* gcc.target/x86_64/abi/avx/asm-support.S (snapshot_ret): Preserve
[official-gcc/alias-decl.git] / gcc / tree-ssa-pre.c
blobcafebe677702f795b6cb27add30d6fbabb04ecad
1 /* SSA-PRE for trees.
2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
3 Free Software Foundation, Inc.
4 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
5 <stevenb@suse.de>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
12 any later version.
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 #include "config.h"
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "ggc.h"
28 #include "tree.h"
29 #include "basic-block.h"
30 #include "diagnostic.h"
31 #include "tree-inline.h"
32 #include "tree-flow.h"
33 #include "gimple.h"
34 #include "tree-dump.h"
35 #include "timevar.h"
36 #include "fibheap.h"
37 #include "hashtab.h"
38 #include "tree-iterator.h"
39 #include "real.h"
40 #include "alloc-pool.h"
41 #include "obstack.h"
42 #include "tree-pass.h"
43 #include "flags.h"
44 #include "bitmap.h"
45 #include "langhooks.h"
46 #include "cfgloop.h"
47 #include "tree-ssa-sccvn.h"
48 #include "tree-scalar-evolution.h"
49 #include "params.h"
50 #include "dbgcnt.h"
52 /* TODO:
54 1. Avail sets can be shared by making an avail_find_leader that
55 walks up the dominator tree and looks in those avail sets.
56 This might affect code optimality, it's unclear right now.
57 2. Strength reduction can be performed by anticipating expressions
58 we can repair later on.
59 3. We can do back-substitution or smarter value numbering to catch
60 commutative expressions split up over multiple statements.
63 /* For ease of terminology, "expression node" in the below refers to
64 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
65 represent the actual statement containing the expressions we care about,
66 and we cache the value number by putting it in the expression. */
68 /* Basic algorithm
70 First we walk the statements to generate the AVAIL sets, the
71 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
72 generation of values/expressions by a given block. We use them
73 when computing the ANTIC sets. The AVAIL sets consist of
74 SSA_NAME's that represent values, so we know what values are
75 available in what blocks. AVAIL is a forward dataflow problem. In
76 SSA, values are never killed, so we don't need a kill set, or a
77 fixpoint iteration, in order to calculate the AVAIL sets. In
78 traditional parlance, AVAIL sets tell us the downsafety of the
79 expressions/values.
81 Next, we generate the ANTIC sets. These sets represent the
82 anticipatable expressions. ANTIC is a backwards dataflow
83 problem. An expression is anticipatable in a given block if it could
84 be generated in that block. This means that if we had to perform
85 an insertion in that block, of the value of that expression, we
86 could. Calculating the ANTIC sets requires phi translation of
87 expressions, because the flow goes backwards through phis. We must
88 iterate to a fixpoint of the ANTIC sets, because we have a kill
89 set. Even in SSA form, values are not live over the entire
90 function, only from their definition point onwards. So we have to
91 remove values from the ANTIC set once we go past the definition
92 point of the leaders that make them up.
93 compute_antic/compute_antic_aux performs this computation.
95 Third, we perform insertions to make partially redundant
96 expressions fully redundant.
98 An expression is partially redundant (excluding partial
99 anticipation) if:
101 1. It is AVAIL in some, but not all, of the predecessors of a
102 given block.
103 2. It is ANTIC in all the predecessors.
105 In order to make it fully redundant, we insert the expression into
106 the predecessors where it is not available, but is ANTIC.
108 For the partial anticipation case, we only perform insertion if it
109 is partially anticipated in some block, and fully available in all
110 of the predecessors.
112 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
113 performs these steps.
115 Fourth, we eliminate fully redundant expressions.
116 This is a simple statement walk that replaces redundant
117 calculations with the now available values. */
119 /* Representations of value numbers:
121 Value numbers are represented by a representative SSA_NAME. We
122 will create fake SSA_NAME's in situations where we need a
123 representative but do not have one (because it is a complex
124 expression). In order to facilitate storing the value numbers in
125 bitmaps, and keep the number of wasted SSA_NAME's down, we also
126 associate a value_id with each value number, and create full blown
127 ssa_name's only where we actually need them (IE in operands of
128 existing expressions).
130 Theoretically you could replace all the value_id's with
131 SSA_NAME_VERSION, but this would allocate a large number of
132 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
133 It would also require an additional indirection at each point we
134 use the value id. */
136 /* Representation of expressions on value numbers:
138 Expressions consisting of value numbers are represented the same
139 way as our VN internally represents them, with an additional
140 "pre_expr" wrapping around them in order to facilitate storing all
141 of the expressions in the same sets. */
143 /* Representation of sets:
145 The dataflow sets do not need to be sorted in any particular order
146 for the majority of their lifetime, are simply represented as two
147 bitmaps, one that keeps track of values present in the set, and one
148 that keeps track of expressions present in the set.
150 When we need them in topological order, we produce it on demand by
151 transforming the bitmap into an array and sorting it into topo
152 order. */
154 /* Type of expression, used to know which member of the PRE_EXPR union
155 is valid. */
157 enum pre_expr_kind
159 NAME,
160 NARY,
161 REFERENCE,
162 CONSTANT
165 typedef union pre_expr_union_d
167 tree name;
168 tree constant;
169 vn_nary_op_t nary;
170 vn_reference_t reference;
171 } pre_expr_union;
173 typedef struct pre_expr_d
175 enum pre_expr_kind kind;
176 unsigned int id;
177 pre_expr_union u;
178 } *pre_expr;
180 #define PRE_EXPR_NAME(e) (e)->u.name
181 #define PRE_EXPR_NARY(e) (e)->u.nary
182 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
183 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
185 static int
186 pre_expr_eq (const void *p1, const void *p2)
188 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
189 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
191 if (e1->kind != e2->kind)
192 return false;
194 switch (e1->kind)
196 case CONSTANT:
197 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
198 PRE_EXPR_CONSTANT (e2));
199 case NAME:
200 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
201 case NARY:
202 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
203 case REFERENCE:
204 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
205 PRE_EXPR_REFERENCE (e2));
206 default:
207 gcc_unreachable ();
211 static hashval_t
212 pre_expr_hash (const void *p1)
214 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
215 switch (e->kind)
217 case CONSTANT:
218 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
219 case NAME:
220 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
221 case NARY:
222 return PRE_EXPR_NARY (e)->hashcode;
223 case REFERENCE:
224 return PRE_EXPR_REFERENCE (e)->hashcode;
225 default:
226 gcc_unreachable ();
231 /* Next global expression id number. */
232 static unsigned int next_expression_id;
234 /* Mapping from expression to id number we can use in bitmap sets. */
235 DEF_VEC_P (pre_expr);
236 DEF_VEC_ALLOC_P (pre_expr, heap);
237 static VEC(pre_expr, heap) *expressions;
238 static htab_t expression_to_id;
239 static VEC(unsigned, heap) *name_to_id;
241 /* Allocate an expression id for EXPR. */
243 static inline unsigned int
244 alloc_expression_id (pre_expr expr)
246 void **slot;
247 /* Make sure we won't overflow. */
248 gcc_assert (next_expression_id + 1 > next_expression_id);
249 expr->id = next_expression_id++;
250 VEC_safe_push (pre_expr, heap, expressions, expr);
251 if (expr->kind == NAME)
253 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
254 /* VEC_safe_grow_cleared allocates no headroom. Avoid frequent
255 re-allocations by using VEC_reserve upfront. There is no
256 VEC_quick_grow_cleared unfortunately. */
257 VEC_reserve (unsigned, heap, name_to_id, num_ssa_names);
258 VEC_safe_grow_cleared (unsigned, heap, name_to_id, num_ssa_names);
259 gcc_assert (VEC_index (unsigned, name_to_id, version) == 0);
260 VEC_replace (unsigned, name_to_id, version, expr->id);
262 else
264 slot = htab_find_slot (expression_to_id, expr, INSERT);
265 gcc_assert (!*slot);
266 *slot = expr;
268 return next_expression_id - 1;
271 /* Return the expression id for tree EXPR. */
273 static inline unsigned int
274 get_expression_id (const pre_expr expr)
276 return expr->id;
279 static inline unsigned int
280 lookup_expression_id (const pre_expr expr)
282 void **slot;
284 if (expr->kind == NAME)
286 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
287 if (VEC_length (unsigned, name_to_id) <= version)
288 return 0;
289 return VEC_index (unsigned, name_to_id, version);
291 else
293 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
294 if (!slot)
295 return 0;
296 return ((pre_expr)*slot)->id;
300 /* Return the existing expression id for EXPR, or create one if one
301 does not exist yet. */
303 static inline unsigned int
304 get_or_alloc_expression_id (pre_expr expr)
306 unsigned int id = lookup_expression_id (expr);
307 if (id == 0)
308 return alloc_expression_id (expr);
309 return expr->id = id;
312 /* Return the expression that has expression id ID */
314 static inline pre_expr
315 expression_for_id (unsigned int id)
317 return VEC_index (pre_expr, expressions, id);
320 /* Free the expression id field in all of our expressions,
321 and then destroy the expressions array. */
323 static void
324 clear_expression_ids (void)
326 VEC_free (pre_expr, heap, expressions);
329 static alloc_pool pre_expr_pool;
331 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
333 static pre_expr
334 get_or_alloc_expr_for_name (tree name)
336 struct pre_expr_d expr;
337 pre_expr result;
338 unsigned int result_id;
340 expr.kind = NAME;
341 expr.id = 0;
342 PRE_EXPR_NAME (&expr) = name;
343 result_id = lookup_expression_id (&expr);
344 if (result_id != 0)
345 return expression_for_id (result_id);
347 result = (pre_expr) pool_alloc (pre_expr_pool);
348 result->kind = NAME;
349 PRE_EXPR_NAME (result) = name;
350 alloc_expression_id (result);
351 return result;
354 static bool in_fre = false;
356 /* An unordered bitmap set. One bitmap tracks values, the other,
357 expressions. */
358 typedef struct bitmap_set
360 bitmap expressions;
361 bitmap values;
362 } *bitmap_set_t;
364 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
365 EXECUTE_IF_SET_IN_BITMAP((set)->expressions, 0, (id), (bi))
367 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
368 EXECUTE_IF_SET_IN_BITMAP((set)->values, 0, (id), (bi))
370 /* Mapping from value id to expressions with that value_id. */
371 DEF_VEC_P (bitmap_set_t);
372 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
373 static VEC(bitmap_set_t, heap) *value_expressions;
375 /* Sets that we need to keep track of. */
376 typedef struct bb_bitmap_sets
378 /* The EXP_GEN set, which represents expressions/values generated in
379 a basic block. */
380 bitmap_set_t exp_gen;
382 /* The PHI_GEN set, which represents PHI results generated in a
383 basic block. */
384 bitmap_set_t phi_gen;
386 /* The TMP_GEN set, which represents results/temporaries generated
387 in a basic block. IE the LHS of an expression. */
388 bitmap_set_t tmp_gen;
390 /* The AVAIL_OUT set, which represents which values are available in
391 a given basic block. */
392 bitmap_set_t avail_out;
394 /* The ANTIC_IN set, which represents which values are anticipatable
395 in a given basic block. */
396 bitmap_set_t antic_in;
398 /* The PA_IN set, which represents which values are
399 partially anticipatable in a given basic block. */
400 bitmap_set_t pa_in;
402 /* The NEW_SETS set, which is used during insertion to augment the
403 AVAIL_OUT set of blocks with the new insertions performed during
404 the current iteration. */
405 bitmap_set_t new_sets;
407 /* A cache for value_dies_in_block_x. */
408 bitmap expr_dies;
410 /* True if we have visited this block during ANTIC calculation. */
411 unsigned int visited : 1;
413 /* True we have deferred processing this block during ANTIC
414 calculation until its successor is processed. */
415 unsigned int deferred : 1;
417 /* True when the block contains a call that might not return. */
418 unsigned int contains_may_not_return_call : 1;
419 } *bb_value_sets_t;
421 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
422 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
423 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
424 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
425 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
426 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
427 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
428 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
429 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
430 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
431 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
434 /* Basic block list in postorder. */
435 static int *postorder;
437 /* This structure is used to keep track of statistics on what
438 optimization PRE was able to perform. */
439 static struct
441 /* The number of RHS computations eliminated by PRE. */
442 int eliminations;
444 /* The number of new expressions/temporaries generated by PRE. */
445 int insertions;
447 /* The number of inserts found due to partial anticipation */
448 int pa_insert;
450 /* The number of new PHI nodes added by PRE. */
451 int phis;
453 /* The number of values found constant. */
454 int constified;
456 } pre_stats;
458 static bool do_partial_partial;
459 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
460 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
461 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
462 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
463 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
464 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
465 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
466 unsigned int, bool);
467 static bitmap_set_t bitmap_set_new (void);
468 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
469 gimple, tree);
470 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
471 gimple);
472 static unsigned int get_expr_value_id (pre_expr);
474 /* We can add and remove elements and entries to and from sets
475 and hash tables, so we use alloc pools for them. */
477 static alloc_pool bitmap_set_pool;
478 static bitmap_obstack grand_bitmap_obstack;
480 /* To avoid adding 300 temporary variables when we only need one, we
481 only create one temporary variable, on demand, and build ssa names
482 off that. We do have to change the variable if the types don't
483 match the current variable's type. */
484 static tree pretemp;
485 static tree storetemp;
486 static tree prephitemp;
488 /* Set of blocks with statements that have had its EH information
489 cleaned up. */
490 static bitmap need_eh_cleanup;
492 /* The phi_translate_table caches phi translations for a given
493 expression and predecessor. */
495 static htab_t phi_translate_table;
497 /* A three tuple {e, pred, v} used to cache phi translations in the
498 phi_translate_table. */
500 typedef struct expr_pred_trans_d
502 /* The expression. */
503 pre_expr e;
505 /* The predecessor block along which we translated the expression. */
506 basic_block pred;
508 /* The value that resulted from the translation. */
509 pre_expr v;
511 /* The hashcode for the expression, pred pair. This is cached for
512 speed reasons. */
513 hashval_t hashcode;
514 } *expr_pred_trans_t;
515 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
517 /* Return the hash value for a phi translation table entry. */
519 static hashval_t
520 expr_pred_trans_hash (const void *p)
522 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
523 return ve->hashcode;
526 /* Return true if two phi translation table entries are the same.
527 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
529 static int
530 expr_pred_trans_eq (const void *p1, const void *p2)
532 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
533 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
534 basic_block b1 = ve1->pred;
535 basic_block b2 = ve2->pred;
537 /* If they are not translations for the same basic block, they can't
538 be equal. */
539 if (b1 != b2)
540 return false;
541 return pre_expr_eq (ve1->e, ve2->e);
544 /* Search in the phi translation table for the translation of
545 expression E in basic block PRED.
546 Return the translated value, if found, NULL otherwise. */
548 static inline pre_expr
549 phi_trans_lookup (pre_expr e, basic_block pred)
551 void **slot;
552 struct expr_pred_trans_d ept;
554 ept.e = e;
555 ept.pred = pred;
556 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
557 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
558 NO_INSERT);
559 if (!slot)
560 return NULL;
561 else
562 return ((expr_pred_trans_t) *slot)->v;
566 /* Add the tuple mapping from {expression E, basic block PRED} to
567 value V, to the phi translation table. */
569 static inline void
570 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
572 void **slot;
573 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
574 new_pair->e = e;
575 new_pair->pred = pred;
576 new_pair->v = v;
577 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
578 pred->index);
580 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
581 new_pair->hashcode, INSERT);
582 if (*slot)
583 free (*slot);
584 *slot = (void *) new_pair;
588 /* Add expression E to the expression set of value id V. */
590 void
591 add_to_value (unsigned int v, pre_expr e)
593 bitmap_set_t set;
595 gcc_assert (get_expr_value_id (e) == v);
597 if (v >= VEC_length (bitmap_set_t, value_expressions))
599 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
600 v + 1);
603 set = VEC_index (bitmap_set_t, value_expressions, v);
604 if (!set)
606 set = bitmap_set_new ();
607 VEC_replace (bitmap_set_t, value_expressions, v, set);
610 bitmap_insert_into_set_1 (set, e, v, true);
613 /* Create a new bitmap set and return it. */
615 static bitmap_set_t
616 bitmap_set_new (void)
618 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
619 ret->expressions = BITMAP_ALLOC (&grand_bitmap_obstack);
620 ret->values = BITMAP_ALLOC (&grand_bitmap_obstack);
621 return ret;
624 /* Return the value id for a PRE expression EXPR. */
626 static unsigned int
627 get_expr_value_id (pre_expr expr)
629 switch (expr->kind)
631 case CONSTANT:
633 unsigned int id;
634 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
635 if (id == 0)
637 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
638 add_to_value (id, expr);
640 return id;
642 case NAME:
643 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
644 case NARY:
645 return PRE_EXPR_NARY (expr)->value_id;
646 case REFERENCE:
647 return PRE_EXPR_REFERENCE (expr)->value_id;
648 default:
649 gcc_unreachable ();
653 /* Remove an expression EXPR from a bitmapped set. */
655 static void
656 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
658 unsigned int val = get_expr_value_id (expr);
659 if (!value_id_constant_p (val))
661 bitmap_clear_bit (set->values, val);
662 bitmap_clear_bit (set->expressions, get_expression_id (expr));
666 static void
667 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
668 unsigned int val, bool allow_constants)
670 if (allow_constants || !value_id_constant_p (val))
672 /* We specifically expect this and only this function to be able to
673 insert constants into a set. */
674 bitmap_set_bit (set->values, val);
675 bitmap_set_bit (set->expressions, get_or_alloc_expression_id (expr));
679 /* Insert an expression EXPR into a bitmapped set. */
681 static void
682 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
684 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
687 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
689 static void
690 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
692 bitmap_copy (dest->expressions, orig->expressions);
693 bitmap_copy (dest->values, orig->values);
697 /* Free memory used up by SET. */
698 static void
699 bitmap_set_free (bitmap_set_t set)
701 BITMAP_FREE (set->expressions);
702 BITMAP_FREE (set->values);
706 /* Generate an topological-ordered array of bitmap set SET. */
708 static VEC(pre_expr, heap) *
709 sorted_array_from_bitmap_set (bitmap_set_t set)
711 unsigned int i, j;
712 bitmap_iterator bi, bj;
713 VEC(pre_expr, heap) *result;
715 /* Pre-allocate roughly enough space for the array. */
716 result = VEC_alloc (pre_expr, heap, bitmap_count_bits (set->values));
718 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
720 /* The number of expressions having a given value is usually
721 relatively small. Thus, rather than making a vector of all
722 the expressions and sorting it by value-id, we walk the values
723 and check in the reverse mapping that tells us what expressions
724 have a given value, to filter those in our set. As a result,
725 the expressions are inserted in value-id order, which means
726 topological order.
728 If this is somehow a significant lose for some cases, we can
729 choose which set to walk based on the set size. */
730 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
731 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
733 if (bitmap_bit_p (set->expressions, j))
734 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
738 return result;
741 /* Perform bitmapped set operation DEST &= ORIG. */
743 static void
744 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
746 bitmap_iterator bi;
747 unsigned int i;
749 if (dest != orig)
751 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
753 bitmap_and_into (dest->values, orig->values);
754 bitmap_copy (temp, dest->expressions);
755 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
757 pre_expr expr = expression_for_id (i);
758 unsigned int value_id = get_expr_value_id (expr);
759 if (!bitmap_bit_p (dest->values, value_id))
760 bitmap_clear_bit (dest->expressions, i);
762 BITMAP_FREE (temp);
766 /* Subtract all values and expressions contained in ORIG from DEST. */
768 static bitmap_set_t
769 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
771 bitmap_set_t result = bitmap_set_new ();
772 bitmap_iterator bi;
773 unsigned int i;
775 bitmap_and_compl (result->expressions, dest->expressions,
776 orig->expressions);
778 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
780 pre_expr expr = expression_for_id (i);
781 unsigned int value_id = get_expr_value_id (expr);
782 bitmap_set_bit (result->values, value_id);
785 return result;
788 /* Subtract all the values in bitmap set B from bitmap set A. */
790 static void
791 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
793 unsigned int i;
794 bitmap_iterator bi;
795 bitmap temp = BITMAP_ALLOC (&grand_bitmap_obstack);
797 bitmap_copy (temp, a->expressions);
798 EXECUTE_IF_SET_IN_BITMAP (temp, 0, i, bi)
800 pre_expr expr = expression_for_id (i);
801 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
802 bitmap_remove_from_set (a, expr);
804 BITMAP_FREE (temp);
808 /* Return true if bitmapped set SET contains the value VALUE_ID. */
810 static bool
811 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
813 if (value_id_constant_p (value_id))
814 return true;
816 if (!set || bitmap_empty_p (set->expressions))
817 return false;
819 return bitmap_bit_p (set->values, value_id);
822 static inline bool
823 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
825 return bitmap_bit_p (set->expressions, get_expression_id (expr));
828 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
830 static void
831 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
832 const pre_expr expr)
834 bitmap_set_t exprset;
835 unsigned int i;
836 bitmap_iterator bi;
838 if (value_id_constant_p (lookfor))
839 return;
841 if (!bitmap_set_contains_value (set, lookfor))
842 return;
844 /* The number of expressions having a given value is usually
845 significantly less than the total number of expressions in SET.
846 Thus, rather than check, for each expression in SET, whether it
847 has the value LOOKFOR, we walk the reverse mapping that tells us
848 what expressions have a given value, and see if any of those
849 expressions are in our set. For large testcases, this is about
850 5-10x faster than walking the bitmap. If this is somehow a
851 significant lose for some cases, we can choose which set to walk
852 based on the set size. */
853 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
854 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
856 if (bitmap_bit_p (set->expressions, i))
858 bitmap_clear_bit (set->expressions, i);
859 bitmap_set_bit (set->expressions, get_expression_id (expr));
860 return;
865 /* Return true if two bitmap sets are equal. */
867 static bool
868 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
870 return bitmap_equal_p (a->values, b->values);
873 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
874 and add it otherwise. */
876 static void
877 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
879 unsigned int val = get_expr_value_id (expr);
881 if (bitmap_set_contains_value (set, val))
882 bitmap_set_replace_value (set, val, expr);
883 else
884 bitmap_insert_into_set (set, expr);
887 /* Insert EXPR into SET if EXPR's value is not already present in
888 SET. */
890 static void
891 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
893 unsigned int val = get_expr_value_id (expr);
895 #ifdef ENABLE_CHECKING
896 gcc_assert (expr->id == get_or_alloc_expression_id (expr));
897 #endif
899 /* Constant values are always considered to be part of the set. */
900 if (value_id_constant_p (val))
901 return;
903 /* If the value membership changed, add the expression. */
904 if (bitmap_set_bit (set->values, val))
905 bitmap_set_bit (set->expressions, expr->id);
908 /* Print out EXPR to outfile. */
910 static void
911 print_pre_expr (FILE *outfile, const pre_expr expr)
913 switch (expr->kind)
915 case CONSTANT:
916 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
917 break;
918 case NAME:
919 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
920 break;
921 case NARY:
923 unsigned int i;
924 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
925 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
926 for (i = 0; i < nary->length; i++)
928 print_generic_expr (outfile, nary->op[i], 0);
929 if (i != (unsigned) nary->length - 1)
930 fprintf (outfile, ",");
932 fprintf (outfile, "}");
934 break;
936 case REFERENCE:
938 vn_reference_op_t vro;
939 unsigned int i;
940 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
941 fprintf (outfile, "{");
942 for (i = 0;
943 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
944 i++)
946 bool closebrace = false;
947 if (vro->opcode != SSA_NAME
948 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
950 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
951 if (vro->op0)
953 fprintf (outfile, "<");
954 closebrace = true;
957 if (vro->op0)
959 print_generic_expr (outfile, vro->op0, 0);
960 if (vro->op1)
962 fprintf (outfile, ",");
963 print_generic_expr (outfile, vro->op1, 0);
965 if (vro->op2)
967 fprintf (outfile, ",");
968 print_generic_expr (outfile, vro->op2, 0);
971 if (closebrace)
972 fprintf (outfile, ">");
973 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
974 fprintf (outfile, ",");
976 fprintf (outfile, "}");
977 if (ref->vuse)
979 fprintf (outfile, "@");
980 print_generic_expr (outfile, ref->vuse, 0);
983 break;
986 void debug_pre_expr (pre_expr);
988 /* Like print_pre_expr but always prints to stderr. */
989 void
990 debug_pre_expr (pre_expr e)
992 print_pre_expr (stderr, e);
993 fprintf (stderr, "\n");
996 /* Print out SET to OUTFILE. */
998 static void
999 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1000 const char *setname, int blockindex)
1002 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1003 if (set)
1005 bool first = true;
1006 unsigned i;
1007 bitmap_iterator bi;
1009 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1011 const pre_expr expr = expression_for_id (i);
1013 if (!first)
1014 fprintf (outfile, ", ");
1015 first = false;
1016 print_pre_expr (outfile, expr);
1018 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1021 fprintf (outfile, " }\n");
1024 void debug_bitmap_set (bitmap_set_t);
1026 void
1027 debug_bitmap_set (bitmap_set_t set)
1029 print_bitmap_set (stderr, set, "debug", 0);
1032 /* Print out the expressions that have VAL to OUTFILE. */
1034 void
1035 print_value_expressions (FILE *outfile, unsigned int val)
1037 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1038 if (set)
1040 char s[10];
1041 sprintf (s, "%04d", val);
1042 print_bitmap_set (outfile, set, s, 0);
1047 void
1048 debug_value_expressions (unsigned int val)
1050 print_value_expressions (stderr, val);
1053 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1054 represent it. */
1056 static pre_expr
1057 get_or_alloc_expr_for_constant (tree constant)
1059 unsigned int result_id;
1060 unsigned int value_id;
1061 struct pre_expr_d expr;
1062 pre_expr newexpr;
1064 expr.kind = CONSTANT;
1065 PRE_EXPR_CONSTANT (&expr) = constant;
1066 result_id = lookup_expression_id (&expr);
1067 if (result_id != 0)
1068 return expression_for_id (result_id);
1070 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1071 newexpr->kind = CONSTANT;
1072 PRE_EXPR_CONSTANT (newexpr) = constant;
1073 alloc_expression_id (newexpr);
1074 value_id = get_or_alloc_constant_value_id (constant);
1075 add_to_value (value_id, newexpr);
1076 return newexpr;
1079 /* Given a value id V, find the actual tree representing the constant
1080 value if there is one, and return it. Return NULL if we can't find
1081 a constant. */
1083 static tree
1084 get_constant_for_value_id (unsigned int v)
1086 if (value_id_constant_p (v))
1088 unsigned int i;
1089 bitmap_iterator bi;
1090 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1092 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1094 pre_expr expr = expression_for_id (i);
1095 if (expr->kind == CONSTANT)
1096 return PRE_EXPR_CONSTANT (expr);
1099 return NULL;
1102 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1103 Currently only supports constants and SSA_NAMES. */
1104 static pre_expr
1105 get_or_alloc_expr_for (tree t)
1107 if (TREE_CODE (t) == SSA_NAME)
1108 return get_or_alloc_expr_for_name (t);
1109 else if (is_gimple_min_invariant (t))
1110 return get_or_alloc_expr_for_constant (t);
1111 else
1113 /* More complex expressions can result from SCCVN expression
1114 simplification that inserts values for them. As they all
1115 do not have VOPs the get handled by the nary ops struct. */
1116 vn_nary_op_t result;
1117 unsigned int result_id;
1118 vn_nary_op_lookup (t, &result);
1119 if (result != NULL)
1121 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1122 e->kind = NARY;
1123 PRE_EXPR_NARY (e) = result;
1124 result_id = lookup_expression_id (e);
1125 if (result_id != 0)
1127 pool_free (pre_expr_pool, e);
1128 e = expression_for_id (result_id);
1129 return e;
1131 alloc_expression_id (e);
1132 return e;
1135 return NULL;
1138 /* Return the folded version of T if T, when folded, is a gimple
1139 min_invariant. Otherwise, return T. */
1141 static pre_expr
1142 fully_constant_expression (pre_expr e)
1144 switch (e->kind)
1146 case CONSTANT:
1147 return e;
1148 case NARY:
1150 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1151 switch (TREE_CODE_CLASS (nary->opcode))
1153 case tcc_expression:
1154 if (nary->opcode == TRUTH_NOT_EXPR)
1155 goto do_unary;
1156 if (nary->opcode != TRUTH_AND_EXPR
1157 && nary->opcode != TRUTH_OR_EXPR
1158 && nary->opcode != TRUTH_XOR_EXPR)
1159 return e;
1160 /* Fallthrough. */
1161 case tcc_binary:
1162 case tcc_comparison:
1164 /* We have to go from trees to pre exprs to value ids to
1165 constants. */
1166 tree naryop0 = nary->op[0];
1167 tree naryop1 = nary->op[1];
1168 tree result;
1169 if (!is_gimple_min_invariant (naryop0))
1171 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1172 unsigned int vrep0 = get_expr_value_id (rep0);
1173 tree const0 = get_constant_for_value_id (vrep0);
1174 if (const0)
1175 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1177 if (!is_gimple_min_invariant (naryop1))
1179 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1180 unsigned int vrep1 = get_expr_value_id (rep1);
1181 tree const1 = get_constant_for_value_id (vrep1);
1182 if (const1)
1183 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1185 result = fold_binary (nary->opcode, nary->type,
1186 naryop0, naryop1);
1187 if (result && is_gimple_min_invariant (result))
1188 return get_or_alloc_expr_for_constant (result);
1189 /* We might have simplified the expression to a
1190 SSA_NAME for example from x_1 * 1. But we cannot
1191 insert a PHI for x_1 unconditionally as x_1 might
1192 not be available readily. */
1193 return e;
1195 case tcc_reference:
1196 if (nary->opcode != REALPART_EXPR
1197 && nary->opcode != IMAGPART_EXPR
1198 && nary->opcode != VIEW_CONVERT_EXPR)
1199 return e;
1200 /* Fallthrough. */
1201 case tcc_unary:
1202 do_unary:
1204 /* We have to go from trees to pre exprs to value ids to
1205 constants. */
1206 tree naryop0 = nary->op[0];
1207 tree const0, result;
1208 if (is_gimple_min_invariant (naryop0))
1209 const0 = naryop0;
1210 else
1212 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1213 unsigned int vrep0 = get_expr_value_id (rep0);
1214 const0 = get_constant_for_value_id (vrep0);
1216 result = NULL;
1217 if (const0)
1219 tree type1 = TREE_TYPE (nary->op[0]);
1220 const0 = fold_convert (type1, const0);
1221 result = fold_unary (nary->opcode, nary->type, const0);
1223 if (result && is_gimple_min_invariant (result))
1224 return get_or_alloc_expr_for_constant (result);
1225 return e;
1227 default:
1228 return e;
1231 case REFERENCE:
1233 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1234 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1235 vn_reference_op_t op;
1237 /* Try to simplify the translated expression if it is
1238 a call to a builtin function with at most two arguments. */
1239 op = VEC_index (vn_reference_op_s, operands, 0);
1240 if (op->opcode == CALL_EXPR
1241 && TREE_CODE (op->op0) == ADDR_EXPR
1242 && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL
1243 && DECL_BUILT_IN (TREE_OPERAND (op->op0, 0))
1244 && VEC_length (vn_reference_op_s, operands) >= 2
1245 && VEC_length (vn_reference_op_s, operands) <= 3)
1247 vn_reference_op_t arg0, arg1 = NULL;
1248 bool anyconst = false;
1249 arg0 = VEC_index (vn_reference_op_s, operands, 1);
1250 if (VEC_length (vn_reference_op_s, operands) > 2)
1251 arg1 = VEC_index (vn_reference_op_s, operands, 2);
1252 if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant
1253 || (arg0->opcode == ADDR_EXPR
1254 && is_gimple_min_invariant (arg0->op0)))
1255 anyconst = true;
1256 if (arg1
1257 && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant
1258 || (arg1->opcode == ADDR_EXPR
1259 && is_gimple_min_invariant (arg1->op0))))
1260 anyconst = true;
1261 if (anyconst)
1263 tree folded = build_call_expr (TREE_OPERAND (op->op0, 0),
1264 arg1 ? 2 : 1,
1265 arg0->op0,
1266 arg1 ? arg1->op0 : NULL);
1267 if (folded
1268 && TREE_CODE (folded) == NOP_EXPR)
1269 folded = TREE_OPERAND (folded, 0);
1270 if (folded
1271 && is_gimple_min_invariant (folded))
1272 return get_or_alloc_expr_for_constant (folded);
1275 return e;
1277 default:
1278 return e;
1280 return e;
1283 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1284 it has the value it would have in BLOCK. Set *SAME_VALID to true
1285 in case the new vuse doesn't change the value id of the OPERANDS. */
1287 static tree
1288 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1289 alias_set_type set, tree type, tree vuse,
1290 basic_block phiblock,
1291 basic_block block, bool *same_valid)
1293 gimple phi = SSA_NAME_DEF_STMT (vuse);
1294 ao_ref ref;
1295 edge e = NULL;
1296 bool use_oracle;
1298 *same_valid = true;
1300 if (gimple_bb (phi) != phiblock)
1301 return vuse;
1303 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1305 /* Use the alias-oracle to find either the PHI node in this block,
1306 the first VUSE used in this block that is equivalent to vuse or
1307 the first VUSE which definition in this block kills the value. */
1308 if (gimple_code (phi) == GIMPLE_PHI)
1309 e = find_edge (block, phiblock);
1310 else if (use_oracle)
1311 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1313 vuse = gimple_vuse (phi);
1314 phi = SSA_NAME_DEF_STMT (vuse);
1315 if (gimple_bb (phi) != phiblock)
1316 return vuse;
1317 if (gimple_code (phi) == GIMPLE_PHI)
1319 e = find_edge (block, phiblock);
1320 break;
1323 else
1324 return NULL_TREE;
1326 if (e)
1328 if (use_oracle)
1330 bitmap visited = NULL;
1331 /* Try to find a vuse that dominates this phi node by skipping
1332 non-clobbering statements. */
1333 vuse = get_continuation_for_phi (phi, &ref, &visited);
1334 if (visited)
1335 BITMAP_FREE (visited);
1337 else
1338 vuse = NULL_TREE;
1339 if (!vuse)
1341 /* If we didn't find any, the value ID can't stay the same,
1342 but return the translated vuse. */
1343 *same_valid = false;
1344 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1346 /* ??? We would like to return vuse here as this is the canonical
1347 upmost vdef that this reference is associated with. But during
1348 insertion of the references into the hash tables we only ever
1349 directly insert with their direct gimple_vuse, hence returning
1350 something else would make us not find the other expression. */
1351 return PHI_ARG_DEF (phi, e->dest_idx);
1354 return NULL_TREE;
1357 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1358 SET2. This is used to avoid making a set consisting of the union
1359 of PA_IN and ANTIC_IN during insert. */
1361 static inline pre_expr
1362 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1364 pre_expr result;
1366 result = bitmap_find_leader (set1, val, NULL);
1367 if (!result && set2)
1368 result = bitmap_find_leader (set2, val, NULL);
1369 return result;
1372 /* Get the tree type for our PRE expression e. */
1374 static tree
1375 get_expr_type (const pre_expr e)
1377 switch (e->kind)
1379 case NAME:
1380 return TREE_TYPE (PRE_EXPR_NAME (e));
1381 case CONSTANT:
1382 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1383 case REFERENCE:
1384 return PRE_EXPR_REFERENCE (e)->type;
1385 case NARY:
1386 return PRE_EXPR_NARY (e)->type;
1388 gcc_unreachable();
1391 /* Get a representative SSA_NAME for a given expression.
1392 Since all of our sub-expressions are treated as values, we require
1393 them to be SSA_NAME's for simplicity.
1394 Prior versions of GVNPRE used to use "value handles" here, so that
1395 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1396 either case, the operands are really values (IE we do not expect
1397 them to be usable without finding leaders). */
1399 static tree
1400 get_representative_for (const pre_expr e)
1402 tree exprtype;
1403 tree name;
1404 unsigned int value_id = get_expr_value_id (e);
1406 switch (e->kind)
1408 case NAME:
1409 return PRE_EXPR_NAME (e);
1410 case CONSTANT:
1411 return PRE_EXPR_CONSTANT (e);
1412 case NARY:
1413 case REFERENCE:
1415 /* Go through all of the expressions representing this value
1416 and pick out an SSA_NAME. */
1417 unsigned int i;
1418 bitmap_iterator bi;
1419 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1420 value_id);
1421 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1423 pre_expr rep = expression_for_id (i);
1424 if (rep->kind == NAME)
1425 return PRE_EXPR_NAME (rep);
1428 break;
1430 /* If we reached here we couldn't find an SSA_NAME. This can
1431 happen when we've discovered a value that has never appeared in
1432 the program as set to an SSA_NAME, most likely as the result of
1433 phi translation. */
1434 if (dump_file)
1436 fprintf (dump_file,
1437 "Could not find SSA_NAME representative for expression:");
1438 print_pre_expr (dump_file, e);
1439 fprintf (dump_file, "\n");
1442 exprtype = get_expr_type (e);
1444 /* Build and insert the assignment of the end result to the temporary
1445 that we will return. */
1446 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1448 pretemp = create_tmp_var (exprtype, "pretmp");
1449 get_var_ann (pretemp);
1452 name = make_ssa_name (pretemp, gimple_build_nop ());
1453 VN_INFO_GET (name)->value_id = value_id;
1454 if (e->kind == CONSTANT)
1455 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1456 else
1457 VN_INFO (name)->valnum = name;
1459 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1460 if (dump_file)
1462 fprintf (dump_file, "Created SSA_NAME representative ");
1463 print_generic_expr (dump_file, name, 0);
1464 fprintf (dump_file, " for expression:");
1465 print_pre_expr (dump_file, e);
1466 fprintf (dump_file, "\n");
1469 return name;
1475 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1476 the phis in PRED. Return NULL if we can't find a leader for each part
1477 of the translated expression. */
1479 static pre_expr
1480 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1481 basic_block pred, basic_block phiblock)
1483 pre_expr oldexpr = expr;
1484 pre_expr phitrans;
1486 if (!expr)
1487 return NULL;
1489 /* Constants contain no values that need translation. */
1490 if (expr->kind == CONSTANT)
1491 return expr;
1493 if (value_id_constant_p (get_expr_value_id (expr)))
1494 return expr;
1496 phitrans = phi_trans_lookup (expr, pred);
1497 if (phitrans)
1498 return phitrans;
1500 switch (expr->kind)
1502 case NARY:
1504 unsigned int i;
1505 bool changed = false;
1506 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1507 struct vn_nary_op_s newnary;
1508 /* The NARY structure is only guaranteed to have been
1509 allocated to the nary->length operands. */
1510 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1511 - sizeof (tree) * (4 - nary->length)));
1513 for (i = 0; i < newnary.length; i++)
1515 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1516 continue;
1517 else
1519 pre_expr leader, result;
1520 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1521 leader = find_leader_in_sets (op_val_id, set1, set2);
1522 result = phi_translate (leader, set1, set2, pred, phiblock);
1523 if (result && result != leader)
1525 tree name = get_representative_for (result);
1526 if (!name)
1527 return NULL;
1528 newnary.op[i] = name;
1530 else if (!result)
1531 return NULL;
1533 changed |= newnary.op[i] != nary->op[i];
1536 if (changed)
1538 pre_expr constant;
1539 unsigned int new_val_id;
1541 tree result = vn_nary_op_lookup_pieces (newnary.length,
1542 newnary.opcode,
1543 newnary.type,
1544 newnary.op[0],
1545 newnary.op[1],
1546 newnary.op[2],
1547 newnary.op[3],
1548 &nary);
1549 if (result && is_gimple_min_invariant (result))
1550 return get_or_alloc_expr_for_constant (result);
1552 expr = (pre_expr) pool_alloc (pre_expr_pool);
1553 expr->kind = NARY;
1554 expr->id = 0;
1555 if (nary)
1557 PRE_EXPR_NARY (expr) = nary;
1558 constant = fully_constant_expression (expr);
1559 if (constant != expr)
1560 return constant;
1562 new_val_id = nary->value_id;
1563 get_or_alloc_expression_id (expr);
1565 else
1567 new_val_id = get_next_value_id ();
1568 VEC_safe_grow_cleared (bitmap_set_t, heap,
1569 value_expressions,
1570 get_max_value_id() + 1);
1571 nary = vn_nary_op_insert_pieces (newnary.length,
1572 newnary.opcode,
1573 newnary.type,
1574 newnary.op[0],
1575 newnary.op[1],
1576 newnary.op[2],
1577 newnary.op[3],
1578 result, new_val_id);
1579 PRE_EXPR_NARY (expr) = nary;
1580 constant = fully_constant_expression (expr);
1581 if (constant != expr)
1582 return constant;
1583 get_or_alloc_expression_id (expr);
1585 add_to_value (new_val_id, expr);
1587 phi_trans_add (oldexpr, expr, pred);
1588 return expr;
1590 break;
1592 case REFERENCE:
1594 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1595 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1596 tree vuse = ref->vuse;
1597 tree newvuse = vuse;
1598 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1599 bool changed = false, same_valid = true;
1600 unsigned int i, j;
1601 vn_reference_op_t operand;
1602 vn_reference_t newref;
1604 for (i = 0, j = 0;
1605 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1607 pre_expr opresult;
1608 pre_expr leader;
1609 tree oldop0 = operand->op0;
1610 tree oldop1 = operand->op1;
1611 tree oldop2 = operand->op2;
1612 tree op0 = oldop0;
1613 tree op1 = oldop1;
1614 tree op2 = oldop2;
1615 tree type = operand->type;
1616 vn_reference_op_s newop = *operand;
1618 if (op0 && TREE_CODE (op0) == SSA_NAME)
1620 unsigned int op_val_id = VN_INFO (op0)->value_id;
1621 leader = find_leader_in_sets (op_val_id, set1, set2);
1622 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1623 if (opresult && opresult != leader)
1625 tree name = get_representative_for (opresult);
1626 if (!name)
1627 break;
1628 op0 = name;
1630 else if (!opresult)
1631 break;
1633 changed |= op0 != oldop0;
1635 if (op1 && TREE_CODE (op1) == SSA_NAME)
1637 unsigned int op_val_id = VN_INFO (op1)->value_id;
1638 leader = find_leader_in_sets (op_val_id, set1, set2);
1639 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1640 if (opresult && opresult != leader)
1642 tree name = get_representative_for (opresult);
1643 if (!name)
1644 break;
1645 op1 = name;
1647 else if (!opresult)
1648 break;
1650 /* We can't possibly insert these. */
1651 else if (op1 && !is_gimple_min_invariant (op1))
1652 break;
1653 changed |= op1 != oldop1;
1654 if (op2 && TREE_CODE (op2) == SSA_NAME)
1656 unsigned int op_val_id = VN_INFO (op2)->value_id;
1657 leader = find_leader_in_sets (op_val_id, set1, set2);
1658 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1659 if (opresult && opresult != leader)
1661 tree name = get_representative_for (opresult);
1662 if (!name)
1663 break;
1664 op2 = name;
1666 else if (!opresult)
1667 break;
1669 /* We can't possibly insert these. */
1670 else if (op2 && !is_gimple_min_invariant (op2))
1671 break;
1672 changed |= op2 != oldop2;
1674 if (!newoperands)
1675 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1676 /* We may have changed from an SSA_NAME to a constant */
1677 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1678 newop.opcode = TREE_CODE (op0);
1679 newop.type = type;
1680 newop.op0 = op0;
1681 newop.op1 = op1;
1682 newop.op2 = op2;
1683 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1684 /* If it transforms from an SSA_NAME to an address, fold with
1685 a preceding indirect reference. */
1686 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1687 && VEC_index (vn_reference_op_s,
1688 newoperands, j - 1)->opcode == INDIRECT_REF)
1689 vn_reference_fold_indirect (&newoperands, &j);
1691 if (i != VEC_length (vn_reference_op_s, operands))
1693 if (newoperands)
1694 VEC_free (vn_reference_op_s, heap, newoperands);
1695 return NULL;
1698 if (vuse)
1700 newvuse = translate_vuse_through_block (newoperands,
1701 ref->set, ref->type,
1702 vuse, phiblock, pred,
1703 &same_valid);
1704 if (newvuse == NULL_TREE)
1706 VEC_free (vn_reference_op_s, heap, newoperands);
1707 return NULL;
1711 if (changed || newvuse != vuse)
1713 unsigned int new_val_id;
1714 pre_expr constant;
1716 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1717 ref->type,
1718 newoperands,
1719 &newref, true);
1720 if (newref)
1721 VEC_free (vn_reference_op_s, heap, newoperands);
1723 if (result && is_gimple_min_invariant (result))
1725 gcc_assert (!newoperands);
1726 return get_or_alloc_expr_for_constant (result);
1729 expr = (pre_expr) pool_alloc (pre_expr_pool);
1730 expr->kind = REFERENCE;
1731 expr->id = 0;
1733 if (newref)
1735 PRE_EXPR_REFERENCE (expr) = newref;
1736 constant = fully_constant_expression (expr);
1737 if (constant != expr)
1738 return constant;
1740 new_val_id = newref->value_id;
1741 get_or_alloc_expression_id (expr);
1743 else
1745 if (changed || !same_valid)
1747 new_val_id = get_next_value_id ();
1748 VEC_safe_grow_cleared (bitmap_set_t, heap,
1749 value_expressions,
1750 get_max_value_id() + 1);
1752 else
1753 new_val_id = ref->value_id;
1754 newref = vn_reference_insert_pieces (newvuse, ref->set,
1755 ref->type,
1756 newoperands,
1757 result, new_val_id);
1758 newoperands = NULL;
1759 PRE_EXPR_REFERENCE (expr) = newref;
1760 constant = fully_constant_expression (expr);
1761 if (constant != expr)
1762 return constant;
1763 get_or_alloc_expression_id (expr);
1765 add_to_value (new_val_id, expr);
1767 VEC_free (vn_reference_op_s, heap, newoperands);
1768 phi_trans_add (oldexpr, expr, pred);
1769 return expr;
1771 break;
1773 case NAME:
1775 gimple phi = NULL;
1776 edge e;
1777 gimple def_stmt;
1778 tree name = PRE_EXPR_NAME (expr);
1780 def_stmt = SSA_NAME_DEF_STMT (name);
1781 if (gimple_code (def_stmt) == GIMPLE_PHI
1782 && gimple_bb (def_stmt) == phiblock)
1783 phi = def_stmt;
1784 else
1785 return expr;
1787 e = find_edge (pred, gimple_bb (phi));
1788 if (e)
1790 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1791 pre_expr newexpr;
1793 if (TREE_CODE (def) == SSA_NAME)
1794 def = VN_INFO (def)->valnum;
1796 /* Handle constant. */
1797 if (is_gimple_min_invariant (def))
1798 return get_or_alloc_expr_for_constant (def);
1800 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1801 return NULL;
1803 newexpr = get_or_alloc_expr_for_name (def);
1804 return newexpr;
1807 return expr;
1809 default:
1810 gcc_unreachable ();
1814 /* For each expression in SET, translate the values through phi nodes
1815 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1816 expressions in DEST. */
1818 static void
1819 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1820 basic_block phiblock)
1822 VEC (pre_expr, heap) *exprs;
1823 pre_expr expr;
1824 int i;
1826 if (!phi_nodes (phiblock))
1828 bitmap_set_copy (dest, set);
1829 return;
1832 exprs = sorted_array_from_bitmap_set (set);
1833 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
1835 pre_expr translated;
1836 translated = phi_translate (expr, set, NULL, pred, phiblock);
1838 /* Don't add empty translations to the cache */
1839 if (translated)
1840 phi_trans_add (expr, translated, pred);
1842 if (translated != NULL)
1843 bitmap_value_insert_into_set (dest, translated);
1845 VEC_free (pre_expr, heap, exprs);
1848 /* Find the leader for a value (i.e., the name representing that
1849 value) in a given set, and return it. If STMT is non-NULL it
1850 makes sure the defining statement for the leader dominates it.
1851 Return NULL if no leader is found. */
1853 static pre_expr
1854 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1856 if (value_id_constant_p (val))
1858 unsigned int i;
1859 bitmap_iterator bi;
1860 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1862 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1864 pre_expr expr = expression_for_id (i);
1865 if (expr->kind == CONSTANT)
1866 return expr;
1869 if (bitmap_set_contains_value (set, val))
1871 /* Rather than walk the entire bitmap of expressions, and see
1872 whether any of them has the value we are looking for, we look
1873 at the reverse mapping, which tells us the set of expressions
1874 that have a given value (IE value->expressions with that
1875 value) and see if any of those expressions are in our set.
1876 The number of expressions per value is usually significantly
1877 less than the number of expressions in the set. In fact, for
1878 large testcases, doing it this way is roughly 5-10x faster
1879 than walking the bitmap.
1880 If this is somehow a significant lose for some cases, we can
1881 choose which set to walk based on which set is smaller. */
1882 unsigned int i;
1883 bitmap_iterator bi;
1884 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1886 EXECUTE_IF_AND_IN_BITMAP (exprset->expressions,
1887 set->expressions, 0, i, bi)
1889 pre_expr val = expression_for_id (i);
1890 /* At the point where stmt is not null, there should always
1891 be an SSA_NAME first in the list of expressions. */
1892 if (stmt)
1894 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1895 if (gimple_code (def_stmt) != GIMPLE_PHI
1896 && gimple_bb (def_stmt) == gimple_bb (stmt)
1897 && gimple_uid (def_stmt) >= gimple_uid (stmt))
1898 continue;
1900 return val;
1903 return NULL;
1906 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1907 BLOCK by seeing if it is not killed in the block. Note that we are
1908 only determining whether there is a store that kills it. Because
1909 of the order in which clean iterates over values, we are guaranteed
1910 that altered operands will have caused us to be eliminated from the
1911 ANTIC_IN set already. */
1913 static bool
1914 value_dies_in_block_x (pre_expr expr, basic_block block)
1916 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1917 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1918 gimple def;
1919 gimple_stmt_iterator gsi;
1920 unsigned id = get_expression_id (expr);
1921 bool res = false;
1922 ao_ref ref;
1924 if (!vuse)
1925 return false;
1927 /* Lookup a previously calculated result. */
1928 if (EXPR_DIES (block)
1929 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1930 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1932 /* A memory expression {e, VUSE} dies in the block if there is a
1933 statement that may clobber e. If, starting statement walk from the
1934 top of the basic block, a statement uses VUSE there can be no kill
1935 inbetween that use and the original statement that loaded {e, VUSE},
1936 so we can stop walking. */
1937 ref.base = NULL_TREE;
1938 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1940 tree def_vuse, def_vdef;
1941 def = gsi_stmt (gsi);
1942 def_vuse = gimple_vuse (def);
1943 def_vdef = gimple_vdef (def);
1945 /* Not a memory statement. */
1946 if (!def_vuse)
1947 continue;
1949 /* Not a may-def. */
1950 if (!def_vdef)
1952 /* A load with the same VUSE, we're done. */
1953 if (def_vuse == vuse)
1954 break;
1956 continue;
1959 /* Init ref only if we really need it. */
1960 if (ref.base == NULL_TREE
1961 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1962 refx->operands))
1964 res = true;
1965 break;
1967 /* If the statement may clobber expr, it dies. */
1968 if (stmt_may_clobber_ref_p_1 (def, &ref))
1970 res = true;
1971 break;
1975 /* Remember the result. */
1976 if (!EXPR_DIES (block))
1977 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1978 bitmap_set_bit (EXPR_DIES (block), id * 2);
1979 if (res)
1980 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1982 return res;
1986 #define union_contains_value(SET1, SET2, VAL) \
1987 (bitmap_set_contains_value ((SET1), (VAL)) \
1988 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
1990 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
1992 static bool
1993 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
1994 vn_reference_op_t vro)
1996 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
1998 struct pre_expr_d temp;
1999 temp.kind = NAME;
2000 temp.id = 0;
2001 PRE_EXPR_NAME (&temp) = vro->op0;
2002 temp.id = lookup_expression_id (&temp);
2003 if (temp.id == 0)
2004 return false;
2005 if (!union_contains_value (set1, set2,
2006 get_expr_value_id (&temp)))
2007 return false;
2009 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2011 struct pre_expr_d temp;
2012 temp.kind = NAME;
2013 temp.id = 0;
2014 PRE_EXPR_NAME (&temp) = vro->op1;
2015 temp.id = lookup_expression_id (&temp);
2016 if (temp.id == 0)
2017 return false;
2018 if (!union_contains_value (set1, set2,
2019 get_expr_value_id (&temp)))
2020 return false;
2023 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2025 struct pre_expr_d temp;
2026 temp.kind = NAME;
2027 temp.id = 0;
2028 PRE_EXPR_NAME (&temp) = vro->op2;
2029 temp.id = lookup_expression_id (&temp);
2030 if (temp.id == 0)
2031 return false;
2032 if (!union_contains_value (set1, set2,
2033 get_expr_value_id (&temp)))
2034 return false;
2037 return true;
2040 /* Determine if the expression EXPR is valid in SET1 U SET2.
2041 ONLY SET2 CAN BE NULL.
2042 This means that we have a leader for each part of the expression
2043 (if it consists of values), or the expression is an SSA_NAME.
2044 For loads/calls, we also see if the vuse is killed in this block. */
2046 static bool
2047 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2048 basic_block block)
2050 switch (expr->kind)
2052 case NAME:
2053 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2054 case NARY:
2056 unsigned int i;
2057 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2058 for (i = 0; i < nary->length; i++)
2060 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2062 struct pre_expr_d temp;
2063 temp.kind = NAME;
2064 temp.id = 0;
2065 PRE_EXPR_NAME (&temp) = nary->op[i];
2066 temp.id = lookup_expression_id (&temp);
2067 if (temp.id == 0)
2068 return false;
2069 if (!union_contains_value (set1, set2,
2070 get_expr_value_id (&temp)))
2071 return false;
2074 /* If the NARY may trap make sure the block does not contain
2075 a possible exit point.
2076 ??? This is overly conservative if we translate AVAIL_OUT
2077 as the available expression might be after the exit point. */
2078 if (BB_MAY_NOTRETURN (block)
2079 && vn_nary_may_trap (nary))
2080 return false;
2081 return true;
2083 break;
2084 case REFERENCE:
2086 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2087 vn_reference_op_t vro;
2088 unsigned int i;
2090 for (i = 0; VEC_iterate (vn_reference_op_s, ref->operands, i, vro); i++)
2092 if (!vro_valid_in_sets (set1, set2, vro))
2093 return false;
2095 if (ref->vuse)
2097 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2098 if (!gimple_nop_p (def_stmt)
2099 && gimple_bb (def_stmt) != block
2100 && !dominated_by_p (CDI_DOMINATORS,
2101 block, gimple_bb (def_stmt)))
2102 return false;
2104 return !value_dies_in_block_x (expr, block);
2106 default:
2107 gcc_unreachable ();
2111 /* Clean the set of expressions that are no longer valid in SET1 or
2112 SET2. This means expressions that are made up of values we have no
2113 leaders for in SET1 or SET2. This version is used for partial
2114 anticipation, which means it is not valid in either ANTIC_IN or
2115 PA_IN. */
2117 static void
2118 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2120 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2121 pre_expr expr;
2122 int i;
2124 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2126 if (!valid_in_sets (set1, set2, expr, block))
2127 bitmap_remove_from_set (set1, expr);
2129 VEC_free (pre_expr, heap, exprs);
2132 /* Clean the set of expressions that are no longer valid in SET. This
2133 means expressions that are made up of values we have no leaders for
2134 in SET. */
2136 static void
2137 clean (bitmap_set_t set, basic_block block)
2139 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2140 pre_expr expr;
2141 int i;
2143 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
2145 if (!valid_in_sets (set, NULL, expr, block))
2146 bitmap_remove_from_set (set, expr);
2148 VEC_free (pre_expr, heap, exprs);
2151 static sbitmap has_abnormal_preds;
2153 /* List of blocks that may have changed during ANTIC computation and
2154 thus need to be iterated over. */
2156 static sbitmap changed_blocks;
2158 /* Decide whether to defer a block for a later iteration, or PHI
2159 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2160 should defer the block, and true if we processed it. */
2162 static bool
2163 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2164 basic_block block, basic_block phiblock)
2166 if (!BB_VISITED (phiblock))
2168 SET_BIT (changed_blocks, block->index);
2169 BB_VISITED (block) = 0;
2170 BB_DEFERRED (block) = 1;
2171 return false;
2173 else
2174 phi_translate_set (dest, source, block, phiblock);
2175 return true;
2178 /* Compute the ANTIC set for BLOCK.
2180 If succs(BLOCK) > 1 then
2181 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2182 else if succs(BLOCK) == 1 then
2183 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2185 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2188 static bool
2189 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2191 bool changed = false;
2192 bitmap_set_t S, old, ANTIC_OUT;
2193 bitmap_iterator bi;
2194 unsigned int bii;
2195 edge e;
2196 edge_iterator ei;
2198 old = ANTIC_OUT = S = NULL;
2199 BB_VISITED (block) = 1;
2201 /* If any edges from predecessors are abnormal, antic_in is empty,
2202 so do nothing. */
2203 if (block_has_abnormal_pred_edge)
2204 goto maybe_dump_sets;
2206 old = ANTIC_IN (block);
2207 ANTIC_OUT = bitmap_set_new ();
2209 /* If the block has no successors, ANTIC_OUT is empty. */
2210 if (EDGE_COUNT (block->succs) == 0)
2212 /* If we have one successor, we could have some phi nodes to
2213 translate through. */
2214 else if (single_succ_p (block))
2216 basic_block succ_bb = single_succ (block);
2218 /* We trade iterations of the dataflow equations for having to
2219 phi translate the maximal set, which is incredibly slow
2220 (since the maximal set often has 300+ members, even when you
2221 have a small number of blocks).
2222 Basically, we defer the computation of ANTIC for this block
2223 until we have processed it's successor, which will inevitably
2224 have a *much* smaller set of values to phi translate once
2225 clean has been run on it.
2226 The cost of doing this is that we technically perform more
2227 iterations, however, they are lower cost iterations.
2229 Timings for PRE on tramp3d-v4:
2230 without maximal set fix: 11 seconds
2231 with maximal set fix/without deferring: 26 seconds
2232 with maximal set fix/with deferring: 11 seconds
2235 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2236 block, succ_bb))
2238 changed = true;
2239 goto maybe_dump_sets;
2242 /* If we have multiple successors, we take the intersection of all of
2243 them. Note that in the case of loop exit phi nodes, we may have
2244 phis to translate through. */
2245 else
2247 VEC(basic_block, heap) * worklist;
2248 size_t i;
2249 basic_block bprime, first = NULL;
2251 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2252 FOR_EACH_EDGE (e, ei, block->succs)
2254 if (!first
2255 && BB_VISITED (e->dest))
2256 first = e->dest;
2257 else if (BB_VISITED (e->dest))
2258 VEC_quick_push (basic_block, worklist, e->dest);
2261 /* Of multiple successors we have to have visited one already. */
2262 if (!first)
2264 SET_BIT (changed_blocks, block->index);
2265 BB_VISITED (block) = 0;
2266 BB_DEFERRED (block) = 1;
2267 changed = true;
2268 VEC_free (basic_block, heap, worklist);
2269 goto maybe_dump_sets;
2272 if (phi_nodes (first))
2273 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2274 else
2275 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2277 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2279 if (phi_nodes (bprime))
2281 bitmap_set_t tmp = bitmap_set_new ();
2282 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2283 bitmap_set_and (ANTIC_OUT, tmp);
2284 bitmap_set_free (tmp);
2286 else
2287 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2289 VEC_free (basic_block, heap, worklist);
2292 /* Generate ANTIC_OUT - TMP_GEN. */
2293 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2295 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2296 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2297 TMP_GEN (block));
2299 /* Then union in the ANTIC_OUT - TMP_GEN values,
2300 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2301 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2302 bitmap_value_insert_into_set (ANTIC_IN (block),
2303 expression_for_id (bii));
2305 clean (ANTIC_IN (block), block);
2307 /* !old->expressions can happen when we deferred a block. */
2308 if (!old->expressions || !bitmap_set_equal (old, ANTIC_IN (block)))
2310 changed = true;
2311 SET_BIT (changed_blocks, block->index);
2312 FOR_EACH_EDGE (e, ei, block->preds)
2313 SET_BIT (changed_blocks, e->src->index);
2315 else
2316 RESET_BIT (changed_blocks, block->index);
2318 maybe_dump_sets:
2319 if (dump_file && (dump_flags & TDF_DETAILS))
2321 if (!BB_DEFERRED (block) || BB_VISITED (block))
2323 if (ANTIC_OUT)
2324 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2326 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2327 block->index);
2329 if (S)
2330 print_bitmap_set (dump_file, S, "S", block->index);
2332 else
2334 fprintf (dump_file,
2335 "Block %d was deferred for a future iteration.\n",
2336 block->index);
2339 if (old)
2340 bitmap_set_free (old);
2341 if (S)
2342 bitmap_set_free (S);
2343 if (ANTIC_OUT)
2344 bitmap_set_free (ANTIC_OUT);
2345 return changed;
2348 /* Compute PARTIAL_ANTIC for BLOCK.
2350 If succs(BLOCK) > 1 then
2351 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2352 in ANTIC_OUT for all succ(BLOCK)
2353 else if succs(BLOCK) == 1 then
2354 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2356 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2357 - ANTIC_IN[BLOCK])
2360 static bool
2361 compute_partial_antic_aux (basic_block block,
2362 bool block_has_abnormal_pred_edge)
2364 bool changed = false;
2365 bitmap_set_t old_PA_IN;
2366 bitmap_set_t PA_OUT;
2367 edge e;
2368 edge_iterator ei;
2369 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2371 old_PA_IN = PA_OUT = NULL;
2373 /* If any edges from predecessors are abnormal, antic_in is empty,
2374 so do nothing. */
2375 if (block_has_abnormal_pred_edge)
2376 goto maybe_dump_sets;
2378 /* If there are too many partially anticipatable values in the
2379 block, phi_translate_set can take an exponential time: stop
2380 before the translation starts. */
2381 if (max_pa
2382 && single_succ_p (block)
2383 && bitmap_count_bits (PA_IN (single_succ (block))->values) > max_pa)
2384 goto maybe_dump_sets;
2386 old_PA_IN = PA_IN (block);
2387 PA_OUT = bitmap_set_new ();
2389 /* If the block has no successors, ANTIC_OUT is empty. */
2390 if (EDGE_COUNT (block->succs) == 0)
2392 /* If we have one successor, we could have some phi nodes to
2393 translate through. Note that we can't phi translate across DFS
2394 back edges in partial antic, because it uses a union operation on
2395 the successors. For recurrences like IV's, we will end up
2396 generating a new value in the set on each go around (i + 3 (VH.1)
2397 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2398 else if (single_succ_p (block))
2400 basic_block succ = single_succ (block);
2401 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2402 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2404 /* If we have multiple successors, we take the union of all of
2405 them. */
2406 else
2408 VEC(basic_block, heap) * worklist;
2409 size_t i;
2410 basic_block bprime;
2412 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2413 FOR_EACH_EDGE (e, ei, block->succs)
2415 if (e->flags & EDGE_DFS_BACK)
2416 continue;
2417 VEC_quick_push (basic_block, worklist, e->dest);
2419 if (VEC_length (basic_block, worklist) > 0)
2421 for (i = 0; VEC_iterate (basic_block, worklist, i, bprime); i++)
2423 unsigned int i;
2424 bitmap_iterator bi;
2426 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2427 bitmap_value_insert_into_set (PA_OUT,
2428 expression_for_id (i));
2429 if (phi_nodes (bprime))
2431 bitmap_set_t pa_in = bitmap_set_new ();
2432 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2433 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2434 bitmap_value_insert_into_set (PA_OUT,
2435 expression_for_id (i));
2436 bitmap_set_free (pa_in);
2438 else
2439 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2440 bitmap_value_insert_into_set (PA_OUT,
2441 expression_for_id (i));
2444 VEC_free (basic_block, heap, worklist);
2447 /* PA_IN starts with PA_OUT - TMP_GEN.
2448 Then we subtract things from ANTIC_IN. */
2449 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2451 /* For partial antic, we want to put back in the phi results, since
2452 we will properly avoid making them partially antic over backedges. */
2453 bitmap_ior_into (PA_IN (block)->values, PHI_GEN (block)->values);
2454 bitmap_ior_into (PA_IN (block)->expressions, PHI_GEN (block)->expressions);
2456 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2457 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2459 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2461 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2463 changed = true;
2464 SET_BIT (changed_blocks, block->index);
2465 FOR_EACH_EDGE (e, ei, block->preds)
2466 SET_BIT (changed_blocks, e->src->index);
2468 else
2469 RESET_BIT (changed_blocks, block->index);
2471 maybe_dump_sets:
2472 if (dump_file && (dump_flags & TDF_DETAILS))
2474 if (PA_OUT)
2475 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2477 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2479 if (old_PA_IN)
2480 bitmap_set_free (old_PA_IN);
2481 if (PA_OUT)
2482 bitmap_set_free (PA_OUT);
2483 return changed;
2486 /* Compute ANTIC and partial ANTIC sets. */
2488 static void
2489 compute_antic (void)
2491 bool changed = true;
2492 int num_iterations = 0;
2493 basic_block block;
2494 int i;
2496 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2497 We pre-build the map of blocks with incoming abnormal edges here. */
2498 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2499 sbitmap_zero (has_abnormal_preds);
2501 FOR_EACH_BB (block)
2503 edge_iterator ei;
2504 edge e;
2506 FOR_EACH_EDGE (e, ei, block->preds)
2508 e->flags &= ~EDGE_DFS_BACK;
2509 if (e->flags & EDGE_ABNORMAL)
2511 SET_BIT (has_abnormal_preds, block->index);
2512 break;
2516 BB_VISITED (block) = 0;
2517 BB_DEFERRED (block) = 0;
2519 /* While we are here, give empty ANTIC_IN sets to each block. */
2520 ANTIC_IN (block) = bitmap_set_new ();
2521 PA_IN (block) = bitmap_set_new ();
2524 /* At the exit block we anticipate nothing. */
2525 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2526 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2527 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2529 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2530 sbitmap_ones (changed_blocks);
2531 while (changed)
2533 if (dump_file && (dump_flags & TDF_DETAILS))
2534 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2535 num_iterations++;
2536 changed = false;
2537 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2539 if (TEST_BIT (changed_blocks, postorder[i]))
2541 basic_block block = BASIC_BLOCK (postorder[i]);
2542 changed |= compute_antic_aux (block,
2543 TEST_BIT (has_abnormal_preds,
2544 block->index));
2547 #ifdef ENABLE_CHECKING
2548 /* Theoretically possible, but *highly* unlikely. */
2549 gcc_assert (num_iterations < 500);
2550 #endif
2553 statistics_histogram_event (cfun, "compute_antic iterations",
2554 num_iterations);
2556 if (do_partial_partial)
2558 sbitmap_ones (changed_blocks);
2559 mark_dfs_back_edges ();
2560 num_iterations = 0;
2561 changed = true;
2562 while (changed)
2564 if (dump_file && (dump_flags & TDF_DETAILS))
2565 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2566 num_iterations++;
2567 changed = false;
2568 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2570 if (TEST_BIT (changed_blocks, postorder[i]))
2572 basic_block block = BASIC_BLOCK (postorder[i]);
2573 changed
2574 |= compute_partial_antic_aux (block,
2575 TEST_BIT (has_abnormal_preds,
2576 block->index));
2579 #ifdef ENABLE_CHECKING
2580 /* Theoretically possible, but *highly* unlikely. */
2581 gcc_assert (num_iterations < 500);
2582 #endif
2584 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2585 num_iterations);
2587 sbitmap_free (has_abnormal_preds);
2588 sbitmap_free (changed_blocks);
2591 /* Return true if we can value number the call in STMT. This is true
2592 if we have a pure or constant call. */
2594 static bool
2595 can_value_number_call (gimple stmt)
2597 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2598 return true;
2599 return false;
2602 /* Return true if OP is a tree which we can perform PRE on.
2603 This may not match the operations we can value number, but in
2604 a perfect world would. */
2606 static bool
2607 can_PRE_operation (tree op)
2609 return UNARY_CLASS_P (op)
2610 || BINARY_CLASS_P (op)
2611 || COMPARISON_CLASS_P (op)
2612 || TREE_CODE (op) == INDIRECT_REF
2613 || TREE_CODE (op) == COMPONENT_REF
2614 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2615 || TREE_CODE (op) == CALL_EXPR
2616 || TREE_CODE (op) == ARRAY_REF;
2620 /* Inserted expressions are placed onto this worklist, which is used
2621 for performing quick dead code elimination of insertions we made
2622 that didn't turn out to be necessary. */
2623 static VEC(gimple,heap) *inserted_exprs;
2624 static bitmap inserted_phi_names;
2626 /* Pool allocated fake store expressions are placed onto this
2627 worklist, which, after performing dead code elimination, is walked
2628 to see which expressions need to be put into GC'able memory */
2629 static VEC(gimple, heap) *need_creation;
2631 /* The actual worker for create_component_ref_by_pieces. */
2633 static tree
2634 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2635 unsigned int *operand, gimple_seq *stmts,
2636 gimple domstmt)
2638 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2639 *operand);
2640 tree genop;
2641 ++*operand;
2642 switch (currop->opcode)
2644 case CALL_EXPR:
2646 tree folded, sc = currop->op1;
2647 unsigned int nargs = 0;
2648 tree *args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2649 ref->operands) - 1);
2650 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2652 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2653 operand, stmts,
2654 domstmt);
2655 nargs++;
2657 folded = build_call_array (currop->type,
2658 TREE_CODE (currop->op0) == FUNCTION_DECL
2659 ? build_fold_addr_expr (currop->op0)
2660 : currop->op0,
2661 nargs, args);
2662 free (args);
2663 if (sc)
2665 pre_expr scexpr = get_or_alloc_expr_for (sc);
2666 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2667 if (!sc)
2668 return NULL_TREE;
2669 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2671 return folded;
2673 break;
2674 case TARGET_MEM_REF:
2676 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2677 *operand);
2678 pre_expr op0expr;
2679 tree genop0 = NULL_TREE;
2680 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2681 stmts, domstmt);
2682 if (!baseop)
2683 return NULL_TREE;
2684 if (currop->op0)
2686 op0expr = get_or_alloc_expr_for (currop->op0);
2687 genop0 = find_or_generate_expression (block, op0expr,
2688 stmts, domstmt);
2689 if (!genop0)
2690 return NULL_TREE;
2692 if (DECL_P (baseop))
2693 return build6 (TARGET_MEM_REF, currop->type,
2694 baseop, NULL_TREE,
2695 genop0, currop->op1, currop->op2,
2696 unshare_expr (nextop->op1));
2697 else
2698 return build6 (TARGET_MEM_REF, currop->type,
2699 NULL_TREE, baseop,
2700 genop0, currop->op1, currop->op2,
2701 unshare_expr (nextop->op1));
2703 break;
2704 case ADDR_EXPR:
2705 if (currop->op0)
2707 gcc_assert (is_gimple_min_invariant (currop->op0));
2708 return currop->op0;
2710 /* Fallthrough. */
2711 case REALPART_EXPR:
2712 case IMAGPART_EXPR:
2713 case VIEW_CONVERT_EXPR:
2715 tree folded;
2716 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2717 operand,
2718 stmts, domstmt);
2719 if (!genop0)
2720 return NULL_TREE;
2721 folded = fold_build1 (currop->opcode, currop->type,
2722 genop0);
2723 return folded;
2725 break;
2726 case ALIGN_INDIRECT_REF:
2727 case MISALIGNED_INDIRECT_REF:
2728 case INDIRECT_REF:
2730 tree folded;
2731 tree genop1 = create_component_ref_by_pieces_1 (block, ref,
2732 operand,
2733 stmts, domstmt);
2734 if (!genop1)
2735 return NULL_TREE;
2736 genop1 = fold_convert (build_pointer_type (currop->type),
2737 genop1);
2739 if (currop->opcode == MISALIGNED_INDIRECT_REF)
2740 folded = fold_build2 (currop->opcode, currop->type,
2741 genop1, currop->op1);
2742 else
2743 folded = fold_build1 (currop->opcode, currop->type,
2744 genop1);
2745 return folded;
2747 break;
2748 case BIT_FIELD_REF:
2750 tree folded;
2751 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2752 stmts, domstmt);
2753 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2754 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2755 tree genop1;
2756 tree genop2;
2758 if (!genop0)
2759 return NULL_TREE;
2760 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2761 if (!genop1)
2762 return NULL_TREE;
2763 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2764 if (!genop2)
2765 return NULL_TREE;
2766 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2767 genop2);
2768 return folded;
2771 /* For array ref vn_reference_op's, operand 1 of the array ref
2772 is op0 of the reference op and operand 3 of the array ref is
2773 op1. */
2774 case ARRAY_RANGE_REF:
2775 case ARRAY_REF:
2777 tree genop0;
2778 tree genop1 = currop->op0;
2779 pre_expr op1expr;
2780 tree genop2 = currop->op1;
2781 pre_expr op2expr;
2782 tree genop3 = currop->op2;
2783 pre_expr op3expr;
2784 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2785 stmts, domstmt);
2786 if (!genop0)
2787 return NULL_TREE;
2788 op1expr = get_or_alloc_expr_for (genop1);
2789 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2790 if (!genop1)
2791 return NULL_TREE;
2792 if (genop2)
2794 op2expr = get_or_alloc_expr_for (genop2);
2795 genop2 = find_or_generate_expression (block, op2expr, stmts,
2796 domstmt);
2797 if (!genop2)
2798 return NULL_TREE;
2800 if (genop3)
2802 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2803 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2804 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2805 op3expr = get_or_alloc_expr_for (genop3);
2806 genop3 = find_or_generate_expression (block, op3expr, stmts,
2807 domstmt);
2808 if (!genop3)
2809 return NULL_TREE;
2811 return build4 (currop->opcode, currop->type, genop0, genop1,
2812 genop2, genop3);
2814 case COMPONENT_REF:
2816 tree op0;
2817 tree op1;
2818 tree genop2 = currop->op1;
2819 pre_expr op2expr;
2820 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2821 stmts, domstmt);
2822 if (!op0)
2823 return NULL_TREE;
2824 /* op1 should be a FIELD_DECL, which are represented by
2825 themselves. */
2826 op1 = currop->op0;
2827 if (genop2)
2829 op2expr = get_or_alloc_expr_for (genop2);
2830 genop2 = find_or_generate_expression (block, op2expr, stmts,
2831 domstmt);
2832 if (!genop2)
2833 return NULL_TREE;
2836 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2837 genop2);
2839 break;
2840 case SSA_NAME:
2842 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2843 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2844 return genop;
2846 case STRING_CST:
2847 case INTEGER_CST:
2848 case COMPLEX_CST:
2849 case VECTOR_CST:
2850 case REAL_CST:
2851 case CONSTRUCTOR:
2852 case VAR_DECL:
2853 case PARM_DECL:
2854 case CONST_DECL:
2855 case RESULT_DECL:
2856 case FUNCTION_DECL:
2857 return currop->op0;
2859 default:
2860 gcc_unreachable ();
2864 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2865 COMPONENT_REF or INDIRECT_REF or ARRAY_REF portion, because we'd end up with
2866 trying to rename aggregates into ssa form directly, which is a no no.
2868 Thus, this routine doesn't create temporaries, it just builds a
2869 single access expression for the array, calling
2870 find_or_generate_expression to build the innermost pieces.
2872 This function is a subroutine of create_expression_by_pieces, and
2873 should not be called on it's own unless you really know what you
2874 are doing. */
2876 static tree
2877 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2878 gimple_seq *stmts, gimple domstmt)
2880 unsigned int op = 0;
2881 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2884 /* Find a leader for an expression, or generate one using
2885 create_expression_by_pieces if it's ANTIC but
2886 complex.
2887 BLOCK is the basic_block we are looking for leaders in.
2888 EXPR is the expression to find a leader or generate for.
2889 STMTS is the statement list to put the inserted expressions on.
2890 Returns the SSA_NAME of the LHS of the generated expression or the
2891 leader.
2892 DOMSTMT if non-NULL is a statement that should be dominated by
2893 all uses in the generated expression. If DOMSTMT is non-NULL this
2894 routine can fail and return NULL_TREE. Otherwise it will assert
2895 on failure. */
2897 static tree
2898 find_or_generate_expression (basic_block block, pre_expr expr,
2899 gimple_seq *stmts, gimple domstmt)
2901 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2902 get_expr_value_id (expr), domstmt);
2903 tree genop = NULL;
2904 if (leader)
2906 if (leader->kind == NAME)
2907 genop = PRE_EXPR_NAME (leader);
2908 else if (leader->kind == CONSTANT)
2909 genop = PRE_EXPR_CONSTANT (leader);
2912 /* If it's still NULL, it must be a complex expression, so generate
2913 it recursively. Not so for FRE though. */
2914 if (genop == NULL
2915 && !in_fre)
2917 bitmap_set_t exprset;
2918 unsigned int lookfor = get_expr_value_id (expr);
2919 bool handled = false;
2920 bitmap_iterator bi;
2921 unsigned int i;
2923 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2924 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2926 pre_expr temp = expression_for_id (i);
2927 if (temp->kind != NAME)
2929 handled = true;
2930 genop = create_expression_by_pieces (block, temp, stmts,
2931 domstmt,
2932 get_expr_type (expr));
2933 break;
2936 if (!handled && domstmt)
2937 return NULL_TREE;
2939 gcc_assert (handled);
2941 return genop;
2944 #define NECESSARY GF_PLF_1
2946 /* Create an expression in pieces, so that we can handle very complex
2947 expressions that may be ANTIC, but not necessary GIMPLE.
2948 BLOCK is the basic block the expression will be inserted into,
2949 EXPR is the expression to insert (in value form)
2950 STMTS is a statement list to append the necessary insertions into.
2952 This function will die if we hit some value that shouldn't be
2953 ANTIC but is (IE there is no leader for it, or its components).
2954 This function may also generate expressions that are themselves
2955 partially or fully redundant. Those that are will be either made
2956 fully redundant during the next iteration of insert (for partially
2957 redundant ones), or eliminated by eliminate (for fully redundant
2958 ones).
2960 If DOMSTMT is non-NULL then we make sure that all uses in the
2961 expressions dominate that statement. In this case the function
2962 can return NULL_TREE to signal failure. */
2964 static tree
2965 create_expression_by_pieces (basic_block block, pre_expr expr,
2966 gimple_seq *stmts, gimple domstmt, tree type)
2968 tree temp, name;
2969 tree folded;
2970 gimple_seq forced_stmts = NULL;
2971 unsigned int value_id;
2972 gimple_stmt_iterator gsi;
2973 tree exprtype = type ? type : get_expr_type (expr);
2974 pre_expr nameexpr;
2975 gimple newstmt;
2977 switch (expr->kind)
2979 /* We may hit the NAME/CONSTANT case if we have to convert types
2980 that value numbering saw through. */
2981 case NAME:
2982 folded = PRE_EXPR_NAME (expr);
2983 break;
2984 case CONSTANT:
2985 folded = PRE_EXPR_CONSTANT (expr);
2986 break;
2987 case REFERENCE:
2989 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2990 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
2992 break;
2993 case NARY:
2995 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2996 switch (nary->length)
2998 case 2:
3000 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3001 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
3002 tree genop1 = find_or_generate_expression (block, op1,
3003 stmts, domstmt);
3004 tree genop2 = find_or_generate_expression (block, op2,
3005 stmts, domstmt);
3006 if (!genop1 || !genop2)
3007 return NULL_TREE;
3008 genop1 = fold_convert (TREE_TYPE (nary->op[0]),
3009 genop1);
3010 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
3011 may be a constant with the wrong type. */
3012 if (nary->opcode == POINTER_PLUS_EXPR)
3013 genop2 = fold_convert (sizetype, genop2);
3014 else
3015 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
3017 folded = fold_build2 (nary->opcode, nary->type,
3018 genop1, genop2);
3020 break;
3021 case 1:
3023 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3024 tree genop1 = find_or_generate_expression (block, op1,
3025 stmts, domstmt);
3026 if (!genop1)
3027 return NULL_TREE;
3028 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3030 folded = fold_build1 (nary->opcode, nary->type,
3031 genop1);
3033 break;
3034 default:
3035 return NULL_TREE;
3038 break;
3039 default:
3040 return NULL_TREE;
3043 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3044 folded = fold_convert (exprtype, folded);
3046 /* Force the generated expression to be a sequence of GIMPLE
3047 statements.
3048 We have to call unshare_expr because force_gimple_operand may
3049 modify the tree we pass to it. */
3050 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3051 false, NULL);
3053 /* If we have any intermediate expressions to the value sets, add them
3054 to the value sets and chain them in the instruction stream. */
3055 if (forced_stmts)
3057 gsi = gsi_start (forced_stmts);
3058 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3060 gimple stmt = gsi_stmt (gsi);
3061 tree forcedname = gimple_get_lhs (stmt);
3062 pre_expr nameexpr;
3064 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3065 if (TREE_CODE (forcedname) == SSA_NAME)
3067 VN_INFO_GET (forcedname)->valnum = forcedname;
3068 VN_INFO (forcedname)->value_id = get_next_value_id ();
3069 nameexpr = get_or_alloc_expr_for_name (forcedname);
3070 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3071 if (!in_fre)
3072 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3073 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3075 mark_symbols_for_renaming (stmt);
3077 gimple_seq_add_seq (stmts, forced_stmts);
3080 /* Build and insert the assignment of the end result to the temporary
3081 that we will return. */
3082 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3084 pretemp = create_tmp_var (exprtype, "pretmp");
3085 get_var_ann (pretemp);
3088 temp = pretemp;
3089 add_referenced_var (temp);
3091 if (TREE_CODE (exprtype) == COMPLEX_TYPE
3092 || TREE_CODE (exprtype) == VECTOR_TYPE)
3093 DECL_GIMPLE_REG_P (temp) = 1;
3095 newstmt = gimple_build_assign (temp, folded);
3096 name = make_ssa_name (temp, newstmt);
3097 gimple_assign_set_lhs (newstmt, name);
3098 gimple_set_plf (newstmt, NECESSARY, false);
3100 gimple_seq_add_stmt (stmts, newstmt);
3101 VEC_safe_push (gimple, heap, inserted_exprs, newstmt);
3103 /* All the symbols in NEWEXPR should be put into SSA form. */
3104 mark_symbols_for_renaming (newstmt);
3106 /* Add a value number to the temporary.
3107 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3108 we are creating the expression by pieces, and this particular piece of
3109 the expression may have been represented. There is no harm in replacing
3110 here. */
3111 VN_INFO_GET (name)->valnum = name;
3112 value_id = get_expr_value_id (expr);
3113 VN_INFO (name)->value_id = value_id;
3114 nameexpr = get_or_alloc_expr_for_name (name);
3115 add_to_value (value_id, nameexpr);
3116 if (!in_fre)
3117 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3118 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3120 pre_stats.insertions++;
3121 if (dump_file && (dump_flags & TDF_DETAILS))
3123 fprintf (dump_file, "Inserted ");
3124 print_gimple_stmt (dump_file, newstmt, 0, 0);
3125 fprintf (dump_file, " in predecessor %d\n", block->index);
3128 return name;
3132 /* Returns true if we want to inhibit the insertions of PHI nodes
3133 for the given EXPR for basic block BB (a member of a loop).
3134 We want to do this, when we fear that the induction variable we
3135 create might inhibit vectorization. */
3137 static bool
3138 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3140 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3141 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3142 vn_reference_op_t op;
3143 unsigned i;
3145 /* If we aren't going to vectorize we don't inhibit anything. */
3146 if (!flag_tree_vectorize)
3147 return false;
3149 /* Otherwise we inhibit the insertion when the address of the
3150 memory reference is a simple induction variable. In other
3151 cases the vectorizer won't do anything anyway (either it's
3152 loop invariant or a complicated expression). */
3153 for (i = 0; VEC_iterate (vn_reference_op_s, ops, i, op); ++i)
3155 switch (op->opcode)
3157 case ARRAY_REF:
3158 case ARRAY_RANGE_REF:
3159 if (TREE_CODE (op->op0) != SSA_NAME)
3160 break;
3161 /* Fallthru. */
3162 case SSA_NAME:
3164 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3165 affine_iv iv;
3166 /* Default defs are loop invariant. */
3167 if (!defbb)
3168 break;
3169 /* Defined outside this loop, also loop invariant. */
3170 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3171 break;
3172 /* If it's a simple induction variable inhibit insertion,
3173 the vectorizer might be interested in this one. */
3174 if (simple_iv (bb->loop_father, bb->loop_father,
3175 op->op0, &iv, true))
3176 return true;
3177 /* No simple IV, vectorizer can't do anything, hence no
3178 reason to inhibit the transformation for this operand. */
3179 break;
3181 default:
3182 break;
3185 return false;
3188 /* Insert the to-be-made-available values of expression EXPRNUM for each
3189 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3190 merge the result with a phi node, given the same value number as
3191 NODE. Return true if we have inserted new stuff. */
3193 static bool
3194 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3195 pre_expr *avail)
3197 pre_expr expr = expression_for_id (exprnum);
3198 pre_expr newphi;
3199 unsigned int val = get_expr_value_id (expr);
3200 edge pred;
3201 bool insertions = false;
3202 bool nophi = false;
3203 basic_block bprime;
3204 pre_expr eprime;
3205 edge_iterator ei;
3206 tree type = get_expr_type (expr);
3207 tree temp;
3208 gimple phi;
3210 if (dump_file && (dump_flags & TDF_DETAILS))
3212 fprintf (dump_file, "Found partial redundancy for expression ");
3213 print_pre_expr (dump_file, expr);
3214 fprintf (dump_file, " (%04d)\n", val);
3217 /* Make sure we aren't creating an induction variable. */
3218 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3220 bool firstinsideloop = false;
3221 bool secondinsideloop = false;
3222 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3223 EDGE_PRED (block, 0)->src);
3224 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3225 EDGE_PRED (block, 1)->src);
3226 /* Induction variables only have one edge inside the loop. */
3227 if ((firstinsideloop ^ secondinsideloop)
3228 && (expr->kind != REFERENCE
3229 || inhibit_phi_insertion (block, expr)))
3231 if (dump_file && (dump_flags & TDF_DETAILS))
3232 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3233 nophi = true;
3237 /* Make the necessary insertions. */
3238 FOR_EACH_EDGE (pred, ei, block->preds)
3240 gimple_seq stmts = NULL;
3241 tree builtexpr;
3242 bprime = pred->src;
3243 eprime = avail[bprime->index];
3245 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3247 builtexpr = create_expression_by_pieces (bprime,
3248 eprime,
3249 &stmts, NULL,
3250 type);
3251 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3252 gsi_insert_seq_on_edge (pred, stmts);
3253 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3254 insertions = true;
3256 else if (eprime->kind == CONSTANT)
3258 /* Constants may not have the right type, fold_convert
3259 should give us back a constant with the right type.
3261 tree constant = PRE_EXPR_CONSTANT (eprime);
3262 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3264 tree builtexpr = fold_convert (type, constant);
3265 if (!is_gimple_min_invariant (builtexpr))
3267 tree forcedexpr = force_gimple_operand (builtexpr,
3268 &stmts, true,
3269 NULL);
3270 if (!is_gimple_min_invariant (forcedexpr))
3272 if (forcedexpr != builtexpr)
3274 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3275 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3277 if (stmts)
3279 gimple_stmt_iterator gsi;
3280 gsi = gsi_start (stmts);
3281 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3283 gimple stmt = gsi_stmt (gsi);
3284 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3285 gimple_set_plf (stmt, NECESSARY, false);
3287 gsi_insert_seq_on_edge (pred, stmts);
3289 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3294 else if (eprime->kind == NAME)
3296 /* We may have to do a conversion because our value
3297 numbering can look through types in certain cases, but
3298 our IL requires all operands of a phi node have the same
3299 type. */
3300 tree name = PRE_EXPR_NAME (eprime);
3301 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3303 tree builtexpr;
3304 tree forcedexpr;
3305 builtexpr = fold_convert (type, name);
3306 forcedexpr = force_gimple_operand (builtexpr,
3307 &stmts, true,
3308 NULL);
3310 if (forcedexpr != name)
3312 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3313 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3316 if (stmts)
3318 gimple_stmt_iterator gsi;
3319 gsi = gsi_start (stmts);
3320 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3322 gimple stmt = gsi_stmt (gsi);
3323 VEC_safe_push (gimple, heap, inserted_exprs, stmt);
3324 gimple_set_plf (stmt, NECESSARY, false);
3326 gsi_insert_seq_on_edge (pred, stmts);
3328 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3332 /* If we didn't want a phi node, and we made insertions, we still have
3333 inserted new stuff, and thus return true. If we didn't want a phi node,
3334 and didn't make insertions, we haven't added anything new, so return
3335 false. */
3336 if (nophi && insertions)
3337 return true;
3338 else if (nophi && !insertions)
3339 return false;
3341 /* Now build a phi for the new variable. */
3342 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3344 prephitemp = create_tmp_var (type, "prephitmp");
3345 get_var_ann (prephitemp);
3348 temp = prephitemp;
3349 add_referenced_var (temp);
3351 if (TREE_CODE (type) == COMPLEX_TYPE
3352 || TREE_CODE (type) == VECTOR_TYPE)
3353 DECL_GIMPLE_REG_P (temp) = 1;
3354 phi = create_phi_node (temp, block);
3356 gimple_set_plf (phi, NECESSARY, false);
3357 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3358 VN_INFO (gimple_phi_result (phi))->value_id = val;
3359 VEC_safe_push (gimple, heap, inserted_exprs, phi);
3360 bitmap_set_bit (inserted_phi_names,
3361 SSA_NAME_VERSION (gimple_phi_result (phi)));
3362 FOR_EACH_EDGE (pred, ei, block->preds)
3364 pre_expr ae = avail[pred->src->index];
3365 gcc_assert (get_expr_type (ae) == type
3366 || useless_type_conversion_p (type, get_expr_type (ae)));
3367 if (ae->kind == CONSTANT)
3368 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3369 else
3370 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3371 UNKNOWN_LOCATION);
3374 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3375 add_to_value (val, newphi);
3377 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3378 this insertion, since we test for the existence of this value in PHI_GEN
3379 before proceeding with the partial redundancy checks in insert_aux.
3381 The value may exist in AVAIL_OUT, in particular, it could be represented
3382 by the expression we are trying to eliminate, in which case we want the
3383 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3384 inserted there.
3386 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3387 this block, because if it did, it would have existed in our dominator's
3388 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3391 bitmap_insert_into_set (PHI_GEN (block), newphi);
3392 bitmap_value_replace_in_set (AVAIL_OUT (block),
3393 newphi);
3394 bitmap_insert_into_set (NEW_SETS (block),
3395 newphi);
3397 if (dump_file && (dump_flags & TDF_DETAILS))
3399 fprintf (dump_file, "Created phi ");
3400 print_gimple_stmt (dump_file, phi, 0, 0);
3401 fprintf (dump_file, " in block %d\n", block->index);
3403 pre_stats.phis++;
3404 return true;
3409 /* Perform insertion of partially redundant values.
3410 For BLOCK, do the following:
3411 1. Propagate the NEW_SETS of the dominator into the current block.
3412 If the block has multiple predecessors,
3413 2a. Iterate over the ANTIC expressions for the block to see if
3414 any of them are partially redundant.
3415 2b. If so, insert them into the necessary predecessors to make
3416 the expression fully redundant.
3417 2c. Insert a new PHI merging the values of the predecessors.
3418 2d. Insert the new PHI, and the new expressions, into the
3419 NEW_SETS set.
3420 3. Recursively call ourselves on the dominator children of BLOCK.
3422 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3423 do_regular_insertion and do_partial_insertion.
3427 static bool
3428 do_regular_insertion (basic_block block, basic_block dom)
3430 bool new_stuff = false;
3431 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3432 pre_expr expr;
3433 int i;
3435 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3437 if (expr->kind != NAME)
3439 pre_expr *avail;
3440 unsigned int val;
3441 bool by_some = false;
3442 bool cant_insert = false;
3443 bool all_same = true;
3444 pre_expr first_s = NULL;
3445 edge pred;
3446 basic_block bprime;
3447 pre_expr eprime = NULL;
3448 edge_iterator ei;
3449 pre_expr edoubleprime = NULL;
3450 bool do_insertion = false;
3452 val = get_expr_value_id (expr);
3453 if (bitmap_set_contains_value (PHI_GEN (block), val))
3454 continue;
3455 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3457 if (dump_file && (dump_flags & TDF_DETAILS))
3458 fprintf (dump_file, "Found fully redundant value\n");
3459 continue;
3462 avail = XCNEWVEC (pre_expr, last_basic_block);
3463 FOR_EACH_EDGE (pred, ei, block->preds)
3465 unsigned int vprime;
3467 /* We should never run insertion for the exit block
3468 and so not come across fake pred edges. */
3469 gcc_assert (!(pred->flags & EDGE_FAKE));
3470 bprime = pred->src;
3471 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3472 bprime, block);
3474 /* eprime will generally only be NULL if the
3475 value of the expression, translated
3476 through the PHI for this predecessor, is
3477 undefined. If that is the case, we can't
3478 make the expression fully redundant,
3479 because its value is undefined along a
3480 predecessor path. We can thus break out
3481 early because it doesn't matter what the
3482 rest of the results are. */
3483 if (eprime == NULL)
3485 cant_insert = true;
3486 break;
3489 eprime = fully_constant_expression (eprime);
3490 vprime = get_expr_value_id (eprime);
3491 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3492 vprime, NULL);
3493 if (edoubleprime == NULL)
3495 avail[bprime->index] = eprime;
3496 all_same = false;
3498 else
3500 avail[bprime->index] = edoubleprime;
3501 by_some = true;
3502 /* We want to perform insertions to remove a redundancy on
3503 a path in the CFG we want to optimize for speed. */
3504 if (optimize_edge_for_speed_p (pred))
3505 do_insertion = true;
3506 if (first_s == NULL)
3507 first_s = edoubleprime;
3508 else if (!pre_expr_eq (first_s, edoubleprime))
3509 all_same = false;
3512 /* If we can insert it, it's not the same value
3513 already existing along every predecessor, and
3514 it's defined by some predecessor, it is
3515 partially redundant. */
3516 if (!cant_insert && !all_same && by_some && do_insertion
3517 && dbg_cnt (treepre_insert))
3519 if (insert_into_preds_of_block (block, get_expression_id (expr),
3520 avail))
3521 new_stuff = true;
3523 /* If all edges produce the same value and that value is
3524 an invariant, then the PHI has the same value on all
3525 edges. Note this. */
3526 else if (!cant_insert && all_same && eprime
3527 && (edoubleprime->kind == CONSTANT
3528 || edoubleprime->kind == NAME)
3529 && !value_id_constant_p (val))
3531 unsigned int j;
3532 bitmap_iterator bi;
3533 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3534 value_expressions, val);
3536 unsigned int new_val = get_expr_value_id (edoubleprime);
3537 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3539 pre_expr expr = expression_for_id (j);
3541 if (expr->kind == NAME)
3543 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3544 /* Just reset the value id and valnum so it is
3545 the same as the constant we have discovered. */
3546 if (edoubleprime->kind == CONSTANT)
3548 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3549 pre_stats.constified++;
3551 else
3552 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3553 info->value_id = new_val;
3557 free (avail);
3561 VEC_free (pre_expr, heap, exprs);
3562 return new_stuff;
3566 /* Perform insertion for partially anticipatable expressions. There
3567 is only one case we will perform insertion for these. This case is
3568 if the expression is partially anticipatable, and fully available.
3569 In this case, we know that putting it earlier will enable us to
3570 remove the later computation. */
3573 static bool
3574 do_partial_partial_insertion (basic_block block, basic_block dom)
3576 bool new_stuff = false;
3577 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3578 pre_expr expr;
3579 int i;
3581 for (i = 0; VEC_iterate (pre_expr, exprs, i, expr); i++)
3583 if (expr->kind != NAME)
3585 pre_expr *avail;
3586 unsigned int val;
3587 bool by_all = true;
3588 bool cant_insert = false;
3589 edge pred;
3590 basic_block bprime;
3591 pre_expr eprime = NULL;
3592 edge_iterator ei;
3594 val = get_expr_value_id (expr);
3595 if (bitmap_set_contains_value (PHI_GEN (block), val))
3596 continue;
3597 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3598 continue;
3600 avail = XCNEWVEC (pre_expr, last_basic_block);
3601 FOR_EACH_EDGE (pred, ei, block->preds)
3603 unsigned int vprime;
3604 pre_expr edoubleprime;
3606 /* We should never run insertion for the exit block
3607 and so not come across fake pred edges. */
3608 gcc_assert (!(pred->flags & EDGE_FAKE));
3609 bprime = pred->src;
3610 eprime = phi_translate (expr, ANTIC_IN (block),
3611 PA_IN (block),
3612 bprime, block);
3614 /* eprime will generally only be NULL if the
3615 value of the expression, translated
3616 through the PHI for this predecessor, is
3617 undefined. If that is the case, we can't
3618 make the expression fully redundant,
3619 because its value is undefined along a
3620 predecessor path. We can thus break out
3621 early because it doesn't matter what the
3622 rest of the results are. */
3623 if (eprime == NULL)
3625 cant_insert = true;
3626 break;
3629 eprime = fully_constant_expression (eprime);
3630 vprime = get_expr_value_id (eprime);
3631 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3632 vprime, NULL);
3633 if (edoubleprime == NULL)
3635 by_all = false;
3636 break;
3638 else
3639 avail[bprime->index] = edoubleprime;
3643 /* If we can insert it, it's not the same value
3644 already existing along every predecessor, and
3645 it's defined by some predecessor, it is
3646 partially redundant. */
3647 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3649 pre_stats.pa_insert++;
3650 if (insert_into_preds_of_block (block, get_expression_id (expr),
3651 avail))
3652 new_stuff = true;
3654 free (avail);
3658 VEC_free (pre_expr, heap, exprs);
3659 return new_stuff;
3662 static bool
3663 insert_aux (basic_block block)
3665 basic_block son;
3666 bool new_stuff = false;
3668 if (block)
3670 basic_block dom;
3671 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3672 if (dom)
3674 unsigned i;
3675 bitmap_iterator bi;
3676 bitmap_set_t newset = NEW_SETS (dom);
3677 if (newset)
3679 /* Note that we need to value_replace both NEW_SETS, and
3680 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3681 represented by some non-simple expression here that we want
3682 to replace it with. */
3683 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3685 pre_expr expr = expression_for_id (i);
3686 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3687 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3690 if (!single_pred_p (block))
3692 new_stuff |= do_regular_insertion (block, dom);
3693 if (do_partial_partial)
3694 new_stuff |= do_partial_partial_insertion (block, dom);
3698 for (son = first_dom_son (CDI_DOMINATORS, block);
3699 son;
3700 son = next_dom_son (CDI_DOMINATORS, son))
3702 new_stuff |= insert_aux (son);
3705 return new_stuff;
3708 /* Perform insertion of partially redundant values. */
3710 static void
3711 insert (void)
3713 bool new_stuff = true;
3714 basic_block bb;
3715 int num_iterations = 0;
3717 FOR_ALL_BB (bb)
3718 NEW_SETS (bb) = bitmap_set_new ();
3720 while (new_stuff)
3722 num_iterations++;
3723 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3725 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3729 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3731 static void
3732 add_to_exp_gen (basic_block block, tree op)
3734 if (!in_fre)
3736 pre_expr result;
3737 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3738 return;
3739 result = get_or_alloc_expr_for_name (op);
3740 bitmap_value_insert_into_set (EXP_GEN (block), result);
3744 /* Create value ids for PHI in BLOCK. */
3746 static void
3747 make_values_for_phi (gimple phi, basic_block block)
3749 tree result = gimple_phi_result (phi);
3751 /* We have no need for virtual phis, as they don't represent
3752 actual computations. */
3753 if (is_gimple_reg (result))
3755 pre_expr e = get_or_alloc_expr_for_name (result);
3756 add_to_value (get_expr_value_id (e), e);
3757 bitmap_insert_into_set (PHI_GEN (block), e);
3758 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3759 if (!in_fre)
3761 unsigned i;
3762 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3764 tree arg = gimple_phi_arg_def (phi, i);
3765 if (TREE_CODE (arg) == SSA_NAME)
3767 e = get_or_alloc_expr_for_name (arg);
3768 add_to_value (get_expr_value_id (e), e);
3775 /* Compute the AVAIL set for all basic blocks.
3777 This function performs value numbering of the statements in each basic
3778 block. The AVAIL sets are built from information we glean while doing
3779 this value numbering, since the AVAIL sets contain only one entry per
3780 value.
3782 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3783 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3785 static void
3786 compute_avail (void)
3789 basic_block block, son;
3790 basic_block *worklist;
3791 size_t sp = 0;
3792 unsigned i;
3794 /* We pretend that default definitions are defined in the entry block.
3795 This includes function arguments and the static chain decl. */
3796 for (i = 1; i < num_ssa_names; ++i)
3798 tree name = ssa_name (i);
3799 pre_expr e;
3800 if (!name
3801 || !SSA_NAME_IS_DEFAULT_DEF (name)
3802 || has_zero_uses (name)
3803 || !is_gimple_reg (name))
3804 continue;
3806 e = get_or_alloc_expr_for_name (name);
3807 add_to_value (get_expr_value_id (e), e);
3808 if (!in_fre)
3809 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3810 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3813 /* Allocate the worklist. */
3814 worklist = XNEWVEC (basic_block, n_basic_blocks);
3816 /* Seed the algorithm by putting the dominator children of the entry
3817 block on the worklist. */
3818 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3819 son;
3820 son = next_dom_son (CDI_DOMINATORS, son))
3821 worklist[sp++] = son;
3823 /* Loop until the worklist is empty. */
3824 while (sp)
3826 gimple_stmt_iterator gsi;
3827 gimple stmt;
3828 basic_block dom;
3829 unsigned int stmt_uid = 1;
3831 /* Pick a block from the worklist. */
3832 block = worklist[--sp];
3834 /* Initially, the set of available values in BLOCK is that of
3835 its immediate dominator. */
3836 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3837 if (dom)
3838 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3840 /* Generate values for PHI nodes. */
3841 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3842 make_values_for_phi (gsi_stmt (gsi), block);
3844 BB_MAY_NOTRETURN (block) = 0;
3846 /* Now compute value numbers and populate value sets with all
3847 the expressions computed in BLOCK. */
3848 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3850 ssa_op_iter iter;
3851 tree op;
3853 stmt = gsi_stmt (gsi);
3854 gimple_set_uid (stmt, stmt_uid++);
3856 /* Cache whether the basic-block has any non-visible side-effect
3857 or control flow.
3858 If this isn't a call or it is the last stmt in the
3859 basic-block then the CFG represents things correctly. */
3860 if (is_gimple_call (stmt)
3861 && !stmt_ends_bb_p (stmt))
3863 /* Non-looping const functions always return normally.
3864 Otherwise the call might not return or have side-effects
3865 that forbids hoisting possibly trapping expressions
3866 before it. */
3867 int flags = gimple_call_flags (stmt);
3868 if (!(flags & ECF_CONST)
3869 || (flags & ECF_LOOPING_CONST_OR_PURE))
3870 BB_MAY_NOTRETURN (block) = 1;
3873 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3875 pre_expr e = get_or_alloc_expr_for_name (op);
3877 add_to_value (get_expr_value_id (e), e);
3878 if (!in_fre)
3879 bitmap_insert_into_set (TMP_GEN (block), e);
3880 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3883 if (gimple_has_volatile_ops (stmt)
3884 || stmt_could_throw_p (stmt))
3885 continue;
3887 switch (gimple_code (stmt))
3889 case GIMPLE_RETURN:
3890 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3891 add_to_exp_gen (block, op);
3892 continue;
3894 case GIMPLE_CALL:
3896 vn_reference_t ref;
3897 unsigned int i;
3898 vn_reference_op_t vro;
3899 pre_expr result = NULL;
3900 VEC(vn_reference_op_s, heap) *ops = NULL;
3902 if (!can_value_number_call (stmt))
3903 continue;
3905 copy_reference_ops_from_call (stmt, &ops);
3906 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3907 gimple_expr_type (stmt),
3908 ops, &ref, false);
3909 VEC_free (vn_reference_op_s, heap, ops);
3910 if (!ref)
3911 continue;
3913 for (i = 0; VEC_iterate (vn_reference_op_s,
3914 ref->operands, i,
3915 vro); i++)
3917 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3918 add_to_exp_gen (block, vro->op0);
3919 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3920 add_to_exp_gen (block, vro->op1);
3921 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3922 add_to_exp_gen (block, vro->op2);
3924 result = (pre_expr) pool_alloc (pre_expr_pool);
3925 result->kind = REFERENCE;
3926 result->id = 0;
3927 PRE_EXPR_REFERENCE (result) = ref;
3929 get_or_alloc_expression_id (result);
3930 add_to_value (get_expr_value_id (result), result);
3931 if (!in_fre)
3932 bitmap_value_insert_into_set (EXP_GEN (block), result);
3933 continue;
3936 case GIMPLE_ASSIGN:
3938 pre_expr result = NULL;
3939 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
3941 case tcc_unary:
3942 case tcc_binary:
3943 case tcc_comparison:
3945 vn_nary_op_t nary;
3946 unsigned int i;
3948 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
3949 gimple_assign_rhs_code (stmt),
3950 gimple_expr_type (stmt),
3951 gimple_assign_rhs1 (stmt),
3952 gimple_assign_rhs2 (stmt),
3953 NULL_TREE, NULL_TREE, &nary);
3955 if (!nary)
3956 continue;
3958 for (i = 0; i < nary->length; i++)
3959 if (TREE_CODE (nary->op[i]) == SSA_NAME)
3960 add_to_exp_gen (block, nary->op[i]);
3962 result = (pre_expr) pool_alloc (pre_expr_pool);
3963 result->kind = NARY;
3964 result->id = 0;
3965 PRE_EXPR_NARY (result) = nary;
3966 break;
3969 case tcc_declaration:
3970 case tcc_reference:
3972 vn_reference_t ref;
3973 unsigned int i;
3974 vn_reference_op_t vro;
3976 vn_reference_lookup (gimple_assign_rhs1 (stmt),
3977 gimple_vuse (stmt),
3978 true, &ref);
3979 if (!ref)
3980 continue;
3982 for (i = 0; VEC_iterate (vn_reference_op_s,
3983 ref->operands, i,
3984 vro); i++)
3986 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
3987 add_to_exp_gen (block, vro->op0);
3988 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
3989 add_to_exp_gen (block, vro->op1);
3990 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
3991 add_to_exp_gen (block, vro->op2);
3993 result = (pre_expr) pool_alloc (pre_expr_pool);
3994 result->kind = REFERENCE;
3995 result->id = 0;
3996 PRE_EXPR_REFERENCE (result) = ref;
3997 break;
4000 default:
4001 /* For any other statement that we don't
4002 recognize, simply add all referenced
4003 SSA_NAMEs to EXP_GEN. */
4004 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4005 add_to_exp_gen (block, op);
4006 continue;
4009 get_or_alloc_expression_id (result);
4010 add_to_value (get_expr_value_id (result), result);
4011 if (!in_fre)
4012 bitmap_value_insert_into_set (EXP_GEN (block), result);
4014 continue;
4016 default:
4017 break;
4021 /* Put the dominator children of BLOCK on the worklist of blocks
4022 to compute available sets for. */
4023 for (son = first_dom_son (CDI_DOMINATORS, block);
4024 son;
4025 son = next_dom_son (CDI_DOMINATORS, son))
4026 worklist[sp++] = son;
4029 free (worklist);
4032 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4033 than the available expressions for it. The insertion point is
4034 right before the first use in STMT. Returns the SSA_NAME that should
4035 be used for replacement. */
4037 static tree
4038 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4040 basic_block bb = gimple_bb (stmt);
4041 gimple_stmt_iterator gsi;
4042 gimple_seq stmts = NULL;
4043 tree expr;
4044 pre_expr e;
4046 /* First create a value expression from the expression we want
4047 to insert and associate it with the value handle for SSA_VN. */
4048 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4049 if (e == NULL)
4050 return NULL_TREE;
4052 /* Then use create_expression_by_pieces to generate a valid
4053 expression to insert at this point of the IL stream. */
4054 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4055 if (expr == NULL_TREE)
4056 return NULL_TREE;
4057 gsi = gsi_for_stmt (stmt);
4058 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4060 return expr;
4063 /* Eliminate fully redundant computations. */
4065 static unsigned int
4066 eliminate (void)
4068 VEC (gimple, heap) *to_remove = NULL;
4069 basic_block b;
4070 unsigned int todo = 0;
4071 gimple_stmt_iterator gsi;
4072 gimple stmt;
4073 unsigned i;
4075 FOR_EACH_BB (b)
4077 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4079 stmt = gsi_stmt (gsi);
4081 /* Lookup the RHS of the expression, see if we have an
4082 available computation for it. If so, replace the RHS with
4083 the available computation. */
4084 if (gimple_has_lhs (stmt)
4085 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4086 && !gimple_assign_ssa_name_copy_p (stmt)
4087 && (!gimple_assign_single_p (stmt)
4088 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4089 && !gimple_has_volatile_ops (stmt)
4090 && !has_zero_uses (gimple_get_lhs (stmt)))
4092 tree lhs = gimple_get_lhs (stmt);
4093 tree rhs = NULL_TREE;
4094 tree sprime = NULL;
4095 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4096 pre_expr sprimeexpr;
4098 if (gimple_assign_single_p (stmt))
4099 rhs = gimple_assign_rhs1 (stmt);
4101 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4102 get_expr_value_id (lhsexpr),
4103 NULL);
4105 if (sprimeexpr)
4107 if (sprimeexpr->kind == CONSTANT)
4108 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4109 else if (sprimeexpr->kind == NAME)
4110 sprime = PRE_EXPR_NAME (sprimeexpr);
4111 else
4112 gcc_unreachable ();
4115 /* If there is no existing leader but SCCVN knows this
4116 value is constant, use that constant. */
4117 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4119 sprime = VN_INFO (lhs)->valnum;
4120 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4121 TREE_TYPE (sprime)))
4122 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4124 if (dump_file && (dump_flags & TDF_DETAILS))
4126 fprintf (dump_file, "Replaced ");
4127 print_gimple_expr (dump_file, stmt, 0, 0);
4128 fprintf (dump_file, " with ");
4129 print_generic_expr (dump_file, sprime, 0);
4130 fprintf (dump_file, " in ");
4131 print_gimple_stmt (dump_file, stmt, 0, 0);
4133 pre_stats.eliminations++;
4134 propagate_tree_value_into_stmt (&gsi, sprime);
4135 stmt = gsi_stmt (gsi);
4136 update_stmt (stmt);
4137 continue;
4140 /* If there is no existing usable leader but SCCVN thinks
4141 it has an expression it wants to use as replacement,
4142 insert that. */
4143 if (!sprime || sprime == lhs)
4145 tree val = VN_INFO (lhs)->valnum;
4146 if (val != VN_TOP
4147 && TREE_CODE (val) == SSA_NAME
4148 && VN_INFO (val)->needs_insertion
4149 && can_PRE_operation (vn_get_expr_for (val)))
4150 sprime = do_SCCVN_insertion (stmt, val);
4152 if (sprime
4153 && sprime != lhs
4154 && (rhs == NULL_TREE
4155 || TREE_CODE (rhs) != SSA_NAME
4156 || may_propagate_copy (rhs, sprime)))
4158 gcc_assert (sprime != rhs);
4160 if (dump_file && (dump_flags & TDF_DETAILS))
4162 fprintf (dump_file, "Replaced ");
4163 print_gimple_expr (dump_file, stmt, 0, 0);
4164 fprintf (dump_file, " with ");
4165 print_generic_expr (dump_file, sprime, 0);
4166 fprintf (dump_file, " in ");
4167 print_gimple_stmt (dump_file, stmt, 0, 0);
4170 if (TREE_CODE (sprime) == SSA_NAME)
4171 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4172 NECESSARY, true);
4173 /* We need to make sure the new and old types actually match,
4174 which may require adding a simple cast, which fold_convert
4175 will do for us. */
4176 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4177 && !useless_type_conversion_p (gimple_expr_type (stmt),
4178 TREE_TYPE (sprime)))
4179 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4181 pre_stats.eliminations++;
4182 propagate_tree_value_into_stmt (&gsi, sprime);
4183 stmt = gsi_stmt (gsi);
4184 update_stmt (stmt);
4186 /* If we removed EH side effects from the statement, clean
4187 its EH information. */
4188 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4190 bitmap_set_bit (need_eh_cleanup,
4191 gimple_bb (stmt)->index);
4192 if (dump_file && (dump_flags & TDF_DETAILS))
4193 fprintf (dump_file, " Removed EH side effects.\n");
4197 /* If the statement is a scalar store, see if the expression
4198 has the same value number as its rhs. If so, the store is
4199 dead. */
4200 else if (gimple_assign_single_p (stmt)
4201 && !is_gimple_reg (gimple_assign_lhs (stmt))
4202 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4203 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4205 tree rhs = gimple_assign_rhs1 (stmt);
4206 tree val;
4207 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4208 gimple_vuse (stmt), true, NULL);
4209 if (TREE_CODE (rhs) == SSA_NAME)
4210 rhs = VN_INFO (rhs)->valnum;
4211 if (val
4212 && operand_equal_p (val, rhs, 0))
4214 if (dump_file && (dump_flags & TDF_DETAILS))
4216 fprintf (dump_file, "Deleted redundant store ");
4217 print_gimple_stmt (dump_file, stmt, 0, 0);
4220 /* Queue stmt for removal. */
4221 VEC_safe_push (gimple, heap, to_remove, stmt);
4224 /* Visit COND_EXPRs and fold the comparison with the
4225 available value-numbers. */
4226 else if (gimple_code (stmt) == GIMPLE_COND)
4228 tree op0 = gimple_cond_lhs (stmt);
4229 tree op1 = gimple_cond_rhs (stmt);
4230 tree result;
4232 if (TREE_CODE (op0) == SSA_NAME)
4233 op0 = VN_INFO (op0)->valnum;
4234 if (TREE_CODE (op1) == SSA_NAME)
4235 op1 = VN_INFO (op1)->valnum;
4236 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4237 op0, op1);
4238 if (result && TREE_CODE (result) == INTEGER_CST)
4240 if (integer_zerop (result))
4241 gimple_cond_make_false (stmt);
4242 else
4243 gimple_cond_make_true (stmt);
4244 update_stmt (stmt);
4245 todo = TODO_cleanup_cfg;
4248 /* Visit indirect calls and turn them into direct calls if
4249 possible. */
4250 if (gimple_code (stmt) == GIMPLE_CALL
4251 && TREE_CODE (gimple_call_fn (stmt)) == SSA_NAME)
4253 tree fn = VN_INFO (gimple_call_fn (stmt))->valnum;
4254 if (TREE_CODE (fn) == ADDR_EXPR
4255 && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
4257 if (dump_file && (dump_flags & TDF_DETAILS))
4259 fprintf (dump_file, "Replacing call target with ");
4260 print_generic_expr (dump_file, fn, 0);
4261 fprintf (dump_file, " in ");
4262 print_gimple_stmt (dump_file, stmt, 0, 0);
4265 gimple_call_set_fn (stmt, fn);
4266 update_stmt (stmt);
4267 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4269 bitmap_set_bit (need_eh_cleanup,
4270 gimple_bb (stmt)->index);
4271 if (dump_file && (dump_flags & TDF_DETAILS))
4272 fprintf (dump_file, " Removed EH side effects.\n");
4275 /* Changing an indirect call to a direct call may
4276 have exposed different semantics. This may
4277 require an SSA update. */
4278 todo |= TODO_update_ssa_only_virtuals;
4283 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4285 gimple stmt, phi = gsi_stmt (gsi);
4286 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4287 pre_expr sprimeexpr, resexpr;
4288 gimple_stmt_iterator gsi2;
4290 /* We want to perform redundant PHI elimination. Do so by
4291 replacing the PHI with a single copy if possible.
4292 Do not touch inserted, single-argument or virtual PHIs. */
4293 if (gimple_phi_num_args (phi) == 1
4294 || !is_gimple_reg (res)
4295 || bitmap_bit_p (inserted_phi_names, SSA_NAME_VERSION (res)))
4297 gsi_next (&gsi);
4298 continue;
4301 resexpr = get_or_alloc_expr_for_name (res);
4302 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4303 get_expr_value_id (resexpr), NULL);
4304 if (sprimeexpr)
4306 if (sprimeexpr->kind == CONSTANT)
4307 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4308 else if (sprimeexpr->kind == NAME)
4309 sprime = PRE_EXPR_NAME (sprimeexpr);
4310 else
4311 gcc_unreachable ();
4313 if (!sprimeexpr
4314 || sprime == res)
4316 gsi_next (&gsi);
4317 continue;
4320 if (dump_file && (dump_flags & TDF_DETAILS))
4322 fprintf (dump_file, "Replaced redundant PHI node defining ");
4323 print_generic_expr (dump_file, res, 0);
4324 fprintf (dump_file, " with ");
4325 print_generic_expr (dump_file, sprime, 0);
4326 fprintf (dump_file, "\n");
4329 remove_phi_node (&gsi, false);
4331 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4332 sprime = fold_convert (TREE_TYPE (res), sprime);
4333 stmt = gimple_build_assign (res, sprime);
4334 SSA_NAME_DEF_STMT (res) = stmt;
4335 if (TREE_CODE (sprime) == SSA_NAME)
4336 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4337 NECESSARY, true);
4338 gsi2 = gsi_after_labels (b);
4339 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4340 /* Queue the copy for eventual removal. */
4341 VEC_safe_push (gimple, heap, to_remove, stmt);
4342 pre_stats.eliminations++;
4346 /* We cannot remove stmts during BB walk, especially not release SSA
4347 names there as this confuses the VN machinery. The stmts ending
4348 up in to_remove are either stores or simple copies. */
4349 for (i = 0; VEC_iterate (gimple, to_remove, i, stmt); ++i)
4351 tree lhs = gimple_assign_lhs (stmt);
4352 use_operand_p use_p;
4353 gimple use_stmt;
4355 /* If there is a single use only, propagate the equivalency
4356 instead of keeping the copy. */
4357 if (TREE_CODE (lhs) == SSA_NAME
4358 && single_imm_use (lhs, &use_p, &use_stmt)
4359 && may_propagate_copy (USE_FROM_PTR (use_p),
4360 gimple_assign_rhs1 (stmt)))
4362 SET_USE (use_p, gimple_assign_rhs1 (stmt));
4363 update_stmt (use_stmt);
4366 /* If this is a store or a now unused copy, remove it. */
4367 if (TREE_CODE (lhs) != SSA_NAME
4368 || has_zero_uses (lhs))
4370 gsi = gsi_for_stmt (stmt);
4371 unlink_stmt_vdef (stmt);
4372 gsi_remove (&gsi, true);
4373 release_defs (stmt);
4376 VEC_free (gimple, heap, to_remove);
4378 return todo;
4381 /* Borrow a bit of tree-ssa-dce.c for the moment.
4382 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4383 this may be a bit faster, and we may want critical edges kept split. */
4385 /* If OP's defining statement has not already been determined to be necessary,
4386 mark that statement necessary. Return the stmt, if it is newly
4387 necessary. */
4389 static inline gimple
4390 mark_operand_necessary (tree op)
4392 gimple stmt;
4394 gcc_assert (op);
4396 if (TREE_CODE (op) != SSA_NAME)
4397 return NULL;
4399 stmt = SSA_NAME_DEF_STMT (op);
4400 gcc_assert (stmt);
4402 if (gimple_plf (stmt, NECESSARY)
4403 || gimple_nop_p (stmt))
4404 return NULL;
4406 gimple_set_plf (stmt, NECESSARY, true);
4407 return stmt;
4410 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4411 to insert PHI nodes sometimes, and because value numbering of casts isn't
4412 perfect, we sometimes end up inserting dead code. This simple DCE-like
4413 pass removes any insertions we made that weren't actually used. */
4415 static void
4416 remove_dead_inserted_code (void)
4418 VEC(gimple,heap) *worklist = NULL;
4419 int i;
4420 gimple t;
4422 worklist = VEC_alloc (gimple, heap, VEC_length (gimple, inserted_exprs));
4423 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4425 if (gimple_plf (t, NECESSARY))
4426 VEC_quick_push (gimple, worklist, t);
4428 while (VEC_length (gimple, worklist) > 0)
4430 t = VEC_pop (gimple, worklist);
4432 /* PHI nodes are somewhat special in that each PHI alternative has
4433 data and control dependencies. All the statements feeding the
4434 PHI node's arguments are always necessary. */
4435 if (gimple_code (t) == GIMPLE_PHI)
4437 unsigned k;
4439 VEC_reserve (gimple, heap, worklist, gimple_phi_num_args (t));
4440 for (k = 0; k < gimple_phi_num_args (t); k++)
4442 tree arg = PHI_ARG_DEF (t, k);
4443 if (TREE_CODE (arg) == SSA_NAME)
4445 gimple n = mark_operand_necessary (arg);
4446 if (n)
4447 VEC_quick_push (gimple, worklist, n);
4451 else
4453 /* Propagate through the operands. Examine all the USE, VUSE and
4454 VDEF operands in this statement. Mark all the statements
4455 which feed this statement's uses as necessary. */
4456 ssa_op_iter iter;
4457 tree use;
4459 /* The operands of VDEF expressions are also needed as they
4460 represent potential definitions that may reach this
4461 statement (VDEF operands allow us to follow def-def
4462 links). */
4464 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4466 gimple n = mark_operand_necessary (use);
4467 if (n)
4468 VEC_safe_push (gimple, heap, worklist, n);
4473 for (i = 0; VEC_iterate (gimple, inserted_exprs, i, t); i++)
4475 if (!gimple_plf (t, NECESSARY))
4477 gimple_stmt_iterator gsi;
4479 if (dump_file && (dump_flags & TDF_DETAILS))
4481 fprintf (dump_file, "Removing unnecessary insertion:");
4482 print_gimple_stmt (dump_file, t, 0, 0);
4485 gsi = gsi_for_stmt (t);
4486 if (gimple_code (t) == GIMPLE_PHI)
4487 remove_phi_node (&gsi, true);
4488 else
4490 gsi_remove (&gsi, true);
4491 release_defs (t);
4495 VEC_free (gimple, heap, worklist);
4498 /* Initialize data structures used by PRE. */
4500 static void
4501 init_pre (bool do_fre)
4503 basic_block bb;
4505 next_expression_id = 1;
4506 expressions = NULL;
4507 VEC_safe_push (pre_expr, heap, expressions, NULL);
4508 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4509 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4510 get_max_value_id() + 1);
4511 name_to_id = NULL;
4513 in_fre = do_fre;
4515 inserted_exprs = NULL;
4516 need_creation = NULL;
4517 pretemp = NULL_TREE;
4518 storetemp = NULL_TREE;
4519 prephitemp = NULL_TREE;
4521 connect_infinite_loops_to_exit ();
4522 memset (&pre_stats, 0, sizeof (pre_stats));
4525 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4526 post_order_compute (postorder, false, false);
4528 FOR_ALL_BB (bb)
4529 bb->aux = XCNEWVEC (struct bb_bitmap_sets, 1);
4531 calculate_dominance_info (CDI_POST_DOMINATORS);
4532 calculate_dominance_info (CDI_DOMINATORS);
4534 bitmap_obstack_initialize (&grand_bitmap_obstack);
4535 inserted_phi_names = BITMAP_ALLOC (&grand_bitmap_obstack);
4536 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4537 expr_pred_trans_eq, free);
4538 expression_to_id = htab_create (num_ssa_names * 3,
4539 pre_expr_hash,
4540 pre_expr_eq, NULL);
4541 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4542 sizeof (struct bitmap_set), 30);
4543 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4544 sizeof (struct pre_expr_d), 30);
4545 FOR_ALL_BB (bb)
4547 EXP_GEN (bb) = bitmap_set_new ();
4548 PHI_GEN (bb) = bitmap_set_new ();
4549 TMP_GEN (bb) = bitmap_set_new ();
4550 AVAIL_OUT (bb) = bitmap_set_new ();
4553 need_eh_cleanup = BITMAP_ALLOC (NULL);
4557 /* Deallocate data structures used by PRE. */
4559 static void
4560 fini_pre (bool do_fre)
4562 basic_block bb;
4564 free (postorder);
4565 VEC_free (bitmap_set_t, heap, value_expressions);
4566 VEC_free (gimple, heap, inserted_exprs);
4567 VEC_free (gimple, heap, need_creation);
4568 bitmap_obstack_release (&grand_bitmap_obstack);
4569 free_alloc_pool (bitmap_set_pool);
4570 free_alloc_pool (pre_expr_pool);
4571 htab_delete (phi_translate_table);
4572 htab_delete (expression_to_id);
4573 VEC_free (unsigned, heap, name_to_id);
4575 FOR_ALL_BB (bb)
4577 free (bb->aux);
4578 bb->aux = NULL;
4581 free_dominance_info (CDI_POST_DOMINATORS);
4583 if (!bitmap_empty_p (need_eh_cleanup))
4585 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4586 cleanup_tree_cfg ();
4589 BITMAP_FREE (need_eh_cleanup);
4591 if (!do_fre)
4592 loop_optimizer_finalize ();
4595 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4596 only wants to do full redundancy elimination. */
4598 static unsigned int
4599 execute_pre (bool do_fre)
4601 unsigned int todo = 0;
4603 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4605 /* This has to happen before SCCVN runs because
4606 loop_optimizer_init may create new phis, etc. */
4607 if (!do_fre)
4608 loop_optimizer_init (LOOPS_NORMAL);
4610 if (!run_scc_vn (do_fre))
4612 if (!do_fre)
4614 remove_dead_inserted_code ();
4615 loop_optimizer_finalize ();
4618 return 0;
4620 init_pre (do_fre);
4621 scev_initialize ();
4624 /* Collect and value number expressions computed in each basic block. */
4625 compute_avail ();
4627 if (dump_file && (dump_flags & TDF_DETAILS))
4629 basic_block bb;
4631 FOR_ALL_BB (bb)
4633 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4634 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4635 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4636 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4640 /* Insert can get quite slow on an incredibly large number of basic
4641 blocks due to some quadratic behavior. Until this behavior is
4642 fixed, don't run it when he have an incredibly large number of
4643 bb's. If we aren't going to run insert, there is no point in
4644 computing ANTIC, either, even though it's plenty fast. */
4645 if (!do_fre && n_basic_blocks < 4000)
4647 compute_antic ();
4648 insert ();
4651 /* Remove all the redundant expressions. */
4652 todo |= eliminate ();
4654 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4655 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4656 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4657 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4658 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4660 /* Make sure to remove fake edges before committing our inserts.
4661 This makes sure we don't end up with extra critical edges that
4662 we would need to split. */
4663 remove_fake_exit_edges ();
4664 gsi_commit_edge_inserts ();
4666 clear_expression_ids ();
4667 free_scc_vn ();
4668 if (!do_fre)
4669 remove_dead_inserted_code ();
4671 scev_finalize ();
4672 fini_pre (do_fre);
4674 return todo;
4677 /* Gate and execute functions for PRE. */
4679 static unsigned int
4680 do_pre (void)
4682 return execute_pre (false);
4685 static bool
4686 gate_pre (void)
4688 return flag_tree_pre != 0;
4691 struct gimple_opt_pass pass_pre =
4694 GIMPLE_PASS,
4695 "pre", /* name */
4696 gate_pre, /* gate */
4697 do_pre, /* execute */
4698 NULL, /* sub */
4699 NULL, /* next */
4700 0, /* static_pass_number */
4701 TV_TREE_PRE, /* tv_id */
4702 PROP_no_crit_edges | PROP_cfg
4703 | PROP_ssa, /* properties_required */
4704 0, /* properties_provided */
4705 0, /* properties_destroyed */
4706 TODO_rebuild_alias, /* todo_flags_start */
4707 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4708 | TODO_verify_ssa /* todo_flags_finish */
4713 /* Gate and execute functions for FRE. */
4715 static unsigned int
4716 execute_fre (void)
4718 return execute_pre (true);
4721 static bool
4722 gate_fre (void)
4724 return flag_tree_fre != 0;
4727 struct gimple_opt_pass pass_fre =
4730 GIMPLE_PASS,
4731 "fre", /* name */
4732 gate_fre, /* gate */
4733 execute_fre, /* execute */
4734 NULL, /* sub */
4735 NULL, /* next */
4736 0, /* static_pass_number */
4737 TV_TREE_FRE, /* tv_id */
4738 PROP_cfg | PROP_ssa, /* properties_required */
4739 0, /* properties_provided */
4740 0, /* properties_destroyed */
4741 0, /* todo_flags_start */
4742 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */