3769 if (insert_into_preds_of_block (block, get_expression_id (expr),
3770 avail))
3771 new_stuff = true;
3773 free (avail);
3777 VEC_free (pre_expr, heap, exprs);
3778 return new_stuff;
3781 static bool
3782 insert_aux (basic_block block)
3784 basic_block son;
3785 bool new_stuff = false;
3787 if (block)
3789 basic_block dom;
3790 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3791 if (dom)
3793 unsigned i;
3794 bitmap_iterator bi;
3795 bitmap_set_t newset = NEW_SETS (dom);
3796 if (newset)
3798 /* Note that we need to value_replace both NEW_SETS, and
3799 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3800 represented by some non-simple expression here that we want
3801 to replace it with. */
3802 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3804 pre_expr expr = expression_for_id (i);
3805 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3806 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3809 if (!single_pred_p (block))
3811 new_stuff |= do_regular_insertion (block, dom);
3812 if (do_partial_partial)
3813 new_stuff |= do_partial_partial_insertion (block, dom);
3817 for (son = first_dom_son (CDI_DOMINATORS, block);
3818 son;
3819 son = next_dom_son (CDI_DOMINATORS, son))
3821 new_stuff |= insert_aux (son);
3824 return new_stuff;
3827 /* Perform insertion of partially redundant values. */
3829 static void
3830 insert (void)
3832 bool new_stuff = true;
3833 basic_block bb;
3834 int num_iterations = 0;
3836 FOR_ALL_BB (bb)
3837 NEW_SETS (bb) = bitmap_set_new ();
3839 while (new_stuff)
3841 num_iterations++;
3842 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3844 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3848 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3850 static void
3851 add_to_exp_gen (basic_block block, tree op)
3853 if (!in_fre)
3855 pre_expr result;
3856 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3857 return;
3858 result = get_or_alloc_expr_for_name (op);
3859 bitmap_value_insert_into_set (EXP_GEN (block), result);
3863 /* Create value ids for PHI in BLOCK. */
3865 static void
3866 make_values_for_phi (gimple phi, basic_block block)
3868 tree result = gimple_phi_result (phi);
3870 /* We have no need for virtual phis, as they don't represent
3871 actual computations. */
3872 if (is_gimple_reg (result))
3874 pre_expr e = get_or_alloc_expr_for_name (result);
3875 add_to_value (get_expr_value_id (e), e);
3876 bitmap_insert_into_set (PHI_GEN (block), e);
3877 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3878 if (!in_fre)
3880 unsigned i;
3881 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3883 tree arg = gimple_phi_arg_def (phi, i);
3884 if (TREE_CODE (arg) == SSA_NAME)
3886 e = get_or_alloc_expr_for_name (arg);
3887 add_to_value (get_expr_value_id (e), e);
3894 /* Compute the AVAIL set for all basic blocks.
3896 This function performs value numbering of the statements in each basic
3897 block. The AVAIL sets are built from information we glean while doing
3898 this value numbering, since the AVAIL sets contain only one entry per
3899 value.
3901 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3902 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3904 static void
3905 compute_avail (void)
3908 basic_block block, son;
3909 basic_block *worklist;
3910 size_t sp = 0;
3911 unsigned i;
3913 /* We pretend that default definitions are defined in the entry block.
3914 This includes function arguments and the static chain decl. */
3915 for (i = 1; i < num_ssa_names; ++i)
3917 tree name = ssa_name (i);
3918 pre_expr e;
3919 if (!name
3920 || !SSA_NAME_IS_DEFAULT_DEF (name)
3921 || has_zero_uses (name)
3922 || !is_gimple_reg (name))
3923 continue;
3925 e = get_or_alloc_expr_for_name (name);
3926 add_to_value (get_expr_value_id (e), e);
3927 if (!in_fre)
3928 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3929 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3932 /* Allocate the worklist. */
3933 worklist = XNEWVEC (basic_block, n_basic_blocks);
3935 /* Seed the algorithm by putting the dominator children of the entry
3936 block on the worklist. */
3937 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3938 son;
3939 son = next_dom_son (CDI_DOMINATORS, son))
3940 worklist[sp++] = son;
3942 /* Loop until the worklist is empty. */
3943 while (sp)
3945 gimple_stmt_iterator gsi;
3946 gimple stmt;
3947 basic_block dom;
3948 unsigned int stmt_uid = 1;
3950 /* Pick a block from the worklist. */
3951 block = worklist[--sp];
3953 /* Initially, the set of available values in BLOCK is that of
3954 its immediate dominator. */
3955 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3956 if (dom)
3957 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3959 /* Generate values for PHI nodes. */
3960 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3961 make_values_for_phi (gsi_stmt (gsi), block);
3963 BB_MAY_NOTRETURN (block) = 0;
3965 /* Now compute value numbers and populate value sets with all
3966 the expressions computed in BLOCK. */
3967 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3969 ssa_op_iter iter;
3970 tree op;
3972 stmt = gsi_stmt (gsi);
3973 gimple_set_uid (stmt, stmt_uid++);
3975 /* Cache whether the basic-block has any non-visible side-effect
3976 or control flow.
3977 If this isn't a call or it is the last stmt in the
3978 basic-block then the CFG represents things correctly. */
3979 if (is_gimple_call (stmt)
3980 && !stmt_ends_bb_p (stmt))
3982 /* Non-looping const functions always return normally.
3983 Otherwise the call might not return or have side-effects
3984 that forbids hoisting possibly trapping expressions
3985 before it. */
3986 int flags = gimple_call_flags (stmt);
3987 if (!(flags & ECF_CONST)
3988 || (flags & ECF_LOOPING_CONST_OR_PURE))
3989 BB_MAY_NOTRETURN (block) = 1;
3992 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3994 pre_expr e = get_or_alloc_expr_for_name (op);
3996 add_to_value (get_expr_value_id (e), e);
3997 if (!in_fre)
3998 bitmap_insert_into_set (TMP_GEN (block), e);
3999 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
4002 if (gimple_has_volatile_ops (stmt)
4003 || stmt_could_throw_p (stmt))
4004 continue;
4006 switch (gimple_code (stmt))
4008 case GIMPLE_RETURN:
4009 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4010 add_to_exp_gen (block, op);
4011 continue;
4013 case GIMPLE_CALL:
4015 vn_reference_t ref;
4016 unsigned int i;
4017 vn_reference_op_t vro;
4018 pre_expr result = NULL;
4019 VEC(vn_reference_op_s, heap) *ops = NULL;
4021 if (!can_value_number_call (stmt))
4022 continue;
4024 copy_reference_ops_from_call (stmt, &ops);
4025 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4026 gimple_expr_type (stmt),
4027 ops, &ref, VN_NOWALK);
4028 VEC_free (vn_reference_op_s, heap, ops);
4029 if (!ref)
4030 continue;
4032 for (i = 0; VEC_iterate (vn_reference_op_s,
4033 ref->operands, i,
4034 vro); i++)
4036 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4037 add_to_exp_gen (block, vro->op0);
4038 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4039 add_to_exp_gen (block, vro->op1);
4040 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4041 add_to_exp_gen (block, vro->op2);
4043 result = (pre_expr) pool_alloc (pre_expr_pool);
4044 result->kind = REFERENCE;
4045 result->id = 0;
4046 PRE_EXPR_REFERENCE (result) = ref;
4048 get_or_alloc_expression_id (result);
4049 add_to_value (get_expr_value_id (result), result);
4050 if (!in_fre)
4051 bitmap_value_insert_into_set (EXP_GEN (block), result);
4052 continue;
4055 case GIMPLE_ASSIGN:
4057 pre_expr result = NULL;
4058 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4060 case tcc_unary:
4061 case tcc_binary:
4062 case tcc_comparison:
4064 vn_nary_op_t nary;
4065 unsigned int i;
4067 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4068 gimple_assign_rhs_code (stmt),
4069 gimple_expr_type (stmt),
4070 gimple_assign_rhs1 (stmt),
4071 gimple_assign_rhs2 (stmt),
4072 NULL_TREE, NULL_TREE, &nary);
4074 if (!nary)
4075 continue;
4077 for (i = 0; i < nary->length; i++)
4078 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4079 add_to_exp_gen (block, nary->op[i]);
4081 result = (pre_expr) pool_alloc (pre_expr_pool);
4082 result->kind = NARY;
4083 result->id = 0;
4084 PRE_EXPR_NARY (result) = nary;
4085 break;
4088 case tcc_declaration:
4089 case tcc_reference:
4091 vn_reference_t ref;
4092 unsigned int i;
4093 vn_reference_op_t vro;
4095 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4096 gimple_vuse (stmt),
4097 VN_WALK, &ref);
4098 if (!ref)
4099 continue;
4101 for (i = 0; VEC_iterate (vn_reference_op_s,
4102 ref->operands, i,
4103 vro); i++)
4105 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4106 add_to_exp_gen (block, vro->op0);
4107 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4108 add_to_exp_gen (block, vro->op1);
4109 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4110 add_to_exp_gen (block, vro->op2);
4112 result = (pre_expr) pool_alloc (pre_expr_pool);
4113 result->kind = REFERENCE;
4114 result->id = 0;
4115 PRE_EXPR_REFERENCE (result) = ref;
4116 break;
4119 default:
4120 /* For any other statement that we don't
4121 recognize, simply add all referenced
4122 SSA_NAMEs to EXP_GEN. */
4123 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4124 add_to_exp_gen (block, op);
4125 continue;
4128 get_or_alloc_expression_id (result);
4129 add_to_value (get_expr_value_id (result), result);
4130 if (!in_fre)
4131 bitmap_value_insert_into_set (EXP_GEN (block), result);
4133 continue;
4135 default:
4136 break;
4140 /* Put the dominator children of BLOCK on the worklist of blocks
4141 to compute available sets for. */
4142 for (son = first_dom_son (CDI_DOMINATORS, block);
4143 son;
4144 son = next_dom_son (CDI_DOMINATORS, son))
4145 worklist[sp++] = son;
4148 free (worklist);
4151 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4152 than the available expressions for it. The insertion point is
4153 right before the first use in STMT. Returns the SSA_NAME that should
4154 be used for replacement. */
4156 static tree
4157 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4159 basic_block bb = gimple_bb (stmt);
4160 gimple_stmt_iterator gsi;
4161 gimple_seq stmts = NULL;
4162 tree expr;
4163 pre_expr e;
4165 /* First create a value expression from the expression we want
4166 to insert and associate it with the value handle for SSA_VN. */
4167 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4168 if (e == NULL)
4169 return NULL_TREE;
4171 /* Then use create_expression_by_pieces to generate a valid
4172 expression to insert at this point of the IL stream. */
4173 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4174 if (expr == NULL_TREE)
4175 return NULL_TREE;
4176 gsi = gsi_for_stmt (stmt);
4177 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4179 return expr;
4182 /* Eliminate fully redundant computations. */
4184 static unsigned int
4185 eliminate (void)
4187 VEC (gimple, heap) *to_remove = NULL;
4188 basic_block b;
4189 unsigned int todo = 0;
4190 gimple_stmt_iterator gsi;
4191 gimple stmt;
4192 unsigned i;
4194 FOR_EACH_BB (b)
4196 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4198 stmt = gsi_stmt (gsi);
4200 /* Lookup the RHS of the expression, see if we have an
4201 available computation for it. If so, replace the RHS with
4202 the available computation. */
4203 if (gimple_has_lhs (stmt)
4204 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4205 && !gimple_assign_ssa_name_copy_p (stmt)
4206 && (!gimple_assign_single_p (stmt)
4207 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4208 && !gimple_has_volatile_ops (stmt)
4209 && !has_zero_uses (gimple_get_lhs (stmt)))
4211 tree lhs = gimple_get_lhs (stmt);
4212 tree rhs = NULL_TREE;
4213 tree sprime = NULL;
4214 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4215 pre_expr sprimeexpr;
4217 if (gimple_assign_single_p (stmt))
4218 rhs = gimple_assign_rhs1 (stmt);
4220 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4221 get_expr_value_id (lhsexpr),
4222 NULL);
4224 if (sprimeexpr)
4226 if (sprimeexpr->kind == CONSTANT)
4227 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4228 else if (sprimeexpr->kind == NAME)
4229 sprime = PRE_EXPR_NAME (sprimeexpr);
4230 else
4231 gcc_unreachable ();
4234 /* If there is no existing leader but SCCVN knows this
4235 value is constant, use that constant. */
4236 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4238 sprime = VN_INFO (lhs)->valnum;
4239 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4240 TREE_TYPE (sprime)))
4241 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4243 if (dump_file && (dump_flags & TDF_DETAILS))
4245 fprintf (dump_file, "Replaced ");
4246 print_gimple_expr (dump_file, stmt, 0, 0);
4247 fprintf (dump_file, " with ");
4248 print_generic_expr (dump_file, sprime, 0);
4249 fprintf (dump_file, " in ");
4250 print_gimple_stmt (dump_file, stmt, 0, 0);
4252 pre_stats.eliminations++;
4253 propagate_tree_value_into_stmt (&gsi, sprime);
4254 stmt = gsi_stmt (gsi);
4255 update_stmt (stmt);
4256 continue;
4259 /* If there is no existing usable leader but SCCVN thinks
4260 it has an expression it wants to use as replacement,
4261 insert that. */
4262 if (!sprime || sprime == lhs)
4264 tree val = VN_INFO (lhs)->valnum;
4265 if (val != VN_TOP
4266 && TREE_CODE (val) == SSA_NAME
4267 && VN_INFO (val)->needs_insertion
4268 && can_PRE_operation (vn_get_expr_for (val)))
4269 sprime = do_SCCVN_insertion (stmt, val);
4271 if (sprime
4272 && sprime != lhs
4273 && (rhs == NULL_TREE
4274 || TREE_CODE (rhs) != SSA_NAME
4275 || may_propagate_copy (rhs, sprime)))
4277 bool can_make_abnormal_goto
4278 = is_gimple_call (stmt)
4279 && stmt_can_make_abnormal_goto (stmt);
4281 gcc_assert (sprime != rhs);
4283 if (dump_file && (dump_flags & TDF_DETAILS))
4285 fprintf (dump_file, "Replaced ");
4286 print_gimple_expr (dump_file, stmt, 0, 0);
4287 fprintf (dump_file, " with ");
4288 print_generic_expr (dump_file, sprime, 0);
4289 fprintf (dump_file, " in ");
4290 print_gimple_stmt (dump_file, stmt, 0, 0);
4293 if (TREE_CODE (sprime) == SSA_NAME)
4294 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4295 NECESSARY, true);
4296 /* We need to make sure the new and old types actually match,
4297 which may require adding a simple cast, which fold_convert
4298 will do for us. */
4299 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4300 && !useless_type_conversion_p (gimple_expr_type (stmt),
4301 TREE_TYPE (sprime)))
4302 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4304 pre_stats.eliminations++;
4305 propagate_tree_value_into_stmt (&gsi, sprime);
4306 stmt = gsi_stmt (gsi);
4307 update_stmt (stmt);
4309 /* If we removed EH side-effects from the statement, clean
4310 its EH information. */
4311 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4313 bitmap_set_bit (need_eh_cleanup,
4314 gimple_bb (stmt)->index);
4315 if (dump_file && (dump_flags & TDF_DETAILS))
4316 fprintf (dump_file, " Removed EH side-effects.\n");
4319 /* Likewise for AB side-effects. */
4320 if (can_make_abnormal_goto
4321 && !stmt_can_make_abnormal_goto (stmt))
4323 bitmap_set_bit (need_ab_cleanup,
4324 gimple_bb (stmt)->index);
4325 if (dump_file && (dump_flags & TDF_DETAILS))
4326 fprintf (dump_file, " Removed AB side-effects.\n");
4330 /* If the statement is a scalar store, see if the expression
4331 has the same value number as its rhs. If so, the store is
4332 dead. */
4333 else if (gimple_assign_single_p (stmt)
4334 && !is_gimple_reg (gimple_assign_lhs (stmt))
4335 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4336 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4338 tree rhs = gimple_assign_rhs1 (stmt);
4339 tree val;
4340 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4341 gimple_vuse (stmt), VN_WALK, NULL);
4342 if (TREE_CODE (rhs) == SSA_NAME)
4343 rhs = VN_INFO (rhs)->valnum;
4344 if (val
4345 && operand_equal_p (val, rhs, 0))
4347 if (dump_file && (dump_flags & TDF_DETAILS))
4349 fprintf (dump_file, "Deleted redundant store ");
4350 print_gimple_stmt (dump_file, stmt, 0, 0);
4353 /* Queue stmt for removal. */
4354 VEC_safe_push (gimple, heap, to_remove, stmt);
4357 /* Visit COND_EXPRs and fold the comparison with the
4358 available value-numbers. */
4359 else if (gimple_code (stmt) == GIMPLE_COND)
4361 tree op0 = gimple_cond_lhs (stmt);
4362 tree op1 = gimple_cond_rhs (stmt);
4363 tree result;
4365 if (TREE_CODE (op0) == SSA_NAME)
4366 op0 = VN_INFO (op0)->valnum;
4367 if (TREE_CODE (op1) == SSA_NAME)
4368 op1 = VN_INFO (op1)->valnum;
4369 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4370 op0, op1);
4371 if (result && TREE_CODE (result) == INTEGER_CST)
4373 if (integer_zerop (result))
4374 gimple_cond_make_false (stmt);
4375 else
4376 gimple_cond_make_true (stmt);
4377 update_stmt (stmt);
4378 todo = TODO_cleanup_cfg;
4381 /* Visit indirect calls and turn them into direct calls if
4382 possible. */
4383 if (is_gimple_call (stmt)
4384 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4386 tree orig_fn = gimple_call_fn (stmt);
4387 tree fn = VN_INFO (orig_fn)->valnum;
4388 if (TREE_CODE (fn) == ADDR_EXPR
4389 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL
4390 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4391 TREE_TYPE (fn)))
4393 bool can_make_abnormal_goto
4394 = stmt_can_make_abnormal_goto (stmt);
4395 bool was_noreturn = gimple_call_noreturn_p (stmt);
4397 if (dump_file && (dump_flags & TDF_DETAILS))
4399 fprintf (dump_file, "Replacing call target with ");
4400 print_generic_expr (dump_file, fn, 0);
4401 fprintf (dump_file, " in ");
4402 print_gimple_stmt (dump_file, stmt, 0, 0);
4405 gimple_call_set_fn (stmt, fn);
4406 update_stmt (stmt);
4408 /* When changing a call into a noreturn call, cfg cleanup
4409 is needed to fix up the noreturn call. */
4410 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4411 todo |= TODO_cleanup_cfg;
4413 /* If we removed EH side-effects from the statement, clean
4414 its EH information. */
4415 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4417 bitmap_set_bit (need_eh_cleanup,
4418 gimple_bb (stmt)->index);
4419 if (dump_file && (dump_flags & TDF_DETAILS))
4420 fprintf (dump_file, " Removed EH side-effects.\n");
4423 /* Likewise for AB side-effects. */
4424 if (can_make_abnormal_goto
4425 && !stmt_can_make_abnormal_goto (stmt))
4427 bitmap_set_bit (need_ab_cleanup,
4428 gimple_bb (stmt)->index);
4429 if (dump_file && (dump_flags & TDF_DETAILS))
4430 fprintf (dump_file, " Removed AB side-effects.\n");
4433 /* Changing an indirect call to a direct call may
4434 have exposed different semantics. This may
4435 require an SSA update. */
4436 todo |= TODO_update_ssa_only_virtuals;
4441 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4443 gimple stmt, phi = gsi_stmt (gsi);
4444 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4445 pre_expr sprimeexpr, resexpr;
4446 gimple_stmt_iterator gsi2;
4448 /* We want to perform redundant PHI elimination. Do so by
4449 replacing the PHI with a single copy if possible.
4450 Do not touch inserted, single-argument or virtual PHIs. */
4451 if (gimple_phi_num_args (phi) == 1
4452 || !is_gimple_reg (res))
4454 gsi_next (&gsi);
4455 continue;
4458 resexpr = get_or_alloc_expr_for_name (res);
4459 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4460 get_expr_value_id (resexpr), NULL);
4461 if (sprimeexpr)
4463 if (sprimeexpr->kind == CONSTANT)
4464 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4465 else if (sprimeexpr->kind == NAME)
4466 sprime = PRE_EXPR_NAME (sprimeexpr);
4467 else
4468 gcc_unreachable ();
4470 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4472 sprime = VN_INFO (res)->valnum;
4473 if (!useless_type_conversion_p (TREE_TYPE (res),
4474 TREE_TYPE (sprime)))
4475 sprime = fold_convert (TREE_TYPE (res), sprime);
4477 if (!sprime
4478 || sprime == res)
4480 gsi_next (&gsi);
4481 continue;
4484 if (dump_file && (dump_flags & TDF_DETAILS))
4486 fprintf (dump_file, "Replaced redundant PHI node defining ");
4487 print_generic_expr (dump_file, res, 0);
4488 fprintf (dump_file, " with ");
4489 print_generic_expr (dump_file, sprime, 0);
4490 fprintf (dump_file, "\n");
4493 remove_phi_node (&gsi, false);
4495 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4496 && TREE_CODE (sprime) == SSA_NAME)
4497 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4499 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4500 sprime = fold_convert (TREE_TYPE (res), sprime);
4501 stmt = gimple_build_assign (res, sprime);
4502 SSA_NAME_DEF_STMT (res) = stmt;
4503 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4505 gsi2 = gsi_after_labels (b);
4506 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4507 /* Queue the copy for eventual removal. */
4508 VEC_safe_push (gimple, heap, to_remove, stmt);
4509 /* If we inserted this PHI node ourself, it's not an elimination. */
4510 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4511 pre_stats.phis--;
4512 else
4513 pre_stats.eliminations++;
4517 /* We cannot remove stmts during BB walk, especially not release SSA
4518 names there as this confuses the VN machinery. The stmts ending
4519 up in to_remove are either stores or simple copies. */
4520 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4522 tree lhs = gimple_assign_lhs (stmt);
4523 tree rhs = gimple_assign_rhs1 (stmt);
4524 use_operand_p use_p;
4525 gimple use_stmt;
4527 /* If there is a single use only, propagate the equivalency
4528 instead of keeping the copy. */
4529 if (TREE_CODE (lhs) == SSA_NAME
4530 && TREE_CODE (rhs) == SSA_NAME
4531 && single_imm_use (lhs, &use_p, &use_stmt)
4532 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4534 SET_USE (use_p, rhs);
4535 update_stmt (use_stmt);
4536 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4537 && TREE_CODE (rhs) == SSA_NAME)
4538 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4541 /* If this is a store or a now unused copy, remove it. */
4542 if (TREE_CODE (lhs) != SSA_NAME
4543 || has_zero_uses (lhs))
4545 basic_block bb = gimple_bb (stmt);
4546 gsi = gsi_for_stmt (stmt);
4547 unlink_stmt_vdef (stmt);
4548 gsi_remove (&gsi, true);
4549 if (gimple_purge_dead_eh_edges (bb))
4550 todo |= TODO_cleanup_cfg;
4551 if (TREE_CODE (lhs) == SSA_NAME)
4552 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4553 release_defs (stmt);
4556 VEC_free (gimple, heap, to_remove);
4558 return todo;
4561 /* Borrow a bit of tree-ssa-dce.c for the moment.
4562 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4563 this may be a bit faster, and we may want critical edges kept split. */
4565 /* If OP's defining statement has not already been determined to be necessary,
4566 mark that statement necessary. Return the stmt, if it is newly
4567 necessary. */
4569 static inline gimple
4570 mark_operand_necessary (tree op)
4572 gimple stmt;
4574 gcc_assert (op);
4576 if (TREE_CODE (op) != SSA_NAME)
4577 return NULL;
4579 stmt = SSA_NAME_DEF_STMT (op);
4580 gcc_assert (stmt);
4582 if (gimple_plf (stmt, NECESSARY)
4583 || gimple_nop_p (stmt))
4584 return NULL;
4586 gimple_set_plf (stmt, NECESSARY, true);
4587 return stmt;
4590 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4591 to insert PHI nodes sometimes, and because value numbering of casts isn't
4592 perfect, we sometimes end up inserting dead code. This simple DCE-like
4593 pass removes any insertions we made that weren't actually used. */
4595 static void
4596 remove_dead_inserted_code (void)
4598 bitmap worklist;
4599 unsigned i;
4600 bitmap_iterator bi;
4601 gimple t;
4603 worklist = BITMAP_ALLOC (NULL);
4604 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4606 t = SSA_NAME_DEF_STMT (ssa_name (i));
4607 if (gimple_plf (t, NECESSARY))
4608 bitmap_set_bit (worklist, i);
4610 while (!bitmap_empty_p (worklist))
4612 i = bitmap_first_set_bit (worklist);
4613 bitmap_clear_bit (worklist, i);
4614 t = SSA_NAME_DEF_STMT (ssa_name (i));
4616 /* PHI nodes are somewhat special in that each PHI alternative has
4617 data and control dependencies. All the statements feeding the
4618 PHI node's arguments are always necessary. */
4619 if (gimple_code (t) == GIMPLE_PHI)
4621 unsigned k;
4623 for (k = 0; k < gimple_phi_num_args (t); k++)
4625 tree arg = PHI_ARG_DEF (t, k);
4626 if (TREE_CODE (arg) == SSA_NAME)
4628 gimple n = mark_operand_necessary (arg);
4629 if (n)
4630 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4634 else
4636 /* Propagate through the operands. Examine all the USE, VUSE and
4637 VDEF operands in this statement. Mark all the statements
4638 which feed this statement's uses as necessary. */
4639 ssa_op_iter iter;
4640 tree use;
4642 /* The operands of VDEF expressions are also needed as they
4643 represent potential definitions that may reach this
4644 statement (VDEF operands allow us to follow def-def
4645 links). */
4647 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4649 gimple n = mark_operand_necessary (use);
4650 if (n)
4651 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4656 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4658 t = SSA_NAME_DEF_STMT (ssa_name (i));
4659 if (!gimple_plf (t, NECESSARY))
4661 gimple_stmt_iterator gsi;
4663 if (dump_file && (dump_flags & TDF_DETAILS))
4665 fprintf (dump_file, "Removing unnecessary insertion:");
4666 print_gimple_stmt (dump_file, t, 0, 0);
4669 gsi = gsi_for_stmt (t);
4670 if (gimple_code (t) == GIMPLE_PHI)
4671 remove_phi_node (&gsi, true);
4672 else
4674 gsi_remove (&gsi, true);
4675 release_defs (t);
4679 BITMAP_FREE (worklist);
4682 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4683 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4684 the number of visited blocks. */
4686 static int
4687 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4689 edge_iterator *stack;
4690 int sp;
4691 int post_order_num = 0;
4692 sbitmap visited;
4694 if (include_entry_exit)
4695 post_order[post_order_num++] = EXIT_BLOCK;
4697 /* Allocate stack for back-tracking up CFG. */
4698 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4699 sp = 0;
4701 /* Allocate bitmap to track nodes that have been visited. */
4702 visited = sbitmap_alloc (last_basic_block);
4704 /* None of the nodes in the CFG have been visited yet. */
4705 sbitmap_zero (visited);
4707 /* Push the last edge on to the stack. */
4708 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4710 while (sp)
4712 edge_iterator ei;
4713 basic_block src;
4714 basic_block dest;
4716 /* Look at the edge on the top of the stack. */
4717 ei = stack[sp - 1];
4718 src = ei_edge (ei)->src;
4719 dest = ei_edge (ei)->dest;
4721 /* Check if the edge destination has been visited yet. */
4722 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4724 /* Mark that we have visited the destination. */
4725 SET_BIT (visited, src->index);
4727 if (EDGE_COUNT (src->preds) > 0)
4728 /* Since the DEST node has been visited for the first
4729 time, check its successors. */
4730 stack[sp++] = ei_start (src->preds);
4731 else
4732 post_order[post_order_num++] = src->index;
4734 else
4736 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4737 post_order[post_order_num++] = dest->index;
4739 if (!ei_one_before_end_p (ei))
4740 ei_next (&stack[sp - 1]);
4741 else
4742 sp--;
4746 if (include_entry_exit)
4747 post_order[post_order_num++] = ENTRY_BLOCK;
4749 free (stack);
4750 sbitmap_free (visited);
4751 return post_order_num;
4755 /* Initialize data structures used by PRE. */
4757 static void
4758 init_pre (bool do_fre)
4760 basic_block bb;
4762 next_expression_id = 1;
4763 expressions = NULL;
4764 VEC_safe_push (pre_expr, heap, expressions, NULL);
4765 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4766 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4767 get_max_value_id() + 1);
4768 name_to_id = NULL;
4770 in_fre = do_fre;
4772 inserted_exprs = BITMAP_ALLOC (NULL);
4773 need_creation = NULL;
4774 pretemp = NULL_TREE;
4775 storetemp = NULL_TREE;
4776 prephitemp = NULL_TREE;
4778 connect_infinite_loops_to_exit ();
4779 memset (&pre_stats, 0, sizeof (pre_stats));
4782 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4783 my_rev_post_order_compute (postorder, false);
4785 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4787 calculate_dominance_info (CDI_POST_DOMINATORS);
4788 calculate_dominance_info (CDI_DOMINATORS);
4790 bitmap_obstack_initialize (&grand_bitmap_obstack);
4791 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4792 expr_pred_trans_eq, free);
4793 expression_to_id = htab_create (num_ssa_names * 3,
4794 pre_expr_hash,
4795 pre_expr_eq, NULL);
4796 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4797 sizeof (struct bitmap_set), 30);
4798 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4799 sizeof (struct pre_expr_d), 30);
4800 FOR_ALL_BB (bb)
4802 EXP_GEN (bb) = bitmap_set_new ();
4803 PHI_GEN (bb) = bitmap_set_new ();
4804 TMP_GEN (bb) = bitmap_set_new ();
4805 AVAIL_OUT (bb) = bitmap_set_new ();
4808 need_eh_cleanup = BITMAP_ALLOC (NULL);
4809 need_ab_cleanup = BITMAP_ALLOC (NULL);
4813 /* Deallocate data structures used by PRE. */
4815 static void
4816 fini_pre (bool do_fre)
4818 free (postorder);
4819 VEC_free (bitmap_set_t, heap, value_expressions);
4820 BITMAP_FREE (inserted_exprs);
4821 VEC_free (gimple, heap, need_creation);
4822 bitmap_obstack_release (&grand_bitmap_obstack);
4823 free_alloc_pool (bitmap_set_pool);
4824 free_alloc_pool (pre_expr_pool);
4825 htab_delete (phi_translate_table);
4826 htab_delete (expression_to_id);
4827 VEC_free (unsigned, heap, name_to_id);
4829 free_aux_for_blocks ();
4831 free_dominance_info (CDI_POST_DOMINATORS);
4833 if (!bitmap_empty_p (need_eh_cleanup))
4835 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4836 cleanup_tree_cfg ();
4839 BITMAP_FREE (need_eh_cleanup);
4841 if (!bitmap_empty_p (need_ab_cleanup))
4843 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4844 cleanup_tree_cfg ();
4847 BITMAP_FREE (need_ab_cleanup);
4849 if (!do_fre)
4850 loop_optimizer_finalize ();
4853 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4854 only wants to do full redundancy elimination. */
4856 static unsigned int
4857 execute_pre (bool do_fre)
4859 unsigned int todo = 0;
4861 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4863 /* This has to happen before SCCVN runs because
4864 loop_optimizer_init may create new phis, etc. */
4865 if (!do_fre)
4866 loop_optimizer_init (LOOPS_NORMAL);
4868 if (!run_scc_vn (do_fre ? VN_WALKREWRITE : VN_WALK))
4870 if (!do_fre)
4871 loop_optimizer_finalize ();
4873 return 0;
4876 init_pre (do_fre);
4877 scev_initialize ();
4879 /* Collect and value number expressions computed in each basic block. */
4880 compute_avail ();
4882 if (dump_file && (dump_flags & TDF_DETAILS))
4884 basic_block bb;
4886 FOR_ALL_BB (bb)
4888 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4889 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4890 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4891 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4895 /* Insert can get quite slow on an incredibly large number of basic
4896 blocks due to some quadratic behavior. Until this behavior is
4897 fixed, don't run it when he have an incredibly large number of
4898 bb's. If we aren't going to run insert, there is no point in
4899 computing ANTIC, either, even though it's plenty fast. */
4900 if (!do_fre && n_basic_blocks < 4000)
4902 compute_antic ();
4903 insert ();
4906 /* Make sure to remove fake edges before committing our inserts.
4907 This makes sure we don't end up with extra critical edges that
4908 we would need to split. */
4909 remove_fake_exit_edges ();
4910 gsi_commit_edge_inserts ();
4912 /* Remove all the redundant expressions. */
4913 todo |= eliminate ();
4915 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4916 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4917 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4918 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4919 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4921 clear_expression_ids ();
4922 free_scc_vn ();
4923 if (!do_fre)
4925 remove_dead_inserted_code ();
4926 todo |= TODO_verify_flow;
4929 scev_finalize ();
4930 fini_pre (do_fre);
4932 return todo;
4935 /* Gate and execute functions for PRE. */
4937 static unsigned int
4938 do_pre (void)
4940 return execute_pre (false);
4943 static bool
4944 gate_pre (void)
4946 return flag_tree_pre != 0;
4949 struct gimple_opt_pass pass_pre =
4952 GIMPLE_PASS,
4953 "pre", /* name */
4954 gate_pre, /* gate */
4955 do_pre, /* execute */
4956 NULL, /* sub */
4957 NULL, /* next */
4958 0, /* static_pass_number */
4959 TV_TREE_PRE, /* tv_id */
4960 PROP_no_crit_edges | PROP_cfg
4961 | PROP_ssa, /* properties_required */
4962 0, /* properties_provided */
4963 0, /* properties_destroyed */
4964 TODO_rebuild_alias, /* todo_flags_start */
4965 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4966 | TODO_verify_ssa /* todo_flags_finish */
4971 /* Gate and execute functions for FRE. */
4973 static unsigned int
4974 execute_fre (void)
4976 return execute_pre (true);
4979 static bool
4980 gate_fre (void)
4982 return flag_tree_fre != 0;
4985 struct gimple_opt_pass pass_fre =
4988 GIMPLE_PASS,
4989 "fre", /* name */
4990 gate_fre, /* gate */
4991 execute_fre, /* execute */
4992 NULL, /* sub */
4993 NULL, /* next */
4994 0, /* static_pass_number */
4995 TV_TREE_FRE, /* tv_id */
4996 PROP_cfg | PROP_ssa, /* properties_required */
4997 0, /* properties_provided */
4998 0, /* properties_destroyed */
4999 0, /* todo_flags_start */
5000 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */