new tests
[official-gcc.git] / gcc / tree-ssa-pre.c
blob51d9dcde9af3b760588ab7dcb6e939184ecb433d
1 /* SSA-PRE for trees.
2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
5 <stevenb@suse.de>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
12 any later version.
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 #include "config.h"
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "tree.h"
28 #include "basic-block.h"
29 #include "tree-pretty-print.h"
30 #include "gimple-pretty-print.h"
31 #include "tree-inline.h"
32 #include "tree-flow.h"
33 #include "gimple.h"
34 #include "tree-dump.h"
35 #include "timevar.h"
36 #include "fibheap.h"
37 #include "hashtab.h"
38 #include "tree-iterator.h"
39 #include "alloc-pool.h"
40 #include "obstack.h"
41 #include "tree-pass.h"
42 #include "flags.h"
43 #include "bitmap.h"
44 #include "langhooks.h"
45 #include "cfgloop.h"
46 #include "tree-ssa-sccvn.h"
47 #include "tree-scalar-evolution.h"
48 #include "params.h"
49 #include "dbgcnt.h"
51 /* TODO:
53 1. Avail sets can be shared by making an avail_find_leader that
54 walks up the dominator tree and looks in those avail sets.
55 This might affect code optimality, it's unclear right now.
56 2. Strength reduction can be performed by anticipating expressions
57 we can repair later on.
58 3. We can do back-substitution or smarter value numbering to catch
59 commutative expressions split up over multiple statements.
62 /* For ease of terminology, "expression node" in the below refers to
63 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
64 represent the actual statement containing the expressions we care about,
65 and we cache the value number by putting it in the expression. */
67 /* Basic algorithm
69 First we walk the statements to generate the AVAIL sets, the
70 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
71 generation of values/expressions by a given block. We use them
72 when computing the ANTIC sets. The AVAIL sets consist of
73 SSA_NAME's that represent values, so we know what values are
74 available in what blocks. AVAIL is a forward dataflow problem. In
75 SSA, values are never killed, so we don't need a kill set, or a
76 fixpoint iteration, in order to calculate the AVAIL sets. In
77 traditional parlance, AVAIL sets tell us the downsafety of the
78 expressions/values.
80 Next, we generate the ANTIC sets. These sets represent the
81 anticipatable expressions. ANTIC is a backwards dataflow
82 problem. An expression is anticipatable in a given block if it could
83 be generated in that block. This means that if we had to perform
84 an insertion in that block, of the value of that expression, we
85 could. Calculating the ANTIC sets requires phi translation of
86 expressions, because the flow goes backwards through phis. We must
87 iterate to a fixpoint of the ANTIC sets, because we have a kill
88 set. Even in SSA form, values are not live over the entire
89 function, only from their definition point onwards. So we have to
90 remove values from the ANTIC set once we go past the definition
91 point of the leaders that make them up.
92 compute_antic/compute_antic_aux performs this computation.
94 Third, we perform insertions to make partially redundant
95 expressions fully redundant.
97 An expression is partially redundant (excluding partial
98 anticipation) if:
100 1. It is AVAIL in some, but not all, of the predecessors of a
101 given block.
102 2. It is ANTIC in all the predecessors.
104 In order to make it fully redundant, we insert the expression into
105 the predecessors where it is not available, but is ANTIC.
107 For the partial anticipation case, we only perform insertion if it
108 is partially anticipated in some block, and fully available in all
109 of the predecessors.
111 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
112 performs these steps.
114 Fourth, we eliminate fully redundant expressions.
115 This is a simple statement walk that replaces redundant
116 calculations with the now available values. */
118 /* Representations of value numbers:
120 Value numbers are represented by a representative SSA_NAME. We
121 will create fake SSA_NAME's in situations where we need a
122 representative but do not have one (because it is a complex
123 expression). In order to facilitate storing the value numbers in
124 bitmaps, and keep the number of wasted SSA_NAME's down, we also
125 associate a value_id with each value number, and create full blown
126 ssa_name's only where we actually need them (IE in operands of
127 existing expressions).
129 Theoretically you could replace all the value_id's with
130 SSA_NAME_VERSION, but this would allocate a large number of
131 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
132 It would also require an additional indirection at each point we
133 use the value id. */
135 /* Representation of expressions on value numbers:
137 Expressions consisting of value numbers are represented the same
138 way as our VN internally represents them, with an additional
139 "pre_expr" wrapping around them in order to facilitate storing all
140 of the expressions in the same sets. */
142 /* Representation of sets:
144 The dataflow sets do not need to be sorted in any particular order
145 for the majority of their lifetime, are simply represented as two
146 bitmaps, one that keeps track of values present in the set, and one
147 that keeps track of expressions present in the set.
149 When we need them in topological order, we produce it on demand by
150 transforming the bitmap into an array and sorting it into topo
151 order. */
153 /* Type of expression, used to know which member of the PRE_EXPR union
154 is valid. */
156 enum pre_expr_kind
158 NAME,
159 NARY,
160 REFERENCE,
161 CONSTANT
164 typedef union pre_expr_union_d
166 tree name;
167 tree constant;
168 vn_nary_op_t nary;
169 vn_reference_t reference;
170 } pre_expr_union;
172 typedef struct pre_expr_d
174 enum pre_expr_kind kind;
175 unsigned int id;
176 pre_expr_union u;
177 } *pre_expr;
179 #define PRE_EXPR_NAME(e) (e)->u.name
180 #define PRE_EXPR_NARY(e) (e)->u.nary
181 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
182 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
184 static int
185 pre_expr_eq (const void *p1, const void *p2)
187 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
188 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
190 if (e1->kind != e2->kind)
191 return false;
193 switch (e1->kind)
195 case CONSTANT:
196 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
197 PRE_EXPR_CONSTANT (e2));
198 case NAME:
199 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
200 case NARY:
201 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
202 case REFERENCE:
203 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
204 PRE_EXPR_REFERENCE (e2));
205 default:
206 gcc_unreachable ();
210 static hashval_t
211 pre_expr_hash (const void *p1)
213 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
214 switch (e->kind)
216 case CONSTANT:
217 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
218 case NAME:
219 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
220 case NARY:
221 return PRE_EXPR_NARY (e)->hashcode;
222 case REFERENCE:
223 return PRE_EXPR_REFERENCE (e)->hashcode;
224 default:
225 gcc_unreachable ();
230 /* Next global expression id number. */
231 static unsigned int next_expression_id;
233 /* Mapping from expression to id number we can use in bitmap sets. */
234 DEF_VEC_P (pre_expr);
235 DEF_VEC_ALLOC_P (pre_expr, heap);
236 static VEC(pre_expr, heap) *expressions;
237 static htab_t expression_to_id;
238 static VEC(unsigned, heap) *name_to_id;
240 /* Allocate an expression id for EXPR. */
242 static inline unsigned int
243 alloc_expression_id (pre_expr expr)
245 void **slot;
246 /* Make sure we won't overflow. */
247 gcc_assert (next_expression_id + 1 > next_expression_id);
248 expr->id = next_expression_id++;
249 VEC_safe_push (pre_expr, heap, expressions, expr);
250 if (expr->kind == NAME)
252 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
253 /* VEC_safe_grow_cleared allocates no headroom. Avoid frequent
254 re-allocations by using VEC_reserve upfront. There is no
255 VEC_quick_grow_cleared unfortunately. */
256 VEC_reserve (unsigned, heap, name_to_id, num_ssa_names);
257 VEC_safe_grow_cleared (unsigned, heap, name_to_id, num_ssa_names);
258 gcc_assert (VEC_index (unsigned, name_to_id, version) == 0);
259 VEC_replace (unsigned, name_to_id, version, expr->id);
261 else
263 slot = htab_find_slot (expression_to_id, expr, INSERT);
264 gcc_assert (!*slot);
265 *slot = expr;
267 return next_expression_id - 1;
270 /* Return the expression id for tree EXPR. */
272 static inline unsigned int
273 get_expression_id (const pre_expr expr)
275 return expr->id;
278 static inline unsigned int
279 lookup_expression_id (const pre_expr expr)
281 void **slot;
283 if (expr->kind == NAME)
285 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
286 if (VEC_length (unsigned, name_to_id) <= version)
287 return 0;
288 return VEC_index (unsigned, name_to_id, version);
290 else
292 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
293 if (!slot)
294 return 0;
295 return ((pre_expr)*slot)->id;
299 /* Return the existing expression id for EXPR, or create one if one
300 does not exist yet. */
302 static inline unsigned int
303 get_or_alloc_expression_id (pre_expr expr)
305 unsigned int id = lookup_expression_id (expr);
306 if (id == 0)
307 return alloc_expression_id (expr);
308 return expr->id = id;
311 /* Return the expression that has expression id ID */
313 static inline pre_expr
314 expression_for_id (unsigned int id)
316 return VEC_index (pre_expr, expressions, id);
319 /* Free the expression id field in all of our expressions,
320 and then destroy the expressions array. */
322 static void
323 clear_expression_ids (void)
325 VEC_free (pre_expr, heap, expressions);
328 static alloc_pool pre_expr_pool;
330 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
332 static pre_expr
333 get_or_alloc_expr_for_name (tree name)
335 struct pre_expr_d expr;
336 pre_expr result;
337 unsigned int result_id;
339 expr.kind = NAME;
340 expr.id = 0;
341 PRE_EXPR_NAME (&expr) = name;
342 result_id = lookup_expression_id (&expr);
343 if (result_id != 0)
344 return expression_for_id (result_id);
346 result = (pre_expr) pool_alloc (pre_expr_pool);
347 result->kind = NAME;
348 PRE_EXPR_NAME (result) = name;
349 alloc_expression_id (result);
350 return result;
353 static bool in_fre = false;
355 /* An unordered bitmap set. One bitmap tracks values, the other,
356 expressions. */
357 typedef struct bitmap_set
359 bitmap_head expressions;
360 bitmap_head values;
361 } *bitmap_set_t;
363 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
364 EXECUTE_IF_SET_IN_BITMAP(&(set)->expressions, 0, (id), (bi))
366 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
367 EXECUTE_IF_SET_IN_BITMAP(&(set)->values, 0, (id), (bi))
369 /* Mapping from value id to expressions with that value_id. */
370 DEF_VEC_P (bitmap_set_t);
371 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
372 static VEC(bitmap_set_t, heap) *value_expressions;
374 /* Sets that we need to keep track of. */
375 typedef struct bb_bitmap_sets
377 /* The EXP_GEN set, which represents expressions/values generated in
378 a basic block. */
379 bitmap_set_t exp_gen;
381 /* The PHI_GEN set, which represents PHI results generated in a
382 basic block. */
383 bitmap_set_t phi_gen;
385 /* The TMP_GEN set, which represents results/temporaries generated
386 in a basic block. IE the LHS of an expression. */
387 bitmap_set_t tmp_gen;
389 /* The AVAIL_OUT set, which represents which values are available in
390 a given basic block. */
391 bitmap_set_t avail_out;
393 /* The ANTIC_IN set, which represents which values are anticipatable
394 in a given basic block. */
395 bitmap_set_t antic_in;
397 /* The PA_IN set, which represents which values are
398 partially anticipatable in a given basic block. */
399 bitmap_set_t pa_in;
401 /* The NEW_SETS set, which is used during insertion to augment the
402 AVAIL_OUT set of blocks with the new insertions performed during
403 the current iteration. */
404 bitmap_set_t new_sets;
406 /* A cache for value_dies_in_block_x. */
407 bitmap expr_dies;
409 /* True if we have visited this block during ANTIC calculation. */
410 unsigned int visited : 1;
412 /* True we have deferred processing this block during ANTIC
413 calculation until its successor is processed. */
414 unsigned int deferred : 1;
416 /* True when the block contains a call that might not return. */
417 unsigned int contains_may_not_return_call : 1;
418 } *bb_value_sets_t;
420 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
421 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
422 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
423 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
424 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
425 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
426 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
427 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
428 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
429 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
430 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
433 /* Basic block list in postorder. */
434 static int *postorder;
436 /* This structure is used to keep track of statistics on what
437 optimization PRE was able to perform. */
438 static struct
440 /* The number of RHS computations eliminated by PRE. */
441 int eliminations;
443 /* The number of new expressions/temporaries generated by PRE. */
444 int insertions;
446 /* The number of inserts found due to partial anticipation */
447 int pa_insert;
449 /* The number of new PHI nodes added by PRE. */
450 int phis;
452 /* The number of values found constant. */
453 int constified;
455 } pre_stats;
457 static bool do_partial_partial;
458 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
459 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
460 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
461 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
462 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
463 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
464 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
465 unsigned int, bool);
466 static bitmap_set_t bitmap_set_new (void);
467 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
468 gimple, tree);
469 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
470 gimple);
471 static unsigned int get_expr_value_id (pre_expr);
473 /* We can add and remove elements and entries to and from sets
474 and hash tables, so we use alloc pools for them. */
476 static alloc_pool bitmap_set_pool;
477 static bitmap_obstack grand_bitmap_obstack;
479 /* To avoid adding 300 temporary variables when we only need one, we
480 only create one temporary variable, on demand, and build ssa names
481 off that. We do have to change the variable if the types don't
482 match the current variable's type. */
483 static tree pretemp;
484 static tree storetemp;
485 static tree prephitemp;
487 /* Set of blocks with statements that have had their EH properties changed. */
488 static bitmap need_eh_cleanup;
490 /* Set of blocks with statements that have had their AB properties changed. */
491 static bitmap need_ab_cleanup;
493 /* The phi_translate_table caches phi translations for a given
494 expression and predecessor. */
496 static htab_t phi_translate_table;
498 /* A three tuple {e, pred, v} used to cache phi translations in the
499 phi_translate_table. */
501 typedef struct expr_pred_trans_d
503 /* The expression. */
504 pre_expr e;
506 /* The predecessor block along which we translated the expression. */
507 basic_block pred;
509 /* The value that resulted from the translation. */
510 pre_expr v;
512 /* The hashcode for the expression, pred pair. This is cached for
513 speed reasons. */
514 hashval_t hashcode;
515 } *expr_pred_trans_t;
516 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
518 /* Return the hash value for a phi translation table entry. */
520 static hashval_t
521 expr_pred_trans_hash (const void *p)
523 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
524 return ve->hashcode;
527 /* Return true if two phi translation table entries are the same.
528 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
530 static int
531 expr_pred_trans_eq (const void *p1, const void *p2)
533 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
534 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
535 basic_block b1 = ve1->pred;
536 basic_block b2 = ve2->pred;
538 /* If they are not translations for the same basic block, they can't
539 be equal. */
540 if (b1 != b2)
541 return false;
542 return pre_expr_eq (ve1->e, ve2->e);
545 /* Search in the phi translation table for the translation of
546 expression E in basic block PRED.
547 Return the translated value, if found, NULL otherwise. */
549 static inline pre_expr
550 phi_trans_lookup (pre_expr e, basic_block pred)
552 void **slot;
553 struct expr_pred_trans_d ept;
555 ept.e = e;
556 ept.pred = pred;
557 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
558 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
559 NO_INSERT);
560 if (!slot)
561 return NULL;
562 else
563 return ((expr_pred_trans_t) *slot)->v;
567 /* Add the tuple mapping from {expression E, basic block PRED} to
568 value V, to the phi translation table. */
570 static inline void
571 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
573 void **slot;
574 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
575 new_pair->e = e;
576 new_pair->pred = pred;
577 new_pair->v = v;
578 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
579 pred->index);
581 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
582 new_pair->hashcode, INSERT);
583 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 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
620 bitmap_initialize (&ret->values, &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_clear (&set->expressions);
702 bitmap_clear (&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_head temp;
752 bitmap_initialize (&temp, &grand_bitmap_obstack);
754 bitmap_and_into (&dest->values, &orig->values);
755 bitmap_copy (&temp, &dest->expressions);
756 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
758 pre_expr expr = expression_for_id (i);
759 unsigned int value_id = get_expr_value_id (expr);
760 if (!bitmap_bit_p (&dest->values, value_id))
761 bitmap_clear_bit (&dest->expressions, i);
763 bitmap_clear (&temp);
767 /* Subtract all values and expressions contained in ORIG from DEST. */
769 static bitmap_set_t
770 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
772 bitmap_set_t result = bitmap_set_new ();
773 bitmap_iterator bi;
774 unsigned int i;
776 bitmap_and_compl (&result->expressions, &dest->expressions,
777 &orig->expressions);
779 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
781 pre_expr expr = expression_for_id (i);
782 unsigned int value_id = get_expr_value_id (expr);
783 bitmap_set_bit (&result->values, value_id);
786 return result;
789 /* Subtract all the values in bitmap set B from bitmap set A. */
791 static void
792 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
794 unsigned int i;
795 bitmap_iterator bi;
796 bitmap_head temp;
798 bitmap_initialize (&temp, &grand_bitmap_obstack);
800 bitmap_copy (&temp, &a->expressions);
801 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
803 pre_expr expr = expression_for_id (i);
804 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
805 bitmap_remove_from_set (a, expr);
807 bitmap_clear (&temp);
811 /* Return true if bitmapped set SET contains the value VALUE_ID. */
813 static bool
814 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
816 if (value_id_constant_p (value_id))
817 return true;
819 if (!set || bitmap_empty_p (&set->expressions))
820 return false;
822 return bitmap_bit_p (&set->values, value_id);
825 static inline bool
826 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
828 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
831 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
833 static void
834 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
835 const pre_expr expr)
837 bitmap_set_t exprset;
838 unsigned int i;
839 bitmap_iterator bi;
841 if (value_id_constant_p (lookfor))
842 return;
844 if (!bitmap_set_contains_value (set, lookfor))
845 return;
847 /* The number of expressions having a given value is usually
848 significantly less than the total number of expressions in SET.
849 Thus, rather than check, for each expression in SET, whether it
850 has the value LOOKFOR, we walk the reverse mapping that tells us
851 what expressions have a given value, and see if any of those
852 expressions are in our set. For large testcases, this is about
853 5-10x faster than walking the bitmap. If this is somehow a
854 significant lose for some cases, we can choose which set to walk
855 based on the set size. */
856 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
857 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
859 if (bitmap_clear_bit (&set->expressions, i))
861 bitmap_set_bit (&set->expressions, get_expression_id (expr));
862 return;
867 /* Return true if two bitmap sets are equal. */
869 static bool
870 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
872 return bitmap_equal_p (&a->values, &b->values);
875 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
876 and add it otherwise. */
878 static void
879 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
881 unsigned int val = get_expr_value_id (expr);
883 if (bitmap_set_contains_value (set, val))
884 bitmap_set_replace_value (set, val, expr);
885 else
886 bitmap_insert_into_set (set, expr);
889 /* Insert EXPR into SET if EXPR's value is not already present in
890 SET. */
892 static void
893 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
895 unsigned int val = get_expr_value_id (expr);
897 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
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 DEBUG_FUNCTION 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 DEBUG_FUNCTION 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 DEBUG_FUNCTION 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 tree folded;
1235 if ((folded = fully_constant_vn_reference_p (ref)))
1236 return get_or_alloc_expr_for_constant (folded);
1237 return e;
1239 default:
1240 return e;
1242 return e;
1245 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1246 it has the value it would have in BLOCK. Set *SAME_VALID to true
1247 in case the new vuse doesn't change the value id of the OPERANDS. */
1249 static tree
1250 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1251 alias_set_type set, tree type, tree vuse,
1252 basic_block phiblock,
1253 basic_block block, bool *same_valid)
1255 gimple phi = SSA_NAME_DEF_STMT (vuse);
1256 ao_ref ref;
1257 edge e = NULL;
1258 bool use_oracle;
1260 *same_valid = true;
1262 if (gimple_bb (phi) != phiblock)
1263 return vuse;
1265 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1267 /* Use the alias-oracle to find either the PHI node in this block,
1268 the first VUSE used in this block that is equivalent to vuse or
1269 the first VUSE which definition in this block kills the value. */
1270 if (gimple_code (phi) == GIMPLE_PHI)
1271 e = find_edge (block, phiblock);
1272 else if (use_oracle)
1273 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1275 vuse = gimple_vuse (phi);
1276 phi = SSA_NAME_DEF_STMT (vuse);
1277 if (gimple_bb (phi) != phiblock)
1278 return vuse;
1279 if (gimple_code (phi) == GIMPLE_PHI)
1281 e = find_edge (block, phiblock);
1282 break;
1285 else
1286 return NULL_TREE;
1288 if (e)
1290 if (use_oracle)
1292 bitmap visited = NULL;
1293 /* Try to find a vuse that dominates this phi node by skipping
1294 non-clobbering statements. */
1295 vuse = get_continuation_for_phi (phi, &ref, &visited);
1296 if (visited)
1297 BITMAP_FREE (visited);
1299 else
1300 vuse = NULL_TREE;
1301 if (!vuse)
1303 /* If we didn't find any, the value ID can't stay the same,
1304 but return the translated vuse. */
1305 *same_valid = false;
1306 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1308 /* ??? We would like to return vuse here as this is the canonical
1309 upmost vdef that this reference is associated with. But during
1310 insertion of the references into the hash tables we only ever
1311 directly insert with their direct gimple_vuse, hence returning
1312 something else would make us not find the other expression. */
1313 return PHI_ARG_DEF (phi, e->dest_idx);
1316 return NULL_TREE;
1319 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1320 SET2. This is used to avoid making a set consisting of the union
1321 of PA_IN and ANTIC_IN during insert. */
1323 static inline pre_expr
1324 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1326 pre_expr result;
1328 result = bitmap_find_leader (set1, val, NULL);
1329 if (!result && set2)
1330 result = bitmap_find_leader (set2, val, NULL);
1331 return result;
1334 /* Get the tree type for our PRE expression e. */
1336 static tree
1337 get_expr_type (const pre_expr e)
1339 switch (e->kind)
1341 case NAME:
1342 return TREE_TYPE (PRE_EXPR_NAME (e));
1343 case CONSTANT:
1344 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1345 case REFERENCE:
1346 return PRE_EXPR_REFERENCE (e)->type;
1347 case NARY:
1348 return PRE_EXPR_NARY (e)->type;
1350 gcc_unreachable();
1353 /* Get a representative SSA_NAME for a given expression.
1354 Since all of our sub-expressions are treated as values, we require
1355 them to be SSA_NAME's for simplicity.
1356 Prior versions of GVNPRE used to use "value handles" here, so that
1357 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1358 either case, the operands are really values (IE we do not expect
1359 them to be usable without finding leaders). */
1361 static tree
1362 get_representative_for (const pre_expr e)
1364 tree exprtype;
1365 tree name;
1366 unsigned int value_id = get_expr_value_id (e);
1368 switch (e->kind)
1370 case NAME:
1371 return PRE_EXPR_NAME (e);
1372 case CONSTANT:
1373 return PRE_EXPR_CONSTANT (e);
1374 case NARY:
1375 case REFERENCE:
1377 /* Go through all of the expressions representing this value
1378 and pick out an SSA_NAME. */
1379 unsigned int i;
1380 bitmap_iterator bi;
1381 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1382 value_id);
1383 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1385 pre_expr rep = expression_for_id (i);
1386 if (rep->kind == NAME)
1387 return PRE_EXPR_NAME (rep);
1390 break;
1392 /* If we reached here we couldn't find an SSA_NAME. This can
1393 happen when we've discovered a value that has never appeared in
1394 the program as set to an SSA_NAME, most likely as the result of
1395 phi translation. */
1396 if (dump_file)
1398 fprintf (dump_file,
1399 "Could not find SSA_NAME representative for expression:");
1400 print_pre_expr (dump_file, e);
1401 fprintf (dump_file, "\n");
1404 exprtype = get_expr_type (e);
1406 /* Build and insert the assignment of the end result to the temporary
1407 that we will return. */
1408 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1410 pretemp = create_tmp_reg (exprtype, "pretmp");
1411 get_var_ann (pretemp);
1414 name = make_ssa_name (pretemp, gimple_build_nop ());
1415 VN_INFO_GET (name)->value_id = value_id;
1416 if (e->kind == CONSTANT)
1417 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1418 else
1419 VN_INFO (name)->valnum = name;
1421 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1422 if (dump_file)
1424 fprintf (dump_file, "Created SSA_NAME representative ");
1425 print_generic_expr (dump_file, name, 0);
1426 fprintf (dump_file, " for expression:");
1427 print_pre_expr (dump_file, e);
1428 fprintf (dump_file, "\n");
1431 return name;
1436 static pre_expr
1437 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1438 basic_block pred, basic_block phiblock);
1440 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1441 the phis in PRED. Return NULL if we can't find a leader for each part
1442 of the translated expression. */
1444 static pre_expr
1445 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1446 basic_block pred, basic_block phiblock)
1448 switch (expr->kind)
1450 case NARY:
1452 unsigned int i;
1453 bool changed = false;
1454 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1455 struct vn_nary_op_s newnary;
1456 /* The NARY structure is only guaranteed to have been
1457 allocated to the nary->length operands. */
1458 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1459 - sizeof (tree) * (4 - nary->length)));
1461 for (i = 0; i < newnary.length; i++)
1463 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1464 continue;
1465 else
1467 pre_expr leader, result;
1468 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1469 leader = find_leader_in_sets (op_val_id, set1, set2);
1470 result = phi_translate (leader, set1, set2, pred, phiblock);
1471 if (result && result != leader)
1473 tree name = get_representative_for (result);
1474 if (!name)
1475 return NULL;
1476 newnary.op[i] = name;
1478 else if (!result)
1479 return NULL;
1481 changed |= newnary.op[i] != nary->op[i];
1484 if (changed)
1486 pre_expr constant;
1487 unsigned int new_val_id;
1489 tree result = vn_nary_op_lookup_pieces (newnary.length,
1490 newnary.opcode,
1491 newnary.type,
1492 newnary.op[0],
1493 newnary.op[1],
1494 newnary.op[2],
1495 newnary.op[3],
1496 &nary);
1497 if (result && is_gimple_min_invariant (result))
1498 return get_or_alloc_expr_for_constant (result);
1500 expr = (pre_expr) pool_alloc (pre_expr_pool);
1501 expr->kind = NARY;
1502 expr->id = 0;
1503 if (nary)
1505 PRE_EXPR_NARY (expr) = nary;
1506 constant = fully_constant_expression (expr);
1507 if (constant != expr)
1508 return constant;
1510 new_val_id = nary->value_id;
1511 get_or_alloc_expression_id (expr);
1513 else
1515 new_val_id = get_next_value_id ();
1516 VEC_safe_grow_cleared (bitmap_set_t, heap,
1517 value_expressions,
1518 get_max_value_id() + 1);
1519 nary = vn_nary_op_insert_pieces (newnary.length,
1520 newnary.opcode,
1521 newnary.type,
1522 newnary.op[0],
1523 newnary.op[1],
1524 newnary.op[2],
1525 newnary.op[3],
1526 result, new_val_id);
1527 PRE_EXPR_NARY (expr) = nary;
1528 constant = fully_constant_expression (expr);
1529 if (constant != expr)
1530 return constant;
1531 get_or_alloc_expression_id (expr);
1533 add_to_value (new_val_id, expr);
1535 return expr;
1537 break;
1539 case REFERENCE:
1541 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1542 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1543 tree vuse = ref->vuse;
1544 tree newvuse = vuse;
1545 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1546 bool changed = false, same_valid = true;
1547 unsigned int i, j;
1548 vn_reference_op_t operand;
1549 vn_reference_t newref;
1551 for (i = 0, j = 0;
1552 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1554 pre_expr opresult;
1555 pre_expr leader;
1556 tree oldop0 = operand->op0;
1557 tree oldop1 = operand->op1;
1558 tree oldop2 = operand->op2;
1559 tree op0 = oldop0;
1560 tree op1 = oldop1;
1561 tree op2 = oldop2;
1562 tree type = operand->type;
1563 vn_reference_op_s newop = *operand;
1565 if (op0 && TREE_CODE (op0) == SSA_NAME)
1567 unsigned int op_val_id = VN_INFO (op0)->value_id;
1568 leader = find_leader_in_sets (op_val_id, set1, set2);
1569 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1570 if (opresult && opresult != leader)
1572 tree name = get_representative_for (opresult);
1573 if (!name)
1574 break;
1575 op0 = name;
1577 else if (!opresult)
1578 break;
1580 changed |= op0 != oldop0;
1582 if (op1 && TREE_CODE (op1) == SSA_NAME)
1584 unsigned int op_val_id = VN_INFO (op1)->value_id;
1585 leader = find_leader_in_sets (op_val_id, set1, set2);
1586 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1587 if (opresult && opresult != leader)
1589 tree name = get_representative_for (opresult);
1590 if (!name)
1591 break;
1592 op1 = name;
1594 else if (!opresult)
1595 break;
1597 /* We can't possibly insert these. */
1598 else if (op1 && !is_gimple_min_invariant (op1))
1599 break;
1600 changed |= op1 != oldop1;
1601 if (op2 && TREE_CODE (op2) == SSA_NAME)
1603 unsigned int op_val_id = VN_INFO (op2)->value_id;
1604 leader = find_leader_in_sets (op_val_id, set1, set2);
1605 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1606 if (opresult && opresult != leader)
1608 tree name = get_representative_for (opresult);
1609 if (!name)
1610 break;
1611 op2 = name;
1613 else if (!opresult)
1614 break;
1616 /* We can't possibly insert these. */
1617 else if (op2 && !is_gimple_min_invariant (op2))
1618 break;
1619 changed |= op2 != oldop2;
1621 if (!newoperands)
1622 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1623 /* We may have changed from an SSA_NAME to a constant */
1624 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1625 newop.opcode = TREE_CODE (op0);
1626 newop.type = type;
1627 newop.op0 = op0;
1628 newop.op1 = op1;
1629 newop.op2 = op2;
1630 /* If it transforms a non-constant ARRAY_REF into a constant
1631 one, adjust the constant offset. */
1632 if (newop.opcode == ARRAY_REF
1633 && newop.off == -1
1634 && TREE_CODE (op0) == INTEGER_CST
1635 && TREE_CODE (op1) == INTEGER_CST
1636 && TREE_CODE (op2) == INTEGER_CST)
1638 double_int off = tree_to_double_int (op0);
1639 off = double_int_add (off,
1640 double_int_neg
1641 (tree_to_double_int (op1)));
1642 off = double_int_mul (off, tree_to_double_int (op2));
1643 if (double_int_fits_in_shwi_p (off))
1644 newop.off = off.low;
1646 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1647 /* If it transforms from an SSA_NAME to an address, fold with
1648 a preceding indirect reference. */
1649 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1650 && VEC_index (vn_reference_op_s,
1651 newoperands, j - 1)->opcode == MEM_REF)
1652 vn_reference_fold_indirect (&newoperands, &j);
1654 if (i != VEC_length (vn_reference_op_s, operands))
1656 if (newoperands)
1657 VEC_free (vn_reference_op_s, heap, newoperands);
1658 return NULL;
1661 if (vuse)
1663 newvuse = translate_vuse_through_block (newoperands,
1664 ref->set, ref->type,
1665 vuse, phiblock, pred,
1666 &same_valid);
1667 if (newvuse == NULL_TREE)
1669 VEC_free (vn_reference_op_s, heap, newoperands);
1670 return NULL;
1674 if (changed || newvuse != vuse)
1676 unsigned int new_val_id;
1677 pre_expr constant;
1678 bool converted = false;
1680 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1681 ref->type,
1682 newoperands,
1683 &newref, VN_WALK);
1684 if (result)
1685 VEC_free (vn_reference_op_s, heap, newoperands);
1687 if (result
1688 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1690 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1691 converted = true;
1693 else if (!result && newref
1694 && !useless_type_conversion_p (ref->type, newref->type))
1696 VEC_free (vn_reference_op_s, heap, newoperands);
1697 return NULL;
1700 if (result && is_gimple_min_invariant (result))
1702 gcc_assert (!newoperands);
1703 return get_or_alloc_expr_for_constant (result);
1706 expr = (pre_expr) pool_alloc (pre_expr_pool);
1707 expr->kind = REFERENCE;
1708 expr->id = 0;
1710 if (converted)
1712 vn_nary_op_t nary;
1713 tree nresult;
1715 gcc_assert (CONVERT_EXPR_P (result)
1716 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1718 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1719 TREE_TYPE (result),
1720 TREE_OPERAND (result, 0),
1721 NULL_TREE, NULL_TREE,
1722 NULL_TREE,
1723 &nary);
1724 if (nresult && is_gimple_min_invariant (nresult))
1725 return get_or_alloc_expr_for_constant (nresult);
1727 expr->kind = NARY;
1728 if (nary)
1730 PRE_EXPR_NARY (expr) = nary;
1731 constant = fully_constant_expression (expr);
1732 if (constant != expr)
1733 return constant;
1735 new_val_id = nary->value_id;
1736 get_or_alloc_expression_id (expr);
1738 else
1740 new_val_id = get_next_value_id ();
1741 VEC_safe_grow_cleared (bitmap_set_t, heap,
1742 value_expressions,
1743 get_max_value_id() + 1);
1744 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1745 TREE_TYPE (result),
1746 TREE_OPERAND (result, 0),
1747 NULL_TREE, NULL_TREE,
1748 NULL_TREE, NULL_TREE,
1749 new_val_id);
1750 PRE_EXPR_NARY (expr) = nary;
1751 constant = fully_constant_expression (expr);
1752 if (constant != expr)
1753 return constant;
1754 get_or_alloc_expression_id (expr);
1757 else if (newref)
1759 PRE_EXPR_REFERENCE (expr) = newref;
1760 constant = fully_constant_expression (expr);
1761 if (constant != expr)
1762 return constant;
1764 new_val_id = newref->value_id;
1765 get_or_alloc_expression_id (expr);
1767 else
1769 if (changed || !same_valid)
1771 new_val_id = get_next_value_id ();
1772 VEC_safe_grow_cleared (bitmap_set_t, heap,
1773 value_expressions,
1774 get_max_value_id() + 1);
1776 else
1777 new_val_id = ref->value_id;
1778 newref = vn_reference_insert_pieces (newvuse, ref->set,
1779 ref->type,
1780 newoperands,
1781 result, new_val_id);
1782 newoperands = NULL;
1783 PRE_EXPR_REFERENCE (expr) = newref;
1784 constant = fully_constant_expression (expr);
1785 if (constant != expr)
1786 return constant;
1787 get_or_alloc_expression_id (expr);
1789 add_to_value (new_val_id, expr);
1791 VEC_free (vn_reference_op_s, heap, newoperands);
1792 return expr;
1794 break;
1796 case NAME:
1798 gimple phi = NULL;
1799 edge e;
1800 gimple def_stmt;
1801 tree name = PRE_EXPR_NAME (expr);
1803 def_stmt = SSA_NAME_DEF_STMT (name);
1804 if (gimple_code (def_stmt) == GIMPLE_PHI
1805 && gimple_bb (def_stmt) == phiblock)
1806 phi = def_stmt;
1807 else
1808 return expr;
1810 e = find_edge (pred, gimple_bb (phi));
1811 if (e)
1813 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1814 pre_expr newexpr;
1816 if (TREE_CODE (def) == SSA_NAME)
1817 def = VN_INFO (def)->valnum;
1819 /* Handle constant. */
1820 if (is_gimple_min_invariant (def))
1821 return get_or_alloc_expr_for_constant (def);
1823 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1824 return NULL;
1826 newexpr = get_or_alloc_expr_for_name (def);
1827 return newexpr;
1830 return expr;
1832 default:
1833 gcc_unreachable ();
1837 /* Wrapper around phi_translate_1 providing caching functionality. */
1839 static pre_expr
1840 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1841 basic_block pred, basic_block phiblock)
1843 pre_expr phitrans;
1845 if (!expr)
1846 return NULL;
1848 /* Constants contain no values that need translation. */
1849 if (expr->kind == CONSTANT)
1850 return expr;
1852 if (value_id_constant_p (get_expr_value_id (expr)))
1853 return expr;
1855 if (expr->kind != NAME)
1857 phitrans = phi_trans_lookup (expr, pred);
1858 if (phitrans)
1859 return phitrans;
1862 /* Translate. */
1863 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1865 /* Don't add empty translations to the cache. Neither add
1866 translations of NAMEs as those are cheap to translate. */
1867 if (phitrans
1868 && expr->kind != NAME)
1869 phi_trans_add (expr, phitrans, pred);
1871 return phitrans;
1875 /* For each expression in SET, translate the values through phi nodes
1876 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1877 expressions in DEST. */
1879 static void
1880 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1881 basic_block phiblock)
1883 VEC (pre_expr, heap) *exprs;
1884 pre_expr expr;
1885 int i;
1887 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1889 bitmap_set_copy (dest, set);
1890 return;
1893 exprs = sorted_array_from_bitmap_set (set);
1894 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1896 pre_expr translated;
1897 translated = phi_translate (expr, set, NULL, pred, phiblock);
1898 if (!translated)
1899 continue;
1901 /* We might end up with multiple expressions from SET being
1902 translated to the same value. In this case we do not want
1903 to retain the NARY or REFERENCE expression but prefer a NAME
1904 which would be the leader. */
1905 if (translated->kind == NAME)
1906 bitmap_value_replace_in_set (dest, translated);
1907 else
1908 bitmap_value_insert_into_set (dest, translated);
1910 VEC_free (pre_expr, heap, exprs);
1913 /* Find the leader for a value (i.e., the name representing that
1914 value) in a given set, and return it. If STMT is non-NULL it
1915 makes sure the defining statement for the leader dominates it.
1916 Return NULL if no leader is found. */
1918 static pre_expr
1919 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1921 if (value_id_constant_p (val))
1923 unsigned int i;
1924 bitmap_iterator bi;
1925 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1927 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1929 pre_expr expr = expression_for_id (i);
1930 if (expr->kind == CONSTANT)
1931 return expr;
1934 if (bitmap_set_contains_value (set, val))
1936 /* Rather than walk the entire bitmap of expressions, and see
1937 whether any of them has the value we are looking for, we look
1938 at the reverse mapping, which tells us the set of expressions
1939 that have a given value (IE value->expressions with that
1940 value) and see if any of those expressions are in our set.
1941 The number of expressions per value is usually significantly
1942 less than the number of expressions in the set. In fact, for
1943 large testcases, doing it this way is roughly 5-10x faster
1944 than walking the bitmap.
1945 If this is somehow a significant lose for some cases, we can
1946 choose which set to walk based on which set is smaller. */
1947 unsigned int i;
1948 bitmap_iterator bi;
1949 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1951 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1952 &set->expressions, 0, i, bi)
1954 pre_expr val = expression_for_id (i);
1955 /* At the point where stmt is not null, there should always
1956 be an SSA_NAME first in the list of expressions. */
1957 if (stmt)
1959 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1960 if (gimple_code (def_stmt) != GIMPLE_PHI
1961 && gimple_bb (def_stmt) == gimple_bb (stmt)
1962 /* PRE insertions are at the end of the basic-block
1963 and have UID 0. */
1964 && (gimple_uid (def_stmt) == 0
1965 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1966 continue;
1968 return val;
1971 return NULL;
1974 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1975 BLOCK by seeing if it is not killed in the block. Note that we are
1976 only determining whether there is a store that kills it. Because
1977 of the order in which clean iterates over values, we are guaranteed
1978 that altered operands will have caused us to be eliminated from the
1979 ANTIC_IN set already. */
1981 static bool
1982 value_dies_in_block_x (pre_expr expr, basic_block block)
1984 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1985 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1986 gimple def;
1987 gimple_stmt_iterator gsi;
1988 unsigned id = get_expression_id (expr);
1989 bool res = false;
1990 ao_ref ref;
1992 if (!vuse)
1993 return false;
1995 /* Lookup a previously calculated result. */
1996 if (EXPR_DIES (block)
1997 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1998 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
2000 /* A memory expression {e, VUSE} dies in the block if there is a
2001 statement that may clobber e. If, starting statement walk from the
2002 top of the basic block, a statement uses VUSE there can be no kill
2003 inbetween that use and the original statement that loaded {e, VUSE},
2004 so we can stop walking. */
2005 ref.base = NULL_TREE;
2006 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
2008 tree def_vuse, def_vdef;
2009 def = gsi_stmt (gsi);
2010 def_vuse = gimple_vuse (def);
2011 def_vdef = gimple_vdef (def);
2013 /* Not a memory statement. */
2014 if (!def_vuse)
2015 continue;
2017 /* Not a may-def. */
2018 if (!def_vdef)
2020 /* A load with the same VUSE, we're done. */
2021 if (def_vuse == vuse)
2022 break;
2024 continue;
2027 /* Init ref only if we really need it. */
2028 if (ref.base == NULL_TREE
2029 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
2030 refx->operands))
2032 res = true;
2033 break;
2035 /* If the statement may clobber expr, it dies. */
2036 if (stmt_may_clobber_ref_p_1 (def, &ref))
2038 res = true;
2039 break;
2043 /* Remember the result. */
2044 if (!EXPR_DIES (block))
2045 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2046 bitmap_set_bit (EXPR_DIES (block), id * 2);
2047 if (res)
2048 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2050 return res;
2054 #define union_contains_value(SET1, SET2, VAL) \
2055 (bitmap_set_contains_value ((SET1), (VAL)) \
2056 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
2058 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2060 static bool
2061 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2062 vn_reference_op_t vro)
2064 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2066 struct pre_expr_d temp;
2067 temp.kind = NAME;
2068 temp.id = 0;
2069 PRE_EXPR_NAME (&temp) = vro->op0;
2070 temp.id = lookup_expression_id (&temp);
2071 if (temp.id == 0)
2072 return false;
2073 if (!union_contains_value (set1, set2,
2074 get_expr_value_id (&temp)))
2075 return false;
2077 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2079 struct pre_expr_d temp;
2080 temp.kind = NAME;
2081 temp.id = 0;
2082 PRE_EXPR_NAME (&temp) = vro->op1;
2083 temp.id = lookup_expression_id (&temp);
2084 if (temp.id == 0)
2085 return false;
2086 if (!union_contains_value (set1, set2,
2087 get_expr_value_id (&temp)))
2088 return false;
2091 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2093 struct pre_expr_d temp;
2094 temp.kind = NAME;
2095 temp.id = 0;
2096 PRE_EXPR_NAME (&temp) = vro->op2;
2097 temp.id = lookup_expression_id (&temp);
2098 if (temp.id == 0)
2099 return false;
2100 if (!union_contains_value (set1, set2,
2101 get_expr_value_id (&temp)))
2102 return false;
2105 return true;
2108 /* Determine if the expression EXPR is valid in SET1 U SET2.
2109 ONLY SET2 CAN BE NULL.
2110 This means that we have a leader for each part of the expression
2111 (if it consists of values), or the expression is an SSA_NAME.
2112 For loads/calls, we also see if the vuse is killed in this block. */
2114 static bool
2115 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2116 basic_block block)
2118 switch (expr->kind)
2120 case NAME:
2121 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2122 case NARY:
2124 unsigned int i;
2125 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2126 for (i = 0; i < nary->length; i++)
2128 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2130 struct pre_expr_d temp;
2131 temp.kind = NAME;
2132 temp.id = 0;
2133 PRE_EXPR_NAME (&temp) = nary->op[i];
2134 temp.id = lookup_expression_id (&temp);
2135 if (temp.id == 0)
2136 return false;
2137 if (!union_contains_value (set1, set2,
2138 get_expr_value_id (&temp)))
2139 return false;
2142 /* If the NARY may trap make sure the block does not contain
2143 a possible exit point.
2144 ??? This is overly conservative if we translate AVAIL_OUT
2145 as the available expression might be after the exit point. */
2146 if (BB_MAY_NOTRETURN (block)
2147 && vn_nary_may_trap (nary))
2148 return false;
2149 return true;
2151 break;
2152 case REFERENCE:
2154 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2155 vn_reference_op_t vro;
2156 unsigned int i;
2158 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2160 if (!vro_valid_in_sets (set1, set2, vro))
2161 return false;
2163 if (ref->vuse)
2165 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2166 if (!gimple_nop_p (def_stmt)
2167 && gimple_bb (def_stmt) != block
2168 && !dominated_by_p (CDI_DOMINATORS,
2169 block, gimple_bb (def_stmt)))
2170 return false;
2172 return !value_dies_in_block_x (expr, block);
2174 default:
2175 gcc_unreachable ();
2179 /* Clean the set of expressions that are no longer valid in SET1 or
2180 SET2. This means expressions that are made up of values we have no
2181 leaders for in SET1 or SET2. This version is used for partial
2182 anticipation, which means it is not valid in either ANTIC_IN or
2183 PA_IN. */
2185 static void
2186 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2188 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2189 pre_expr expr;
2190 int i;
2192 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2194 if (!valid_in_sets (set1, set2, expr, block))
2195 bitmap_remove_from_set (set1, expr);
2197 VEC_free (pre_expr, heap, exprs);
2200 /* Clean the set of expressions that are no longer valid in SET. This
2201 means expressions that are made up of values we have no leaders for
2202 in SET. */
2204 static void
2205 clean (bitmap_set_t set, basic_block block)
2207 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2208 pre_expr expr;
2209 int i;
2211 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2213 if (!valid_in_sets (set, NULL, expr, block))
2214 bitmap_remove_from_set (set, expr);
2216 VEC_free (pre_expr, heap, exprs);
2219 static sbitmap has_abnormal_preds;
2221 /* List of blocks that may have changed during ANTIC computation and
2222 thus need to be iterated over. */
2224 static sbitmap changed_blocks;
2226 /* Decide whether to defer a block for a later iteration, or PHI
2227 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2228 should defer the block, and true if we processed it. */
2230 static bool
2231 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2232 basic_block block, basic_block phiblock)
2234 if (!BB_VISITED (phiblock))
2236 SET_BIT (changed_blocks, block->index);
2237 BB_VISITED (block) = 0;
2238 BB_DEFERRED (block) = 1;
2239 return false;
2241 else
2242 phi_translate_set (dest, source, block, phiblock);
2243 return true;
2246 /* Compute the ANTIC set for BLOCK.
2248 If succs(BLOCK) > 1 then
2249 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2250 else if succs(BLOCK) == 1 then
2251 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2253 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2256 static bool
2257 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2259 bool changed = false;
2260 bitmap_set_t S, old, ANTIC_OUT;
2261 bitmap_iterator bi;
2262 unsigned int bii;
2263 edge e;
2264 edge_iterator ei;
2266 old = ANTIC_OUT = S = NULL;
2267 BB_VISITED (block) = 1;
2269 /* If any edges from predecessors are abnormal, antic_in is empty,
2270 so do nothing. */
2271 if (block_has_abnormal_pred_edge)
2272 goto maybe_dump_sets;
2274 old = ANTIC_IN (block);
2275 ANTIC_OUT = bitmap_set_new ();
2277 /* If the block has no successors, ANTIC_OUT is empty. */
2278 if (EDGE_COUNT (block->succs) == 0)
2280 /* If we have one successor, we could have some phi nodes to
2281 translate through. */
2282 else if (single_succ_p (block))
2284 basic_block succ_bb = single_succ (block);
2286 /* We trade iterations of the dataflow equations for having to
2287 phi translate the maximal set, which is incredibly slow
2288 (since the maximal set often has 300+ members, even when you
2289 have a small number of blocks).
2290 Basically, we defer the computation of ANTIC for this block
2291 until we have processed it's successor, which will inevitably
2292 have a *much* smaller set of values to phi translate once
2293 clean has been run on it.
2294 The cost of doing this is that we technically perform more
2295 iterations, however, they are lower cost iterations.
2297 Timings for PRE on tramp3d-v4:
2298 without maximal set fix: 11 seconds
2299 with maximal set fix/without deferring: 26 seconds
2300 with maximal set fix/with deferring: 11 seconds
2303 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2304 block, succ_bb))
2306 changed = true;
2307 goto maybe_dump_sets;
2310 /* If we have multiple successors, we take the intersection of all of
2311 them. Note that in the case of loop exit phi nodes, we may have
2312 phis to translate through. */
2313 else
2315 VEC(basic_block, heap) * worklist;
2316 size_t i;
2317 basic_block bprime, first = NULL;
2319 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2320 FOR_EACH_EDGE (e, ei, block->succs)
2322 if (!first
2323 && BB_VISITED (e->dest))
2324 first = e->dest;
2325 else if (BB_VISITED (e->dest))
2326 VEC_quick_push (basic_block, worklist, e->dest);
2329 /* Of multiple successors we have to have visited one already. */
2330 if (!first)
2332 SET_BIT (changed_blocks, block->index);
2333 BB_VISITED (block) = 0;
2334 BB_DEFERRED (block) = 1;
2335 changed = true;
2336 VEC_free (basic_block, heap, worklist);
2337 goto maybe_dump_sets;
2340 if (!gimple_seq_empty_p (phi_nodes (first)))
2341 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2342 else
2343 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2345 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2347 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2349 bitmap_set_t tmp = bitmap_set_new ();
2350 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2351 bitmap_set_and (ANTIC_OUT, tmp);
2352 bitmap_set_free (tmp);
2354 else
2355 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2357 VEC_free (basic_block, heap, worklist);
2360 /* Generate ANTIC_OUT - TMP_GEN. */
2361 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2363 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2364 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2365 TMP_GEN (block));
2367 /* Then union in the ANTIC_OUT - TMP_GEN values,
2368 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2369 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2370 bitmap_value_insert_into_set (ANTIC_IN (block),
2371 expression_for_id (bii));
2373 clean (ANTIC_IN (block), block);
2375 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2377 changed = true;
2378 SET_BIT (changed_blocks, block->index);
2379 FOR_EACH_EDGE (e, ei, block->preds)
2380 SET_BIT (changed_blocks, e->src->index);
2382 else
2383 RESET_BIT (changed_blocks, block->index);
2385 maybe_dump_sets:
2386 if (dump_file && (dump_flags & TDF_DETAILS))
2388 if (!BB_DEFERRED (block) || BB_VISITED (block))
2390 if (ANTIC_OUT)
2391 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2393 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2394 block->index);
2396 if (S)
2397 print_bitmap_set (dump_file, S, "S", block->index);
2399 else
2401 fprintf (dump_file,
2402 "Block %d was deferred for a future iteration.\n",
2403 block->index);
2406 if (old)
2407 bitmap_set_free (old);
2408 if (S)
2409 bitmap_set_free (S);
2410 if (ANTIC_OUT)
2411 bitmap_set_free (ANTIC_OUT);
2412 return changed;
2415 /* Compute PARTIAL_ANTIC for BLOCK.
2417 If succs(BLOCK) > 1 then
2418 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2419 in ANTIC_OUT for all succ(BLOCK)
2420 else if succs(BLOCK) == 1 then
2421 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2423 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2424 - ANTIC_IN[BLOCK])
2427 static bool
2428 compute_partial_antic_aux (basic_block block,
2429 bool block_has_abnormal_pred_edge)
2431 bool changed = false;
2432 bitmap_set_t old_PA_IN;
2433 bitmap_set_t PA_OUT;
2434 edge e;
2435 edge_iterator ei;
2436 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2438 old_PA_IN = PA_OUT = NULL;
2440 /* If any edges from predecessors are abnormal, antic_in is empty,
2441 so do nothing. */
2442 if (block_has_abnormal_pred_edge)
2443 goto maybe_dump_sets;
2445 /* If there are too many partially anticipatable values in the
2446 block, phi_translate_set can take an exponential time: stop
2447 before the translation starts. */
2448 if (max_pa
2449 && single_succ_p (block)
2450 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2451 goto maybe_dump_sets;
2453 old_PA_IN = PA_IN (block);
2454 PA_OUT = bitmap_set_new ();
2456 /* If the block has no successors, ANTIC_OUT is empty. */
2457 if (EDGE_COUNT (block->succs) == 0)
2459 /* If we have one successor, we could have some phi nodes to
2460 translate through. Note that we can't phi translate across DFS
2461 back edges in partial antic, because it uses a union operation on
2462 the successors. For recurrences like IV's, we will end up
2463 generating a new value in the set on each go around (i + 3 (VH.1)
2464 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2465 else if (single_succ_p (block))
2467 basic_block succ = single_succ (block);
2468 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2469 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2471 /* If we have multiple successors, we take the union of all of
2472 them. */
2473 else
2475 VEC(basic_block, heap) * worklist;
2476 size_t i;
2477 basic_block bprime;
2479 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2480 FOR_EACH_EDGE (e, ei, block->succs)
2482 if (e->flags & EDGE_DFS_BACK)
2483 continue;
2484 VEC_quick_push (basic_block, worklist, e->dest);
2486 if (VEC_length (basic_block, worklist) > 0)
2488 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2490 unsigned int i;
2491 bitmap_iterator bi;
2493 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2494 bitmap_value_insert_into_set (PA_OUT,
2495 expression_for_id (i));
2496 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2498 bitmap_set_t pa_in = bitmap_set_new ();
2499 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2500 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2501 bitmap_value_insert_into_set (PA_OUT,
2502 expression_for_id (i));
2503 bitmap_set_free (pa_in);
2505 else
2506 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2507 bitmap_value_insert_into_set (PA_OUT,
2508 expression_for_id (i));
2511 VEC_free (basic_block, heap, worklist);
2514 /* PA_IN starts with PA_OUT - TMP_GEN.
2515 Then we subtract things from ANTIC_IN. */
2516 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2518 /* For partial antic, we want to put back in the phi results, since
2519 we will properly avoid making them partially antic over backedges. */
2520 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2521 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2523 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2524 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2526 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2528 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2530 changed = true;
2531 SET_BIT (changed_blocks, block->index);
2532 FOR_EACH_EDGE (e, ei, block->preds)
2533 SET_BIT (changed_blocks, e->src->index);
2535 else
2536 RESET_BIT (changed_blocks, block->index);
2538 maybe_dump_sets:
2539 if (dump_file && (dump_flags & TDF_DETAILS))
2541 if (PA_OUT)
2542 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2544 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2546 if (old_PA_IN)
2547 bitmap_set_free (old_PA_IN);
2548 if (PA_OUT)
2549 bitmap_set_free (PA_OUT);
2550 return changed;
2553 /* Compute ANTIC and partial ANTIC sets. */
2555 static void
2556 compute_antic (void)
2558 bool changed = true;
2559 int num_iterations = 0;
2560 basic_block block;
2561 int i;
2563 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2564 We pre-build the map of blocks with incoming abnormal edges here. */
2565 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2566 sbitmap_zero (has_abnormal_preds);
2568 FOR_EACH_BB (block)
2570 edge_iterator ei;
2571 edge e;
2573 FOR_EACH_EDGE (e, ei, block->preds)
2575 e->flags &= ~EDGE_DFS_BACK;
2576 if (e->flags & EDGE_ABNORMAL)
2578 SET_BIT (has_abnormal_preds, block->index);
2579 break;
2583 BB_VISITED (block) = 0;
2584 BB_DEFERRED (block) = 0;
2586 /* While we are here, give empty ANTIC_IN sets to each block. */
2587 ANTIC_IN (block) = bitmap_set_new ();
2588 PA_IN (block) = bitmap_set_new ();
2591 /* At the exit block we anticipate nothing. */
2592 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2593 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2594 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2596 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2597 sbitmap_ones (changed_blocks);
2598 while (changed)
2600 if (dump_file && (dump_flags & TDF_DETAILS))
2601 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2602 /* ??? We need to clear our PHI translation cache here as the
2603 ANTIC sets shrink and we restrict valid translations to
2604 those having operands with leaders in ANTIC. Same below
2605 for PA ANTIC computation. */
2606 num_iterations++;
2607 changed = false;
2608 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2610 if (TEST_BIT (changed_blocks, postorder[i]))
2612 basic_block block = BASIC_BLOCK (postorder[i]);
2613 changed |= compute_antic_aux (block,
2614 TEST_BIT (has_abnormal_preds,
2615 block->index));
2618 /* Theoretically possible, but *highly* unlikely. */
2619 gcc_checking_assert (num_iterations < 500);
2622 statistics_histogram_event (cfun, "compute_antic iterations",
2623 num_iterations);
2625 if (do_partial_partial)
2627 sbitmap_ones (changed_blocks);
2628 mark_dfs_back_edges ();
2629 num_iterations = 0;
2630 changed = true;
2631 while (changed)
2633 if (dump_file && (dump_flags & TDF_DETAILS))
2634 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2635 num_iterations++;
2636 changed = false;
2637 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2639 if (TEST_BIT (changed_blocks, postorder[i]))
2641 basic_block block = BASIC_BLOCK (postorder[i]);
2642 changed
2643 |= compute_partial_antic_aux (block,
2644 TEST_BIT (has_abnormal_preds,
2645 block->index));
2648 /* Theoretically possible, but *highly* unlikely. */
2649 gcc_checking_assert (num_iterations < 500);
2651 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2652 num_iterations);
2654 sbitmap_free (has_abnormal_preds);
2655 sbitmap_free (changed_blocks);
2658 /* Return true if we can value number the call in STMT. This is true
2659 if we have a pure or constant call to a real function. */
2661 static bool
2662 can_value_number_call (gimple stmt)
2664 if (gimple_call_internal_p (stmt))
2665 return false;
2666 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2667 return true;
2668 return false;
2671 /* Return true if OP is a tree which we can perform PRE on.
2672 This may not match the operations we can value number, but in
2673 a perfect world would. */
2675 static bool
2676 can_PRE_operation (tree op)
2678 return UNARY_CLASS_P (op)
2679 || BINARY_CLASS_P (op)
2680 || COMPARISON_CLASS_P (op)
2681 || TREE_CODE (op) == MEM_REF
2682 || TREE_CODE (op) == COMPONENT_REF
2683 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2684 || TREE_CODE (op) == CALL_EXPR
2685 || TREE_CODE (op) == ARRAY_REF;
2689 /* Inserted expressions are placed onto this worklist, which is used
2690 for performing quick dead code elimination of insertions we made
2691 that didn't turn out to be necessary. */
2692 static bitmap inserted_exprs;
2694 /* Pool allocated fake store expressions are placed onto this
2695 worklist, which, after performing dead code elimination, is walked
2696 to see which expressions need to be put into GC'able memory */
2697 static VEC(gimple, heap) *need_creation;
2699 /* The actual worker for create_component_ref_by_pieces. */
2701 static tree
2702 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2703 unsigned int *operand, gimple_seq *stmts,
2704 gimple domstmt)
2706 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2707 *operand);
2708 tree genop;
2709 ++*operand;
2710 switch (currop->opcode)
2712 case CALL_EXPR:
2714 tree folded, sc = NULL_TREE;
2715 unsigned int nargs = 0;
2716 tree fn, *args;
2717 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2718 fn = currop->op0;
2719 else
2721 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2722 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2723 if (!fn)
2724 return NULL_TREE;
2726 if (currop->op1)
2728 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2729 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2730 if (!sc)
2731 return NULL_TREE;
2733 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2734 ref->operands) - 1);
2735 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2737 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2738 operand, stmts,
2739 domstmt);
2740 if (!args[nargs])
2742 free (args);
2743 return NULL_TREE;
2745 nargs++;
2747 folded = build_call_array (currop->type,
2748 (TREE_CODE (fn) == FUNCTION_DECL
2749 ? build_fold_addr_expr (fn) : fn),
2750 nargs, args);
2751 free (args);
2752 if (sc)
2753 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2754 return folded;
2756 break;
2757 case MEM_REF:
2759 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2760 stmts, domstmt);
2761 tree offset = currop->op0;
2762 if (!baseop)
2763 return NULL_TREE;
2764 if (TREE_CODE (baseop) == ADDR_EXPR
2765 && handled_component_p (TREE_OPERAND (baseop, 0)))
2767 HOST_WIDE_INT off;
2768 tree base;
2769 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2770 &off);
2771 gcc_assert (base);
2772 offset = int_const_binop (PLUS_EXPR, offset,
2773 build_int_cst (TREE_TYPE (offset),
2774 off), 0);
2775 baseop = build_fold_addr_expr (base);
2777 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2779 break;
2780 case TARGET_MEM_REF:
2782 pre_expr op0expr, op1expr;
2783 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2784 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2785 ++*operand);
2786 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2787 stmts, domstmt);
2788 if (!baseop)
2789 return NULL_TREE;
2790 if (currop->op0)
2792 op0expr = get_or_alloc_expr_for (currop->op0);
2793 genop0 = find_or_generate_expression (block, op0expr,
2794 stmts, domstmt);
2795 if (!genop0)
2796 return NULL_TREE;
2798 if (nextop->op0)
2800 op1expr = get_or_alloc_expr_for (nextop->op0);
2801 genop1 = find_or_generate_expression (block, op1expr,
2802 stmts, domstmt);
2803 if (!genop1)
2804 return NULL_TREE;
2806 return build5 (TARGET_MEM_REF, currop->type,
2807 baseop, currop->op2, genop0, currop->op1, genop1);
2809 break;
2810 case ADDR_EXPR:
2811 if (currop->op0)
2813 gcc_assert (is_gimple_min_invariant (currop->op0));
2814 return currop->op0;
2816 /* Fallthrough. */
2817 case REALPART_EXPR:
2818 case IMAGPART_EXPR:
2819 case VIEW_CONVERT_EXPR:
2821 tree folded;
2822 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2823 operand,
2824 stmts, domstmt);
2825 if (!genop0)
2826 return NULL_TREE;
2827 folded = fold_build1 (currop->opcode, currop->type,
2828 genop0);
2829 return folded;
2831 break;
2832 case BIT_FIELD_REF:
2834 tree folded;
2835 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2836 stmts, domstmt);
2837 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2838 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2839 tree genop1;
2840 tree genop2;
2842 if (!genop0)
2843 return NULL_TREE;
2844 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2845 if (!genop1)
2846 return NULL_TREE;
2847 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2848 if (!genop2)
2849 return NULL_TREE;
2850 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2851 genop2);
2852 return folded;
2855 /* For array ref vn_reference_op's, operand 1 of the array ref
2856 is op0 of the reference op and operand 3 of the array ref is
2857 op1. */
2858 case ARRAY_RANGE_REF:
2859 case ARRAY_REF:
2861 tree genop0;
2862 tree genop1 = currop->op0;
2863 pre_expr op1expr;
2864 tree genop2 = currop->op1;
2865 pre_expr op2expr;
2866 tree genop3 = currop->op2;
2867 pre_expr op3expr;
2868 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2869 stmts, domstmt);
2870 if (!genop0)
2871 return NULL_TREE;
2872 op1expr = get_or_alloc_expr_for (genop1);
2873 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2874 if (!genop1)
2875 return NULL_TREE;
2876 if (genop2)
2878 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2879 /* Drop zero minimum index if redundant. */
2880 if (integer_zerop (genop2)
2881 && (!domain_type
2882 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2883 genop2 = NULL_TREE;
2884 else
2886 op2expr = get_or_alloc_expr_for (genop2);
2887 genop2 = find_or_generate_expression (block, op2expr, stmts,
2888 domstmt);
2889 if (!genop2)
2890 return NULL_TREE;
2893 if (genop3)
2895 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2896 /* We can't always put a size in units of the element alignment
2897 here as the element alignment may be not visible. See
2898 PR43783. Simply drop the element size for constant
2899 sizes. */
2900 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2901 genop3 = NULL_TREE;
2902 else
2904 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2905 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2906 op3expr = get_or_alloc_expr_for (genop3);
2907 genop3 = find_or_generate_expression (block, op3expr, stmts,
2908 domstmt);
2909 if (!genop3)
2910 return NULL_TREE;
2913 return build4 (currop->opcode, currop->type, genop0, genop1,
2914 genop2, genop3);
2916 case COMPONENT_REF:
2918 tree op0;
2919 tree op1;
2920 tree genop2 = currop->op1;
2921 pre_expr op2expr;
2922 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2923 stmts, domstmt);
2924 if (!op0)
2925 return NULL_TREE;
2926 /* op1 should be a FIELD_DECL, which are represented by
2927 themselves. */
2928 op1 = currop->op0;
2929 if (genop2)
2931 op2expr = get_or_alloc_expr_for (genop2);
2932 genop2 = find_or_generate_expression (block, op2expr, stmts,
2933 domstmt);
2934 if (!genop2)
2935 return NULL_TREE;
2938 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2939 genop2);
2941 break;
2942 case SSA_NAME:
2944 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2945 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2946 return genop;
2948 case STRING_CST:
2949 case INTEGER_CST:
2950 case COMPLEX_CST:
2951 case VECTOR_CST:
2952 case REAL_CST:
2953 case CONSTRUCTOR:
2954 case VAR_DECL:
2955 case PARM_DECL:
2956 case CONST_DECL:
2957 case RESULT_DECL:
2958 case FUNCTION_DECL:
2959 return currop->op0;
2961 default:
2962 gcc_unreachable ();
2966 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2967 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2968 trying to rename aggregates into ssa form directly, which is a no no.
2970 Thus, this routine doesn't create temporaries, it just builds a
2971 single access expression for the array, calling
2972 find_or_generate_expression to build the innermost pieces.
2974 This function is a subroutine of create_expression_by_pieces, and
2975 should not be called on it's own unless you really know what you
2976 are doing. */
2978 static tree
2979 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2980 gimple_seq *stmts, gimple domstmt)
2982 unsigned int op = 0;
2983 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2986 /* Find a leader for an expression, or generate one using
2987 create_expression_by_pieces if it's ANTIC but
2988 complex.
2989 BLOCK is the basic_block we are looking for leaders in.
2990 EXPR is the expression to find a leader or generate for.
2991 STMTS is the statement list to put the inserted expressions on.
2992 Returns the SSA_NAME of the LHS of the generated expression or the
2993 leader.
2994 DOMSTMT if non-NULL is a statement that should be dominated by
2995 all uses in the generated expression. If DOMSTMT is non-NULL this
2996 routine can fail and return NULL_TREE. Otherwise it will assert
2997 on failure. */
2999 static tree
3000 find_or_generate_expression (basic_block block, pre_expr expr,
3001 gimple_seq *stmts, gimple domstmt)
3003 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
3004 get_expr_value_id (expr), domstmt);
3005 tree genop = NULL;
3006 if (leader)
3008 if (leader->kind == NAME)
3009 genop = PRE_EXPR_NAME (leader);
3010 else if (leader->kind == CONSTANT)
3011 genop = PRE_EXPR_CONSTANT (leader);
3014 /* If it's still NULL, it must be a complex expression, so generate
3015 it recursively. Not so if inserting expressions for values generated
3016 by SCCVN. */
3017 if (genop == NULL
3018 && !domstmt)
3020 bitmap_set_t exprset;
3021 unsigned int lookfor = get_expr_value_id (expr);
3022 bool handled = false;
3023 bitmap_iterator bi;
3024 unsigned int i;
3026 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3027 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3029 pre_expr temp = expression_for_id (i);
3030 if (temp->kind != NAME)
3032 handled = true;
3033 genop = create_expression_by_pieces (block, temp, stmts,
3034 domstmt,
3035 get_expr_type (expr));
3036 break;
3039 if (!handled && domstmt)
3040 return NULL_TREE;
3042 gcc_assert (handled);
3044 return genop;
3047 #define NECESSARY GF_PLF_1
3049 /* Create an expression in pieces, so that we can handle very complex
3050 expressions that may be ANTIC, but not necessary GIMPLE.
3051 BLOCK is the basic block the expression will be inserted into,
3052 EXPR is the expression to insert (in value form)
3053 STMTS is a statement list to append the necessary insertions into.
3055 This function will die if we hit some value that shouldn't be
3056 ANTIC but is (IE there is no leader for it, or its components).
3057 This function may also generate expressions that are themselves
3058 partially or fully redundant. Those that are will be either made
3059 fully redundant during the next iteration of insert (for partially
3060 redundant ones), or eliminated by eliminate (for fully redundant
3061 ones).
3063 If DOMSTMT is non-NULL then we make sure that all uses in the
3064 expressions dominate that statement. In this case the function
3065 can return NULL_TREE to signal failure. */
3067 static tree
3068 create_expression_by_pieces (basic_block block, pre_expr expr,
3069 gimple_seq *stmts, gimple domstmt, tree type)
3071 tree temp, name;
3072 tree folded;
3073 gimple_seq forced_stmts = NULL;
3074 unsigned int value_id;
3075 gimple_stmt_iterator gsi;
3076 tree exprtype = type ? type : get_expr_type (expr);
3077 pre_expr nameexpr;
3078 gimple newstmt;
3080 switch (expr->kind)
3082 /* We may hit the NAME/CONSTANT case if we have to convert types
3083 that value numbering saw through. */
3084 case NAME:
3085 folded = PRE_EXPR_NAME (expr);
3086 break;
3087 case CONSTANT:
3088 folded = PRE_EXPR_CONSTANT (expr);
3089 break;
3090 case REFERENCE:
3092 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3093 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3095 break;
3096 case NARY:
3098 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3099 switch (nary->length)
3101 case 2:
3103 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3104 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
3105 tree genop1 = find_or_generate_expression (block, op1,
3106 stmts, domstmt);
3107 tree genop2 = find_or_generate_expression (block, op2,
3108 stmts, domstmt);
3109 if (!genop1 || !genop2)
3110 return NULL_TREE;
3111 /* Ensure op2 is a sizetype for POINTER_PLUS_EXPR. It
3112 may be a constant with the wrong type. */
3113 if (nary->opcode == POINTER_PLUS_EXPR)
3115 genop1 = fold_convert (nary->type, genop1);
3116 genop2 = fold_convert (sizetype, genop2);
3118 else
3120 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3121 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
3124 folded = fold_build2 (nary->opcode, nary->type,
3125 genop1, genop2);
3127 break;
3128 case 1:
3130 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3131 tree genop1 = find_or_generate_expression (block, op1,
3132 stmts, domstmt);
3133 if (!genop1)
3134 return NULL_TREE;
3135 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3137 folded = fold_build1 (nary->opcode, nary->type,
3138 genop1);
3140 break;
3141 default:
3142 return NULL_TREE;
3145 break;
3146 default:
3147 return NULL_TREE;
3150 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3151 folded = fold_convert (exprtype, folded);
3153 /* Force the generated expression to be a sequence of GIMPLE
3154 statements.
3155 We have to call unshare_expr because force_gimple_operand may
3156 modify the tree we pass to it. */
3157 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3158 false, NULL);
3160 /* If we have any intermediate expressions to the value sets, add them
3161 to the value sets and chain them in the instruction stream. */
3162 if (forced_stmts)
3164 gsi = gsi_start (forced_stmts);
3165 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3167 gimple stmt = gsi_stmt (gsi);
3168 tree forcedname = gimple_get_lhs (stmt);
3169 pre_expr nameexpr;
3171 if (TREE_CODE (forcedname) == SSA_NAME)
3173 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3174 VN_INFO_GET (forcedname)->valnum = forcedname;
3175 VN_INFO (forcedname)->value_id = get_next_value_id ();
3176 nameexpr = get_or_alloc_expr_for_name (forcedname);
3177 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3178 if (!in_fre)
3179 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3180 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3182 mark_symbols_for_renaming (stmt);
3184 gimple_seq_add_seq (stmts, forced_stmts);
3187 /* Build and insert the assignment of the end result to the temporary
3188 that we will return. */
3189 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3191 pretemp = create_tmp_reg (exprtype, "pretmp");
3192 get_var_ann (pretemp);
3195 temp = pretemp;
3196 add_referenced_var (temp);
3198 newstmt = gimple_build_assign (temp, folded);
3199 name = make_ssa_name (temp, newstmt);
3200 gimple_assign_set_lhs (newstmt, name);
3201 gimple_set_plf (newstmt, NECESSARY, false);
3203 gimple_seq_add_stmt (stmts, newstmt);
3204 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3206 /* All the symbols in NEWEXPR should be put into SSA form. */
3207 mark_symbols_for_renaming (newstmt);
3209 /* Add a value number to the temporary.
3210 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3211 we are creating the expression by pieces, and this particular piece of
3212 the expression may have been represented. There is no harm in replacing
3213 here. */
3214 VN_INFO_GET (name)->valnum = name;
3215 value_id = get_expr_value_id (expr);
3216 VN_INFO (name)->value_id = value_id;
3217 nameexpr = get_or_alloc_expr_for_name (name);
3218 add_to_value (value_id, nameexpr);
3219 if (NEW_SETS (block))
3220 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3221 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3223 pre_stats.insertions++;
3224 if (dump_file && (dump_flags & TDF_DETAILS))
3226 fprintf (dump_file, "Inserted ");
3227 print_gimple_stmt (dump_file, newstmt, 0, 0);
3228 fprintf (dump_file, " in predecessor %d\n", block->index);
3231 return name;
3235 /* Returns true if we want to inhibit the insertions of PHI nodes
3236 for the given EXPR for basic block BB (a member of a loop).
3237 We want to do this, when we fear that the induction variable we
3238 create might inhibit vectorization. */
3240 static bool
3241 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3243 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3244 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3245 vn_reference_op_t op;
3246 unsigned i;
3248 /* If we aren't going to vectorize we don't inhibit anything. */
3249 if (!flag_tree_vectorize)
3250 return false;
3252 /* Otherwise we inhibit the insertion when the address of the
3253 memory reference is a simple induction variable. In other
3254 cases the vectorizer won't do anything anyway (either it's
3255 loop invariant or a complicated expression). */
3256 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3258 switch (op->opcode)
3260 case ARRAY_REF:
3261 case ARRAY_RANGE_REF:
3262 if (TREE_CODE (op->op0) != SSA_NAME)
3263 break;
3264 /* Fallthru. */
3265 case SSA_NAME:
3267 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3268 affine_iv iv;
3269 /* Default defs are loop invariant. */
3270 if (!defbb)
3271 break;
3272 /* Defined outside this loop, also loop invariant. */
3273 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3274 break;
3275 /* If it's a simple induction variable inhibit insertion,
3276 the vectorizer might be interested in this one. */
3277 if (simple_iv (bb->loop_father, bb->loop_father,
3278 op->op0, &iv, true))
3279 return true;
3280 /* No simple IV, vectorizer can't do anything, hence no
3281 reason to inhibit the transformation for this operand. */
3282 break;
3284 default:
3285 break;
3288 return false;
3291 /* Insert the to-be-made-available values of expression EXPRNUM for each
3292 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3293 merge the result with a phi node, given the same value number as
3294 NODE. Return true if we have inserted new stuff. */
3296 static bool
3297 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3298 pre_expr *avail)
3300 pre_expr expr = expression_for_id (exprnum);
3301 pre_expr newphi;
3302 unsigned int val = get_expr_value_id (expr);
3303 edge pred;
3304 bool insertions = false;
3305 bool nophi = false;
3306 basic_block bprime;
3307 pre_expr eprime;
3308 edge_iterator ei;
3309 tree type = get_expr_type (expr);
3310 tree temp;
3311 gimple phi;
3313 if (dump_file && (dump_flags & TDF_DETAILS))
3315 fprintf (dump_file, "Found partial redundancy for expression ");
3316 print_pre_expr (dump_file, expr);
3317 fprintf (dump_file, " (%04d)\n", val);
3320 /* Make sure we aren't creating an induction variable. */
3321 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3323 bool firstinsideloop = false;
3324 bool secondinsideloop = false;
3325 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3326 EDGE_PRED (block, 0)->src);
3327 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3328 EDGE_PRED (block, 1)->src);
3329 /* Induction variables only have one edge inside the loop. */
3330 if ((firstinsideloop ^ secondinsideloop)
3331 && (expr->kind != REFERENCE
3332 || inhibit_phi_insertion (block, expr)))
3334 if (dump_file && (dump_flags & TDF_DETAILS))
3335 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3336 nophi = true;
3340 /* Make the necessary insertions. */
3341 FOR_EACH_EDGE (pred, ei, block->preds)
3343 gimple_seq stmts = NULL;
3344 tree builtexpr;
3345 bprime = pred->src;
3346 eprime = avail[bprime->index];
3348 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3350 builtexpr = create_expression_by_pieces (bprime,
3351 eprime,
3352 &stmts, NULL,
3353 type);
3354 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3355 gsi_insert_seq_on_edge (pred, stmts);
3356 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3357 insertions = true;
3359 else if (eprime->kind == CONSTANT)
3361 /* Constants may not have the right type, fold_convert
3362 should give us back a constant with the right type.
3364 tree constant = PRE_EXPR_CONSTANT (eprime);
3365 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3367 tree builtexpr = fold_convert (type, constant);
3368 if (!is_gimple_min_invariant (builtexpr))
3370 tree forcedexpr = force_gimple_operand (builtexpr,
3371 &stmts, true,
3372 NULL);
3373 if (!is_gimple_min_invariant (forcedexpr))
3375 if (forcedexpr != builtexpr)
3377 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3378 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3380 if (stmts)
3382 gimple_stmt_iterator gsi;
3383 gsi = gsi_start (stmts);
3384 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3386 gimple stmt = gsi_stmt (gsi);
3387 tree lhs = gimple_get_lhs (stmt);
3388 if (TREE_CODE (lhs) == SSA_NAME)
3389 bitmap_set_bit (inserted_exprs,
3390 SSA_NAME_VERSION (lhs));
3391 gimple_set_plf (stmt, NECESSARY, false);
3393 gsi_insert_seq_on_edge (pred, stmts);
3395 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3398 else
3399 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3402 else if (eprime->kind == NAME)
3404 /* We may have to do a conversion because our value
3405 numbering can look through types in certain cases, but
3406 our IL requires all operands of a phi node have the same
3407 type. */
3408 tree name = PRE_EXPR_NAME (eprime);
3409 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3411 tree builtexpr;
3412 tree forcedexpr;
3413 builtexpr = fold_convert (type, name);
3414 forcedexpr = force_gimple_operand (builtexpr,
3415 &stmts, true,
3416 NULL);
3418 if (forcedexpr != name)
3420 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3421 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3424 if (stmts)
3426 gimple_stmt_iterator gsi;
3427 gsi = gsi_start (stmts);
3428 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3430 gimple stmt = gsi_stmt (gsi);
3431 tree lhs = gimple_get_lhs (stmt);
3432 if (TREE_CODE (lhs) == SSA_NAME)
3433 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3434 gimple_set_plf (stmt, NECESSARY, false);
3436 gsi_insert_seq_on_edge (pred, stmts);
3438 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3442 /* If we didn't want a phi node, and we made insertions, we still have
3443 inserted new stuff, and thus return true. If we didn't want a phi node,
3444 and didn't make insertions, we haven't added anything new, so return
3445 false. */
3446 if (nophi && insertions)
3447 return true;
3448 else if (nophi && !insertions)
3449 return false;
3451 /* Now build a phi for the new variable. */
3452 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3454 prephitemp = create_tmp_var (type, "prephitmp");
3455 get_var_ann (prephitemp);
3458 temp = prephitemp;
3459 add_referenced_var (temp);
3461 if (TREE_CODE (type) == COMPLEX_TYPE
3462 || TREE_CODE (type) == VECTOR_TYPE)
3463 DECL_GIMPLE_REG_P (temp) = 1;
3464 phi = create_phi_node (temp, block);
3466 gimple_set_plf (phi, NECESSARY, false);
3467 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3468 VN_INFO (gimple_phi_result (phi))->value_id = val;
3469 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3470 FOR_EACH_EDGE (pred, ei, block->preds)
3472 pre_expr ae = avail[pred->src->index];
3473 gcc_assert (get_expr_type (ae) == type
3474 || useless_type_conversion_p (type, get_expr_type (ae)));
3475 if (ae->kind == CONSTANT)
3476 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3477 else
3478 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3479 UNKNOWN_LOCATION);
3482 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3483 add_to_value (val, newphi);
3485 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3486 this insertion, since we test for the existence of this value in PHI_GEN
3487 before proceeding with the partial redundancy checks in insert_aux.
3489 The value may exist in AVAIL_OUT, in particular, it could be represented
3490 by the expression we are trying to eliminate, in which case we want the
3491 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3492 inserted there.
3494 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3495 this block, because if it did, it would have existed in our dominator's
3496 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3499 bitmap_insert_into_set (PHI_GEN (block), newphi);
3500 bitmap_value_replace_in_set (AVAIL_OUT (block),
3501 newphi);
3502 bitmap_insert_into_set (NEW_SETS (block),
3503 newphi);
3505 if (dump_file && (dump_flags & TDF_DETAILS))
3507 fprintf (dump_file, "Created phi ");
3508 print_gimple_stmt (dump_file, phi, 0, 0);
3509 fprintf (dump_file, " in block %d\n", block->index);
3511 pre_stats.phis++;
3512 return true;
3517 /* Perform insertion of partially redundant values.
3518 For BLOCK, do the following:
3519 1. Propagate the NEW_SETS of the dominator into the current block.
3520 If the block has multiple predecessors,
3521 2a. Iterate over the ANTIC expressions for the block to see if
3522 any of them are partially redundant.
3523 2b. If so, insert them into the necessary predecessors to make
3524 the expression fully redundant.
3525 2c. Insert a new PHI merging the values of the predecessors.
3526 2d. Insert the new PHI, and the new expressions, into the
3527 NEW_SETS set.
3528 3. Recursively call ourselves on the dominator children of BLOCK.
3530 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3531 do_regular_insertion and do_partial_insertion.
3535 static bool
3536 do_regular_insertion (basic_block block, basic_block dom)
3538 bool new_stuff = false;
3539 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3540 pre_expr expr;
3541 int i;
3543 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3545 if (expr->kind != NAME)
3547 pre_expr *avail;
3548 unsigned int val;
3549 bool by_some = false;
3550 bool cant_insert = false;
3551 bool all_same = true;
3552 pre_expr first_s = NULL;
3553 edge pred;
3554 basic_block bprime;
3555 pre_expr eprime = NULL;
3556 edge_iterator ei;
3557 pre_expr edoubleprime = NULL;
3558 bool do_insertion = false;
3560 val = get_expr_value_id (expr);
3561 if (bitmap_set_contains_value (PHI_GEN (block), val))
3562 continue;
3563 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3565 if (dump_file && (dump_flags & TDF_DETAILS))
3566 fprintf (dump_file, "Found fully redundant value\n");
3567 continue;
3570 avail = XCNEWVEC (pre_expr, last_basic_block);
3571 FOR_EACH_EDGE (pred, ei, block->preds)
3573 unsigned int vprime;
3575 /* We should never run insertion for the exit block
3576 and so not come across fake pred edges. */
3577 gcc_assert (!(pred->flags & EDGE_FAKE));
3578 bprime = pred->src;
3579 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3580 bprime, block);
3582 /* eprime will generally only be NULL if the
3583 value of the expression, translated
3584 through the PHI for this predecessor, is
3585 undefined. If that is the case, we can't
3586 make the expression fully redundant,
3587 because its value is undefined along a
3588 predecessor path. We can thus break out
3589 early because it doesn't matter what the
3590 rest of the results are. */
3591 if (eprime == NULL)
3593 cant_insert = true;
3594 break;
3597 eprime = fully_constant_expression (eprime);
3598 vprime = get_expr_value_id (eprime);
3599 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3600 vprime, NULL);
3601 if (edoubleprime == NULL)
3603 avail[bprime->index] = eprime;
3604 all_same = false;
3606 else
3608 avail[bprime->index] = edoubleprime;
3609 by_some = true;
3610 /* We want to perform insertions to remove a redundancy on
3611 a path in the CFG we want to optimize for speed. */
3612 if (optimize_edge_for_speed_p (pred))
3613 do_insertion = true;
3614 if (first_s == NULL)
3615 first_s = edoubleprime;
3616 else if (!pre_expr_eq (first_s, edoubleprime))
3617 all_same = false;
3620 /* If we can insert it, it's not the same value
3621 already existing along every predecessor, and
3622 it's defined by some predecessor, it is
3623 partially redundant. */
3624 if (!cant_insert && !all_same && by_some)
3626 if (!do_insertion)
3628 if (dump_file && (dump_flags & TDF_DETAILS))
3630 fprintf (dump_file, "Skipping partial redundancy for "
3631 "expression ");
3632 print_pre_expr (dump_file, expr);
3633 fprintf (dump_file, " (%04d), no redundancy on to be "
3634 "optimized for speed edge\n", val);
3637 else if (dbg_cnt (treepre_insert)
3638 && insert_into_preds_of_block (block,
3639 get_expression_id (expr),
3640 avail))
3641 new_stuff = true;
3643 /* If all edges produce the same value and that value is
3644 an invariant, then the PHI has the same value on all
3645 edges. Note this. */
3646 else if (!cant_insert && all_same && eprime
3647 && (edoubleprime->kind == CONSTANT
3648 || edoubleprime->kind == NAME)
3649 && !value_id_constant_p (val))
3651 unsigned int j;
3652 bitmap_iterator bi;
3653 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3654 value_expressions, val);
3656 unsigned int new_val = get_expr_value_id (edoubleprime);
3657 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3659 pre_expr expr = expression_for_id (j);
3661 if (expr->kind == NAME)
3663 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3664 /* Just reset the value id and valnum so it is
3665 the same as the constant we have discovered. */
3666 if (edoubleprime->kind == CONSTANT)
3668 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3669 pre_stats.constified++;
3671 else
3672 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3673 info->value_id = new_val;
3677 free (avail);
3681 VEC_free (pre_expr, heap, exprs);
3682 return new_stuff;
3686 /* Perform insertion for partially anticipatable expressions. There
3687 is only one case we will perform insertion for these. This case is
3688 if the expression is partially anticipatable, and fully available.
3689 In this case, we know that putting it earlier will enable us to
3690 remove the later computation. */
3693 static bool
3694 do_partial_partial_insertion (basic_block block, basic_block dom)
3696 bool new_stuff = false;
3697 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3698 pre_expr expr;
3699 int i;
3701 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3703 if (expr->kind != NAME)
3705 pre_expr *avail;
3706 unsigned int val;
3707 bool by_all = true;
3708 bool cant_insert = false;
3709 edge pred;
3710 basic_block bprime;
3711 pre_expr eprime = NULL;
3712 edge_iterator ei;
3714 val = get_expr_value_id (expr);
3715 if (bitmap_set_contains_value (PHI_GEN (block), val))
3716 continue;
3717 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3718 continue;
3720 avail = XCNEWVEC (pre_expr, last_basic_block);
3721 FOR_EACH_EDGE (pred, ei, block->preds)
3723 unsigned int vprime;
3724 pre_expr edoubleprime;
3726 /* We should never run insertion for the exit block
3727 and so not come across fake pred edges. */
3728 gcc_assert (!(pred->flags & EDGE_FAKE));
3729 bprime = pred->src;
3730 eprime = phi_translate (expr, ANTIC_IN (block),
3731 PA_IN (block),
3732 bprime, block);
3734 /* eprime will generally only be NULL if the
3735 value of the expression, translated
3736 through the PHI for this predecessor, is
3737 undefined. If that is the case, we can't
3738 make the expression fully redundant,
3739 because its value is undefined along a
3740 predecessor path. We can thus break out
3741 early because it doesn't matter what the
3742 rest of the results are. */
3743 if (eprime == NULL)
3745 cant_insert = true;
3746 break;
3749 eprime = fully_constant_expression (eprime);
3750 vprime = get_expr_value_id (eprime);
3751 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3752 vprime, NULL);
3753 if (edoubleprime == NULL)
3755 by_all = false;
3756 break;
3758 else
3759 avail[bprime->index] = edoubleprime;
3763 /* If we can insert it, it's not the same value
3764 already existing along every predecessor, and
3765 it's defined by some predecessor, it is
3766 partially redundant. */
3767 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3769 pre_stats.pa_insert++;
3770 if (insert_into_preds_of_block (block, get_expression_id (expr),
3771 avail))
3772 new_stuff = true;
3774 free (avail);
3778 VEC_free (pre_expr, heap, exprs);
3779 return new_stuff;
3782 static bool
3783 insert_aux (basic_block block)
3785 basic_block son;
3786 bool new_stuff = false;
3788 if (block)
3790 basic_block dom;
3791 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3792 if (dom)
3794 unsigned i;
3795 bitmap_iterator bi;
3796 bitmap_set_t newset = NEW_SETS (dom);
3797 if (newset)
3799 /* Note that we need to value_replace both NEW_SETS, and
3800 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3801 represented by some non-simple expression here that we want
3802 to replace it with. */
3803 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3805 pre_expr expr = expression_for_id (i);
3806 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3807 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3810 if (!single_pred_p (block))
3812 new_stuff |= do_regular_insertion (block, dom);
3813 if (do_partial_partial)
3814 new_stuff |= do_partial_partial_insertion (block, dom);
3818 for (son = first_dom_son (CDI_DOMINATORS, block);
3819 son;
3820 son = next_dom_son (CDI_DOMINATORS, son))
3822 new_stuff |= insert_aux (son);
3825 return new_stuff;
3828 /* Perform insertion of partially redundant values. */
3830 static void
3831 insert (void)
3833 bool new_stuff = true;
3834 basic_block bb;
3835 int num_iterations = 0;
3837 FOR_ALL_BB (bb)
3838 NEW_SETS (bb) = bitmap_set_new ();
3840 while (new_stuff)
3842 num_iterations++;
3843 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3845 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3849 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3851 static void
3852 add_to_exp_gen (basic_block block, tree op)
3854 if (!in_fre)
3856 pre_expr result;
3857 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3858 return;
3859 result = get_or_alloc_expr_for_name (op);
3860 bitmap_value_insert_into_set (EXP_GEN (block), result);
3864 /* Create value ids for PHI in BLOCK. */
3866 static void
3867 make_values_for_phi (gimple phi, basic_block block)
3869 tree result = gimple_phi_result (phi);
3871 /* We have no need for virtual phis, as they don't represent
3872 actual computations. */
3873 if (is_gimple_reg (result))
3875 pre_expr e = get_or_alloc_expr_for_name (result);
3876 add_to_value (get_expr_value_id (e), e);
3877 bitmap_insert_into_set (PHI_GEN (block), e);
3878 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3879 if (!in_fre)
3881 unsigned i;
3882 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3884 tree arg = gimple_phi_arg_def (phi, i);
3885 if (TREE_CODE (arg) == SSA_NAME)
3887 e = get_or_alloc_expr_for_name (arg);
3888 add_to_value (get_expr_value_id (e), e);
3895 /* Compute the AVAIL set for all basic blocks.
3897 This function performs value numbering of the statements in each basic
3898 block. The AVAIL sets are built from information we glean while doing
3899 this value numbering, since the AVAIL sets contain only one entry per
3900 value.
3902 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3903 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3905 static void
3906 compute_avail (void)
3909 basic_block block, son;
3910 basic_block *worklist;
3911 size_t sp = 0;
3912 unsigned i;
3914 /* We pretend that default definitions are defined in the entry block.
3915 This includes function arguments and the static chain decl. */
3916 for (i = 1; i < num_ssa_names; ++i)
3918 tree name = ssa_name (i);
3919 pre_expr e;
3920 if (!name
3921 || !SSA_NAME_IS_DEFAULT_DEF (name)
3922 || has_zero_uses (name)
3923 || !is_gimple_reg (name))
3924 continue;
3926 e = get_or_alloc_expr_for_name (name);
3927 add_to_value (get_expr_value_id (e), e);
3928 if (!in_fre)
3929 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3930 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3933 /* Allocate the worklist. */
3934 worklist = XNEWVEC (basic_block, n_basic_blocks);
3936 /* Seed the algorithm by putting the dominator children of the entry
3937 block on the worklist. */
3938 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3939 son;
3940 son = next_dom_son (CDI_DOMINATORS, son))
3941 worklist[sp++] = son;
3943 /* Loop until the worklist is empty. */
3944 while (sp)
3946 gimple_stmt_iterator gsi;
3947 gimple stmt;
3948 basic_block dom;
3949 unsigned int stmt_uid = 1;
3951 /* Pick a block from the worklist. */
3952 block = worklist[--sp];
3954 /* Initially, the set of available values in BLOCK is that of
3955 its immediate dominator. */
3956 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3957 if (dom)
3958 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3960 /* Generate values for PHI nodes. */
3961 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3962 make_values_for_phi (gsi_stmt (gsi), block);
3964 BB_MAY_NOTRETURN (block) = 0;
3966 /* Now compute value numbers and populate value sets with all
3967 the expressions computed in BLOCK. */
3968 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3970 ssa_op_iter iter;
3971 tree op;
3973 stmt = gsi_stmt (gsi);
3974 gimple_set_uid (stmt, stmt_uid++);
3976 /* Cache whether the basic-block has any non-visible side-effect
3977 or control flow.
3978 If this isn't a call or it is the last stmt in the
3979 basic-block then the CFG represents things correctly. */
3980 if (is_gimple_call (stmt)
3981 && !stmt_ends_bb_p (stmt))
3983 /* Non-looping const functions always return normally.
3984 Otherwise the call might not return or have side-effects
3985 that forbids hoisting possibly trapping expressions
3986 before it. */
3987 int flags = gimple_call_flags (stmt);
3988 if (!(flags & ECF_CONST)
3989 || (flags & ECF_LOOPING_CONST_OR_PURE))
3990 BB_MAY_NOTRETURN (block) = 1;
3993 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3995 pre_expr e = get_or_alloc_expr_for_name (op);
3997 add_to_value (get_expr_value_id (e), e);
3998 if (!in_fre)
3999 bitmap_insert_into_set (TMP_GEN (block), e);
4000 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
4003 if (gimple_has_volatile_ops (stmt)
4004 || stmt_could_throw_p (stmt))
4005 continue;
4007 switch (gimple_code (stmt))
4009 case GIMPLE_RETURN:
4010 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4011 add_to_exp_gen (block, op);
4012 continue;
4014 case GIMPLE_CALL:
4016 vn_reference_t ref;
4017 unsigned int i;
4018 vn_reference_op_t vro;
4019 pre_expr result = NULL;
4020 VEC(vn_reference_op_s, heap) *ops = NULL;
4022 if (!can_value_number_call (stmt))
4023 continue;
4025 copy_reference_ops_from_call (stmt, &ops);
4026 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4027 gimple_expr_type (stmt),
4028 ops, &ref, VN_NOWALK);
4029 VEC_free (vn_reference_op_s, heap, ops);
4030 if (!ref)
4031 continue;
4033 for (i = 0; VEC_iterate (vn_reference_op_s,
4034 ref->operands, i,
4035 vro); i++)
4037 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4038 add_to_exp_gen (block, vro->op0);
4039 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4040 add_to_exp_gen (block, vro->op1);
4041 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4042 add_to_exp_gen (block, vro->op2);
4044 result = (pre_expr) pool_alloc (pre_expr_pool);
4045 result->kind = REFERENCE;
4046 result->id = 0;
4047 PRE_EXPR_REFERENCE (result) = ref;
4049 get_or_alloc_expression_id (result);
4050 add_to_value (get_expr_value_id (result), result);
4051 if (!in_fre)
4052 bitmap_value_insert_into_set (EXP_GEN (block), result);
4053 continue;
4056 case GIMPLE_ASSIGN:
4058 pre_expr result = NULL;
4059 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4061 case tcc_unary:
4062 case tcc_binary:
4063 case tcc_comparison:
4065 vn_nary_op_t nary;
4066 unsigned int i;
4068 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4069 gimple_assign_rhs_code (stmt),
4070 gimple_expr_type (stmt),
4071 gimple_assign_rhs1 (stmt),
4072 gimple_assign_rhs2 (stmt),
4073 NULL_TREE, NULL_TREE, &nary);
4075 if (!nary)
4076 continue;
4078 for (i = 0; i < nary->length; i++)
4079 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4080 add_to_exp_gen (block, nary->op[i]);
4082 result = (pre_expr) pool_alloc (pre_expr_pool);
4083 result->kind = NARY;
4084 result->id = 0;
4085 PRE_EXPR_NARY (result) = nary;
4086 break;
4089 case tcc_declaration:
4090 case tcc_reference:
4092 vn_reference_t ref;
4093 unsigned int i;
4094 vn_reference_op_t vro;
4096 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4097 gimple_vuse (stmt),
4098 VN_WALK, &ref);
4099 if (!ref)
4100 continue;
4102 for (i = 0; VEC_iterate (vn_reference_op_s,
4103 ref->operands, i,
4104 vro); i++)
4106 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4107 add_to_exp_gen (block, vro->op0);
4108 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4109 add_to_exp_gen (block, vro->op1);
4110 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4111 add_to_exp_gen (block, vro->op2);
4113 result = (pre_expr) pool_alloc (pre_expr_pool);
4114 result->kind = REFERENCE;
4115 result->id = 0;
4116 PRE_EXPR_REFERENCE (result) = ref;
4117 break;
4120 default:
4121 /* For any other statement that we don't
4122 recognize, simply add all referenced
4123 SSA_NAMEs to EXP_GEN. */
4124 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4125 add_to_exp_gen (block, op);
4126 continue;
4129 get_or_alloc_expression_id (result);
4130 add_to_value (get_expr_value_id (result), result);
4131 if (!in_fre)
4132 bitmap_value_insert_into_set (EXP_GEN (block), result);
4134 continue;
4136 default:
4137 break;
4141 /* Put the dominator children of BLOCK on the worklist of blocks
4142 to compute available sets for. */
4143 for (son = first_dom_son (CDI_DOMINATORS, block);
4144 son;
4145 son = next_dom_son (CDI_DOMINATORS, son))
4146 worklist[sp++] = son;
4149 free (worklist);
4152 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4153 than the available expressions for it. The insertion point is
4154 right before the first use in STMT. Returns the SSA_NAME that should
4155 be used for replacement. */
4157 static tree
4158 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4160 basic_block bb = gimple_bb (stmt);
4161 gimple_stmt_iterator gsi;
4162 gimple_seq stmts = NULL;
4163 tree expr;
4164 pre_expr e;
4166 /* First create a value expression from the expression we want
4167 to insert and associate it with the value handle for SSA_VN. */
4168 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4169 if (e == NULL)
4170 return NULL_TREE;
4172 /* Then use create_expression_by_pieces to generate a valid
4173 expression to insert at this point of the IL stream. */
4174 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4175 if (expr == NULL_TREE)
4176 return NULL_TREE;
4177 gsi = gsi_for_stmt (stmt);
4178 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4180 return expr;
4183 /* Eliminate fully redundant computations. */
4185 static unsigned int
4186 eliminate (void)
4188 VEC (gimple, heap) *to_remove = NULL;
4189 VEC (gimple, heap) *to_update = NULL;
4190 basic_block b;
4191 unsigned int todo = 0;
4192 gimple_stmt_iterator gsi;
4193 gimple stmt;
4194 unsigned i;
4196 FOR_EACH_BB (b)
4198 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4200 stmt = gsi_stmt (gsi);
4202 /* Lookup the RHS of the expression, see if we have an
4203 available computation for it. If so, replace the RHS with
4204 the available computation. */
4205 if (gimple_has_lhs (stmt)
4206 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4207 && !gimple_assign_ssa_name_copy_p (stmt)
4208 && (!gimple_assign_single_p (stmt)
4209 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4210 && !gimple_has_volatile_ops (stmt)
4211 && !has_zero_uses (gimple_get_lhs (stmt)))
4213 tree lhs = gimple_get_lhs (stmt);
4214 tree rhs = NULL_TREE;
4215 tree sprime = NULL;
4216 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4217 pre_expr sprimeexpr;
4219 if (gimple_assign_single_p (stmt))
4220 rhs = gimple_assign_rhs1 (stmt);
4222 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4223 get_expr_value_id (lhsexpr),
4224 NULL);
4226 if (sprimeexpr)
4228 if (sprimeexpr->kind == CONSTANT)
4229 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4230 else if (sprimeexpr->kind == NAME)
4231 sprime = PRE_EXPR_NAME (sprimeexpr);
4232 else
4233 gcc_unreachable ();
4236 /* If there is no existing leader but SCCVN knows this
4237 value is constant, use that constant. */
4238 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4240 sprime = VN_INFO (lhs)->valnum;
4241 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4242 TREE_TYPE (sprime)))
4243 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4245 if (dump_file && (dump_flags & TDF_DETAILS))
4247 fprintf (dump_file, "Replaced ");
4248 print_gimple_expr (dump_file, stmt, 0, 0);
4249 fprintf (dump_file, " with ");
4250 print_generic_expr (dump_file, sprime, 0);
4251 fprintf (dump_file, " in ");
4252 print_gimple_stmt (dump_file, stmt, 0, 0);
4254 pre_stats.eliminations++;
4255 propagate_tree_value_into_stmt (&gsi, sprime);
4256 stmt = gsi_stmt (gsi);
4257 update_stmt (stmt);
4258 continue;
4261 /* If there is no existing usable leader but SCCVN thinks
4262 it has an expression it wants to use as replacement,
4263 insert that. */
4264 if (!sprime || sprime == lhs)
4266 tree val = VN_INFO (lhs)->valnum;
4267 if (val != VN_TOP
4268 && TREE_CODE (val) == SSA_NAME
4269 && VN_INFO (val)->needs_insertion
4270 && can_PRE_operation (vn_get_expr_for (val)))
4271 sprime = do_SCCVN_insertion (stmt, val);
4273 if (sprime
4274 && sprime != lhs
4275 && (rhs == NULL_TREE
4276 || TREE_CODE (rhs) != SSA_NAME
4277 || may_propagate_copy (rhs, sprime)))
4279 bool can_make_abnormal_goto
4280 = is_gimple_call (stmt)
4281 && stmt_can_make_abnormal_goto (stmt);
4283 gcc_assert (sprime != rhs);
4285 if (dump_file && (dump_flags & TDF_DETAILS))
4287 fprintf (dump_file, "Replaced ");
4288 print_gimple_expr (dump_file, stmt, 0, 0);
4289 fprintf (dump_file, " with ");
4290 print_generic_expr (dump_file, sprime, 0);
4291 fprintf (dump_file, " in ");
4292 print_gimple_stmt (dump_file, stmt, 0, 0);
4295 if (TREE_CODE (sprime) == SSA_NAME)
4296 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4297 NECESSARY, true);
4298 /* We need to make sure the new and old types actually match,
4299 which may require adding a simple cast, which fold_convert
4300 will do for us. */
4301 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4302 && !useless_type_conversion_p (gimple_expr_type (stmt),
4303 TREE_TYPE (sprime)))
4304 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4306 pre_stats.eliminations++;
4307 propagate_tree_value_into_stmt (&gsi, sprime);
4308 stmt = gsi_stmt (gsi);
4309 update_stmt (stmt);
4311 /* If we removed EH side-effects from the statement, clean
4312 its EH information. */
4313 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4315 bitmap_set_bit (need_eh_cleanup,
4316 gimple_bb (stmt)->index);
4317 if (dump_file && (dump_flags & TDF_DETAILS))
4318 fprintf (dump_file, " Removed EH side-effects.\n");
4321 /* Likewise for AB side-effects. */
4322 if (can_make_abnormal_goto
4323 && !stmt_can_make_abnormal_goto (stmt))
4325 bitmap_set_bit (need_ab_cleanup,
4326 gimple_bb (stmt)->index);
4327 if (dump_file && (dump_flags & TDF_DETAILS))
4328 fprintf (dump_file, " Removed AB side-effects.\n");
4332 /* If the statement is a scalar store, see if the expression
4333 has the same value number as its rhs. If so, the store is
4334 dead. */
4335 else if (gimple_assign_single_p (stmt)
4336 && !is_gimple_reg (gimple_assign_lhs (stmt))
4337 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4338 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4340 tree rhs = gimple_assign_rhs1 (stmt);
4341 tree val;
4342 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4343 gimple_vuse (stmt), VN_WALK, NULL);
4344 if (TREE_CODE (rhs) == SSA_NAME)
4345 rhs = VN_INFO (rhs)->valnum;
4346 if (val
4347 && operand_equal_p (val, rhs, 0))
4349 if (dump_file && (dump_flags & TDF_DETAILS))
4351 fprintf (dump_file, "Deleted redundant store ");
4352 print_gimple_stmt (dump_file, stmt, 0, 0);
4355 /* Queue stmt for removal. */
4356 VEC_safe_push (gimple, heap, to_remove, stmt);
4359 /* Visit COND_EXPRs and fold the comparison with the
4360 available value-numbers. */
4361 else if (gimple_code (stmt) == GIMPLE_COND)
4363 tree op0 = gimple_cond_lhs (stmt);
4364 tree op1 = gimple_cond_rhs (stmt);
4365 tree result;
4367 if (TREE_CODE (op0) == SSA_NAME)
4368 op0 = VN_INFO (op0)->valnum;
4369 if (TREE_CODE (op1) == SSA_NAME)
4370 op1 = VN_INFO (op1)->valnum;
4371 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4372 op0, op1);
4373 if (result && TREE_CODE (result) == INTEGER_CST)
4375 if (integer_zerop (result))
4376 gimple_cond_make_false (stmt);
4377 else
4378 gimple_cond_make_true (stmt);
4379 update_stmt (stmt);
4380 todo = TODO_cleanup_cfg;
4383 /* Visit indirect calls and turn them into direct calls if
4384 possible. */
4385 if (is_gimple_call (stmt))
4387 tree orig_fn = gimple_call_fn (stmt);
4388 tree fn;
4389 if (!orig_fn)
4390 continue;
4391 if (TREE_CODE (orig_fn) == SSA_NAME)
4392 fn = VN_INFO (orig_fn)->valnum;
4393 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4394 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4395 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4396 else
4397 continue;
4398 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4399 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4400 TREE_TYPE (fn)))
4402 bool can_make_abnormal_goto
4403 = stmt_can_make_abnormal_goto (stmt);
4404 bool was_noreturn = gimple_call_noreturn_p (stmt);
4406 if (dump_file && (dump_flags & TDF_DETAILS))
4408 fprintf (dump_file, "Replacing call target with ");
4409 print_generic_expr (dump_file, fn, 0);
4410 fprintf (dump_file, " in ");
4411 print_gimple_stmt (dump_file, stmt, 0, 0);
4414 gimple_call_set_fn (stmt, fn);
4415 VEC_safe_push (gimple, heap, to_update, stmt);
4417 /* When changing a call into a noreturn call, cfg cleanup
4418 is needed to fix up the noreturn call. */
4419 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4420 todo |= TODO_cleanup_cfg;
4422 /* If we removed EH side-effects from the statement, clean
4423 its EH information. */
4424 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4426 bitmap_set_bit (need_eh_cleanup,
4427 gimple_bb (stmt)->index);
4428 if (dump_file && (dump_flags & TDF_DETAILS))
4429 fprintf (dump_file, " Removed EH side-effects.\n");
4432 /* Likewise for AB side-effects. */
4433 if (can_make_abnormal_goto
4434 && !stmt_can_make_abnormal_goto (stmt))
4436 bitmap_set_bit (need_ab_cleanup,
4437 gimple_bb (stmt)->index);
4438 if (dump_file && (dump_flags & TDF_DETAILS))
4439 fprintf (dump_file, " Removed AB side-effects.\n");
4442 /* Changing an indirect call to a direct call may
4443 have exposed different semantics. This may
4444 require an SSA update. */
4445 todo |= TODO_update_ssa_only_virtuals;
4450 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4452 gimple stmt, phi = gsi_stmt (gsi);
4453 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4454 pre_expr sprimeexpr, resexpr;
4455 gimple_stmt_iterator gsi2;
4457 /* We want to perform redundant PHI elimination. Do so by
4458 replacing the PHI with a single copy if possible.
4459 Do not touch inserted, single-argument or virtual PHIs. */
4460 if (gimple_phi_num_args (phi) == 1
4461 || !is_gimple_reg (res))
4463 gsi_next (&gsi);
4464 continue;
4467 resexpr = get_or_alloc_expr_for_name (res);
4468 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4469 get_expr_value_id (resexpr), NULL);
4470 if (sprimeexpr)
4472 if (sprimeexpr->kind == CONSTANT)
4473 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4474 else if (sprimeexpr->kind == NAME)
4475 sprime = PRE_EXPR_NAME (sprimeexpr);
4476 else
4477 gcc_unreachable ();
4479 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4481 sprime = VN_INFO (res)->valnum;
4482 if (!useless_type_conversion_p (TREE_TYPE (res),
4483 TREE_TYPE (sprime)))
4484 sprime = fold_convert (TREE_TYPE (res), sprime);
4486 if (!sprime
4487 || sprime == res)
4489 gsi_next (&gsi);
4490 continue;
4493 if (dump_file && (dump_flags & TDF_DETAILS))
4495 fprintf (dump_file, "Replaced redundant PHI node defining ");
4496 print_generic_expr (dump_file, res, 0);
4497 fprintf (dump_file, " with ");
4498 print_generic_expr (dump_file, sprime, 0);
4499 fprintf (dump_file, "\n");
4502 remove_phi_node (&gsi, false);
4504 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4505 && TREE_CODE (sprime) == SSA_NAME)
4506 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4508 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4509 sprime = fold_convert (TREE_TYPE (res), sprime);
4510 stmt = gimple_build_assign (res, sprime);
4511 SSA_NAME_DEF_STMT (res) = stmt;
4512 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4514 gsi2 = gsi_after_labels (b);
4515 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4516 /* Queue the copy for eventual removal. */
4517 VEC_safe_push (gimple, heap, to_remove, stmt);
4518 /* If we inserted this PHI node ourself, it's not an elimination. */
4519 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4520 pre_stats.phis--;
4521 else
4522 pre_stats.eliminations++;
4526 /* We cannot remove stmts during BB walk, especially not release SSA
4527 names there as this confuses the VN machinery. The stmts ending
4528 up in to_remove are either stores or simple copies. */
4529 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4531 tree lhs = gimple_assign_lhs (stmt);
4532 tree rhs = gimple_assign_rhs1 (stmt);
4533 use_operand_p use_p;
4534 gimple use_stmt;
4536 /* If there is a single use only, propagate the equivalency
4537 instead of keeping the copy. */
4538 if (TREE_CODE (lhs) == SSA_NAME
4539 && TREE_CODE (rhs) == SSA_NAME
4540 && single_imm_use (lhs, &use_p, &use_stmt)
4541 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4543 SET_USE (use_p, rhs);
4544 update_stmt (use_stmt);
4545 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4546 && TREE_CODE (rhs) == SSA_NAME)
4547 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4550 /* If this is a store or a now unused copy, remove it. */
4551 if (TREE_CODE (lhs) != SSA_NAME
4552 || has_zero_uses (lhs))
4554 basic_block bb = gimple_bb (stmt);
4555 gsi = gsi_for_stmt (stmt);
4556 unlink_stmt_vdef (stmt);
4557 gsi_remove (&gsi, true);
4558 if (gimple_purge_dead_eh_edges (bb))
4559 todo |= TODO_cleanup_cfg;
4560 if (TREE_CODE (lhs) == SSA_NAME)
4561 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4562 release_defs (stmt);
4565 VEC_free (gimple, heap, to_remove);
4567 /* We cannot update call statements with virtual operands during
4568 SSA walk. This might remove them which in turn makes our
4569 VN lattice invalid. */
4570 FOR_EACH_VEC_ELT (gimple, to_update, i, stmt)
4571 update_stmt (stmt);
4572 VEC_free (gimple, heap, to_update);
4574 return todo;
4577 /* Borrow a bit of tree-ssa-dce.c for the moment.
4578 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4579 this may be a bit faster, and we may want critical edges kept split. */
4581 /* If OP's defining statement has not already been determined to be necessary,
4582 mark that statement necessary. Return the stmt, if it is newly
4583 necessary. */
4585 static inline gimple
4586 mark_operand_necessary (tree op)
4588 gimple stmt;
4590 gcc_assert (op);
4592 if (TREE_CODE (op) != SSA_NAME)
4593 return NULL;
4595 stmt = SSA_NAME_DEF_STMT (op);
4596 gcc_assert (stmt);
4598 if (gimple_plf (stmt, NECESSARY)
4599 || gimple_nop_p (stmt))
4600 return NULL;
4602 gimple_set_plf (stmt, NECESSARY, true);
4603 return stmt;
4606 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4607 to insert PHI nodes sometimes, and because value numbering of casts isn't
4608 perfect, we sometimes end up inserting dead code. This simple DCE-like
4609 pass removes any insertions we made that weren't actually used. */
4611 static void
4612 remove_dead_inserted_code (void)
4614 bitmap worklist;
4615 unsigned i;
4616 bitmap_iterator bi;
4617 gimple t;
4619 worklist = BITMAP_ALLOC (NULL);
4620 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4622 t = SSA_NAME_DEF_STMT (ssa_name (i));
4623 if (gimple_plf (t, NECESSARY))
4624 bitmap_set_bit (worklist, i);
4626 while (!bitmap_empty_p (worklist))
4628 i = bitmap_first_set_bit (worklist);
4629 bitmap_clear_bit (worklist, i);
4630 t = SSA_NAME_DEF_STMT (ssa_name (i));
4632 /* PHI nodes are somewhat special in that each PHI alternative has
4633 data and control dependencies. All the statements feeding the
4634 PHI node's arguments are always necessary. */
4635 if (gimple_code (t) == GIMPLE_PHI)
4637 unsigned k;
4639 for (k = 0; k < gimple_phi_num_args (t); k++)
4641 tree arg = PHI_ARG_DEF (t, k);
4642 if (TREE_CODE (arg) == SSA_NAME)
4644 gimple n = mark_operand_necessary (arg);
4645 if (n)
4646 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4650 else
4652 /* Propagate through the operands. Examine all the USE, VUSE and
4653 VDEF operands in this statement. Mark all the statements
4654 which feed this statement's uses as necessary. */
4655 ssa_op_iter iter;
4656 tree use;
4658 /* The operands of VDEF expressions are also needed as they
4659 represent potential definitions that may reach this
4660 statement (VDEF operands allow us to follow def-def
4661 links). */
4663 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4665 gimple n = mark_operand_necessary (use);
4666 if (n)
4667 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4672 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4674 t = SSA_NAME_DEF_STMT (ssa_name (i));
4675 if (!gimple_plf (t, NECESSARY))
4677 gimple_stmt_iterator gsi;
4679 if (dump_file && (dump_flags & TDF_DETAILS))
4681 fprintf (dump_file, "Removing unnecessary insertion:");
4682 print_gimple_stmt (dump_file, t, 0, 0);
4685 gsi = gsi_for_stmt (t);
4686 if (gimple_code (t) == GIMPLE_PHI)
4687 remove_phi_node (&gsi, true);
4688 else
4690 gsi_remove (&gsi, true);
4691 release_defs (t);
4695 BITMAP_FREE (worklist);
4698 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4699 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4700 the number of visited blocks. */
4702 static int
4703 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4705 edge_iterator *stack;
4706 int sp;
4707 int post_order_num = 0;
4708 sbitmap visited;
4710 if (include_entry_exit)
4711 post_order[post_order_num++] = EXIT_BLOCK;
4713 /* Allocate stack for back-tracking up CFG. */
4714 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4715 sp = 0;
4717 /* Allocate bitmap to track nodes that have been visited. */
4718 visited = sbitmap_alloc (last_basic_block);
4720 /* None of the nodes in the CFG have been visited yet. */
4721 sbitmap_zero (visited);
4723 /* Push the last edge on to the stack. */
4724 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4726 while (sp)
4728 edge_iterator ei;
4729 basic_block src;
4730 basic_block dest;
4732 /* Look at the edge on the top of the stack. */
4733 ei = stack[sp - 1];
4734 src = ei_edge (ei)->src;
4735 dest = ei_edge (ei)->dest;
4737 /* Check if the edge destination has been visited yet. */
4738 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4740 /* Mark that we have visited the destination. */
4741 SET_BIT (visited, src->index);
4743 if (EDGE_COUNT (src->preds) > 0)
4744 /* Since the DEST node has been visited for the first
4745 time, check its successors. */
4746 stack[sp++] = ei_start (src->preds);
4747 else
4748 post_order[post_order_num++] = src->index;
4750 else
4752 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4753 post_order[post_order_num++] = dest->index;
4755 if (!ei_one_before_end_p (ei))
4756 ei_next (&stack[sp - 1]);
4757 else
4758 sp--;
4762 if (include_entry_exit)
4763 post_order[post_order_num++] = ENTRY_BLOCK;
4765 free (stack);
4766 sbitmap_free (visited);
4767 return post_order_num;
4771 /* Initialize data structures used by PRE. */
4773 static void
4774 init_pre (bool do_fre)
4776 basic_block bb;
4778 next_expression_id = 1;
4779 expressions = NULL;
4780 VEC_safe_push (pre_expr, heap, expressions, NULL);
4781 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4782 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4783 get_max_value_id() + 1);
4784 name_to_id = NULL;
4786 in_fre = do_fre;
4788 inserted_exprs = BITMAP_ALLOC (NULL);
4789 need_creation = NULL;
4790 pretemp = NULL_TREE;
4791 storetemp = NULL_TREE;
4792 prephitemp = NULL_TREE;
4794 connect_infinite_loops_to_exit ();
4795 memset (&pre_stats, 0, sizeof (pre_stats));
4798 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4799 my_rev_post_order_compute (postorder, false);
4801 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4803 calculate_dominance_info (CDI_POST_DOMINATORS);
4804 calculate_dominance_info (CDI_DOMINATORS);
4806 bitmap_obstack_initialize (&grand_bitmap_obstack);
4807 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4808 expr_pred_trans_eq, free);
4809 expression_to_id = htab_create (num_ssa_names * 3,
4810 pre_expr_hash,
4811 pre_expr_eq, NULL);
4812 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4813 sizeof (struct bitmap_set), 30);
4814 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4815 sizeof (struct pre_expr_d), 30);
4816 FOR_ALL_BB (bb)
4818 EXP_GEN (bb) = bitmap_set_new ();
4819 PHI_GEN (bb) = bitmap_set_new ();
4820 TMP_GEN (bb) = bitmap_set_new ();
4821 AVAIL_OUT (bb) = bitmap_set_new ();
4824 need_eh_cleanup = BITMAP_ALLOC (NULL);
4825 need_ab_cleanup = BITMAP_ALLOC (NULL);
4829 /* Deallocate data structures used by PRE. */
4831 static void
4832 fini_pre (bool do_fre)
4834 free (postorder);
4835 VEC_free (bitmap_set_t, heap, value_expressions);
4836 BITMAP_FREE (inserted_exprs);
4837 VEC_free (gimple, heap, need_creation);
4838 bitmap_obstack_release (&grand_bitmap_obstack);
4839 free_alloc_pool (bitmap_set_pool);
4840 free_alloc_pool (pre_expr_pool);
4841 htab_delete (phi_translate_table);
4842 htab_delete (expression_to_id);
4843 VEC_free (unsigned, heap, name_to_id);
4845 free_aux_for_blocks ();
4847 free_dominance_info (CDI_POST_DOMINATORS);
4849 if (!bitmap_empty_p (need_eh_cleanup))
4851 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4852 cleanup_tree_cfg ();
4855 BITMAP_FREE (need_eh_cleanup);
4857 if (!bitmap_empty_p (need_ab_cleanup))
4859 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4860 cleanup_tree_cfg ();
4863 BITMAP_FREE (need_ab_cleanup);
4865 if (!do_fre)
4866 loop_optimizer_finalize ();
4869 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4870 only wants to do full redundancy elimination. */
4872 static unsigned int
4873 execute_pre (bool do_fre)
4875 unsigned int todo = 0;
4877 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4879 /* This has to happen before SCCVN runs because
4880 loop_optimizer_init may create new phis, etc. */
4881 if (!do_fre)
4882 loop_optimizer_init (LOOPS_NORMAL);
4884 if (!run_scc_vn (do_fre ? VN_WALKREWRITE : VN_WALK))
4886 if (!do_fre)
4887 loop_optimizer_finalize ();
4889 return 0;
4892 init_pre (do_fre);
4893 scev_initialize ();
4895 /* Collect and value number expressions computed in each basic block. */
4896 compute_avail ();
4898 if (dump_file && (dump_flags & TDF_DETAILS))
4900 basic_block bb;
4902 FOR_ALL_BB (bb)
4904 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4905 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4906 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4907 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4911 /* Insert can get quite slow on an incredibly large number of basic
4912 blocks due to some quadratic behavior. Until this behavior is
4913 fixed, don't run it when he have an incredibly large number of
4914 bb's. If we aren't going to run insert, there is no point in
4915 computing ANTIC, either, even though it's plenty fast. */
4916 if (!do_fre && n_basic_blocks < 4000)
4918 compute_antic ();
4919 insert ();
4922 /* Make sure to remove fake edges before committing our inserts.
4923 This makes sure we don't end up with extra critical edges that
4924 we would need to split. */
4925 remove_fake_exit_edges ();
4926 gsi_commit_edge_inserts ();
4928 /* Remove all the redundant expressions. */
4929 todo |= eliminate ();
4931 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4932 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4933 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4934 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4935 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4937 clear_expression_ids ();
4938 free_scc_vn ();
4939 if (!do_fre)
4941 remove_dead_inserted_code ();
4942 todo |= TODO_verify_flow;
4945 scev_finalize ();
4946 fini_pre (do_fre);
4948 return todo;
4951 /* Gate and execute functions for PRE. */
4953 static unsigned int
4954 do_pre (void)
4956 return execute_pre (false);
4959 static bool
4960 gate_pre (void)
4962 return flag_tree_pre != 0;
4965 struct gimple_opt_pass pass_pre =
4968 GIMPLE_PASS,
4969 "pre", /* name */
4970 gate_pre, /* gate */
4971 do_pre, /* execute */
4972 NULL, /* sub */
4973 NULL, /* next */
4974 0, /* static_pass_number */
4975 TV_TREE_PRE, /* tv_id */
4976 PROP_no_crit_edges | PROP_cfg
4977 | PROP_ssa, /* properties_required */
4978 0, /* properties_provided */
4979 0, /* properties_destroyed */
4980 TODO_rebuild_alias, /* todo_flags_start */
4981 TODO_update_ssa_only_virtuals | TODO_dump_func | TODO_ggc_collect
4982 | TODO_verify_ssa /* todo_flags_finish */
4987 /* Gate and execute functions for FRE. */
4989 static unsigned int
4990 execute_fre (void)
4992 return execute_pre (true);
4995 static bool
4996 gate_fre (void)
4998 return flag_tree_fre != 0;
5001 struct gimple_opt_pass pass_fre =
5004 GIMPLE_PASS,
5005 "fre", /* name */
5006 gate_fre, /* gate */
5007 execute_fre, /* execute */
5008 NULL, /* sub */
5009 NULL, /* next */
5010 0, /* static_pass_number */
5011 TV_TREE_FRE, /* tv_id */
5012 PROP_cfg | PROP_ssa, /* properties_required */
5013 0, /* properties_provided */
5014 0, /* properties_destroyed */
5015 0, /* todo_flags_start */
5016 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */