PR gcov-profile/51113
[official-gcc.git] / gcc / tree-ssa-pre.c
blob557b56a110d320409267bbbbcfecb35945280c73
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_binary:
1154 case tcc_comparison:
1156 /* We have to go from trees to pre exprs to value ids to
1157 constants. */
1158 tree naryop0 = nary->op[0];
1159 tree naryop1 = nary->op[1];
1160 tree result;
1161 if (!is_gimple_min_invariant (naryop0))
1163 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1164 unsigned int vrep0 = get_expr_value_id (rep0);
1165 tree const0 = get_constant_for_value_id (vrep0);
1166 if (const0)
1167 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1169 if (!is_gimple_min_invariant (naryop1))
1171 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1172 unsigned int vrep1 = get_expr_value_id (rep1);
1173 tree const1 = get_constant_for_value_id (vrep1);
1174 if (const1)
1175 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1177 result = fold_binary (nary->opcode, nary->type,
1178 naryop0, naryop1);
1179 if (result && is_gimple_min_invariant (result))
1180 return get_or_alloc_expr_for_constant (result);
1181 /* We might have simplified the expression to a
1182 SSA_NAME for example from x_1 * 1. But we cannot
1183 insert a PHI for x_1 unconditionally as x_1 might
1184 not be available readily. */
1185 return e;
1187 case tcc_reference:
1188 if (nary->opcode != REALPART_EXPR
1189 && nary->opcode != IMAGPART_EXPR
1190 && nary->opcode != VIEW_CONVERT_EXPR)
1191 return e;
1192 /* Fallthrough. */
1193 case tcc_unary:
1195 /* We have to go from trees to pre exprs to value ids to
1196 constants. */
1197 tree naryop0 = nary->op[0];
1198 tree const0, result;
1199 if (is_gimple_min_invariant (naryop0))
1200 const0 = naryop0;
1201 else
1203 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1204 unsigned int vrep0 = get_expr_value_id (rep0);
1205 const0 = get_constant_for_value_id (vrep0);
1207 result = NULL;
1208 if (const0)
1210 tree type1 = TREE_TYPE (nary->op[0]);
1211 const0 = fold_convert (type1, const0);
1212 result = fold_unary (nary->opcode, nary->type, const0);
1214 if (result && is_gimple_min_invariant (result))
1215 return get_or_alloc_expr_for_constant (result);
1216 return e;
1218 default:
1219 return e;
1222 case REFERENCE:
1224 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1225 tree folded;
1226 if ((folded = fully_constant_vn_reference_p (ref)))
1227 return get_or_alloc_expr_for_constant (folded);
1228 return e;
1230 default:
1231 return e;
1233 return e;
1236 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1237 it has the value it would have in BLOCK. Set *SAME_VALID to true
1238 in case the new vuse doesn't change the value id of the OPERANDS. */
1240 static tree
1241 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1242 alias_set_type set, tree type, tree vuse,
1243 basic_block phiblock,
1244 basic_block block, bool *same_valid)
1246 gimple phi = SSA_NAME_DEF_STMT (vuse);
1247 ao_ref ref;
1248 edge e = NULL;
1249 bool use_oracle;
1251 *same_valid = true;
1253 if (gimple_bb (phi) != phiblock)
1254 return vuse;
1256 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1258 /* Use the alias-oracle to find either the PHI node in this block,
1259 the first VUSE used in this block that is equivalent to vuse or
1260 the first VUSE which definition in this block kills the value. */
1261 if (gimple_code (phi) == GIMPLE_PHI)
1262 e = find_edge (block, phiblock);
1263 else if (use_oracle)
1264 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1266 vuse = gimple_vuse (phi);
1267 phi = SSA_NAME_DEF_STMT (vuse);
1268 if (gimple_bb (phi) != phiblock)
1269 return vuse;
1270 if (gimple_code (phi) == GIMPLE_PHI)
1272 e = find_edge (block, phiblock);
1273 break;
1276 else
1277 return NULL_TREE;
1279 if (e)
1281 if (use_oracle)
1283 bitmap visited = NULL;
1284 /* Try to find a vuse that dominates this phi node by skipping
1285 non-clobbering statements. */
1286 vuse = get_continuation_for_phi (phi, &ref, &visited);
1287 if (visited)
1288 BITMAP_FREE (visited);
1290 else
1291 vuse = NULL_TREE;
1292 if (!vuse)
1294 /* If we didn't find any, the value ID can't stay the same,
1295 but return the translated vuse. */
1296 *same_valid = false;
1297 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1299 /* ??? We would like to return vuse here as this is the canonical
1300 upmost vdef that this reference is associated with. But during
1301 insertion of the references into the hash tables we only ever
1302 directly insert with their direct gimple_vuse, hence returning
1303 something else would make us not find the other expression. */
1304 return PHI_ARG_DEF (phi, e->dest_idx);
1307 return NULL_TREE;
1310 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1311 SET2. This is used to avoid making a set consisting of the union
1312 of PA_IN and ANTIC_IN during insert. */
1314 static inline pre_expr
1315 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1317 pre_expr result;
1319 result = bitmap_find_leader (set1, val, NULL);
1320 if (!result && set2)
1321 result = bitmap_find_leader (set2, val, NULL);
1322 return result;
1325 /* Get the tree type for our PRE expression e. */
1327 static tree
1328 get_expr_type (const pre_expr e)
1330 switch (e->kind)
1332 case NAME:
1333 return TREE_TYPE (PRE_EXPR_NAME (e));
1334 case CONSTANT:
1335 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1336 case REFERENCE:
1337 return PRE_EXPR_REFERENCE (e)->type;
1338 case NARY:
1339 return PRE_EXPR_NARY (e)->type;
1341 gcc_unreachable();
1344 /* Get a representative SSA_NAME for a given expression.
1345 Since all of our sub-expressions are treated as values, we require
1346 them to be SSA_NAME's for simplicity.
1347 Prior versions of GVNPRE used to use "value handles" here, so that
1348 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1349 either case, the operands are really values (IE we do not expect
1350 them to be usable without finding leaders). */
1352 static tree
1353 get_representative_for (const pre_expr e)
1355 tree exprtype;
1356 tree name;
1357 unsigned int value_id = get_expr_value_id (e);
1359 switch (e->kind)
1361 case NAME:
1362 return PRE_EXPR_NAME (e);
1363 case CONSTANT:
1364 return PRE_EXPR_CONSTANT (e);
1365 case NARY:
1366 case REFERENCE:
1368 /* Go through all of the expressions representing this value
1369 and pick out an SSA_NAME. */
1370 unsigned int i;
1371 bitmap_iterator bi;
1372 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1373 value_id);
1374 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1376 pre_expr rep = expression_for_id (i);
1377 if (rep->kind == NAME)
1378 return PRE_EXPR_NAME (rep);
1381 break;
1383 /* If we reached here we couldn't find an SSA_NAME. This can
1384 happen when we've discovered a value that has never appeared in
1385 the program as set to an SSA_NAME, most likely as the result of
1386 phi translation. */
1387 if (dump_file)
1389 fprintf (dump_file,
1390 "Could not find SSA_NAME representative for expression:");
1391 print_pre_expr (dump_file, e);
1392 fprintf (dump_file, "\n");
1395 exprtype = get_expr_type (e);
1397 /* Build and insert the assignment of the end result to the temporary
1398 that we will return. */
1399 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1401 pretemp = create_tmp_reg (exprtype, "pretmp");
1402 add_referenced_var (pretemp);
1405 name = make_ssa_name (pretemp, gimple_build_nop ());
1406 VN_INFO_GET (name)->value_id = value_id;
1407 if (e->kind == CONSTANT)
1408 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1409 else
1410 VN_INFO (name)->valnum = name;
1412 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1413 if (dump_file)
1415 fprintf (dump_file, "Created SSA_NAME representative ");
1416 print_generic_expr (dump_file, name, 0);
1417 fprintf (dump_file, " for expression:");
1418 print_pre_expr (dump_file, e);
1419 fprintf (dump_file, "\n");
1422 return name;
1427 static pre_expr
1428 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1429 basic_block pred, basic_block phiblock);
1431 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1432 the phis in PRED. Return NULL if we can't find a leader for each part
1433 of the translated expression. */
1435 static pre_expr
1436 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1437 basic_block pred, basic_block phiblock)
1439 switch (expr->kind)
1441 case NARY:
1443 unsigned int i;
1444 bool changed = false;
1445 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1446 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1447 sizeof_vn_nary_op (nary->length));
1448 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1450 for (i = 0; i < newnary->length; i++)
1452 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1453 continue;
1454 else
1456 pre_expr leader, result;
1457 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1458 leader = find_leader_in_sets (op_val_id, set1, set2);
1459 result = phi_translate (leader, set1, set2, pred, phiblock);
1460 if (result && result != leader)
1462 tree name = get_representative_for (result);
1463 if (!name)
1464 return NULL;
1465 newnary->op[i] = name;
1467 else if (!result)
1468 return NULL;
1470 changed |= newnary->op[i] != nary->op[i];
1473 if (changed)
1475 pre_expr constant;
1476 unsigned int new_val_id;
1478 tree result = vn_nary_op_lookup_pieces (newnary->length,
1479 newnary->opcode,
1480 newnary->type,
1481 &newnary->op[0],
1482 &nary);
1483 if (result && is_gimple_min_invariant (result))
1484 return get_or_alloc_expr_for_constant (result);
1486 expr = (pre_expr) pool_alloc (pre_expr_pool);
1487 expr->kind = NARY;
1488 expr->id = 0;
1489 if (nary)
1491 PRE_EXPR_NARY (expr) = nary;
1492 constant = fully_constant_expression (expr);
1493 if (constant != expr)
1494 return constant;
1496 new_val_id = nary->value_id;
1497 get_or_alloc_expression_id (expr);
1499 else
1501 new_val_id = get_next_value_id ();
1502 VEC_safe_grow_cleared (bitmap_set_t, heap,
1503 value_expressions,
1504 get_max_value_id() + 1);
1505 nary = vn_nary_op_insert_pieces (newnary->length,
1506 newnary->opcode,
1507 newnary->type,
1508 &newnary->op[0],
1509 result, new_val_id);
1510 PRE_EXPR_NARY (expr) = nary;
1511 constant = fully_constant_expression (expr);
1512 if (constant != expr)
1513 return constant;
1514 get_or_alloc_expression_id (expr);
1516 add_to_value (new_val_id, expr);
1518 return expr;
1520 break;
1522 case REFERENCE:
1524 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1525 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1526 tree vuse = ref->vuse;
1527 tree newvuse = vuse;
1528 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1529 bool changed = false, same_valid = true;
1530 unsigned int i, j, n;
1531 vn_reference_op_t operand;
1532 vn_reference_t newref;
1534 for (i = 0, j = 0;
1535 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1537 pre_expr opresult;
1538 pre_expr leader;
1539 tree op[3];
1540 tree type = operand->type;
1541 vn_reference_op_s newop = *operand;
1542 op[0] = operand->op0;
1543 op[1] = operand->op1;
1544 op[2] = operand->op2;
1545 for (n = 0; n < 3; ++n)
1547 unsigned int op_val_id;
1548 if (!op[n])
1549 continue;
1550 if (TREE_CODE (op[n]) != SSA_NAME)
1552 /* We can't possibly insert these. */
1553 if (n != 0
1554 && !is_gimple_min_invariant (op[n]))
1555 break;
1556 continue;
1558 op_val_id = VN_INFO (op[n])->value_id;
1559 leader = find_leader_in_sets (op_val_id, set1, set2);
1560 if (!leader)
1561 break;
1562 /* Make sure we do not recursively translate ourselves
1563 like for translating a[n_1] with the leader for
1564 n_1 being a[n_1]. */
1565 if (get_expression_id (leader) != get_expression_id (expr))
1567 opresult = phi_translate (leader, set1, set2,
1568 pred, phiblock);
1569 if (!opresult)
1570 break;
1571 if (opresult != leader)
1573 tree name = get_representative_for (opresult);
1574 if (!name)
1575 break;
1576 changed |= name != op[n];
1577 op[n] = name;
1581 if (n != 3)
1583 if (newoperands)
1584 VEC_free (vn_reference_op_s, heap, newoperands);
1585 return NULL;
1587 if (!newoperands)
1588 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1589 /* We may have changed from an SSA_NAME to a constant */
1590 if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1591 newop.opcode = TREE_CODE (op[0]);
1592 newop.type = type;
1593 newop.op0 = op[0];
1594 newop.op1 = op[1];
1595 newop.op2 = op[2];
1596 /* If it transforms a non-constant ARRAY_REF into a constant
1597 one, adjust the constant offset. */
1598 if (newop.opcode == ARRAY_REF
1599 && newop.off == -1
1600 && TREE_CODE (op[0]) == INTEGER_CST
1601 && TREE_CODE (op[1]) == INTEGER_CST
1602 && TREE_CODE (op[2]) == INTEGER_CST)
1604 double_int off = tree_to_double_int (op[0]);
1605 off = double_int_add (off,
1606 double_int_neg
1607 (tree_to_double_int (op[1])));
1608 off = double_int_mul (off, tree_to_double_int (op[2]));
1609 if (double_int_fits_in_shwi_p (off))
1610 newop.off = off.low;
1612 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1613 /* If it transforms from an SSA_NAME to an address, fold with
1614 a preceding indirect reference. */
1615 if (j > 0 && op[0] && TREE_CODE (op[0]) == ADDR_EXPR
1616 && VEC_index (vn_reference_op_s,
1617 newoperands, j - 1)->opcode == MEM_REF)
1618 vn_reference_fold_indirect (&newoperands, &j);
1620 if (i != VEC_length (vn_reference_op_s, operands))
1622 if (newoperands)
1623 VEC_free (vn_reference_op_s, heap, newoperands);
1624 return NULL;
1627 if (vuse)
1629 newvuse = translate_vuse_through_block (newoperands,
1630 ref->set, ref->type,
1631 vuse, phiblock, pred,
1632 &same_valid);
1633 if (newvuse == NULL_TREE)
1635 VEC_free (vn_reference_op_s, heap, newoperands);
1636 return NULL;
1640 if (changed || newvuse != vuse)
1642 unsigned int new_val_id;
1643 pre_expr constant;
1644 bool converted = false;
1646 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1647 ref->type,
1648 newoperands,
1649 &newref, VN_WALK);
1650 if (result)
1651 VEC_free (vn_reference_op_s, heap, newoperands);
1653 if (result
1654 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1656 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1657 converted = true;
1659 else if (!result && newref
1660 && !useless_type_conversion_p (ref->type, newref->type))
1662 VEC_free (vn_reference_op_s, heap, newoperands);
1663 return NULL;
1666 if (result && is_gimple_min_invariant (result))
1668 gcc_assert (!newoperands);
1669 return get_or_alloc_expr_for_constant (result);
1672 expr = (pre_expr) pool_alloc (pre_expr_pool);
1673 expr->kind = REFERENCE;
1674 expr->id = 0;
1676 if (converted)
1678 vn_nary_op_t nary;
1679 tree nresult;
1681 gcc_assert (CONVERT_EXPR_P (result)
1682 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1684 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1685 TREE_TYPE (result),
1686 &TREE_OPERAND (result, 0),
1687 &nary);
1688 if (nresult && is_gimple_min_invariant (nresult))
1689 return get_or_alloc_expr_for_constant (nresult);
1691 expr->kind = NARY;
1692 if (nary)
1694 PRE_EXPR_NARY (expr) = nary;
1695 constant = fully_constant_expression (expr);
1696 if (constant != expr)
1697 return constant;
1699 new_val_id = nary->value_id;
1700 get_or_alloc_expression_id (expr);
1702 else
1704 new_val_id = get_next_value_id ();
1705 VEC_safe_grow_cleared (bitmap_set_t, heap,
1706 value_expressions,
1707 get_max_value_id() + 1);
1708 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1709 TREE_TYPE (result),
1710 &TREE_OPERAND (result, 0),
1711 NULL_TREE,
1712 new_val_id);
1713 PRE_EXPR_NARY (expr) = nary;
1714 constant = fully_constant_expression (expr);
1715 if (constant != expr)
1716 return constant;
1717 get_or_alloc_expression_id (expr);
1720 else if (newref)
1722 PRE_EXPR_REFERENCE (expr) = newref;
1723 constant = fully_constant_expression (expr);
1724 if (constant != expr)
1725 return constant;
1727 new_val_id = newref->value_id;
1728 get_or_alloc_expression_id (expr);
1730 else
1732 if (changed || !same_valid)
1734 new_val_id = get_next_value_id ();
1735 VEC_safe_grow_cleared (bitmap_set_t, heap,
1736 value_expressions,
1737 get_max_value_id() + 1);
1739 else
1740 new_val_id = ref->value_id;
1741 newref = vn_reference_insert_pieces (newvuse, ref->set,
1742 ref->type,
1743 newoperands,
1744 result, new_val_id);
1745 newoperands = NULL;
1746 PRE_EXPR_REFERENCE (expr) = newref;
1747 constant = fully_constant_expression (expr);
1748 if (constant != expr)
1749 return constant;
1750 get_or_alloc_expression_id (expr);
1752 add_to_value (new_val_id, expr);
1754 VEC_free (vn_reference_op_s, heap, newoperands);
1755 return expr;
1757 break;
1759 case NAME:
1761 gimple phi = NULL;
1762 edge e;
1763 gimple def_stmt;
1764 tree name = PRE_EXPR_NAME (expr);
1766 def_stmt = SSA_NAME_DEF_STMT (name);
1767 if (gimple_code (def_stmt) == GIMPLE_PHI
1768 && gimple_bb (def_stmt) == phiblock)
1769 phi = def_stmt;
1770 else
1771 return expr;
1773 e = find_edge (pred, gimple_bb (phi));
1774 if (e)
1776 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1777 pre_expr newexpr;
1779 if (TREE_CODE (def) == SSA_NAME)
1780 def = VN_INFO (def)->valnum;
1782 /* Handle constant. */
1783 if (is_gimple_min_invariant (def))
1784 return get_or_alloc_expr_for_constant (def);
1786 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1787 return NULL;
1789 newexpr = get_or_alloc_expr_for_name (def);
1790 return newexpr;
1793 return expr;
1795 default:
1796 gcc_unreachable ();
1800 /* Wrapper around phi_translate_1 providing caching functionality. */
1802 static pre_expr
1803 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1804 basic_block pred, basic_block phiblock)
1806 pre_expr phitrans;
1808 if (!expr)
1809 return NULL;
1811 /* Constants contain no values that need translation. */
1812 if (expr->kind == CONSTANT)
1813 return expr;
1815 if (value_id_constant_p (get_expr_value_id (expr)))
1816 return expr;
1818 if (expr->kind != NAME)
1820 phitrans = phi_trans_lookup (expr, pred);
1821 if (phitrans)
1822 return phitrans;
1825 /* Translate. */
1826 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1828 /* Don't add empty translations to the cache. Neither add
1829 translations of NAMEs as those are cheap to translate. */
1830 if (phitrans
1831 && expr->kind != NAME)
1832 phi_trans_add (expr, phitrans, pred);
1834 return phitrans;
1838 /* For each expression in SET, translate the values through phi nodes
1839 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1840 expressions in DEST. */
1842 static void
1843 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1844 basic_block phiblock)
1846 VEC (pre_expr, heap) *exprs;
1847 pre_expr expr;
1848 int i;
1850 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1852 bitmap_set_copy (dest, set);
1853 return;
1856 exprs = sorted_array_from_bitmap_set (set);
1857 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1859 pre_expr translated;
1860 translated = phi_translate (expr, set, NULL, pred, phiblock);
1861 if (!translated)
1862 continue;
1864 /* We might end up with multiple expressions from SET being
1865 translated to the same value. In this case we do not want
1866 to retain the NARY or REFERENCE expression but prefer a NAME
1867 which would be the leader. */
1868 if (translated->kind == NAME)
1869 bitmap_value_replace_in_set (dest, translated);
1870 else
1871 bitmap_value_insert_into_set (dest, translated);
1873 VEC_free (pre_expr, heap, exprs);
1876 /* Find the leader for a value (i.e., the name representing that
1877 value) in a given set, and return it. If STMT is non-NULL it
1878 makes sure the defining statement for the leader dominates it.
1879 Return NULL if no leader is found. */
1881 static pre_expr
1882 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1884 if (value_id_constant_p (val))
1886 unsigned int i;
1887 bitmap_iterator bi;
1888 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1890 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1892 pre_expr expr = expression_for_id (i);
1893 if (expr->kind == CONSTANT)
1894 return expr;
1897 if (bitmap_set_contains_value (set, val))
1899 /* Rather than walk the entire bitmap of expressions, and see
1900 whether any of them has the value we are looking for, we look
1901 at the reverse mapping, which tells us the set of expressions
1902 that have a given value (IE value->expressions with that
1903 value) and see if any of those expressions are in our set.
1904 The number of expressions per value is usually significantly
1905 less than the number of expressions in the set. In fact, for
1906 large testcases, doing it this way is roughly 5-10x faster
1907 than walking the bitmap.
1908 If this is somehow a significant lose for some cases, we can
1909 choose which set to walk based on which set is smaller. */
1910 unsigned int i;
1911 bitmap_iterator bi;
1912 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1914 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1915 &set->expressions, 0, i, bi)
1917 pre_expr val = expression_for_id (i);
1918 /* At the point where stmt is not null, there should always
1919 be an SSA_NAME first in the list of expressions. */
1920 if (stmt)
1922 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1923 if (gimple_code (def_stmt) != GIMPLE_PHI
1924 && gimple_bb (def_stmt) == gimple_bb (stmt)
1925 /* PRE insertions are at the end of the basic-block
1926 and have UID 0. */
1927 && (gimple_uid (def_stmt) == 0
1928 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1929 continue;
1931 return val;
1934 return NULL;
1937 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1938 BLOCK by seeing if it is not killed in the block. Note that we are
1939 only determining whether there is a store that kills it. Because
1940 of the order in which clean iterates over values, we are guaranteed
1941 that altered operands will have caused us to be eliminated from the
1942 ANTIC_IN set already. */
1944 static bool
1945 value_dies_in_block_x (pre_expr expr, basic_block block)
1947 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1948 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1949 gimple def;
1950 gimple_stmt_iterator gsi;
1951 unsigned id = get_expression_id (expr);
1952 bool res = false;
1953 ao_ref ref;
1955 if (!vuse)
1956 return false;
1958 /* Lookup a previously calculated result. */
1959 if (EXPR_DIES (block)
1960 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1961 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1963 /* A memory expression {e, VUSE} dies in the block if there is a
1964 statement that may clobber e. If, starting statement walk from the
1965 top of the basic block, a statement uses VUSE there can be no kill
1966 inbetween that use and the original statement that loaded {e, VUSE},
1967 so we can stop walking. */
1968 ref.base = NULL_TREE;
1969 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1971 tree def_vuse, def_vdef;
1972 def = gsi_stmt (gsi);
1973 def_vuse = gimple_vuse (def);
1974 def_vdef = gimple_vdef (def);
1976 /* Not a memory statement. */
1977 if (!def_vuse)
1978 continue;
1980 /* Not a may-def. */
1981 if (!def_vdef)
1983 /* A load with the same VUSE, we're done. */
1984 if (def_vuse == vuse)
1985 break;
1987 continue;
1990 /* Init ref only if we really need it. */
1991 if (ref.base == NULL_TREE
1992 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
1993 refx->operands))
1995 res = true;
1996 break;
1998 /* If the statement may clobber expr, it dies. */
1999 if (stmt_may_clobber_ref_p_1 (def, &ref))
2001 res = true;
2002 break;
2006 /* Remember the result. */
2007 if (!EXPR_DIES (block))
2008 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2009 bitmap_set_bit (EXPR_DIES (block), id * 2);
2010 if (res)
2011 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2013 return res;
2017 #define union_contains_value(SET1, SET2, VAL) \
2018 (bitmap_set_contains_value ((SET1), (VAL)) \
2019 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
2021 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2023 static bool
2024 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2025 vn_reference_op_t vro)
2027 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2029 struct pre_expr_d temp;
2030 temp.kind = NAME;
2031 temp.id = 0;
2032 PRE_EXPR_NAME (&temp) = vro->op0;
2033 temp.id = lookup_expression_id (&temp);
2034 if (temp.id == 0)
2035 return false;
2036 if (!union_contains_value (set1, set2,
2037 get_expr_value_id (&temp)))
2038 return false;
2040 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2042 struct pre_expr_d temp;
2043 temp.kind = NAME;
2044 temp.id = 0;
2045 PRE_EXPR_NAME (&temp) = vro->op1;
2046 temp.id = lookup_expression_id (&temp);
2047 if (temp.id == 0)
2048 return false;
2049 if (!union_contains_value (set1, set2,
2050 get_expr_value_id (&temp)))
2051 return false;
2054 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2056 struct pre_expr_d temp;
2057 temp.kind = NAME;
2058 temp.id = 0;
2059 PRE_EXPR_NAME (&temp) = vro->op2;
2060 temp.id = lookup_expression_id (&temp);
2061 if (temp.id == 0)
2062 return false;
2063 if (!union_contains_value (set1, set2,
2064 get_expr_value_id (&temp)))
2065 return false;
2068 return true;
2071 /* Determine if the expression EXPR is valid in SET1 U SET2.
2072 ONLY SET2 CAN BE NULL.
2073 This means that we have a leader for each part of the expression
2074 (if it consists of values), or the expression is an SSA_NAME.
2075 For loads/calls, we also see if the vuse is killed in this block. */
2077 static bool
2078 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2079 basic_block block)
2081 switch (expr->kind)
2083 case NAME:
2084 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2085 case NARY:
2087 unsigned int i;
2088 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2089 for (i = 0; i < nary->length; i++)
2091 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2093 struct pre_expr_d temp;
2094 temp.kind = NAME;
2095 temp.id = 0;
2096 PRE_EXPR_NAME (&temp) = nary->op[i];
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 /* If the NARY may trap make sure the block does not contain
2106 a possible exit point.
2107 ??? This is overly conservative if we translate AVAIL_OUT
2108 as the available expression might be after the exit point. */
2109 if (BB_MAY_NOTRETURN (block)
2110 && vn_nary_may_trap (nary))
2111 return false;
2112 return true;
2114 break;
2115 case REFERENCE:
2117 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2118 vn_reference_op_t vro;
2119 unsigned int i;
2121 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2123 if (!vro_valid_in_sets (set1, set2, vro))
2124 return false;
2126 if (ref->vuse)
2128 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2129 if (!gimple_nop_p (def_stmt)
2130 && gimple_bb (def_stmt) != block
2131 && !dominated_by_p (CDI_DOMINATORS,
2132 block, gimple_bb (def_stmt)))
2133 return false;
2135 return !value_dies_in_block_x (expr, block);
2137 default:
2138 gcc_unreachable ();
2142 /* Clean the set of expressions that are no longer valid in SET1 or
2143 SET2. This means expressions that are made up of values we have no
2144 leaders for in SET1 or SET2. This version is used for partial
2145 anticipation, which means it is not valid in either ANTIC_IN or
2146 PA_IN. */
2148 static void
2149 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2151 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2152 pre_expr expr;
2153 int i;
2155 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2157 if (!valid_in_sets (set1, set2, expr, block))
2158 bitmap_remove_from_set (set1, expr);
2160 VEC_free (pre_expr, heap, exprs);
2163 /* Clean the set of expressions that are no longer valid in SET. This
2164 means expressions that are made up of values we have no leaders for
2165 in SET. */
2167 static void
2168 clean (bitmap_set_t set, basic_block block)
2170 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2171 pre_expr expr;
2172 int i;
2174 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2176 if (!valid_in_sets (set, NULL, expr, block))
2177 bitmap_remove_from_set (set, expr);
2179 VEC_free (pre_expr, heap, exprs);
2182 static sbitmap has_abnormal_preds;
2184 /* List of blocks that may have changed during ANTIC computation and
2185 thus need to be iterated over. */
2187 static sbitmap changed_blocks;
2189 /* Decide whether to defer a block for a later iteration, or PHI
2190 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2191 should defer the block, and true if we processed it. */
2193 static bool
2194 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2195 basic_block block, basic_block phiblock)
2197 if (!BB_VISITED (phiblock))
2199 SET_BIT (changed_blocks, block->index);
2200 BB_VISITED (block) = 0;
2201 BB_DEFERRED (block) = 1;
2202 return false;
2204 else
2205 phi_translate_set (dest, source, block, phiblock);
2206 return true;
2209 /* Compute the ANTIC set for BLOCK.
2211 If succs(BLOCK) > 1 then
2212 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2213 else if succs(BLOCK) == 1 then
2214 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2216 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2219 static bool
2220 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2222 bool changed = false;
2223 bitmap_set_t S, old, ANTIC_OUT;
2224 bitmap_iterator bi;
2225 unsigned int bii;
2226 edge e;
2227 edge_iterator ei;
2229 old = ANTIC_OUT = S = NULL;
2230 BB_VISITED (block) = 1;
2232 /* If any edges from predecessors are abnormal, antic_in is empty,
2233 so do nothing. */
2234 if (block_has_abnormal_pred_edge)
2235 goto maybe_dump_sets;
2237 old = ANTIC_IN (block);
2238 ANTIC_OUT = bitmap_set_new ();
2240 /* If the block has no successors, ANTIC_OUT is empty. */
2241 if (EDGE_COUNT (block->succs) == 0)
2243 /* If we have one successor, we could have some phi nodes to
2244 translate through. */
2245 else if (single_succ_p (block))
2247 basic_block succ_bb = single_succ (block);
2249 /* We trade iterations of the dataflow equations for having to
2250 phi translate the maximal set, which is incredibly slow
2251 (since the maximal set often has 300+ members, even when you
2252 have a small number of blocks).
2253 Basically, we defer the computation of ANTIC for this block
2254 until we have processed it's successor, which will inevitably
2255 have a *much* smaller set of values to phi translate once
2256 clean has been run on it.
2257 The cost of doing this is that we technically perform more
2258 iterations, however, they are lower cost iterations.
2260 Timings for PRE on tramp3d-v4:
2261 without maximal set fix: 11 seconds
2262 with maximal set fix/without deferring: 26 seconds
2263 with maximal set fix/with deferring: 11 seconds
2266 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2267 block, succ_bb))
2269 changed = true;
2270 goto maybe_dump_sets;
2273 /* If we have multiple successors, we take the intersection of all of
2274 them. Note that in the case of loop exit phi nodes, we may have
2275 phis to translate through. */
2276 else
2278 VEC(basic_block, heap) * worklist;
2279 size_t i;
2280 basic_block bprime, first = NULL;
2282 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2283 FOR_EACH_EDGE (e, ei, block->succs)
2285 if (!first
2286 && BB_VISITED (e->dest))
2287 first = e->dest;
2288 else if (BB_VISITED (e->dest))
2289 VEC_quick_push (basic_block, worklist, e->dest);
2292 /* Of multiple successors we have to have visited one already. */
2293 if (!first)
2295 SET_BIT (changed_blocks, block->index);
2296 BB_VISITED (block) = 0;
2297 BB_DEFERRED (block) = 1;
2298 changed = true;
2299 VEC_free (basic_block, heap, worklist);
2300 goto maybe_dump_sets;
2303 if (!gimple_seq_empty_p (phi_nodes (first)))
2304 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2305 else
2306 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2308 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2310 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2312 bitmap_set_t tmp = bitmap_set_new ();
2313 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2314 bitmap_set_and (ANTIC_OUT, tmp);
2315 bitmap_set_free (tmp);
2317 else
2318 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2320 VEC_free (basic_block, heap, worklist);
2323 /* Generate ANTIC_OUT - TMP_GEN. */
2324 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2326 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2327 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2328 TMP_GEN (block));
2330 /* Then union in the ANTIC_OUT - TMP_GEN values,
2331 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2332 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2333 bitmap_value_insert_into_set (ANTIC_IN (block),
2334 expression_for_id (bii));
2336 clean (ANTIC_IN (block), block);
2338 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2340 changed = true;
2341 SET_BIT (changed_blocks, block->index);
2342 FOR_EACH_EDGE (e, ei, block->preds)
2343 SET_BIT (changed_blocks, e->src->index);
2345 else
2346 RESET_BIT (changed_blocks, block->index);
2348 maybe_dump_sets:
2349 if (dump_file && (dump_flags & TDF_DETAILS))
2351 if (!BB_DEFERRED (block) || BB_VISITED (block))
2353 if (ANTIC_OUT)
2354 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2356 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2357 block->index);
2359 if (S)
2360 print_bitmap_set (dump_file, S, "S", block->index);
2362 else
2364 fprintf (dump_file,
2365 "Block %d was deferred for a future iteration.\n",
2366 block->index);
2369 if (old)
2370 bitmap_set_free (old);
2371 if (S)
2372 bitmap_set_free (S);
2373 if (ANTIC_OUT)
2374 bitmap_set_free (ANTIC_OUT);
2375 return changed;
2378 /* Compute PARTIAL_ANTIC for BLOCK.
2380 If succs(BLOCK) > 1 then
2381 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2382 in ANTIC_OUT for all succ(BLOCK)
2383 else if succs(BLOCK) == 1 then
2384 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2386 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2387 - ANTIC_IN[BLOCK])
2390 static bool
2391 compute_partial_antic_aux (basic_block block,
2392 bool block_has_abnormal_pred_edge)
2394 bool changed = false;
2395 bitmap_set_t old_PA_IN;
2396 bitmap_set_t PA_OUT;
2397 edge e;
2398 edge_iterator ei;
2399 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2401 old_PA_IN = PA_OUT = NULL;
2403 /* If any edges from predecessors are abnormal, antic_in is empty,
2404 so do nothing. */
2405 if (block_has_abnormal_pred_edge)
2406 goto maybe_dump_sets;
2408 /* If there are too many partially anticipatable values in the
2409 block, phi_translate_set can take an exponential time: stop
2410 before the translation starts. */
2411 if (max_pa
2412 && single_succ_p (block)
2413 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2414 goto maybe_dump_sets;
2416 old_PA_IN = PA_IN (block);
2417 PA_OUT = bitmap_set_new ();
2419 /* If the block has no successors, ANTIC_OUT is empty. */
2420 if (EDGE_COUNT (block->succs) == 0)
2422 /* If we have one successor, we could have some phi nodes to
2423 translate through. Note that we can't phi translate across DFS
2424 back edges in partial antic, because it uses a union operation on
2425 the successors. For recurrences like IV's, we will end up
2426 generating a new value in the set on each go around (i + 3 (VH.1)
2427 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2428 else if (single_succ_p (block))
2430 basic_block succ = single_succ (block);
2431 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2432 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2434 /* If we have multiple successors, we take the union of all of
2435 them. */
2436 else
2438 VEC(basic_block, heap) * worklist;
2439 size_t i;
2440 basic_block bprime;
2442 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2443 FOR_EACH_EDGE (e, ei, block->succs)
2445 if (e->flags & EDGE_DFS_BACK)
2446 continue;
2447 VEC_quick_push (basic_block, worklist, e->dest);
2449 if (VEC_length (basic_block, worklist) > 0)
2451 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2453 unsigned int i;
2454 bitmap_iterator bi;
2456 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2457 bitmap_value_insert_into_set (PA_OUT,
2458 expression_for_id (i));
2459 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2461 bitmap_set_t pa_in = bitmap_set_new ();
2462 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2463 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2464 bitmap_value_insert_into_set (PA_OUT,
2465 expression_for_id (i));
2466 bitmap_set_free (pa_in);
2468 else
2469 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2470 bitmap_value_insert_into_set (PA_OUT,
2471 expression_for_id (i));
2474 VEC_free (basic_block, heap, worklist);
2477 /* PA_IN starts with PA_OUT - TMP_GEN.
2478 Then we subtract things from ANTIC_IN. */
2479 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2481 /* For partial antic, we want to put back in the phi results, since
2482 we will properly avoid making them partially antic over backedges. */
2483 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2484 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2486 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2487 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2489 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2491 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2493 changed = true;
2494 SET_BIT (changed_blocks, block->index);
2495 FOR_EACH_EDGE (e, ei, block->preds)
2496 SET_BIT (changed_blocks, e->src->index);
2498 else
2499 RESET_BIT (changed_blocks, block->index);
2501 maybe_dump_sets:
2502 if (dump_file && (dump_flags & TDF_DETAILS))
2504 if (PA_OUT)
2505 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2507 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2509 if (old_PA_IN)
2510 bitmap_set_free (old_PA_IN);
2511 if (PA_OUT)
2512 bitmap_set_free (PA_OUT);
2513 return changed;
2516 /* Compute ANTIC and partial ANTIC sets. */
2518 static void
2519 compute_antic (void)
2521 bool changed = true;
2522 int num_iterations = 0;
2523 basic_block block;
2524 int i;
2526 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2527 We pre-build the map of blocks with incoming abnormal edges here. */
2528 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2529 sbitmap_zero (has_abnormal_preds);
2531 FOR_EACH_BB (block)
2533 edge_iterator ei;
2534 edge e;
2536 FOR_EACH_EDGE (e, ei, block->preds)
2538 e->flags &= ~EDGE_DFS_BACK;
2539 if (e->flags & EDGE_ABNORMAL)
2541 SET_BIT (has_abnormal_preds, block->index);
2542 break;
2546 BB_VISITED (block) = 0;
2547 BB_DEFERRED (block) = 0;
2549 /* While we are here, give empty ANTIC_IN sets to each block. */
2550 ANTIC_IN (block) = bitmap_set_new ();
2551 PA_IN (block) = bitmap_set_new ();
2554 /* At the exit block we anticipate nothing. */
2555 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2556 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2557 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2559 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2560 sbitmap_ones (changed_blocks);
2561 while (changed)
2563 if (dump_file && (dump_flags & TDF_DETAILS))
2564 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2565 /* ??? We need to clear our PHI translation cache here as the
2566 ANTIC sets shrink and we restrict valid translations to
2567 those having operands with leaders in ANTIC. Same below
2568 for PA ANTIC computation. */
2569 num_iterations++;
2570 changed = false;
2571 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2573 if (TEST_BIT (changed_blocks, postorder[i]))
2575 basic_block block = BASIC_BLOCK (postorder[i]);
2576 changed |= compute_antic_aux (block,
2577 TEST_BIT (has_abnormal_preds,
2578 block->index));
2581 /* Theoretically possible, but *highly* unlikely. */
2582 gcc_checking_assert (num_iterations < 500);
2585 statistics_histogram_event (cfun, "compute_antic iterations",
2586 num_iterations);
2588 if (do_partial_partial)
2590 sbitmap_ones (changed_blocks);
2591 mark_dfs_back_edges ();
2592 num_iterations = 0;
2593 changed = true;
2594 while (changed)
2596 if (dump_file && (dump_flags & TDF_DETAILS))
2597 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2598 num_iterations++;
2599 changed = false;
2600 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2602 if (TEST_BIT (changed_blocks, postorder[i]))
2604 basic_block block = BASIC_BLOCK (postorder[i]);
2605 changed
2606 |= compute_partial_antic_aux (block,
2607 TEST_BIT (has_abnormal_preds,
2608 block->index));
2611 /* Theoretically possible, but *highly* unlikely. */
2612 gcc_checking_assert (num_iterations < 500);
2614 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2615 num_iterations);
2617 sbitmap_free (has_abnormal_preds);
2618 sbitmap_free (changed_blocks);
2621 /* Return true if we can value number the call in STMT. This is true
2622 if we have a pure or constant call to a real function. */
2624 static bool
2625 can_value_number_call (gimple stmt)
2627 if (gimple_call_internal_p (stmt))
2628 return false;
2629 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2630 return true;
2631 return false;
2634 /* Return true if OP is a tree which we can perform PRE on.
2635 This may not match the operations we can value number, but in
2636 a perfect world would. */
2638 static bool
2639 can_PRE_operation (tree op)
2641 return UNARY_CLASS_P (op)
2642 || BINARY_CLASS_P (op)
2643 || COMPARISON_CLASS_P (op)
2644 || TREE_CODE (op) == MEM_REF
2645 || TREE_CODE (op) == COMPONENT_REF
2646 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2647 || TREE_CODE (op) == CALL_EXPR
2648 || TREE_CODE (op) == ARRAY_REF;
2652 /* Inserted expressions are placed onto this worklist, which is used
2653 for performing quick dead code elimination of insertions we made
2654 that didn't turn out to be necessary. */
2655 static bitmap inserted_exprs;
2657 /* Pool allocated fake store expressions are placed onto this
2658 worklist, which, after performing dead code elimination, is walked
2659 to see which expressions need to be put into GC'able memory */
2660 static VEC(gimple, heap) *need_creation;
2662 /* The actual worker for create_component_ref_by_pieces. */
2664 static tree
2665 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2666 unsigned int *operand, gimple_seq *stmts,
2667 gimple domstmt)
2669 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2670 *operand);
2671 tree genop;
2672 ++*operand;
2673 switch (currop->opcode)
2675 case CALL_EXPR:
2677 tree folded, sc = NULL_TREE;
2678 unsigned int nargs = 0;
2679 tree fn, *args;
2680 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2681 fn = currop->op0;
2682 else
2684 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2685 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2686 if (!fn)
2687 return NULL_TREE;
2689 if (currop->op1)
2691 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2692 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2693 if (!sc)
2694 return NULL_TREE;
2696 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2697 ref->operands) - 1);
2698 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2700 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2701 operand, stmts,
2702 domstmt);
2703 if (!args[nargs])
2705 free (args);
2706 return NULL_TREE;
2708 nargs++;
2710 folded = build_call_array (currop->type,
2711 (TREE_CODE (fn) == FUNCTION_DECL
2712 ? build_fold_addr_expr (fn) : fn),
2713 nargs, args);
2714 free (args);
2715 if (sc)
2716 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2717 return folded;
2719 break;
2720 case MEM_REF:
2722 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2723 stmts, domstmt);
2724 tree offset = currop->op0;
2725 if (!baseop)
2726 return NULL_TREE;
2727 if (TREE_CODE (baseop) == ADDR_EXPR
2728 && handled_component_p (TREE_OPERAND (baseop, 0)))
2730 HOST_WIDE_INT off;
2731 tree base;
2732 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2733 &off);
2734 gcc_assert (base);
2735 offset = int_const_binop (PLUS_EXPR, offset,
2736 build_int_cst (TREE_TYPE (offset),
2737 off));
2738 baseop = build_fold_addr_expr (base);
2740 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2742 break;
2743 case TARGET_MEM_REF:
2745 pre_expr op0expr, op1expr;
2746 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2747 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2748 ++*operand);
2749 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2750 stmts, domstmt);
2751 if (!baseop)
2752 return NULL_TREE;
2753 if (currop->op0)
2755 op0expr = get_or_alloc_expr_for (currop->op0);
2756 genop0 = find_or_generate_expression (block, op0expr,
2757 stmts, domstmt);
2758 if (!genop0)
2759 return NULL_TREE;
2761 if (nextop->op0)
2763 op1expr = get_or_alloc_expr_for (nextop->op0);
2764 genop1 = find_or_generate_expression (block, op1expr,
2765 stmts, domstmt);
2766 if (!genop1)
2767 return NULL_TREE;
2769 return build5 (TARGET_MEM_REF, currop->type,
2770 baseop, currop->op2, genop0, currop->op1, genop1);
2772 break;
2773 case ADDR_EXPR:
2774 if (currop->op0)
2776 gcc_assert (is_gimple_min_invariant (currop->op0));
2777 return currop->op0;
2779 /* Fallthrough. */
2780 case REALPART_EXPR:
2781 case IMAGPART_EXPR:
2782 case VIEW_CONVERT_EXPR:
2784 tree folded;
2785 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2786 operand,
2787 stmts, domstmt);
2788 if (!genop0)
2789 return NULL_TREE;
2790 folded = fold_build1 (currop->opcode, currop->type,
2791 genop0);
2792 return folded;
2794 break;
2795 case BIT_FIELD_REF:
2797 tree folded;
2798 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2799 stmts, domstmt);
2800 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2801 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2802 tree genop1;
2803 tree genop2;
2805 if (!genop0)
2806 return NULL_TREE;
2807 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2808 if (!genop1)
2809 return NULL_TREE;
2810 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2811 if (!genop2)
2812 return NULL_TREE;
2813 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2814 genop2);
2815 return folded;
2818 /* For array ref vn_reference_op's, operand 1 of the array ref
2819 is op0 of the reference op and operand 3 of the array ref is
2820 op1. */
2821 case ARRAY_RANGE_REF:
2822 case ARRAY_REF:
2824 tree genop0;
2825 tree genop1 = currop->op0;
2826 pre_expr op1expr;
2827 tree genop2 = currop->op1;
2828 pre_expr op2expr;
2829 tree genop3 = currop->op2;
2830 pre_expr op3expr;
2831 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2832 stmts, domstmt);
2833 if (!genop0)
2834 return NULL_TREE;
2835 op1expr = get_or_alloc_expr_for (genop1);
2836 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2837 if (!genop1)
2838 return NULL_TREE;
2839 if (genop2)
2841 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2842 /* Drop zero minimum index if redundant. */
2843 if (integer_zerop (genop2)
2844 && (!domain_type
2845 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2846 genop2 = NULL_TREE;
2847 else
2849 op2expr = get_or_alloc_expr_for (genop2);
2850 genop2 = find_or_generate_expression (block, op2expr, stmts,
2851 domstmt);
2852 if (!genop2)
2853 return NULL_TREE;
2856 if (genop3)
2858 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2859 /* We can't always put a size in units of the element alignment
2860 here as the element alignment may be not visible. See
2861 PR43783. Simply drop the element size for constant
2862 sizes. */
2863 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2864 genop3 = NULL_TREE;
2865 else
2867 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2868 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2869 op3expr = get_or_alloc_expr_for (genop3);
2870 genop3 = find_or_generate_expression (block, op3expr, stmts,
2871 domstmt);
2872 if (!genop3)
2873 return NULL_TREE;
2876 return build4 (currop->opcode, currop->type, genop0, genop1,
2877 genop2, genop3);
2879 case COMPONENT_REF:
2881 tree op0;
2882 tree op1;
2883 tree genop2 = currop->op1;
2884 pre_expr op2expr;
2885 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2886 stmts, domstmt);
2887 if (!op0)
2888 return NULL_TREE;
2889 /* op1 should be a FIELD_DECL, which are represented by
2890 themselves. */
2891 op1 = currop->op0;
2892 if (genop2)
2894 op2expr = get_or_alloc_expr_for (genop2);
2895 genop2 = find_or_generate_expression (block, op2expr, stmts,
2896 domstmt);
2897 if (!genop2)
2898 return NULL_TREE;
2901 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2902 genop2);
2904 break;
2905 case SSA_NAME:
2907 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2908 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2909 return genop;
2911 case STRING_CST:
2912 case INTEGER_CST:
2913 case COMPLEX_CST:
2914 case VECTOR_CST:
2915 case REAL_CST:
2916 case CONSTRUCTOR:
2917 case VAR_DECL:
2918 case PARM_DECL:
2919 case CONST_DECL:
2920 case RESULT_DECL:
2921 case FUNCTION_DECL:
2922 return currop->op0;
2924 default:
2925 gcc_unreachable ();
2929 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2930 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2931 trying to rename aggregates into ssa form directly, which is a no no.
2933 Thus, this routine doesn't create temporaries, it just builds a
2934 single access expression for the array, calling
2935 find_or_generate_expression to build the innermost pieces.
2937 This function is a subroutine of create_expression_by_pieces, and
2938 should not be called on it's own unless you really know what you
2939 are doing. */
2941 static tree
2942 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2943 gimple_seq *stmts, gimple domstmt)
2945 unsigned int op = 0;
2946 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2949 /* Find a leader for an expression, or generate one using
2950 create_expression_by_pieces if it's ANTIC but
2951 complex.
2952 BLOCK is the basic_block we are looking for leaders in.
2953 EXPR is the expression to find a leader or generate for.
2954 STMTS is the statement list to put the inserted expressions on.
2955 Returns the SSA_NAME of the LHS of the generated expression or the
2956 leader.
2957 DOMSTMT if non-NULL is a statement that should be dominated by
2958 all uses in the generated expression. If DOMSTMT is non-NULL this
2959 routine can fail and return NULL_TREE. Otherwise it will assert
2960 on failure. */
2962 static tree
2963 find_or_generate_expression (basic_block block, pre_expr expr,
2964 gimple_seq *stmts, gimple domstmt)
2966 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2967 get_expr_value_id (expr), domstmt);
2968 tree genop = NULL;
2969 if (leader)
2971 if (leader->kind == NAME)
2972 genop = PRE_EXPR_NAME (leader);
2973 else if (leader->kind == CONSTANT)
2974 genop = PRE_EXPR_CONSTANT (leader);
2977 /* If it's still NULL, it must be a complex expression, so generate
2978 it recursively. Not so if inserting expressions for values generated
2979 by SCCVN. */
2980 if (genop == NULL
2981 && !domstmt)
2983 bitmap_set_t exprset;
2984 unsigned int lookfor = get_expr_value_id (expr);
2985 bool handled = false;
2986 bitmap_iterator bi;
2987 unsigned int i;
2989 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
2990 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
2992 pre_expr temp = expression_for_id (i);
2993 if (temp->kind != NAME)
2995 handled = true;
2996 genop = create_expression_by_pieces (block, temp, stmts,
2997 domstmt,
2998 get_expr_type (expr));
2999 break;
3002 if (!handled && domstmt)
3003 return NULL_TREE;
3005 gcc_assert (handled);
3007 return genop;
3010 #define NECESSARY GF_PLF_1
3012 /* Create an expression in pieces, so that we can handle very complex
3013 expressions that may be ANTIC, but not necessary GIMPLE.
3014 BLOCK is the basic block the expression will be inserted into,
3015 EXPR is the expression to insert (in value form)
3016 STMTS is a statement list to append the necessary insertions into.
3018 This function will die if we hit some value that shouldn't be
3019 ANTIC but is (IE there is no leader for it, or its components).
3020 This function may also generate expressions that are themselves
3021 partially or fully redundant. Those that are will be either made
3022 fully redundant during the next iteration of insert (for partially
3023 redundant ones), or eliminated by eliminate (for fully redundant
3024 ones).
3026 If DOMSTMT is non-NULL then we make sure that all uses in the
3027 expressions dominate that statement. In this case the function
3028 can return NULL_TREE to signal failure. */
3030 static tree
3031 create_expression_by_pieces (basic_block block, pre_expr expr,
3032 gimple_seq *stmts, gimple domstmt, tree type)
3034 tree temp, name;
3035 tree folded;
3036 gimple_seq forced_stmts = NULL;
3037 unsigned int value_id;
3038 gimple_stmt_iterator gsi;
3039 tree exprtype = type ? type : get_expr_type (expr);
3040 pre_expr nameexpr;
3041 gimple newstmt;
3043 switch (expr->kind)
3045 /* We may hit the NAME/CONSTANT case if we have to convert types
3046 that value numbering saw through. */
3047 case NAME:
3048 folded = PRE_EXPR_NAME (expr);
3049 break;
3050 case CONSTANT:
3051 folded = PRE_EXPR_CONSTANT (expr);
3052 break;
3053 case REFERENCE:
3055 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3056 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3058 break;
3059 case NARY:
3061 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3062 tree genop[4];
3063 unsigned i;
3064 for (i = 0; i < nary->length; ++i)
3066 pre_expr op = get_or_alloc_expr_for (nary->op[i]);
3067 genop[i] = find_or_generate_expression (block, op,
3068 stmts, domstmt);
3069 if (!genop[i])
3070 return NULL_TREE;
3071 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
3072 may have conversions stripped. */
3073 if (nary->opcode == POINTER_PLUS_EXPR)
3075 if (i == 0)
3076 genop[i] = fold_convert (nary->type, genop[i]);
3077 else if (i == 1)
3078 genop[i] = convert_to_ptrofftype (genop[i]);
3080 else
3081 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]);
3083 if (nary->opcode == CONSTRUCTOR)
3085 VEC(constructor_elt,gc) *elts = NULL;
3086 for (i = 0; i < nary->length; ++i)
3087 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
3088 folded = build_constructor (nary->type, elts);
3090 else
3092 switch (nary->length)
3094 case 1:
3095 folded = fold_build1 (nary->opcode, nary->type,
3096 genop[0]);
3097 break;
3098 case 2:
3099 folded = fold_build2 (nary->opcode, nary->type,
3100 genop[0], genop[1]);
3101 break;
3102 case 3:
3103 folded = fold_build3 (nary->opcode, nary->type,
3104 genop[0], genop[1], genop[3]);
3105 break;
3106 default:
3107 gcc_unreachable ();
3111 break;
3112 default:
3113 return NULL_TREE;
3116 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3117 folded = fold_convert (exprtype, folded);
3119 /* Force the generated expression to be a sequence of GIMPLE
3120 statements.
3121 We have to call unshare_expr because force_gimple_operand may
3122 modify the tree we pass to it. */
3123 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3124 false, NULL);
3126 /* If we have any intermediate expressions to the value sets, add them
3127 to the value sets and chain them in the instruction stream. */
3128 if (forced_stmts)
3130 gsi = gsi_start (forced_stmts);
3131 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3133 gimple stmt = gsi_stmt (gsi);
3134 tree forcedname = gimple_get_lhs (stmt);
3135 pre_expr nameexpr;
3137 if (TREE_CODE (forcedname) == SSA_NAME)
3139 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3140 VN_INFO_GET (forcedname)->valnum = forcedname;
3141 VN_INFO (forcedname)->value_id = get_next_value_id ();
3142 nameexpr = get_or_alloc_expr_for_name (forcedname);
3143 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3144 if (!in_fre)
3145 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3146 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3148 mark_symbols_for_renaming (stmt);
3150 gimple_seq_add_seq (stmts, forced_stmts);
3153 /* Build and insert the assignment of the end result to the temporary
3154 that we will return. */
3155 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3156 pretemp = create_tmp_reg (exprtype, "pretmp");
3158 temp = pretemp;
3159 add_referenced_var (temp);
3161 newstmt = gimple_build_assign (temp, folded);
3162 name = make_ssa_name (temp, newstmt);
3163 gimple_assign_set_lhs (newstmt, name);
3164 gimple_set_plf (newstmt, NECESSARY, false);
3166 gimple_seq_add_stmt (stmts, newstmt);
3167 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3169 /* All the symbols in NEWEXPR should be put into SSA form. */
3170 mark_symbols_for_renaming (newstmt);
3172 /* Fold the last statement. */
3173 gsi = gsi_last (*stmts);
3174 if (fold_stmt_inplace (&gsi))
3175 update_stmt (gsi_stmt (gsi));
3177 /* Add a value number to the temporary.
3178 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3179 we are creating the expression by pieces, and this particular piece of
3180 the expression may have been represented. There is no harm in replacing
3181 here. */
3182 VN_INFO_GET (name)->valnum = name;
3183 value_id = get_expr_value_id (expr);
3184 VN_INFO (name)->value_id = value_id;
3185 nameexpr = get_or_alloc_expr_for_name (name);
3186 add_to_value (value_id, nameexpr);
3187 if (NEW_SETS (block))
3188 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3189 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3191 pre_stats.insertions++;
3192 if (dump_file && (dump_flags & TDF_DETAILS))
3194 fprintf (dump_file, "Inserted ");
3195 print_gimple_stmt (dump_file, newstmt, 0, 0);
3196 fprintf (dump_file, " in predecessor %d\n", block->index);
3199 return name;
3203 /* Returns true if we want to inhibit the insertions of PHI nodes
3204 for the given EXPR for basic block BB (a member of a loop).
3205 We want to do this, when we fear that the induction variable we
3206 create might inhibit vectorization. */
3208 static bool
3209 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3211 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3212 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3213 vn_reference_op_t op;
3214 unsigned i;
3216 /* If we aren't going to vectorize we don't inhibit anything. */
3217 if (!flag_tree_vectorize)
3218 return false;
3220 /* Otherwise we inhibit the insertion when the address of the
3221 memory reference is a simple induction variable. In other
3222 cases the vectorizer won't do anything anyway (either it's
3223 loop invariant or a complicated expression). */
3224 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3226 switch (op->opcode)
3228 case ARRAY_REF:
3229 case ARRAY_RANGE_REF:
3230 if (TREE_CODE (op->op0) != SSA_NAME)
3231 break;
3232 /* Fallthru. */
3233 case SSA_NAME:
3235 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3236 affine_iv iv;
3237 /* Default defs are loop invariant. */
3238 if (!defbb)
3239 break;
3240 /* Defined outside this loop, also loop invariant. */
3241 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3242 break;
3243 /* If it's a simple induction variable inhibit insertion,
3244 the vectorizer might be interested in this one. */
3245 if (simple_iv (bb->loop_father, bb->loop_father,
3246 op->op0, &iv, true))
3247 return true;
3248 /* No simple IV, vectorizer can't do anything, hence no
3249 reason to inhibit the transformation for this operand. */
3250 break;
3252 default:
3253 break;
3256 return false;
3259 /* Insert the to-be-made-available values of expression EXPRNUM for each
3260 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3261 merge the result with a phi node, given the same value number as
3262 NODE. Return true if we have inserted new stuff. */
3264 static bool
3265 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3266 pre_expr *avail)
3268 pre_expr expr = expression_for_id (exprnum);
3269 pre_expr newphi;
3270 unsigned int val = get_expr_value_id (expr);
3271 edge pred;
3272 bool insertions = false;
3273 bool nophi = false;
3274 basic_block bprime;
3275 pre_expr eprime;
3276 edge_iterator ei;
3277 tree type = get_expr_type (expr);
3278 tree temp;
3279 gimple phi;
3281 if (dump_file && (dump_flags & TDF_DETAILS))
3283 fprintf (dump_file, "Found partial redundancy for expression ");
3284 print_pre_expr (dump_file, expr);
3285 fprintf (dump_file, " (%04d)\n", val);
3288 /* Make sure we aren't creating an induction variable. */
3289 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3291 bool firstinsideloop = false;
3292 bool secondinsideloop = false;
3293 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3294 EDGE_PRED (block, 0)->src);
3295 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3296 EDGE_PRED (block, 1)->src);
3297 /* Induction variables only have one edge inside the loop. */
3298 if ((firstinsideloop ^ secondinsideloop)
3299 && (expr->kind != REFERENCE
3300 || inhibit_phi_insertion (block, expr)))
3302 if (dump_file && (dump_flags & TDF_DETAILS))
3303 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3304 nophi = true;
3308 /* Make the necessary insertions. */
3309 FOR_EACH_EDGE (pred, ei, block->preds)
3311 gimple_seq stmts = NULL;
3312 tree builtexpr;
3313 bprime = pred->src;
3314 eprime = avail[bprime->index];
3316 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3318 builtexpr = create_expression_by_pieces (bprime,
3319 eprime,
3320 &stmts, NULL,
3321 type);
3322 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3323 gsi_insert_seq_on_edge (pred, stmts);
3324 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3325 insertions = true;
3327 else if (eprime->kind == CONSTANT)
3329 /* Constants may not have the right type, fold_convert
3330 should give us back a constant with the right type.
3332 tree constant = PRE_EXPR_CONSTANT (eprime);
3333 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3335 tree builtexpr = fold_convert (type, constant);
3336 if (!is_gimple_min_invariant (builtexpr))
3338 tree forcedexpr = force_gimple_operand (builtexpr,
3339 &stmts, true,
3340 NULL);
3341 if (!is_gimple_min_invariant (forcedexpr))
3343 if (forcedexpr != builtexpr)
3345 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3346 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3348 if (stmts)
3350 gimple_stmt_iterator gsi;
3351 gsi = gsi_start (stmts);
3352 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3354 gimple stmt = gsi_stmt (gsi);
3355 tree lhs = gimple_get_lhs (stmt);
3356 if (TREE_CODE (lhs) == SSA_NAME)
3357 bitmap_set_bit (inserted_exprs,
3358 SSA_NAME_VERSION (lhs));
3359 gimple_set_plf (stmt, NECESSARY, false);
3361 gsi_insert_seq_on_edge (pred, stmts);
3363 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3366 else
3367 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3370 else if (eprime->kind == NAME)
3372 /* We may have to do a conversion because our value
3373 numbering can look through types in certain cases, but
3374 our IL requires all operands of a phi node have the same
3375 type. */
3376 tree name = PRE_EXPR_NAME (eprime);
3377 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3379 tree builtexpr;
3380 tree forcedexpr;
3381 builtexpr = fold_convert (type, name);
3382 forcedexpr = force_gimple_operand (builtexpr,
3383 &stmts, true,
3384 NULL);
3386 if (forcedexpr != name)
3388 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3389 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3392 if (stmts)
3394 gimple_stmt_iterator gsi;
3395 gsi = gsi_start (stmts);
3396 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3398 gimple stmt = gsi_stmt (gsi);
3399 tree lhs = gimple_get_lhs (stmt);
3400 if (TREE_CODE (lhs) == SSA_NAME)
3401 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3402 gimple_set_plf (stmt, NECESSARY, false);
3404 gsi_insert_seq_on_edge (pred, stmts);
3406 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3410 /* If we didn't want a phi node, and we made insertions, we still have
3411 inserted new stuff, and thus return true. If we didn't want a phi node,
3412 and didn't make insertions, we haven't added anything new, so return
3413 false. */
3414 if (nophi && insertions)
3415 return true;
3416 else if (nophi && !insertions)
3417 return false;
3419 /* Now build a phi for the new variable. */
3420 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3421 prephitemp = create_tmp_var (type, "prephitmp");
3423 temp = prephitemp;
3424 add_referenced_var (temp);
3426 if (TREE_CODE (type) == COMPLEX_TYPE
3427 || TREE_CODE (type) == VECTOR_TYPE)
3428 DECL_GIMPLE_REG_P (temp) = 1;
3429 phi = create_phi_node (temp, block);
3431 gimple_set_plf (phi, NECESSARY, false);
3432 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3433 VN_INFO (gimple_phi_result (phi))->value_id = val;
3434 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3435 FOR_EACH_EDGE (pred, ei, block->preds)
3437 pre_expr ae = avail[pred->src->index];
3438 gcc_assert (get_expr_type (ae) == type
3439 || useless_type_conversion_p (type, get_expr_type (ae)));
3440 if (ae->kind == CONSTANT)
3441 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3442 else
3443 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3444 UNKNOWN_LOCATION);
3447 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3448 add_to_value (val, newphi);
3450 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3451 this insertion, since we test for the existence of this value in PHI_GEN
3452 before proceeding with the partial redundancy checks in insert_aux.
3454 The value may exist in AVAIL_OUT, in particular, it could be represented
3455 by the expression we are trying to eliminate, in which case we want the
3456 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3457 inserted there.
3459 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3460 this block, because if it did, it would have existed in our dominator's
3461 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3464 bitmap_insert_into_set (PHI_GEN (block), newphi);
3465 bitmap_value_replace_in_set (AVAIL_OUT (block),
3466 newphi);
3467 bitmap_insert_into_set (NEW_SETS (block),
3468 newphi);
3470 if (dump_file && (dump_flags & TDF_DETAILS))
3472 fprintf (dump_file, "Created phi ");
3473 print_gimple_stmt (dump_file, phi, 0, 0);
3474 fprintf (dump_file, " in block %d\n", block->index);
3476 pre_stats.phis++;
3477 return true;
3482 /* Perform insertion of partially redundant values.
3483 For BLOCK, do the following:
3484 1. Propagate the NEW_SETS of the dominator into the current block.
3485 If the block has multiple predecessors,
3486 2a. Iterate over the ANTIC expressions for the block to see if
3487 any of them are partially redundant.
3488 2b. If so, insert them into the necessary predecessors to make
3489 the expression fully redundant.
3490 2c. Insert a new PHI merging the values of the predecessors.
3491 2d. Insert the new PHI, and the new expressions, into the
3492 NEW_SETS set.
3493 3. Recursively call ourselves on the dominator children of BLOCK.
3495 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3496 do_regular_insertion and do_partial_insertion.
3500 static bool
3501 do_regular_insertion (basic_block block, basic_block dom)
3503 bool new_stuff = false;
3504 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3505 pre_expr expr;
3506 int i;
3508 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3510 if (expr->kind != NAME)
3512 pre_expr *avail;
3513 unsigned int val;
3514 bool by_some = false;
3515 bool cant_insert = false;
3516 bool all_same = true;
3517 pre_expr first_s = NULL;
3518 edge pred;
3519 basic_block bprime;
3520 pre_expr eprime = NULL;
3521 edge_iterator ei;
3522 pre_expr edoubleprime = NULL;
3523 bool do_insertion = false;
3525 val = get_expr_value_id (expr);
3526 if (bitmap_set_contains_value (PHI_GEN (block), val))
3527 continue;
3528 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3530 if (dump_file && (dump_flags & TDF_DETAILS))
3531 fprintf (dump_file, "Found fully redundant value\n");
3532 continue;
3535 avail = XCNEWVEC (pre_expr, last_basic_block);
3536 FOR_EACH_EDGE (pred, ei, block->preds)
3538 unsigned int vprime;
3540 /* We should never run insertion for the exit block
3541 and so not come across fake pred edges. */
3542 gcc_assert (!(pred->flags & EDGE_FAKE));
3543 bprime = pred->src;
3544 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3545 bprime, block);
3547 /* eprime will generally only be NULL if the
3548 value of the expression, translated
3549 through the PHI for this predecessor, is
3550 undefined. If that is the case, we can't
3551 make the expression fully redundant,
3552 because its value is undefined along a
3553 predecessor path. We can thus break out
3554 early because it doesn't matter what the
3555 rest of the results are. */
3556 if (eprime == NULL)
3558 cant_insert = true;
3559 break;
3562 eprime = fully_constant_expression (eprime);
3563 vprime = get_expr_value_id (eprime);
3564 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3565 vprime, NULL);
3566 if (edoubleprime == NULL)
3568 avail[bprime->index] = eprime;
3569 all_same = false;
3571 else
3573 avail[bprime->index] = edoubleprime;
3574 by_some = true;
3575 /* We want to perform insertions to remove a redundancy on
3576 a path in the CFG we want to optimize for speed. */
3577 if (optimize_edge_for_speed_p (pred))
3578 do_insertion = true;
3579 if (first_s == NULL)
3580 first_s = edoubleprime;
3581 else if (!pre_expr_eq (first_s, edoubleprime))
3582 all_same = false;
3585 /* If we can insert it, it's not the same value
3586 already existing along every predecessor, and
3587 it's defined by some predecessor, it is
3588 partially redundant. */
3589 if (!cant_insert && !all_same && by_some)
3591 if (!do_insertion)
3593 if (dump_file && (dump_flags & TDF_DETAILS))
3595 fprintf (dump_file, "Skipping partial redundancy for "
3596 "expression ");
3597 print_pre_expr (dump_file, expr);
3598 fprintf (dump_file, " (%04d), no redundancy on to be "
3599 "optimized for speed edge\n", val);
3602 else if (dbg_cnt (treepre_insert)
3603 && insert_into_preds_of_block (block,
3604 get_expression_id (expr),
3605 avail))
3606 new_stuff = true;
3608 /* If all edges produce the same value and that value is
3609 an invariant, then the PHI has the same value on all
3610 edges. Note this. */
3611 else if (!cant_insert && all_same && eprime
3612 && (edoubleprime->kind == CONSTANT
3613 || edoubleprime->kind == NAME)
3614 && !value_id_constant_p (val))
3616 unsigned int j;
3617 bitmap_iterator bi;
3618 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3619 value_expressions, val);
3621 unsigned int new_val = get_expr_value_id (edoubleprime);
3622 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3624 pre_expr expr = expression_for_id (j);
3626 if (expr->kind == NAME)
3628 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3629 /* Just reset the value id and valnum so it is
3630 the same as the constant we have discovered. */
3631 if (edoubleprime->kind == CONSTANT)
3633 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3634 pre_stats.constified++;
3636 else
3637 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3638 info->value_id = new_val;
3642 free (avail);
3646 VEC_free (pre_expr, heap, exprs);
3647 return new_stuff;
3651 /* Perform insertion for partially anticipatable expressions. There
3652 is only one case we will perform insertion for these. This case is
3653 if the expression is partially anticipatable, and fully available.
3654 In this case, we know that putting it earlier will enable us to
3655 remove the later computation. */
3658 static bool
3659 do_partial_partial_insertion (basic_block block, basic_block dom)
3661 bool new_stuff = false;
3662 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3663 pre_expr expr;
3664 int i;
3666 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3668 if (expr->kind != NAME)
3670 pre_expr *avail;
3671 unsigned int val;
3672 bool by_all = true;
3673 bool cant_insert = false;
3674 edge pred;
3675 basic_block bprime;
3676 pre_expr eprime = NULL;
3677 edge_iterator ei;
3679 val = get_expr_value_id (expr);
3680 if (bitmap_set_contains_value (PHI_GEN (block), val))
3681 continue;
3682 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3683 continue;
3685 avail = XCNEWVEC (pre_expr, last_basic_block);
3686 FOR_EACH_EDGE (pred, ei, block->preds)
3688 unsigned int vprime;
3689 pre_expr edoubleprime;
3691 /* We should never run insertion for the exit block
3692 and so not come across fake pred edges. */
3693 gcc_assert (!(pred->flags & EDGE_FAKE));
3694 bprime = pred->src;
3695 eprime = phi_translate (expr, ANTIC_IN (block),
3696 PA_IN (block),
3697 bprime, block);
3699 /* eprime will generally only be NULL if the
3700 value of the expression, translated
3701 through the PHI for this predecessor, is
3702 undefined. If that is the case, we can't
3703 make the expression fully redundant,
3704 because its value is undefined along a
3705 predecessor path. We can thus break out
3706 early because it doesn't matter what the
3707 rest of the results are. */
3708 if (eprime == NULL)
3710 cant_insert = true;
3711 break;
3714 eprime = fully_constant_expression (eprime);
3715 vprime = get_expr_value_id (eprime);
3716 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3717 vprime, NULL);
3718 if (edoubleprime == NULL)
3720 by_all = false;
3721 break;
3723 else
3724 avail[bprime->index] = edoubleprime;
3728 /* If we can insert it, it's not the same value
3729 already existing along every predecessor, and
3730 it's defined by some predecessor, it is
3731 partially redundant. */
3732 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3734 pre_stats.pa_insert++;
3735 if (insert_into_preds_of_block (block, get_expression_id (expr),
3736 avail))
3737 new_stuff = true;
3739 free (avail);
3743 VEC_free (pre_expr, heap, exprs);
3744 return new_stuff;
3747 static bool
3748 insert_aux (basic_block block)
3750 basic_block son;
3751 bool new_stuff = false;
3753 if (block)
3755 basic_block dom;
3756 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3757 if (dom)
3759 unsigned i;
3760 bitmap_iterator bi;
3761 bitmap_set_t newset = NEW_SETS (dom);
3762 if (newset)
3764 /* Note that we need to value_replace both NEW_SETS, and
3765 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3766 represented by some non-simple expression here that we want
3767 to replace it with. */
3768 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3770 pre_expr expr = expression_for_id (i);
3771 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3772 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3775 if (!single_pred_p (block))
3777 new_stuff |= do_regular_insertion (block, dom);
3778 if (do_partial_partial)
3779 new_stuff |= do_partial_partial_insertion (block, dom);
3783 for (son = first_dom_son (CDI_DOMINATORS, block);
3784 son;
3785 son = next_dom_son (CDI_DOMINATORS, son))
3787 new_stuff |= insert_aux (son);
3790 return new_stuff;
3793 /* Perform insertion of partially redundant values. */
3795 static void
3796 insert (void)
3798 bool new_stuff = true;
3799 basic_block bb;
3800 int num_iterations = 0;
3802 FOR_ALL_BB (bb)
3803 NEW_SETS (bb) = bitmap_set_new ();
3805 while (new_stuff)
3807 num_iterations++;
3808 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3810 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3814 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3816 static void
3817 add_to_exp_gen (basic_block block, tree op)
3819 if (!in_fre)
3821 pre_expr result;
3822 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3823 return;
3824 result = get_or_alloc_expr_for_name (op);
3825 bitmap_value_insert_into_set (EXP_GEN (block), result);
3829 /* Create value ids for PHI in BLOCK. */
3831 static void
3832 make_values_for_phi (gimple phi, basic_block block)
3834 tree result = gimple_phi_result (phi);
3836 /* We have no need for virtual phis, as they don't represent
3837 actual computations. */
3838 if (is_gimple_reg (result))
3840 pre_expr e = get_or_alloc_expr_for_name (result);
3841 add_to_value (get_expr_value_id (e), e);
3842 bitmap_insert_into_set (PHI_GEN (block), e);
3843 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3844 if (!in_fre)
3846 unsigned i;
3847 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3849 tree arg = gimple_phi_arg_def (phi, i);
3850 if (TREE_CODE (arg) == SSA_NAME)
3852 e = get_or_alloc_expr_for_name (arg);
3853 add_to_value (get_expr_value_id (e), e);
3860 /* Compute the AVAIL set for all basic blocks.
3862 This function performs value numbering of the statements in each basic
3863 block. The AVAIL sets are built from information we glean while doing
3864 this value numbering, since the AVAIL sets contain only one entry per
3865 value.
3867 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3868 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3870 static void
3871 compute_avail (void)
3874 basic_block block, son;
3875 basic_block *worklist;
3876 size_t sp = 0;
3877 unsigned i;
3879 /* We pretend that default definitions are defined in the entry block.
3880 This includes function arguments and the static chain decl. */
3881 for (i = 1; i < num_ssa_names; ++i)
3883 tree name = ssa_name (i);
3884 pre_expr e;
3885 if (!name
3886 || !SSA_NAME_IS_DEFAULT_DEF (name)
3887 || has_zero_uses (name)
3888 || !is_gimple_reg (name))
3889 continue;
3891 e = get_or_alloc_expr_for_name (name);
3892 add_to_value (get_expr_value_id (e), e);
3893 if (!in_fre)
3894 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3895 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3898 /* Allocate the worklist. */
3899 worklist = XNEWVEC (basic_block, n_basic_blocks);
3901 /* Seed the algorithm by putting the dominator children of the entry
3902 block on the worklist. */
3903 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3904 son;
3905 son = next_dom_son (CDI_DOMINATORS, son))
3906 worklist[sp++] = son;
3908 /* Loop until the worklist is empty. */
3909 while (sp)
3911 gimple_stmt_iterator gsi;
3912 gimple stmt;
3913 basic_block dom;
3914 unsigned int stmt_uid = 1;
3916 /* Pick a block from the worklist. */
3917 block = worklist[--sp];
3919 /* Initially, the set of available values in BLOCK is that of
3920 its immediate dominator. */
3921 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3922 if (dom)
3923 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3925 /* Generate values for PHI nodes. */
3926 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3927 make_values_for_phi (gsi_stmt (gsi), block);
3929 BB_MAY_NOTRETURN (block) = 0;
3931 /* Now compute value numbers and populate value sets with all
3932 the expressions computed in BLOCK. */
3933 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3935 ssa_op_iter iter;
3936 tree op;
3938 stmt = gsi_stmt (gsi);
3939 gimple_set_uid (stmt, stmt_uid++);
3941 /* Cache whether the basic-block has any non-visible side-effect
3942 or control flow.
3943 If this isn't a call or it is the last stmt in the
3944 basic-block then the CFG represents things correctly. */
3945 if (is_gimple_call (stmt)
3946 && !stmt_ends_bb_p (stmt))
3948 /* Non-looping const functions always return normally.
3949 Otherwise the call might not return or have side-effects
3950 that forbids hoisting possibly trapping expressions
3951 before it. */
3952 int flags = gimple_call_flags (stmt);
3953 if (!(flags & ECF_CONST)
3954 || (flags & ECF_LOOPING_CONST_OR_PURE))
3955 BB_MAY_NOTRETURN (block) = 1;
3958 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3960 pre_expr e = get_or_alloc_expr_for_name (op);
3962 add_to_value (get_expr_value_id (e), e);
3963 if (!in_fre)
3964 bitmap_insert_into_set (TMP_GEN (block), e);
3965 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3968 if (gimple_has_volatile_ops (stmt)
3969 || stmt_could_throw_p (stmt))
3970 continue;
3972 switch (gimple_code (stmt))
3974 case GIMPLE_RETURN:
3975 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3976 add_to_exp_gen (block, op);
3977 continue;
3979 case GIMPLE_CALL:
3981 vn_reference_t ref;
3982 unsigned int i;
3983 vn_reference_op_t vro;
3984 pre_expr result = NULL;
3985 VEC(vn_reference_op_s, heap) *ops = NULL;
3987 if (!can_value_number_call (stmt))
3988 continue;
3990 copy_reference_ops_from_call (stmt, &ops);
3991 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
3992 gimple_expr_type (stmt),
3993 ops, &ref, VN_NOWALK);
3994 VEC_free (vn_reference_op_s, heap, ops);
3995 if (!ref)
3996 continue;
3998 for (i = 0; VEC_iterate (vn_reference_op_s,
3999 ref->operands, i,
4000 vro); i++)
4002 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4003 add_to_exp_gen (block, vro->op0);
4004 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4005 add_to_exp_gen (block, vro->op1);
4006 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4007 add_to_exp_gen (block, vro->op2);
4009 result = (pre_expr) pool_alloc (pre_expr_pool);
4010 result->kind = REFERENCE;
4011 result->id = 0;
4012 PRE_EXPR_REFERENCE (result) = ref;
4014 get_or_alloc_expression_id (result);
4015 add_to_value (get_expr_value_id (result), result);
4016 if (!in_fre)
4017 bitmap_value_insert_into_set (EXP_GEN (block), result);
4018 continue;
4021 case GIMPLE_ASSIGN:
4023 pre_expr result = NULL;
4024 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4026 case tcc_unary:
4027 case tcc_binary:
4028 case tcc_comparison:
4030 vn_nary_op_t nary;
4031 unsigned int i;
4033 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4034 gimple_assign_rhs_code (stmt),
4035 gimple_expr_type (stmt),
4036 gimple_assign_rhs1_ptr (stmt),
4037 &nary);
4039 if (!nary)
4040 continue;
4042 for (i = 0; i < nary->length; i++)
4043 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4044 add_to_exp_gen (block, nary->op[i]);
4046 result = (pre_expr) pool_alloc (pre_expr_pool);
4047 result->kind = NARY;
4048 result->id = 0;
4049 PRE_EXPR_NARY (result) = nary;
4050 break;
4053 case tcc_declaration:
4054 case tcc_reference:
4056 vn_reference_t ref;
4057 unsigned int i;
4058 vn_reference_op_t vro;
4060 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4061 gimple_vuse (stmt),
4062 VN_WALK, &ref);
4063 if (!ref)
4064 continue;
4066 for (i = 0; VEC_iterate (vn_reference_op_s,
4067 ref->operands, i,
4068 vro); i++)
4070 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4071 add_to_exp_gen (block, vro->op0);
4072 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4073 add_to_exp_gen (block, vro->op1);
4074 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4075 add_to_exp_gen (block, vro->op2);
4077 result = (pre_expr) pool_alloc (pre_expr_pool);
4078 result->kind = REFERENCE;
4079 result->id = 0;
4080 PRE_EXPR_REFERENCE (result) = ref;
4081 break;
4084 default:
4085 /* For any other statement that we don't
4086 recognize, simply add all referenced
4087 SSA_NAMEs to EXP_GEN. */
4088 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4089 add_to_exp_gen (block, op);
4090 continue;
4093 get_or_alloc_expression_id (result);
4094 add_to_value (get_expr_value_id (result), result);
4095 if (!in_fre)
4096 bitmap_value_insert_into_set (EXP_GEN (block), result);
4098 continue;
4100 default:
4101 break;
4105 /* Put the dominator children of BLOCK on the worklist of blocks
4106 to compute available sets for. */
4107 for (son = first_dom_son (CDI_DOMINATORS, block);
4108 son;
4109 son = next_dom_son (CDI_DOMINATORS, son))
4110 worklist[sp++] = son;
4113 free (worklist);
4116 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4117 than the available expressions for it. The insertion point is
4118 right before the first use in STMT. Returns the SSA_NAME that should
4119 be used for replacement. */
4121 static tree
4122 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4124 basic_block bb = gimple_bb (stmt);
4125 gimple_stmt_iterator gsi;
4126 gimple_seq stmts = NULL;
4127 tree expr;
4128 pre_expr e;
4130 /* First create a value expression from the expression we want
4131 to insert and associate it with the value handle for SSA_VN. */
4132 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4133 if (e == NULL)
4134 return NULL_TREE;
4136 /* Then use create_expression_by_pieces to generate a valid
4137 expression to insert at this point of the IL stream. */
4138 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4139 if (expr == NULL_TREE)
4140 return NULL_TREE;
4141 gsi = gsi_for_stmt (stmt);
4142 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4144 return expr;
4147 /* Eliminate fully redundant computations. */
4149 static unsigned int
4150 eliminate (void)
4152 VEC (gimple, heap) *to_remove = NULL;
4153 VEC (gimple, heap) *to_update = NULL;
4154 basic_block b;
4155 unsigned int todo = 0;
4156 gimple_stmt_iterator gsi;
4157 gimple stmt;
4158 unsigned i;
4160 FOR_EACH_BB (b)
4162 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4164 stmt = gsi_stmt (gsi);
4166 /* Lookup the RHS of the expression, see if we have an
4167 available computation for it. If so, replace the RHS with
4168 the available computation. */
4169 if (gimple_has_lhs (stmt)
4170 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4171 && !gimple_assign_ssa_name_copy_p (stmt)
4172 && (!gimple_assign_single_p (stmt)
4173 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4174 && !gimple_has_volatile_ops (stmt)
4175 && !has_zero_uses (gimple_get_lhs (stmt)))
4177 tree lhs = gimple_get_lhs (stmt);
4178 tree rhs = NULL_TREE;
4179 tree sprime = NULL;
4180 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4181 pre_expr sprimeexpr;
4183 if (gimple_assign_single_p (stmt))
4184 rhs = gimple_assign_rhs1 (stmt);
4186 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4187 get_expr_value_id (lhsexpr),
4188 NULL);
4190 if (sprimeexpr)
4192 if (sprimeexpr->kind == CONSTANT)
4193 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4194 else if (sprimeexpr->kind == NAME)
4195 sprime = PRE_EXPR_NAME (sprimeexpr);
4196 else
4197 gcc_unreachable ();
4200 /* If there is no existing leader but SCCVN knows this
4201 value is constant, use that constant. */
4202 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4204 sprime = VN_INFO (lhs)->valnum;
4205 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4206 TREE_TYPE (sprime)))
4207 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4209 if (dump_file && (dump_flags & TDF_DETAILS))
4211 fprintf (dump_file, "Replaced ");
4212 print_gimple_expr (dump_file, stmt, 0, 0);
4213 fprintf (dump_file, " with ");
4214 print_generic_expr (dump_file, sprime, 0);
4215 fprintf (dump_file, " in ");
4216 print_gimple_stmt (dump_file, stmt, 0, 0);
4218 pre_stats.eliminations++;
4219 propagate_tree_value_into_stmt (&gsi, sprime);
4220 stmt = gsi_stmt (gsi);
4221 update_stmt (stmt);
4222 continue;
4225 /* If there is no existing usable leader but SCCVN thinks
4226 it has an expression it wants to use as replacement,
4227 insert that. */
4228 if (!sprime || sprime == lhs)
4230 tree val = VN_INFO (lhs)->valnum;
4231 if (val != VN_TOP
4232 && TREE_CODE (val) == SSA_NAME
4233 && VN_INFO (val)->needs_insertion
4234 && can_PRE_operation (vn_get_expr_for (val)))
4235 sprime = do_SCCVN_insertion (stmt, val);
4237 if (sprime
4238 && sprime != lhs
4239 && (rhs == NULL_TREE
4240 || TREE_CODE (rhs) != SSA_NAME
4241 || may_propagate_copy (rhs, sprime)))
4243 bool can_make_abnormal_goto
4244 = is_gimple_call (stmt)
4245 && stmt_can_make_abnormal_goto (stmt);
4247 gcc_assert (sprime != rhs);
4249 if (dump_file && (dump_flags & TDF_DETAILS))
4251 fprintf (dump_file, "Replaced ");
4252 print_gimple_expr (dump_file, stmt, 0, 0);
4253 fprintf (dump_file, " with ");
4254 print_generic_expr (dump_file, sprime, 0);
4255 fprintf (dump_file, " in ");
4256 print_gimple_stmt (dump_file, stmt, 0, 0);
4259 if (TREE_CODE (sprime) == SSA_NAME)
4260 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4261 NECESSARY, true);
4262 /* We need to make sure the new and old types actually match,
4263 which may require adding a simple cast, which fold_convert
4264 will do for us. */
4265 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4266 && !useless_type_conversion_p (gimple_expr_type (stmt),
4267 TREE_TYPE (sprime)))
4268 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4270 pre_stats.eliminations++;
4271 propagate_tree_value_into_stmt (&gsi, sprime);
4272 stmt = gsi_stmt (gsi);
4273 update_stmt (stmt);
4275 /* If we removed EH side-effects from the statement, clean
4276 its EH information. */
4277 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4279 bitmap_set_bit (need_eh_cleanup,
4280 gimple_bb (stmt)->index);
4281 if (dump_file && (dump_flags & TDF_DETAILS))
4282 fprintf (dump_file, " Removed EH side-effects.\n");
4285 /* Likewise for AB side-effects. */
4286 if (can_make_abnormal_goto
4287 && !stmt_can_make_abnormal_goto (stmt))
4289 bitmap_set_bit (need_ab_cleanup,
4290 gimple_bb (stmt)->index);
4291 if (dump_file && (dump_flags & TDF_DETAILS))
4292 fprintf (dump_file, " Removed AB side-effects.\n");
4296 /* If the statement is a scalar store, see if the expression
4297 has the same value number as its rhs. If so, the store is
4298 dead. */
4299 else if (gimple_assign_single_p (stmt)
4300 && !is_gimple_reg (gimple_assign_lhs (stmt))
4301 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4302 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4304 tree rhs = gimple_assign_rhs1 (stmt);
4305 tree val;
4306 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4307 gimple_vuse (stmt), VN_WALK, NULL);
4308 if (TREE_CODE (rhs) == SSA_NAME)
4309 rhs = VN_INFO (rhs)->valnum;
4310 if (val
4311 && operand_equal_p (val, rhs, 0))
4313 if (dump_file && (dump_flags & TDF_DETAILS))
4315 fprintf (dump_file, "Deleted redundant store ");
4316 print_gimple_stmt (dump_file, stmt, 0, 0);
4319 /* Queue stmt for removal. */
4320 VEC_safe_push (gimple, heap, to_remove, stmt);
4323 /* Visit COND_EXPRs and fold the comparison with the
4324 available value-numbers. */
4325 else if (gimple_code (stmt) == GIMPLE_COND)
4327 tree op0 = gimple_cond_lhs (stmt);
4328 tree op1 = gimple_cond_rhs (stmt);
4329 tree result;
4331 if (TREE_CODE (op0) == SSA_NAME)
4332 op0 = VN_INFO (op0)->valnum;
4333 if (TREE_CODE (op1) == SSA_NAME)
4334 op1 = VN_INFO (op1)->valnum;
4335 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4336 op0, op1);
4337 if (result && TREE_CODE (result) == INTEGER_CST)
4339 if (integer_zerop (result))
4340 gimple_cond_make_false (stmt);
4341 else
4342 gimple_cond_make_true (stmt);
4343 update_stmt (stmt);
4344 todo = TODO_cleanup_cfg;
4347 /* Visit indirect calls and turn them into direct calls if
4348 possible. */
4349 if (is_gimple_call (stmt))
4351 tree orig_fn = gimple_call_fn (stmt);
4352 tree fn;
4353 if (!orig_fn)
4354 continue;
4355 if (TREE_CODE (orig_fn) == SSA_NAME)
4356 fn = VN_INFO (orig_fn)->valnum;
4357 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4358 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4359 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4360 else
4361 continue;
4362 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4363 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4364 TREE_TYPE (fn)))
4366 bool can_make_abnormal_goto
4367 = stmt_can_make_abnormal_goto (stmt);
4368 bool was_noreturn = gimple_call_noreturn_p (stmt);
4370 if (dump_file && (dump_flags & TDF_DETAILS))
4372 fprintf (dump_file, "Replacing call target with ");
4373 print_generic_expr (dump_file, fn, 0);
4374 fprintf (dump_file, " in ");
4375 print_gimple_stmt (dump_file, stmt, 0, 0);
4378 gimple_call_set_fn (stmt, fn);
4379 VEC_safe_push (gimple, heap, to_update, stmt);
4381 /* When changing a call into a noreturn call, cfg cleanup
4382 is needed to fix up the noreturn call. */
4383 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4384 todo |= TODO_cleanup_cfg;
4386 /* If we removed EH side-effects from the statement, clean
4387 its EH information. */
4388 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4390 bitmap_set_bit (need_eh_cleanup,
4391 gimple_bb (stmt)->index);
4392 if (dump_file && (dump_flags & TDF_DETAILS))
4393 fprintf (dump_file, " Removed EH side-effects.\n");
4396 /* Likewise for AB side-effects. */
4397 if (can_make_abnormal_goto
4398 && !stmt_can_make_abnormal_goto (stmt))
4400 bitmap_set_bit (need_ab_cleanup,
4401 gimple_bb (stmt)->index);
4402 if (dump_file && (dump_flags & TDF_DETAILS))
4403 fprintf (dump_file, " Removed AB side-effects.\n");
4406 /* Changing an indirect call to a direct call may
4407 have exposed different semantics. This may
4408 require an SSA update. */
4409 todo |= TODO_update_ssa_only_virtuals;
4414 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4416 gimple stmt, phi = gsi_stmt (gsi);
4417 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4418 pre_expr sprimeexpr, resexpr;
4419 gimple_stmt_iterator gsi2;
4421 /* We want to perform redundant PHI elimination. Do so by
4422 replacing the PHI with a single copy if possible.
4423 Do not touch inserted, single-argument or virtual PHIs. */
4424 if (gimple_phi_num_args (phi) == 1
4425 || !is_gimple_reg (res))
4427 gsi_next (&gsi);
4428 continue;
4431 resexpr = get_or_alloc_expr_for_name (res);
4432 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4433 get_expr_value_id (resexpr), NULL);
4434 if (sprimeexpr)
4436 if (sprimeexpr->kind == CONSTANT)
4437 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4438 else if (sprimeexpr->kind == NAME)
4439 sprime = PRE_EXPR_NAME (sprimeexpr);
4440 else
4441 gcc_unreachable ();
4443 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4445 sprime = VN_INFO (res)->valnum;
4446 if (!useless_type_conversion_p (TREE_TYPE (res),
4447 TREE_TYPE (sprime)))
4448 sprime = fold_convert (TREE_TYPE (res), sprime);
4450 if (!sprime
4451 || sprime == res)
4453 gsi_next (&gsi);
4454 continue;
4457 if (dump_file && (dump_flags & TDF_DETAILS))
4459 fprintf (dump_file, "Replaced redundant PHI node defining ");
4460 print_generic_expr (dump_file, res, 0);
4461 fprintf (dump_file, " with ");
4462 print_generic_expr (dump_file, sprime, 0);
4463 fprintf (dump_file, "\n");
4466 remove_phi_node (&gsi, false);
4468 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4469 && TREE_CODE (sprime) == SSA_NAME)
4470 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4472 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4473 sprime = fold_convert (TREE_TYPE (res), sprime);
4474 stmt = gimple_build_assign (res, sprime);
4475 SSA_NAME_DEF_STMT (res) = stmt;
4476 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4478 gsi2 = gsi_after_labels (b);
4479 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4480 /* Queue the copy for eventual removal. */
4481 VEC_safe_push (gimple, heap, to_remove, stmt);
4482 /* If we inserted this PHI node ourself, it's not an elimination. */
4483 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4484 pre_stats.phis--;
4485 else
4486 pre_stats.eliminations++;
4490 /* We cannot remove stmts during BB walk, especially not release SSA
4491 names there as this confuses the VN machinery. The stmts ending
4492 up in to_remove are either stores or simple copies. */
4493 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4495 tree lhs = gimple_assign_lhs (stmt);
4496 tree rhs = gimple_assign_rhs1 (stmt);
4497 use_operand_p use_p;
4498 gimple use_stmt;
4500 /* If there is a single use only, propagate the equivalency
4501 instead of keeping the copy. */
4502 if (TREE_CODE (lhs) == SSA_NAME
4503 && TREE_CODE (rhs) == SSA_NAME
4504 && single_imm_use (lhs, &use_p, &use_stmt)
4505 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4507 SET_USE (use_p, rhs);
4508 update_stmt (use_stmt);
4509 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4510 && TREE_CODE (rhs) == SSA_NAME)
4511 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4514 /* If this is a store or a now unused copy, remove it. */
4515 if (TREE_CODE (lhs) != SSA_NAME
4516 || has_zero_uses (lhs))
4518 basic_block bb = gimple_bb (stmt);
4519 gsi = gsi_for_stmt (stmt);
4520 unlink_stmt_vdef (stmt);
4521 gsi_remove (&gsi, true);
4522 if (gimple_purge_dead_eh_edges (bb))
4523 todo |= TODO_cleanup_cfg;
4524 if (TREE_CODE (lhs) == SSA_NAME)
4525 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4526 release_defs (stmt);
4529 VEC_free (gimple, heap, to_remove);
4531 /* We cannot update call statements with virtual operands during
4532 SSA walk. This might remove them which in turn makes our
4533 VN lattice invalid. */
4534 FOR_EACH_VEC_ELT (gimple, to_update, i, stmt)
4535 update_stmt (stmt);
4536 VEC_free (gimple, heap, to_update);
4538 return todo;
4541 /* Borrow a bit of tree-ssa-dce.c for the moment.
4542 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4543 this may be a bit faster, and we may want critical edges kept split. */
4545 /* If OP's defining statement has not already been determined to be necessary,
4546 mark that statement necessary. Return the stmt, if it is newly
4547 necessary. */
4549 static inline gimple
4550 mark_operand_necessary (tree op)
4552 gimple stmt;
4554 gcc_assert (op);
4556 if (TREE_CODE (op) != SSA_NAME)
4557 return NULL;
4559 stmt = SSA_NAME_DEF_STMT (op);
4560 gcc_assert (stmt);
4562 if (gimple_plf (stmt, NECESSARY)
4563 || gimple_nop_p (stmt))
4564 return NULL;
4566 gimple_set_plf (stmt, NECESSARY, true);
4567 return stmt;
4570 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4571 to insert PHI nodes sometimes, and because value numbering of casts isn't
4572 perfect, we sometimes end up inserting dead code. This simple DCE-like
4573 pass removes any insertions we made that weren't actually used. */
4575 static void
4576 remove_dead_inserted_code (void)
4578 bitmap worklist;
4579 unsigned i;
4580 bitmap_iterator bi;
4581 gimple t;
4583 worklist = BITMAP_ALLOC (NULL);
4584 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4586 t = SSA_NAME_DEF_STMT (ssa_name (i));
4587 if (gimple_plf (t, NECESSARY))
4588 bitmap_set_bit (worklist, i);
4590 while (!bitmap_empty_p (worklist))
4592 i = bitmap_first_set_bit (worklist);
4593 bitmap_clear_bit (worklist, i);
4594 t = SSA_NAME_DEF_STMT (ssa_name (i));
4596 /* PHI nodes are somewhat special in that each PHI alternative has
4597 data and control dependencies. All the statements feeding the
4598 PHI node's arguments are always necessary. */
4599 if (gimple_code (t) == GIMPLE_PHI)
4601 unsigned k;
4603 for (k = 0; k < gimple_phi_num_args (t); k++)
4605 tree arg = PHI_ARG_DEF (t, k);
4606 if (TREE_CODE (arg) == SSA_NAME)
4608 gimple n = mark_operand_necessary (arg);
4609 if (n)
4610 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4614 else
4616 /* Propagate through the operands. Examine all the USE, VUSE and
4617 VDEF operands in this statement. Mark all the statements
4618 which feed this statement's uses as necessary. */
4619 ssa_op_iter iter;
4620 tree use;
4622 /* The operands of VDEF expressions are also needed as they
4623 represent potential definitions that may reach this
4624 statement (VDEF operands allow us to follow def-def
4625 links). */
4627 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4629 gimple n = mark_operand_necessary (use);
4630 if (n)
4631 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4636 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4638 t = SSA_NAME_DEF_STMT (ssa_name (i));
4639 if (!gimple_plf (t, NECESSARY))
4641 gimple_stmt_iterator gsi;
4643 if (dump_file && (dump_flags & TDF_DETAILS))
4645 fprintf (dump_file, "Removing unnecessary insertion:");
4646 print_gimple_stmt (dump_file, t, 0, 0);
4649 gsi = gsi_for_stmt (t);
4650 if (gimple_code (t) == GIMPLE_PHI)
4651 remove_phi_node (&gsi, true);
4652 else
4654 gsi_remove (&gsi, true);
4655 release_defs (t);
4659 BITMAP_FREE (worklist);
4662 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4663 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4664 the number of visited blocks. */
4666 static int
4667 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4669 edge_iterator *stack;
4670 int sp;
4671 int post_order_num = 0;
4672 sbitmap visited;
4674 if (include_entry_exit)
4675 post_order[post_order_num++] = EXIT_BLOCK;
4677 /* Allocate stack for back-tracking up CFG. */
4678 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4679 sp = 0;
4681 /* Allocate bitmap to track nodes that have been visited. */
4682 visited = sbitmap_alloc (last_basic_block);
4684 /* None of the nodes in the CFG have been visited yet. */
4685 sbitmap_zero (visited);
4687 /* Push the last edge on to the stack. */
4688 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4690 while (sp)
4692 edge_iterator ei;
4693 basic_block src;
4694 basic_block dest;
4696 /* Look at the edge on the top of the stack. */
4697 ei = stack[sp - 1];
4698 src = ei_edge (ei)->src;
4699 dest = ei_edge (ei)->dest;
4701 /* Check if the edge destination has been visited yet. */
4702 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4704 /* Mark that we have visited the destination. */
4705 SET_BIT (visited, src->index);
4707 if (EDGE_COUNT (src->preds) > 0)
4708 /* Since the DEST node has been visited for the first
4709 time, check its successors. */
4710 stack[sp++] = ei_start (src->preds);
4711 else
4712 post_order[post_order_num++] = src->index;
4714 else
4716 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4717 post_order[post_order_num++] = dest->index;
4719 if (!ei_one_before_end_p (ei))
4720 ei_next (&stack[sp - 1]);
4721 else
4722 sp--;
4726 if (include_entry_exit)
4727 post_order[post_order_num++] = ENTRY_BLOCK;
4729 free (stack);
4730 sbitmap_free (visited);
4731 return post_order_num;
4735 /* Initialize data structures used by PRE. */
4737 static void
4738 init_pre (bool do_fre)
4740 basic_block bb;
4742 next_expression_id = 1;
4743 expressions = NULL;
4744 VEC_safe_push (pre_expr, heap, expressions, NULL);
4745 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4746 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4747 get_max_value_id() + 1);
4748 name_to_id = NULL;
4750 in_fre = do_fre;
4752 inserted_exprs = BITMAP_ALLOC (NULL);
4753 need_creation = NULL;
4754 pretemp = NULL_TREE;
4755 storetemp = NULL_TREE;
4756 prephitemp = NULL_TREE;
4758 connect_infinite_loops_to_exit ();
4759 memset (&pre_stats, 0, sizeof (pre_stats));
4762 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4763 my_rev_post_order_compute (postorder, false);
4765 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4767 calculate_dominance_info (CDI_POST_DOMINATORS);
4768 calculate_dominance_info (CDI_DOMINATORS);
4770 bitmap_obstack_initialize (&grand_bitmap_obstack);
4771 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4772 expr_pred_trans_eq, free);
4773 expression_to_id = htab_create (num_ssa_names * 3,
4774 pre_expr_hash,
4775 pre_expr_eq, NULL);
4776 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4777 sizeof (struct bitmap_set), 30);
4778 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4779 sizeof (struct pre_expr_d), 30);
4780 FOR_ALL_BB (bb)
4782 EXP_GEN (bb) = bitmap_set_new ();
4783 PHI_GEN (bb) = bitmap_set_new ();
4784 TMP_GEN (bb) = bitmap_set_new ();
4785 AVAIL_OUT (bb) = bitmap_set_new ();
4788 need_eh_cleanup = BITMAP_ALLOC (NULL);
4789 need_ab_cleanup = BITMAP_ALLOC (NULL);
4793 /* Deallocate data structures used by PRE. */
4795 static void
4796 fini_pre (bool do_fre)
4798 free (postorder);
4799 VEC_free (bitmap_set_t, heap, value_expressions);
4800 BITMAP_FREE (inserted_exprs);
4801 VEC_free (gimple, heap, need_creation);
4802 bitmap_obstack_release (&grand_bitmap_obstack);
4803 free_alloc_pool (bitmap_set_pool);
4804 free_alloc_pool (pre_expr_pool);
4805 htab_delete (phi_translate_table);
4806 htab_delete (expression_to_id);
4807 VEC_free (unsigned, heap, name_to_id);
4809 free_aux_for_blocks ();
4811 free_dominance_info (CDI_POST_DOMINATORS);
4813 if (!bitmap_empty_p (need_eh_cleanup))
4815 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4816 cleanup_tree_cfg ();
4819 BITMAP_FREE (need_eh_cleanup);
4821 if (!bitmap_empty_p (need_ab_cleanup))
4823 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4824 cleanup_tree_cfg ();
4827 BITMAP_FREE (need_ab_cleanup);
4829 if (!do_fre)
4830 loop_optimizer_finalize ();
4833 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4834 only wants to do full redundancy elimination. */
4836 static unsigned int
4837 execute_pre (bool do_fre)
4839 unsigned int todo = 0;
4841 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4843 /* This has to happen before SCCVN runs because
4844 loop_optimizer_init may create new phis, etc. */
4845 if (!do_fre)
4846 loop_optimizer_init (LOOPS_NORMAL);
4848 if (!run_scc_vn (do_fre ? VN_WALKREWRITE : VN_WALK))
4850 if (!do_fre)
4851 loop_optimizer_finalize ();
4853 return 0;
4856 init_pre (do_fre);
4857 scev_initialize ();
4859 /* Collect and value number expressions computed in each basic block. */
4860 compute_avail ();
4862 if (dump_file && (dump_flags & TDF_DETAILS))
4864 basic_block bb;
4866 FOR_ALL_BB (bb)
4868 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4869 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4870 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4871 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4875 /* Insert can get quite slow on an incredibly large number of basic
4876 blocks due to some quadratic behavior. Until this behavior is
4877 fixed, don't run it when he have an incredibly large number of
4878 bb's. If we aren't going to run insert, there is no point in
4879 computing ANTIC, either, even though it's plenty fast. */
4880 if (!do_fre && n_basic_blocks < 4000)
4882 compute_antic ();
4883 insert ();
4886 /* Make sure to remove fake edges before committing our inserts.
4887 This makes sure we don't end up with extra critical edges that
4888 we would need to split. */
4889 remove_fake_exit_edges ();
4890 gsi_commit_edge_inserts ();
4892 /* Remove all the redundant expressions. */
4893 todo |= eliminate ();
4895 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4896 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4897 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4898 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4899 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4901 clear_expression_ids ();
4902 if (!do_fre)
4904 remove_dead_inserted_code ();
4905 todo |= TODO_verify_flow;
4908 scev_finalize ();
4909 fini_pre (do_fre);
4911 if (!do_fre)
4912 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4913 case we can merge the block with the remaining predecessor of the block.
4914 It should either:
4915 - call merge_blocks after each tail merge iteration
4916 - call merge_blocks after all tail merge iterations
4917 - mark TODO_cleanup_cfg when necessary
4918 - share the cfg cleanup with fini_pre. */
4919 todo |= tail_merge_optimize (todo);
4920 free_scc_vn ();
4922 return todo;
4925 /* Gate and execute functions for PRE. */
4927 static unsigned int
4928 do_pre (void)
4930 return execute_pre (false);
4933 static bool
4934 gate_pre (void)
4936 return flag_tree_pre != 0;
4939 struct gimple_opt_pass pass_pre =
4942 GIMPLE_PASS,
4943 "pre", /* name */
4944 gate_pre, /* gate */
4945 do_pre, /* execute */
4946 NULL, /* sub */
4947 NULL, /* next */
4948 0, /* static_pass_number */
4949 TV_TREE_PRE, /* tv_id */
4950 PROP_no_crit_edges | PROP_cfg
4951 | PROP_ssa, /* properties_required */
4952 0, /* properties_provided */
4953 0, /* properties_destroyed */
4954 TODO_rebuild_alias, /* todo_flags_start */
4955 TODO_update_ssa_only_virtuals | TODO_ggc_collect
4956 | TODO_verify_ssa /* todo_flags_finish */
4961 /* Gate and execute functions for FRE. */
4963 static unsigned int
4964 execute_fre (void)
4966 return execute_pre (true);
4969 static bool
4970 gate_fre (void)
4972 return flag_tree_fre != 0;
4975 struct gimple_opt_pass pass_fre =
4978 GIMPLE_PASS,
4979 "fre", /* name */
4980 gate_fre, /* gate */
4981 execute_fre, /* execute */
4982 NULL, /* sub */
4983 NULL, /* next */
4984 0, /* static_pass_number */
4985 TV_TREE_FRE, /* tv_id */
4986 PROP_cfg | PROP_ssa, /* properties_required */
4987 0, /* properties_provided */
4988 0, /* properties_destroyed */
4989 0, /* todo_flags_start */
4990 TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */