2011-09-15 Tom de Vries <tom@codesourcery.com>
[official-gcc.git] / gcc / tree-ssa-pre.c
blobbee6ef52c6ced0c927afc7d863390104125d618a
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;
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 oldop0 = operand->op0;
1540 tree oldop1 = operand->op1;
1541 tree oldop2 = operand->op2;
1542 tree op0 = oldop0;
1543 tree op1 = oldop1;
1544 tree op2 = oldop2;
1545 tree type = operand->type;
1546 vn_reference_op_s newop = *operand;
1548 if (op0 && TREE_CODE (op0) == SSA_NAME)
1550 unsigned int op_val_id = VN_INFO (op0)->value_id;
1551 leader = find_leader_in_sets (op_val_id, set1, set2);
1552 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1553 if (opresult && opresult != leader)
1555 tree name = get_representative_for (opresult);
1556 if (!name)
1557 break;
1558 op0 = name;
1560 else if (!opresult)
1561 break;
1563 changed |= op0 != oldop0;
1565 if (op1 && TREE_CODE (op1) == SSA_NAME)
1567 unsigned int op_val_id = VN_INFO (op1)->value_id;
1568 leader = find_leader_in_sets (op_val_id, set1, set2);
1569 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1570 if (opresult && opresult != leader)
1572 tree name = get_representative_for (opresult);
1573 if (!name)
1574 break;
1575 op1 = name;
1577 else if (!opresult)
1578 break;
1580 /* We can't possibly insert these. */
1581 else if (op1 && !is_gimple_min_invariant (op1))
1582 break;
1583 changed |= op1 != oldop1;
1584 if (op2 && TREE_CODE (op2) == SSA_NAME)
1586 unsigned int op_val_id = VN_INFO (op2)->value_id;
1587 leader = find_leader_in_sets (op_val_id, set1, set2);
1588 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1589 if (opresult && opresult != leader)
1591 tree name = get_representative_for (opresult);
1592 if (!name)
1593 break;
1594 op2 = name;
1596 else if (!opresult)
1597 break;
1599 /* We can't possibly insert these. */
1600 else if (op2 && !is_gimple_min_invariant (op2))
1601 break;
1602 changed |= op2 != oldop2;
1604 if (!newoperands)
1605 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1606 /* We may have changed from an SSA_NAME to a constant */
1607 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1608 newop.opcode = TREE_CODE (op0);
1609 newop.type = type;
1610 newop.op0 = op0;
1611 newop.op1 = op1;
1612 newop.op2 = op2;
1613 /* If it transforms a non-constant ARRAY_REF into a constant
1614 one, adjust the constant offset. */
1615 if (newop.opcode == ARRAY_REF
1616 && newop.off == -1
1617 && TREE_CODE (op0) == INTEGER_CST
1618 && TREE_CODE (op1) == INTEGER_CST
1619 && TREE_CODE (op2) == INTEGER_CST)
1621 double_int off = tree_to_double_int (op0);
1622 off = double_int_add (off,
1623 double_int_neg
1624 (tree_to_double_int (op1)));
1625 off = double_int_mul (off, tree_to_double_int (op2));
1626 if (double_int_fits_in_shwi_p (off))
1627 newop.off = off.low;
1629 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1630 /* If it transforms from an SSA_NAME to an address, fold with
1631 a preceding indirect reference. */
1632 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1633 && VEC_index (vn_reference_op_s,
1634 newoperands, j - 1)->opcode == MEM_REF)
1635 vn_reference_fold_indirect (&newoperands, &j);
1637 if (i != VEC_length (vn_reference_op_s, operands))
1639 if (newoperands)
1640 VEC_free (vn_reference_op_s, heap, newoperands);
1641 return NULL;
1644 if (vuse)
1646 newvuse = translate_vuse_through_block (newoperands,
1647 ref->set, ref->type,
1648 vuse, phiblock, pred,
1649 &same_valid);
1650 if (newvuse == NULL_TREE)
1652 VEC_free (vn_reference_op_s, heap, newoperands);
1653 return NULL;
1657 if (changed || newvuse != vuse)
1659 unsigned int new_val_id;
1660 pre_expr constant;
1661 bool converted = false;
1663 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1664 ref->type,
1665 newoperands,
1666 &newref, VN_WALK);
1667 if (result)
1668 VEC_free (vn_reference_op_s, heap, newoperands);
1670 if (result
1671 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1673 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1674 converted = true;
1676 else if (!result && newref
1677 && !useless_type_conversion_p (ref->type, newref->type))
1679 VEC_free (vn_reference_op_s, heap, newoperands);
1680 return NULL;
1683 if (result && is_gimple_min_invariant (result))
1685 gcc_assert (!newoperands);
1686 return get_or_alloc_expr_for_constant (result);
1689 expr = (pre_expr) pool_alloc (pre_expr_pool);
1690 expr->kind = REFERENCE;
1691 expr->id = 0;
1693 if (converted)
1695 vn_nary_op_t nary;
1696 tree nresult;
1698 gcc_assert (CONVERT_EXPR_P (result)
1699 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1701 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1702 TREE_TYPE (result),
1703 &TREE_OPERAND (result, 0),
1704 &nary);
1705 if (nresult && is_gimple_min_invariant (nresult))
1706 return get_or_alloc_expr_for_constant (nresult);
1708 expr->kind = NARY;
1709 if (nary)
1711 PRE_EXPR_NARY (expr) = nary;
1712 constant = fully_constant_expression (expr);
1713 if (constant != expr)
1714 return constant;
1716 new_val_id = nary->value_id;
1717 get_or_alloc_expression_id (expr);
1719 else
1721 new_val_id = get_next_value_id ();
1722 VEC_safe_grow_cleared (bitmap_set_t, heap,
1723 value_expressions,
1724 get_max_value_id() + 1);
1725 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1726 TREE_TYPE (result),
1727 &TREE_OPERAND (result, 0),
1728 NULL_TREE,
1729 new_val_id);
1730 PRE_EXPR_NARY (expr) = nary;
1731 constant = fully_constant_expression (expr);
1732 if (constant != expr)
1733 return constant;
1734 get_or_alloc_expression_id (expr);
1737 else if (newref)
1739 PRE_EXPR_REFERENCE (expr) = newref;
1740 constant = fully_constant_expression (expr);
1741 if (constant != expr)
1742 return constant;
1744 new_val_id = newref->value_id;
1745 get_or_alloc_expression_id (expr);
1747 else
1749 if (changed || !same_valid)
1751 new_val_id = get_next_value_id ();
1752 VEC_safe_grow_cleared (bitmap_set_t, heap,
1753 value_expressions,
1754 get_max_value_id() + 1);
1756 else
1757 new_val_id = ref->value_id;
1758 newref = vn_reference_insert_pieces (newvuse, ref->set,
1759 ref->type,
1760 newoperands,
1761 result, new_val_id);
1762 newoperands = NULL;
1763 PRE_EXPR_REFERENCE (expr) = newref;
1764 constant = fully_constant_expression (expr);
1765 if (constant != expr)
1766 return constant;
1767 get_or_alloc_expression_id (expr);
1769 add_to_value (new_val_id, expr);
1771 VEC_free (vn_reference_op_s, heap, newoperands);
1772 return expr;
1774 break;
1776 case NAME:
1778 gimple phi = NULL;
1779 edge e;
1780 gimple def_stmt;
1781 tree name = PRE_EXPR_NAME (expr);
1783 def_stmt = SSA_NAME_DEF_STMT (name);
1784 if (gimple_code (def_stmt) == GIMPLE_PHI
1785 && gimple_bb (def_stmt) == phiblock)
1786 phi = def_stmt;
1787 else
1788 return expr;
1790 e = find_edge (pred, gimple_bb (phi));
1791 if (e)
1793 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1794 pre_expr newexpr;
1796 if (TREE_CODE (def) == SSA_NAME)
1797 def = VN_INFO (def)->valnum;
1799 /* Handle constant. */
1800 if (is_gimple_min_invariant (def))
1801 return get_or_alloc_expr_for_constant (def);
1803 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1804 return NULL;
1806 newexpr = get_or_alloc_expr_for_name (def);
1807 return newexpr;
1810 return expr;
1812 default:
1813 gcc_unreachable ();
1817 /* Wrapper around phi_translate_1 providing caching functionality. */
1819 static pre_expr
1820 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1821 basic_block pred, basic_block phiblock)
1823 pre_expr phitrans;
1825 if (!expr)
1826 return NULL;
1828 /* Constants contain no values that need translation. */
1829 if (expr->kind == CONSTANT)
1830 return expr;
1832 if (value_id_constant_p (get_expr_value_id (expr)))
1833 return expr;
1835 if (expr->kind != NAME)
1837 phitrans = phi_trans_lookup (expr, pred);
1838 if (phitrans)
1839 return phitrans;
1842 /* Translate. */
1843 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1845 /* Don't add empty translations to the cache. Neither add
1846 translations of NAMEs as those are cheap to translate. */
1847 if (phitrans
1848 && expr->kind != NAME)
1849 phi_trans_add (expr, phitrans, pred);
1851 return phitrans;
1855 /* For each expression in SET, translate the values through phi nodes
1856 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1857 expressions in DEST. */
1859 static void
1860 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1861 basic_block phiblock)
1863 VEC (pre_expr, heap) *exprs;
1864 pre_expr expr;
1865 int i;
1867 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1869 bitmap_set_copy (dest, set);
1870 return;
1873 exprs = sorted_array_from_bitmap_set (set);
1874 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1876 pre_expr translated;
1877 translated = phi_translate (expr, set, NULL, pred, phiblock);
1878 if (!translated)
1879 continue;
1881 /* We might end up with multiple expressions from SET being
1882 translated to the same value. In this case we do not want
1883 to retain the NARY or REFERENCE expression but prefer a NAME
1884 which would be the leader. */
1885 if (translated->kind == NAME)
1886 bitmap_value_replace_in_set (dest, translated);
1887 else
1888 bitmap_value_insert_into_set (dest, translated);
1890 VEC_free (pre_expr, heap, exprs);
1893 /* Find the leader for a value (i.e., the name representing that
1894 value) in a given set, and return it. If STMT is non-NULL it
1895 makes sure the defining statement for the leader dominates it.
1896 Return NULL if no leader is found. */
1898 static pre_expr
1899 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1901 if (value_id_constant_p (val))
1903 unsigned int i;
1904 bitmap_iterator bi;
1905 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1907 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1909 pre_expr expr = expression_for_id (i);
1910 if (expr->kind == CONSTANT)
1911 return expr;
1914 if (bitmap_set_contains_value (set, val))
1916 /* Rather than walk the entire bitmap of expressions, and see
1917 whether any of them has the value we are looking for, we look
1918 at the reverse mapping, which tells us the set of expressions
1919 that have a given value (IE value->expressions with that
1920 value) and see if any of those expressions are in our set.
1921 The number of expressions per value is usually significantly
1922 less than the number of expressions in the set. In fact, for
1923 large testcases, doing it this way is roughly 5-10x faster
1924 than walking the bitmap.
1925 If this is somehow a significant lose for some cases, we can
1926 choose which set to walk based on which set is smaller. */
1927 unsigned int i;
1928 bitmap_iterator bi;
1929 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1931 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1932 &set->expressions, 0, i, bi)
1934 pre_expr val = expression_for_id (i);
1935 /* At the point where stmt is not null, there should always
1936 be an SSA_NAME first in the list of expressions. */
1937 if (stmt)
1939 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1940 if (gimple_code (def_stmt) != GIMPLE_PHI
1941 && gimple_bb (def_stmt) == gimple_bb (stmt)
1942 /* PRE insertions are at the end of the basic-block
1943 and have UID 0. */
1944 && (gimple_uid (def_stmt) == 0
1945 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1946 continue;
1948 return val;
1951 return NULL;
1954 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1955 BLOCK by seeing if it is not killed in the block. Note that we are
1956 only determining whether there is a store that kills it. Because
1957 of the order in which clean iterates over values, we are guaranteed
1958 that altered operands will have caused us to be eliminated from the
1959 ANTIC_IN set already. */
1961 static bool
1962 value_dies_in_block_x (pre_expr expr, basic_block block)
1964 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1965 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1966 gimple def;
1967 gimple_stmt_iterator gsi;
1968 unsigned id = get_expression_id (expr);
1969 bool res = false;
1970 ao_ref ref;
1972 if (!vuse)
1973 return false;
1975 /* Lookup a previously calculated result. */
1976 if (EXPR_DIES (block)
1977 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1978 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1980 /* A memory expression {e, VUSE} dies in the block if there is a
1981 statement that may clobber e. If, starting statement walk from the
1982 top of the basic block, a statement uses VUSE there can be no kill
1983 inbetween that use and the original statement that loaded {e, VUSE},
1984 so we can stop walking. */
1985 ref.base = NULL_TREE;
1986 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1988 tree def_vuse, def_vdef;
1989 def = gsi_stmt (gsi);
1990 def_vuse = gimple_vuse (def);
1991 def_vdef = gimple_vdef (def);
1993 /* Not a memory statement. */
1994 if (!def_vuse)
1995 continue;
1997 /* Not a may-def. */
1998 if (!def_vdef)
2000 /* A load with the same VUSE, we're done. */
2001 if (def_vuse == vuse)
2002 break;
2004 continue;
2007 /* Init ref only if we really need it. */
2008 if (ref.base == NULL_TREE
2009 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
2010 refx->operands))
2012 res = true;
2013 break;
2015 /* If the statement may clobber expr, it dies. */
2016 if (stmt_may_clobber_ref_p_1 (def, &ref))
2018 res = true;
2019 break;
2023 /* Remember the result. */
2024 if (!EXPR_DIES (block))
2025 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2026 bitmap_set_bit (EXPR_DIES (block), id * 2);
2027 if (res)
2028 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2030 return res;
2034 #define union_contains_value(SET1, SET2, VAL) \
2035 (bitmap_set_contains_value ((SET1), (VAL)) \
2036 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
2038 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2040 static bool
2041 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2042 vn_reference_op_t vro)
2044 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2046 struct pre_expr_d temp;
2047 temp.kind = NAME;
2048 temp.id = 0;
2049 PRE_EXPR_NAME (&temp) = vro->op0;
2050 temp.id = lookup_expression_id (&temp);
2051 if (temp.id == 0)
2052 return false;
2053 if (!union_contains_value (set1, set2,
2054 get_expr_value_id (&temp)))
2055 return false;
2057 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2059 struct pre_expr_d temp;
2060 temp.kind = NAME;
2061 temp.id = 0;
2062 PRE_EXPR_NAME (&temp) = vro->op1;
2063 temp.id = lookup_expression_id (&temp);
2064 if (temp.id == 0)
2065 return false;
2066 if (!union_contains_value (set1, set2,
2067 get_expr_value_id (&temp)))
2068 return false;
2071 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2073 struct pre_expr_d temp;
2074 temp.kind = NAME;
2075 temp.id = 0;
2076 PRE_EXPR_NAME (&temp) = vro->op2;
2077 temp.id = lookup_expression_id (&temp);
2078 if (temp.id == 0)
2079 return false;
2080 if (!union_contains_value (set1, set2,
2081 get_expr_value_id (&temp)))
2082 return false;
2085 return true;
2088 /* Determine if the expression EXPR is valid in SET1 U SET2.
2089 ONLY SET2 CAN BE NULL.
2090 This means that we have a leader for each part of the expression
2091 (if it consists of values), or the expression is an SSA_NAME.
2092 For loads/calls, we also see if the vuse is killed in this block. */
2094 static bool
2095 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2096 basic_block block)
2098 switch (expr->kind)
2100 case NAME:
2101 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2102 case NARY:
2104 unsigned int i;
2105 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2106 for (i = 0; i < nary->length; i++)
2108 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2110 struct pre_expr_d temp;
2111 temp.kind = NAME;
2112 temp.id = 0;
2113 PRE_EXPR_NAME (&temp) = nary->op[i];
2114 temp.id = lookup_expression_id (&temp);
2115 if (temp.id == 0)
2116 return false;
2117 if (!union_contains_value (set1, set2,
2118 get_expr_value_id (&temp)))
2119 return false;
2122 /* If the NARY may trap make sure the block does not contain
2123 a possible exit point.
2124 ??? This is overly conservative if we translate AVAIL_OUT
2125 as the available expression might be after the exit point. */
2126 if (BB_MAY_NOTRETURN (block)
2127 && vn_nary_may_trap (nary))
2128 return false;
2129 return true;
2131 break;
2132 case REFERENCE:
2134 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2135 vn_reference_op_t vro;
2136 unsigned int i;
2138 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2140 if (!vro_valid_in_sets (set1, set2, vro))
2141 return false;
2143 if (ref->vuse)
2145 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2146 if (!gimple_nop_p (def_stmt)
2147 && gimple_bb (def_stmt) != block
2148 && !dominated_by_p (CDI_DOMINATORS,
2149 block, gimple_bb (def_stmt)))
2150 return false;
2152 return !value_dies_in_block_x (expr, block);
2154 default:
2155 gcc_unreachable ();
2159 /* Clean the set of expressions that are no longer valid in SET1 or
2160 SET2. This means expressions that are made up of values we have no
2161 leaders for in SET1 or SET2. This version is used for partial
2162 anticipation, which means it is not valid in either ANTIC_IN or
2163 PA_IN. */
2165 static void
2166 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2168 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2169 pre_expr expr;
2170 int i;
2172 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2174 if (!valid_in_sets (set1, set2, expr, block))
2175 bitmap_remove_from_set (set1, expr);
2177 VEC_free (pre_expr, heap, exprs);
2180 /* Clean the set of expressions that are no longer valid in SET. This
2181 means expressions that are made up of values we have no leaders for
2182 in SET. */
2184 static void
2185 clean (bitmap_set_t set, basic_block block)
2187 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2188 pre_expr expr;
2189 int i;
2191 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2193 if (!valid_in_sets (set, NULL, expr, block))
2194 bitmap_remove_from_set (set, expr);
2196 VEC_free (pre_expr, heap, exprs);
2199 static sbitmap has_abnormal_preds;
2201 /* List of blocks that may have changed during ANTIC computation and
2202 thus need to be iterated over. */
2204 static sbitmap changed_blocks;
2206 /* Decide whether to defer a block for a later iteration, or PHI
2207 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2208 should defer the block, and true if we processed it. */
2210 static bool
2211 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2212 basic_block block, basic_block phiblock)
2214 if (!BB_VISITED (phiblock))
2216 SET_BIT (changed_blocks, block->index);
2217 BB_VISITED (block) = 0;
2218 BB_DEFERRED (block) = 1;
2219 return false;
2221 else
2222 phi_translate_set (dest, source, block, phiblock);
2223 return true;
2226 /* Compute the ANTIC set for BLOCK.
2228 If succs(BLOCK) > 1 then
2229 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2230 else if succs(BLOCK) == 1 then
2231 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2233 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2236 static bool
2237 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2239 bool changed = false;
2240 bitmap_set_t S, old, ANTIC_OUT;
2241 bitmap_iterator bi;
2242 unsigned int bii;
2243 edge e;
2244 edge_iterator ei;
2246 old = ANTIC_OUT = S = NULL;
2247 BB_VISITED (block) = 1;
2249 /* If any edges from predecessors are abnormal, antic_in is empty,
2250 so do nothing. */
2251 if (block_has_abnormal_pred_edge)
2252 goto maybe_dump_sets;
2254 old = ANTIC_IN (block);
2255 ANTIC_OUT = bitmap_set_new ();
2257 /* If the block has no successors, ANTIC_OUT is empty. */
2258 if (EDGE_COUNT (block->succs) == 0)
2260 /* If we have one successor, we could have some phi nodes to
2261 translate through. */
2262 else if (single_succ_p (block))
2264 basic_block succ_bb = single_succ (block);
2266 /* We trade iterations of the dataflow equations for having to
2267 phi translate the maximal set, which is incredibly slow
2268 (since the maximal set often has 300+ members, even when you
2269 have a small number of blocks).
2270 Basically, we defer the computation of ANTIC for this block
2271 until we have processed it's successor, which will inevitably
2272 have a *much* smaller set of values to phi translate once
2273 clean has been run on it.
2274 The cost of doing this is that we technically perform more
2275 iterations, however, they are lower cost iterations.
2277 Timings for PRE on tramp3d-v4:
2278 without maximal set fix: 11 seconds
2279 with maximal set fix/without deferring: 26 seconds
2280 with maximal set fix/with deferring: 11 seconds
2283 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2284 block, succ_bb))
2286 changed = true;
2287 goto maybe_dump_sets;
2290 /* If we have multiple successors, we take the intersection of all of
2291 them. Note that in the case of loop exit phi nodes, we may have
2292 phis to translate through. */
2293 else
2295 VEC(basic_block, heap) * worklist;
2296 size_t i;
2297 basic_block bprime, first = NULL;
2299 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2300 FOR_EACH_EDGE (e, ei, block->succs)
2302 if (!first
2303 && BB_VISITED (e->dest))
2304 first = e->dest;
2305 else if (BB_VISITED (e->dest))
2306 VEC_quick_push (basic_block, worklist, e->dest);
2309 /* Of multiple successors we have to have visited one already. */
2310 if (!first)
2312 SET_BIT (changed_blocks, block->index);
2313 BB_VISITED (block) = 0;
2314 BB_DEFERRED (block) = 1;
2315 changed = true;
2316 VEC_free (basic_block, heap, worklist);
2317 goto maybe_dump_sets;
2320 if (!gimple_seq_empty_p (phi_nodes (first)))
2321 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2322 else
2323 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2325 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2327 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2329 bitmap_set_t tmp = bitmap_set_new ();
2330 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2331 bitmap_set_and (ANTIC_OUT, tmp);
2332 bitmap_set_free (tmp);
2334 else
2335 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2337 VEC_free (basic_block, heap, worklist);
2340 /* Generate ANTIC_OUT - TMP_GEN. */
2341 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2343 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2344 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2345 TMP_GEN (block));
2347 /* Then union in the ANTIC_OUT - TMP_GEN values,
2348 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2349 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2350 bitmap_value_insert_into_set (ANTIC_IN (block),
2351 expression_for_id (bii));
2353 clean (ANTIC_IN (block), block);
2355 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2357 changed = true;
2358 SET_BIT (changed_blocks, block->index);
2359 FOR_EACH_EDGE (e, ei, block->preds)
2360 SET_BIT (changed_blocks, e->src->index);
2362 else
2363 RESET_BIT (changed_blocks, block->index);
2365 maybe_dump_sets:
2366 if (dump_file && (dump_flags & TDF_DETAILS))
2368 if (!BB_DEFERRED (block) || BB_VISITED (block))
2370 if (ANTIC_OUT)
2371 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2373 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2374 block->index);
2376 if (S)
2377 print_bitmap_set (dump_file, S, "S", block->index);
2379 else
2381 fprintf (dump_file,
2382 "Block %d was deferred for a future iteration.\n",
2383 block->index);
2386 if (old)
2387 bitmap_set_free (old);
2388 if (S)
2389 bitmap_set_free (S);
2390 if (ANTIC_OUT)
2391 bitmap_set_free (ANTIC_OUT);
2392 return changed;
2395 /* Compute PARTIAL_ANTIC for BLOCK.
2397 If succs(BLOCK) > 1 then
2398 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2399 in ANTIC_OUT for all succ(BLOCK)
2400 else if succs(BLOCK) == 1 then
2401 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2403 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2404 - ANTIC_IN[BLOCK])
2407 static bool
2408 compute_partial_antic_aux (basic_block block,
2409 bool block_has_abnormal_pred_edge)
2411 bool changed = false;
2412 bitmap_set_t old_PA_IN;
2413 bitmap_set_t PA_OUT;
2414 edge e;
2415 edge_iterator ei;
2416 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2418 old_PA_IN = PA_OUT = NULL;
2420 /* If any edges from predecessors are abnormal, antic_in is empty,
2421 so do nothing. */
2422 if (block_has_abnormal_pred_edge)
2423 goto maybe_dump_sets;
2425 /* If there are too many partially anticipatable values in the
2426 block, phi_translate_set can take an exponential time: stop
2427 before the translation starts. */
2428 if (max_pa
2429 && single_succ_p (block)
2430 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2431 goto maybe_dump_sets;
2433 old_PA_IN = PA_IN (block);
2434 PA_OUT = bitmap_set_new ();
2436 /* If the block has no successors, ANTIC_OUT is empty. */
2437 if (EDGE_COUNT (block->succs) == 0)
2439 /* If we have one successor, we could have some phi nodes to
2440 translate through. Note that we can't phi translate across DFS
2441 back edges in partial antic, because it uses a union operation on
2442 the successors. For recurrences like IV's, we will end up
2443 generating a new value in the set on each go around (i + 3 (VH.1)
2444 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2445 else if (single_succ_p (block))
2447 basic_block succ = single_succ (block);
2448 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2449 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2451 /* If we have multiple successors, we take the union of all of
2452 them. */
2453 else
2455 VEC(basic_block, heap) * worklist;
2456 size_t i;
2457 basic_block bprime;
2459 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2460 FOR_EACH_EDGE (e, ei, block->succs)
2462 if (e->flags & EDGE_DFS_BACK)
2463 continue;
2464 VEC_quick_push (basic_block, worklist, e->dest);
2466 if (VEC_length (basic_block, worklist) > 0)
2468 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2470 unsigned int i;
2471 bitmap_iterator bi;
2473 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2474 bitmap_value_insert_into_set (PA_OUT,
2475 expression_for_id (i));
2476 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2478 bitmap_set_t pa_in = bitmap_set_new ();
2479 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2480 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2481 bitmap_value_insert_into_set (PA_OUT,
2482 expression_for_id (i));
2483 bitmap_set_free (pa_in);
2485 else
2486 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2487 bitmap_value_insert_into_set (PA_OUT,
2488 expression_for_id (i));
2491 VEC_free (basic_block, heap, worklist);
2494 /* PA_IN starts with PA_OUT - TMP_GEN.
2495 Then we subtract things from ANTIC_IN. */
2496 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2498 /* For partial antic, we want to put back in the phi results, since
2499 we will properly avoid making them partially antic over backedges. */
2500 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2501 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2503 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2504 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2506 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2508 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2510 changed = true;
2511 SET_BIT (changed_blocks, block->index);
2512 FOR_EACH_EDGE (e, ei, block->preds)
2513 SET_BIT (changed_blocks, e->src->index);
2515 else
2516 RESET_BIT (changed_blocks, block->index);
2518 maybe_dump_sets:
2519 if (dump_file && (dump_flags & TDF_DETAILS))
2521 if (PA_OUT)
2522 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2524 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2526 if (old_PA_IN)
2527 bitmap_set_free (old_PA_IN);
2528 if (PA_OUT)
2529 bitmap_set_free (PA_OUT);
2530 return changed;
2533 /* Compute ANTIC and partial ANTIC sets. */
2535 static void
2536 compute_antic (void)
2538 bool changed = true;
2539 int num_iterations = 0;
2540 basic_block block;
2541 int i;
2543 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2544 We pre-build the map of blocks with incoming abnormal edges here. */
2545 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2546 sbitmap_zero (has_abnormal_preds);
2548 FOR_EACH_BB (block)
2550 edge_iterator ei;
2551 edge e;
2553 FOR_EACH_EDGE (e, ei, block->preds)
2555 e->flags &= ~EDGE_DFS_BACK;
2556 if (e->flags & EDGE_ABNORMAL)
2558 SET_BIT (has_abnormal_preds, block->index);
2559 break;
2563 BB_VISITED (block) = 0;
2564 BB_DEFERRED (block) = 0;
2566 /* While we are here, give empty ANTIC_IN sets to each block. */
2567 ANTIC_IN (block) = bitmap_set_new ();
2568 PA_IN (block) = bitmap_set_new ();
2571 /* At the exit block we anticipate nothing. */
2572 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2573 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2574 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2576 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2577 sbitmap_ones (changed_blocks);
2578 while (changed)
2580 if (dump_file && (dump_flags & TDF_DETAILS))
2581 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2582 /* ??? We need to clear our PHI translation cache here as the
2583 ANTIC sets shrink and we restrict valid translations to
2584 those having operands with leaders in ANTIC. Same below
2585 for PA ANTIC computation. */
2586 num_iterations++;
2587 changed = false;
2588 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2590 if (TEST_BIT (changed_blocks, postorder[i]))
2592 basic_block block = BASIC_BLOCK (postorder[i]);
2593 changed |= compute_antic_aux (block,
2594 TEST_BIT (has_abnormal_preds,
2595 block->index));
2598 /* Theoretically possible, but *highly* unlikely. */
2599 gcc_checking_assert (num_iterations < 500);
2602 statistics_histogram_event (cfun, "compute_antic iterations",
2603 num_iterations);
2605 if (do_partial_partial)
2607 sbitmap_ones (changed_blocks);
2608 mark_dfs_back_edges ();
2609 num_iterations = 0;
2610 changed = true;
2611 while (changed)
2613 if (dump_file && (dump_flags & TDF_DETAILS))
2614 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2615 num_iterations++;
2616 changed = false;
2617 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2619 if (TEST_BIT (changed_blocks, postorder[i]))
2621 basic_block block = BASIC_BLOCK (postorder[i]);
2622 changed
2623 |= compute_partial_antic_aux (block,
2624 TEST_BIT (has_abnormal_preds,
2625 block->index));
2628 /* Theoretically possible, but *highly* unlikely. */
2629 gcc_checking_assert (num_iterations < 500);
2631 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2632 num_iterations);
2634 sbitmap_free (has_abnormal_preds);
2635 sbitmap_free (changed_blocks);
2638 /* Return true if we can value number the call in STMT. This is true
2639 if we have a pure or constant call to a real function. */
2641 static bool
2642 can_value_number_call (gimple stmt)
2644 if (gimple_call_internal_p (stmt))
2645 return false;
2646 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2647 return true;
2648 return false;
2651 /* Return true if OP is a tree which we can perform PRE on.
2652 This may not match the operations we can value number, but in
2653 a perfect world would. */
2655 static bool
2656 can_PRE_operation (tree op)
2658 return UNARY_CLASS_P (op)
2659 || BINARY_CLASS_P (op)
2660 || COMPARISON_CLASS_P (op)
2661 || TREE_CODE (op) == MEM_REF
2662 || TREE_CODE (op) == COMPONENT_REF
2663 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2664 || TREE_CODE (op) == CALL_EXPR
2665 || TREE_CODE (op) == ARRAY_REF;
2669 /* Inserted expressions are placed onto this worklist, which is used
2670 for performing quick dead code elimination of insertions we made
2671 that didn't turn out to be necessary. */
2672 static bitmap inserted_exprs;
2674 /* Pool allocated fake store expressions are placed onto this
2675 worklist, which, after performing dead code elimination, is walked
2676 to see which expressions need to be put into GC'able memory */
2677 static VEC(gimple, heap) *need_creation;
2679 /* The actual worker for create_component_ref_by_pieces. */
2681 static tree
2682 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2683 unsigned int *operand, gimple_seq *stmts,
2684 gimple domstmt)
2686 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2687 *operand);
2688 tree genop;
2689 ++*operand;
2690 switch (currop->opcode)
2692 case CALL_EXPR:
2694 tree folded, sc = NULL_TREE;
2695 unsigned int nargs = 0;
2696 tree fn, *args;
2697 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2698 fn = currop->op0;
2699 else
2701 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2702 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2703 if (!fn)
2704 return NULL_TREE;
2706 if (currop->op1)
2708 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2709 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2710 if (!sc)
2711 return NULL_TREE;
2713 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2714 ref->operands) - 1);
2715 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2717 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2718 operand, stmts,
2719 domstmt);
2720 if (!args[nargs])
2722 free (args);
2723 return NULL_TREE;
2725 nargs++;
2727 folded = build_call_array (currop->type,
2728 (TREE_CODE (fn) == FUNCTION_DECL
2729 ? build_fold_addr_expr (fn) : fn),
2730 nargs, args);
2731 free (args);
2732 if (sc)
2733 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2734 return folded;
2736 break;
2737 case MEM_REF:
2739 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2740 stmts, domstmt);
2741 tree offset = currop->op0;
2742 if (!baseop)
2743 return NULL_TREE;
2744 if (TREE_CODE (baseop) == ADDR_EXPR
2745 && handled_component_p (TREE_OPERAND (baseop, 0)))
2747 HOST_WIDE_INT off;
2748 tree base;
2749 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2750 &off);
2751 gcc_assert (base);
2752 offset = int_const_binop (PLUS_EXPR, offset,
2753 build_int_cst (TREE_TYPE (offset),
2754 off));
2755 baseop = build_fold_addr_expr (base);
2757 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2759 break;
2760 case TARGET_MEM_REF:
2762 pre_expr op0expr, op1expr;
2763 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2764 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2765 ++*operand);
2766 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2767 stmts, domstmt);
2768 if (!baseop)
2769 return NULL_TREE;
2770 if (currop->op0)
2772 op0expr = get_or_alloc_expr_for (currop->op0);
2773 genop0 = find_or_generate_expression (block, op0expr,
2774 stmts, domstmt);
2775 if (!genop0)
2776 return NULL_TREE;
2778 if (nextop->op0)
2780 op1expr = get_or_alloc_expr_for (nextop->op0);
2781 genop1 = find_or_generate_expression (block, op1expr,
2782 stmts, domstmt);
2783 if (!genop1)
2784 return NULL_TREE;
2786 return build5 (TARGET_MEM_REF, currop->type,
2787 baseop, currop->op2, genop0, currop->op1, genop1);
2789 break;
2790 case ADDR_EXPR:
2791 if (currop->op0)
2793 gcc_assert (is_gimple_min_invariant (currop->op0));
2794 return currop->op0;
2796 /* Fallthrough. */
2797 case REALPART_EXPR:
2798 case IMAGPART_EXPR:
2799 case VIEW_CONVERT_EXPR:
2801 tree folded;
2802 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2803 operand,
2804 stmts, domstmt);
2805 if (!genop0)
2806 return NULL_TREE;
2807 folded = fold_build1 (currop->opcode, currop->type,
2808 genop0);
2809 return folded;
2811 break;
2812 case BIT_FIELD_REF:
2814 tree folded;
2815 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2816 stmts, domstmt);
2817 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2818 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2819 tree genop1;
2820 tree genop2;
2822 if (!genop0)
2823 return NULL_TREE;
2824 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2825 if (!genop1)
2826 return NULL_TREE;
2827 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2828 if (!genop2)
2829 return NULL_TREE;
2830 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2831 genop2);
2832 return folded;
2835 /* For array ref vn_reference_op's, operand 1 of the array ref
2836 is op0 of the reference op and operand 3 of the array ref is
2837 op1. */
2838 case ARRAY_RANGE_REF:
2839 case ARRAY_REF:
2841 tree genop0;
2842 tree genop1 = currop->op0;
2843 pre_expr op1expr;
2844 tree genop2 = currop->op1;
2845 pre_expr op2expr;
2846 tree genop3 = currop->op2;
2847 pre_expr op3expr;
2848 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2849 stmts, domstmt);
2850 if (!genop0)
2851 return NULL_TREE;
2852 op1expr = get_or_alloc_expr_for (genop1);
2853 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2854 if (!genop1)
2855 return NULL_TREE;
2856 if (genop2)
2858 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2859 /* Drop zero minimum index if redundant. */
2860 if (integer_zerop (genop2)
2861 && (!domain_type
2862 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2863 genop2 = NULL_TREE;
2864 else
2866 op2expr = get_or_alloc_expr_for (genop2);
2867 genop2 = find_or_generate_expression (block, op2expr, stmts,
2868 domstmt);
2869 if (!genop2)
2870 return NULL_TREE;
2873 if (genop3)
2875 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2876 /* We can't always put a size in units of the element alignment
2877 here as the element alignment may be not visible. See
2878 PR43783. Simply drop the element size for constant
2879 sizes. */
2880 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2881 genop3 = NULL_TREE;
2882 else
2884 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2885 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2886 op3expr = get_or_alloc_expr_for (genop3);
2887 genop3 = find_or_generate_expression (block, op3expr, stmts,
2888 domstmt);
2889 if (!genop3)
2890 return NULL_TREE;
2893 return build4 (currop->opcode, currop->type, genop0, genop1,
2894 genop2, genop3);
2896 case COMPONENT_REF:
2898 tree op0;
2899 tree op1;
2900 tree genop2 = currop->op1;
2901 pre_expr op2expr;
2902 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2903 stmts, domstmt);
2904 if (!op0)
2905 return NULL_TREE;
2906 /* op1 should be a FIELD_DECL, which are represented by
2907 themselves. */
2908 op1 = currop->op0;
2909 if (genop2)
2911 op2expr = get_or_alloc_expr_for (genop2);
2912 genop2 = find_or_generate_expression (block, op2expr, stmts,
2913 domstmt);
2914 if (!genop2)
2915 return NULL_TREE;
2918 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2919 genop2);
2921 break;
2922 case SSA_NAME:
2924 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2925 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2926 return genop;
2928 case STRING_CST:
2929 case INTEGER_CST:
2930 case COMPLEX_CST:
2931 case VECTOR_CST:
2932 case REAL_CST:
2933 case CONSTRUCTOR:
2934 case VAR_DECL:
2935 case PARM_DECL:
2936 case CONST_DECL:
2937 case RESULT_DECL:
2938 case FUNCTION_DECL:
2939 return currop->op0;
2941 default:
2942 gcc_unreachable ();
2946 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2947 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2948 trying to rename aggregates into ssa form directly, which is a no no.
2950 Thus, this routine doesn't create temporaries, it just builds a
2951 single access expression for the array, calling
2952 find_or_generate_expression to build the innermost pieces.
2954 This function is a subroutine of create_expression_by_pieces, and
2955 should not be called on it's own unless you really know what you
2956 are doing. */
2958 static tree
2959 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2960 gimple_seq *stmts, gimple domstmt)
2962 unsigned int op = 0;
2963 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2966 /* Find a leader for an expression, or generate one using
2967 create_expression_by_pieces if it's ANTIC but
2968 complex.
2969 BLOCK is the basic_block we are looking for leaders in.
2970 EXPR is the expression to find a leader or generate for.
2971 STMTS is the statement list to put the inserted expressions on.
2972 Returns the SSA_NAME of the LHS of the generated expression or the
2973 leader.
2974 DOMSTMT if non-NULL is a statement that should be dominated by
2975 all uses in the generated expression. If DOMSTMT is non-NULL this
2976 routine can fail and return NULL_TREE. Otherwise it will assert
2977 on failure. */
2979 static tree
2980 find_or_generate_expression (basic_block block, pre_expr expr,
2981 gimple_seq *stmts, gimple domstmt)
2983 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2984 get_expr_value_id (expr), domstmt);
2985 tree genop = NULL;
2986 if (leader)
2988 if (leader->kind == NAME)
2989 genop = PRE_EXPR_NAME (leader);
2990 else if (leader->kind == CONSTANT)
2991 genop = PRE_EXPR_CONSTANT (leader);
2994 /* If it's still NULL, it must be a complex expression, so generate
2995 it recursively. Not so if inserting expressions for values generated
2996 by SCCVN. */
2997 if (genop == NULL
2998 && !domstmt)
3000 bitmap_set_t exprset;
3001 unsigned int lookfor = get_expr_value_id (expr);
3002 bool handled = false;
3003 bitmap_iterator bi;
3004 unsigned int i;
3006 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3007 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3009 pre_expr temp = expression_for_id (i);
3010 if (temp->kind != NAME)
3012 handled = true;
3013 genop = create_expression_by_pieces (block, temp, stmts,
3014 domstmt,
3015 get_expr_type (expr));
3016 break;
3019 if (!handled && domstmt)
3020 return NULL_TREE;
3022 gcc_assert (handled);
3024 return genop;
3027 #define NECESSARY GF_PLF_1
3029 /* Create an expression in pieces, so that we can handle very complex
3030 expressions that may be ANTIC, but not necessary GIMPLE.
3031 BLOCK is the basic block the expression will be inserted into,
3032 EXPR is the expression to insert (in value form)
3033 STMTS is a statement list to append the necessary insertions into.
3035 This function will die if we hit some value that shouldn't be
3036 ANTIC but is (IE there is no leader for it, or its components).
3037 This function may also generate expressions that are themselves
3038 partially or fully redundant. Those that are will be either made
3039 fully redundant during the next iteration of insert (for partially
3040 redundant ones), or eliminated by eliminate (for fully redundant
3041 ones).
3043 If DOMSTMT is non-NULL then we make sure that all uses in the
3044 expressions dominate that statement. In this case the function
3045 can return NULL_TREE to signal failure. */
3047 static tree
3048 create_expression_by_pieces (basic_block block, pre_expr expr,
3049 gimple_seq *stmts, gimple domstmt, tree type)
3051 tree temp, name;
3052 tree folded;
3053 gimple_seq forced_stmts = NULL;
3054 unsigned int value_id;
3055 gimple_stmt_iterator gsi;
3056 tree exprtype = type ? type : get_expr_type (expr);
3057 pre_expr nameexpr;
3058 gimple newstmt;
3060 switch (expr->kind)
3062 /* We may hit the NAME/CONSTANT case if we have to convert types
3063 that value numbering saw through. */
3064 case NAME:
3065 folded = PRE_EXPR_NAME (expr);
3066 break;
3067 case CONSTANT:
3068 folded = PRE_EXPR_CONSTANT (expr);
3069 break;
3070 case REFERENCE:
3072 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3073 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3075 break;
3076 case NARY:
3078 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3079 tree genop[4];
3080 unsigned i;
3081 for (i = 0; i < nary->length; ++i)
3083 pre_expr op = get_or_alloc_expr_for (nary->op[i]);
3084 genop[i] = find_or_generate_expression (block, op,
3085 stmts, domstmt);
3086 if (!genop[i])
3087 return NULL_TREE;
3088 /* Ensure genop[1] is a ptrofftype for POINTER_PLUS_EXPR. It
3089 may be a constant with the wrong type. */
3090 if (i == 1
3091 && nary->opcode == POINTER_PLUS_EXPR)
3092 genop[i] = convert_to_ptrofftype (genop[i]);
3093 else
3094 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]);
3096 if (nary->opcode == CONSTRUCTOR)
3098 VEC(constructor_elt,gc) *elts = NULL;
3099 for (i = 0; i < nary->length; ++i)
3100 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
3101 folded = build_constructor (nary->type, elts);
3103 else
3105 switch (nary->length)
3107 case 1:
3108 folded = fold_build1 (nary->opcode, nary->type,
3109 genop[0]);
3110 break;
3111 case 2:
3112 folded = fold_build2 (nary->opcode, nary->type,
3113 genop[0], genop[1]);
3114 break;
3115 case 3:
3116 folded = fold_build3 (nary->opcode, nary->type,
3117 genop[0], genop[1], genop[3]);
3118 break;
3119 default:
3120 gcc_unreachable ();
3124 break;
3125 default:
3126 return NULL_TREE;
3129 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3130 folded = fold_convert (exprtype, folded);
3132 /* Force the generated expression to be a sequence of GIMPLE
3133 statements.
3134 We have to call unshare_expr because force_gimple_operand may
3135 modify the tree we pass to it. */
3136 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3137 false, NULL);
3139 /* If we have any intermediate expressions to the value sets, add them
3140 to the value sets and chain them in the instruction stream. */
3141 if (forced_stmts)
3143 gsi = gsi_start (forced_stmts);
3144 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3146 gimple stmt = gsi_stmt (gsi);
3147 tree forcedname = gimple_get_lhs (stmt);
3148 pre_expr nameexpr;
3150 if (TREE_CODE (forcedname) == SSA_NAME)
3152 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3153 VN_INFO_GET (forcedname)->valnum = forcedname;
3154 VN_INFO (forcedname)->value_id = get_next_value_id ();
3155 nameexpr = get_or_alloc_expr_for_name (forcedname);
3156 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3157 if (!in_fre)
3158 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3159 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3161 mark_symbols_for_renaming (stmt);
3163 gimple_seq_add_seq (stmts, forced_stmts);
3166 /* Build and insert the assignment of the end result to the temporary
3167 that we will return. */
3168 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3169 pretemp = create_tmp_reg (exprtype, "pretmp");
3171 temp = pretemp;
3172 add_referenced_var (temp);
3174 newstmt = gimple_build_assign (temp, folded);
3175 name = make_ssa_name (temp, newstmt);
3176 gimple_assign_set_lhs (newstmt, name);
3177 gimple_set_plf (newstmt, NECESSARY, false);
3179 gimple_seq_add_stmt (stmts, newstmt);
3180 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3182 /* All the symbols in NEWEXPR should be put into SSA form. */
3183 mark_symbols_for_renaming (newstmt);
3185 /* Fold the last statement. */
3186 gsi = gsi_last (*stmts);
3187 fold_stmt_inplace (&gsi);
3189 /* Add a value number to the temporary.
3190 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3191 we are creating the expression by pieces, and this particular piece of
3192 the expression may have been represented. There is no harm in replacing
3193 here. */
3194 VN_INFO_GET (name)->valnum = name;
3195 value_id = get_expr_value_id (expr);
3196 VN_INFO (name)->value_id = value_id;
3197 nameexpr = get_or_alloc_expr_for_name (name);
3198 add_to_value (value_id, nameexpr);
3199 if (NEW_SETS (block))
3200 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3201 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3203 pre_stats.insertions++;
3204 if (dump_file && (dump_flags & TDF_DETAILS))
3206 fprintf (dump_file, "Inserted ");
3207 print_gimple_stmt (dump_file, newstmt, 0, 0);
3208 fprintf (dump_file, " in predecessor %d\n", block->index);
3211 return name;
3215 /* Returns true if we want to inhibit the insertions of PHI nodes
3216 for the given EXPR for basic block BB (a member of a loop).
3217 We want to do this, when we fear that the induction variable we
3218 create might inhibit vectorization. */
3220 static bool
3221 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3223 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3224 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3225 vn_reference_op_t op;
3226 unsigned i;
3228 /* If we aren't going to vectorize we don't inhibit anything. */
3229 if (!flag_tree_vectorize)
3230 return false;
3232 /* Otherwise we inhibit the insertion when the address of the
3233 memory reference is a simple induction variable. In other
3234 cases the vectorizer won't do anything anyway (either it's
3235 loop invariant or a complicated expression). */
3236 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3238 switch (op->opcode)
3240 case ARRAY_REF:
3241 case ARRAY_RANGE_REF:
3242 if (TREE_CODE (op->op0) != SSA_NAME)
3243 break;
3244 /* Fallthru. */
3245 case SSA_NAME:
3247 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3248 affine_iv iv;
3249 /* Default defs are loop invariant. */
3250 if (!defbb)
3251 break;
3252 /* Defined outside this loop, also loop invariant. */
3253 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3254 break;
3255 /* If it's a simple induction variable inhibit insertion,
3256 the vectorizer might be interested in this one. */
3257 if (simple_iv (bb->loop_father, bb->loop_father,
3258 op->op0, &iv, true))
3259 return true;
3260 /* No simple IV, vectorizer can't do anything, hence no
3261 reason to inhibit the transformation for this operand. */
3262 break;
3264 default:
3265 break;
3268 return false;
3271 /* Insert the to-be-made-available values of expression EXPRNUM for each
3272 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3273 merge the result with a phi node, given the same value number as
3274 NODE. Return true if we have inserted new stuff. */
3276 static bool
3277 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3278 pre_expr *avail)
3280 pre_expr expr = expression_for_id (exprnum);
3281 pre_expr newphi;
3282 unsigned int val = get_expr_value_id (expr);
3283 edge pred;
3284 bool insertions = false;
3285 bool nophi = false;
3286 basic_block bprime;
3287 pre_expr eprime;
3288 edge_iterator ei;
3289 tree type = get_expr_type (expr);
3290 tree temp;
3291 gimple phi;
3293 if (dump_file && (dump_flags & TDF_DETAILS))
3295 fprintf (dump_file, "Found partial redundancy for expression ");
3296 print_pre_expr (dump_file, expr);
3297 fprintf (dump_file, " (%04d)\n", val);
3300 /* Make sure we aren't creating an induction variable. */
3301 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3303 bool firstinsideloop = false;
3304 bool secondinsideloop = false;
3305 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3306 EDGE_PRED (block, 0)->src);
3307 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3308 EDGE_PRED (block, 1)->src);
3309 /* Induction variables only have one edge inside the loop. */
3310 if ((firstinsideloop ^ secondinsideloop)
3311 && (expr->kind != REFERENCE
3312 || inhibit_phi_insertion (block, expr)))
3314 if (dump_file && (dump_flags & TDF_DETAILS))
3315 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3316 nophi = true;
3320 /* Make the necessary insertions. */
3321 FOR_EACH_EDGE (pred, ei, block->preds)
3323 gimple_seq stmts = NULL;
3324 tree builtexpr;
3325 bprime = pred->src;
3326 eprime = avail[bprime->index];
3328 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3330 builtexpr = create_expression_by_pieces (bprime,
3331 eprime,
3332 &stmts, NULL,
3333 type);
3334 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3335 gsi_insert_seq_on_edge (pred, stmts);
3336 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3337 insertions = true;
3339 else if (eprime->kind == CONSTANT)
3341 /* Constants may not have the right type, fold_convert
3342 should give us back a constant with the right type.
3344 tree constant = PRE_EXPR_CONSTANT (eprime);
3345 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3347 tree builtexpr = fold_convert (type, constant);
3348 if (!is_gimple_min_invariant (builtexpr))
3350 tree forcedexpr = force_gimple_operand (builtexpr,
3351 &stmts, true,
3352 NULL);
3353 if (!is_gimple_min_invariant (forcedexpr))
3355 if (forcedexpr != builtexpr)
3357 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3358 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3360 if (stmts)
3362 gimple_stmt_iterator gsi;
3363 gsi = gsi_start (stmts);
3364 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3366 gimple stmt = gsi_stmt (gsi);
3367 tree lhs = gimple_get_lhs (stmt);
3368 if (TREE_CODE (lhs) == SSA_NAME)
3369 bitmap_set_bit (inserted_exprs,
3370 SSA_NAME_VERSION (lhs));
3371 gimple_set_plf (stmt, NECESSARY, false);
3373 gsi_insert_seq_on_edge (pred, stmts);
3375 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3378 else
3379 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3382 else if (eprime->kind == NAME)
3384 /* We may have to do a conversion because our value
3385 numbering can look through types in certain cases, but
3386 our IL requires all operands of a phi node have the same
3387 type. */
3388 tree name = PRE_EXPR_NAME (eprime);
3389 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3391 tree builtexpr;
3392 tree forcedexpr;
3393 builtexpr = fold_convert (type, name);
3394 forcedexpr = force_gimple_operand (builtexpr,
3395 &stmts, true,
3396 NULL);
3398 if (forcedexpr != name)
3400 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3401 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3404 if (stmts)
3406 gimple_stmt_iterator gsi;
3407 gsi = gsi_start (stmts);
3408 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3410 gimple stmt = gsi_stmt (gsi);
3411 tree lhs = gimple_get_lhs (stmt);
3412 if (TREE_CODE (lhs) == SSA_NAME)
3413 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3414 gimple_set_plf (stmt, NECESSARY, false);
3416 gsi_insert_seq_on_edge (pred, stmts);
3418 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3422 /* If we didn't want a phi node, and we made insertions, we still have
3423 inserted new stuff, and thus return true. If we didn't want a phi node,
3424 and didn't make insertions, we haven't added anything new, so return
3425 false. */
3426 if (nophi && insertions)
3427 return true;
3428 else if (nophi && !insertions)
3429 return false;
3431 /* Now build a phi for the new variable. */
3432 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3433 prephitemp = create_tmp_var (type, "prephitmp");
3435 temp = prephitemp;
3436 add_referenced_var (temp);
3438 if (TREE_CODE (type) == COMPLEX_TYPE
3439 || TREE_CODE (type) == VECTOR_TYPE)
3440 DECL_GIMPLE_REG_P (temp) = 1;
3441 phi = create_phi_node (temp, block);
3443 gimple_set_plf (phi, NECESSARY, false);
3444 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3445 VN_INFO (gimple_phi_result (phi))->value_id = val;
3446 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3447 FOR_EACH_EDGE (pred, ei, block->preds)
3449 pre_expr ae = avail[pred->src->index];
3450 gcc_assert (get_expr_type (ae) == type
3451 || useless_type_conversion_p (type, get_expr_type (ae)));
3452 if (ae->kind == CONSTANT)
3453 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3454 else
3455 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3456 UNKNOWN_LOCATION);
3459 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3460 add_to_value (val, newphi);
3462 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3463 this insertion, since we test for the existence of this value in PHI_GEN
3464 before proceeding with the partial redundancy checks in insert_aux.
3466 The value may exist in AVAIL_OUT, in particular, it could be represented
3467 by the expression we are trying to eliminate, in which case we want the
3468 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3469 inserted there.
3471 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3472 this block, because if it did, it would have existed in our dominator's
3473 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3476 bitmap_insert_into_set (PHI_GEN (block), newphi);
3477 bitmap_value_replace_in_set (AVAIL_OUT (block),
3478 newphi);
3479 bitmap_insert_into_set (NEW_SETS (block),
3480 newphi);
3482 if (dump_file && (dump_flags & TDF_DETAILS))
3484 fprintf (dump_file, "Created phi ");
3485 print_gimple_stmt (dump_file, phi, 0, 0);
3486 fprintf (dump_file, " in block %d\n", block->index);
3488 pre_stats.phis++;
3489 return true;
3494 /* Perform insertion of partially redundant values.
3495 For BLOCK, do the following:
3496 1. Propagate the NEW_SETS of the dominator into the current block.
3497 If the block has multiple predecessors,
3498 2a. Iterate over the ANTIC expressions for the block to see if
3499 any of them are partially redundant.
3500 2b. If so, insert them into the necessary predecessors to make
3501 the expression fully redundant.
3502 2c. Insert a new PHI merging the values of the predecessors.
3503 2d. Insert the new PHI, and the new expressions, into the
3504 NEW_SETS set.
3505 3. Recursively call ourselves on the dominator children of BLOCK.
3507 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3508 do_regular_insertion and do_partial_insertion.
3512 static bool
3513 do_regular_insertion (basic_block block, basic_block dom)
3515 bool new_stuff = false;
3516 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3517 pre_expr expr;
3518 int i;
3520 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3522 if (expr->kind != NAME)
3524 pre_expr *avail;
3525 unsigned int val;
3526 bool by_some = false;
3527 bool cant_insert = false;
3528 bool all_same = true;
3529 pre_expr first_s = NULL;
3530 edge pred;
3531 basic_block bprime;
3532 pre_expr eprime = NULL;
3533 edge_iterator ei;
3534 pre_expr edoubleprime = NULL;
3535 bool do_insertion = false;
3537 val = get_expr_value_id (expr);
3538 if (bitmap_set_contains_value (PHI_GEN (block), val))
3539 continue;
3540 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3542 if (dump_file && (dump_flags & TDF_DETAILS))
3543 fprintf (dump_file, "Found fully redundant value\n");
3544 continue;
3547 avail = XCNEWVEC (pre_expr, last_basic_block);
3548 FOR_EACH_EDGE (pred, ei, block->preds)
3550 unsigned int vprime;
3552 /* We should never run insertion for the exit block
3553 and so not come across fake pred edges. */
3554 gcc_assert (!(pred->flags & EDGE_FAKE));
3555 bprime = pred->src;
3556 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3557 bprime, block);
3559 /* eprime will generally only be NULL if the
3560 value of the expression, translated
3561 through the PHI for this predecessor, is
3562 undefined. If that is the case, we can't
3563 make the expression fully redundant,
3564 because its value is undefined along a
3565 predecessor path. We can thus break out
3566 early because it doesn't matter what the
3567 rest of the results are. */
3568 if (eprime == NULL)
3570 cant_insert = true;
3571 break;
3574 eprime = fully_constant_expression (eprime);
3575 vprime = get_expr_value_id (eprime);
3576 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3577 vprime, NULL);
3578 if (edoubleprime == NULL)
3580 avail[bprime->index] = eprime;
3581 all_same = false;
3583 else
3585 avail[bprime->index] = edoubleprime;
3586 by_some = true;
3587 /* We want to perform insertions to remove a redundancy on
3588 a path in the CFG we want to optimize for speed. */
3589 if (optimize_edge_for_speed_p (pred))
3590 do_insertion = true;
3591 if (first_s == NULL)
3592 first_s = edoubleprime;
3593 else if (!pre_expr_eq (first_s, edoubleprime))
3594 all_same = false;
3597 /* If we can insert it, it's not the same value
3598 already existing along every predecessor, and
3599 it's defined by some predecessor, it is
3600 partially redundant. */
3601 if (!cant_insert && !all_same && by_some)
3603 if (!do_insertion)
3605 if (dump_file && (dump_flags & TDF_DETAILS))
3607 fprintf (dump_file, "Skipping partial redundancy for "
3608 "expression ");
3609 print_pre_expr (dump_file, expr);
3610 fprintf (dump_file, " (%04d), no redundancy on to be "
3611 "optimized for speed edge\n", val);
3614 else if (dbg_cnt (treepre_insert)
3615 && insert_into_preds_of_block (block,
3616 get_expression_id (expr),
3617 avail))
3618 new_stuff = true;
3620 /* If all edges produce the same value and that value is
3621 an invariant, then the PHI has the same value on all
3622 edges. Note this. */
3623 else if (!cant_insert && all_same && eprime
3624 && (edoubleprime->kind == CONSTANT
3625 || edoubleprime->kind == NAME)
3626 && !value_id_constant_p (val))
3628 unsigned int j;
3629 bitmap_iterator bi;
3630 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3631 value_expressions, val);
3633 unsigned int new_val = get_expr_value_id (edoubleprime);
3634 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3636 pre_expr expr = expression_for_id (j);
3638 if (expr->kind == NAME)
3640 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3641 /* Just reset the value id and valnum so it is
3642 the same as the constant we have discovered. */
3643 if (edoubleprime->kind == CONSTANT)
3645 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3646 pre_stats.constified++;
3648 else
3649 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3650 info->value_id = new_val;
3654 free (avail);
3658 VEC_free (pre_expr, heap, exprs);
3659 return new_stuff;
3663 /* Perform insertion for partially anticipatable expressions. There
3664 is only one case we will perform insertion for these. This case is
3665 if the expression is partially anticipatable, and fully available.
3666 In this case, we know that putting it earlier will enable us to
3667 remove the later computation. */
3670 static bool
3671 do_partial_partial_insertion (basic_block block, basic_block dom)
3673 bool new_stuff = false;
3674 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3675 pre_expr expr;
3676 int i;
3678 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3680 if (expr->kind != NAME)
3682 pre_expr *avail;
3683 unsigned int val;
3684 bool by_all = true;
3685 bool cant_insert = false;
3686 edge pred;
3687 basic_block bprime;
3688 pre_expr eprime = NULL;
3689 edge_iterator ei;
3691 val = get_expr_value_id (expr);
3692 if (bitmap_set_contains_value (PHI_GEN (block), val))
3693 continue;
3694 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3695 continue;
3697 avail = XCNEWVEC (pre_expr, last_basic_block);
3698 FOR_EACH_EDGE (pred, ei, block->preds)
3700 unsigned int vprime;
3701 pre_expr edoubleprime;
3703 /* We should never run insertion for the exit block
3704 and so not come across fake pred edges. */
3705 gcc_assert (!(pred->flags & EDGE_FAKE));
3706 bprime = pred->src;
3707 eprime = phi_translate (expr, ANTIC_IN (block),
3708 PA_IN (block),
3709 bprime, block);
3711 /* eprime will generally only be NULL if the
3712 value of the expression, translated
3713 through the PHI for this predecessor, is
3714 undefined. If that is the case, we can't
3715 make the expression fully redundant,
3716 because its value is undefined along a
3717 predecessor path. We can thus break out
3718 early because it doesn't matter what the
3719 rest of the results are. */
3720 if (eprime == NULL)
3722 cant_insert = true;
3723 break;
3726 eprime = fully_constant_expression (eprime);
3727 vprime = get_expr_value_id (eprime);
3728 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3729 vprime, NULL);
3730 if (edoubleprime == NULL)
3732 by_all = false;
3733 break;
3735 else
3736 avail[bprime->index] = edoubleprime;
3740 /* If we can insert it, it's not the same value
3741 already existing along every predecessor, and
3742 it's defined by some predecessor, it is
3743 partially redundant. */
3744 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3746 pre_stats.pa_insert++;
3747 if (insert_into_preds_of_block (block, get_expression_id (expr),
3748 avail))
3749 new_stuff = true;
3751 free (avail);
3755 VEC_free (pre_expr, heap, exprs);
3756 return new_stuff;
3759 static bool
3760 insert_aux (basic_block block)
3762 basic_block son;
3763 bool new_stuff = false;
3765 if (block)
3767 basic_block dom;
3768 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3769 if (dom)
3771 unsigned i;
3772 bitmap_iterator bi;
3773 bitmap_set_t newset = NEW_SETS (dom);
3774 if (newset)
3776 /* Note that we need to value_replace both NEW_SETS, and
3777 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3778 represented by some non-simple expression here that we want
3779 to replace it with. */
3780 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3782 pre_expr expr = expression_for_id (i);
3783 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3784 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3787 if (!single_pred_p (block))
3789 new_stuff |= do_regular_insertion (block, dom);
3790 if (do_partial_partial)
3791 new_stuff |= do_partial_partial_insertion (block, dom);
3795 for (son = first_dom_son (CDI_DOMINATORS, block);
3796 son;
3797 son = next_dom_son (CDI_DOMINATORS, son))
3799 new_stuff |= insert_aux (son);
3802 return new_stuff;
3805 /* Perform insertion of partially redundant values. */
3807 static void
3808 insert (void)
3810 bool new_stuff = true;
3811 basic_block bb;
3812 int num_iterations = 0;
3814 FOR_ALL_BB (bb)
3815 NEW_SETS (bb) = bitmap_set_new ();
3817 while (new_stuff)
3819 num_iterations++;
3820 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3822 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3826 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3828 static void
3829 add_to_exp_gen (basic_block block, tree op)
3831 if (!in_fre)
3833 pre_expr result;
3834 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3835 return;
3836 result = get_or_alloc_expr_for_name (op);
3837 bitmap_value_insert_into_set (EXP_GEN (block), result);
3841 /* Create value ids for PHI in BLOCK. */
3843 static void
3844 make_values_for_phi (gimple phi, basic_block block)
3846 tree result = gimple_phi_result (phi);
3848 /* We have no need for virtual phis, as they don't represent
3849 actual computations. */
3850 if (is_gimple_reg (result))
3852 pre_expr e = get_or_alloc_expr_for_name (result);
3853 add_to_value (get_expr_value_id (e), e);
3854 bitmap_insert_into_set (PHI_GEN (block), e);
3855 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3856 if (!in_fre)
3858 unsigned i;
3859 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3861 tree arg = gimple_phi_arg_def (phi, i);
3862 if (TREE_CODE (arg) == SSA_NAME)
3864 e = get_or_alloc_expr_for_name (arg);
3865 add_to_value (get_expr_value_id (e), e);
3872 /* Compute the AVAIL set for all basic blocks.
3874 This function performs value numbering of the statements in each basic
3875 block. The AVAIL sets are built from information we glean while doing
3876 this value numbering, since the AVAIL sets contain only one entry per
3877 value.
3879 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3880 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3882 static void
3883 compute_avail (void)
3886 basic_block block, son;
3887 basic_block *worklist;
3888 size_t sp = 0;
3889 unsigned i;
3891 /* We pretend that default definitions are defined in the entry block.
3892 This includes function arguments and the static chain decl. */
3893 for (i = 1; i < num_ssa_names; ++i)
3895 tree name = ssa_name (i);
3896 pre_expr e;
3897 if (!name
3898 || !SSA_NAME_IS_DEFAULT_DEF (name)
3899 || has_zero_uses (name)
3900 || !is_gimple_reg (name))
3901 continue;
3903 e = get_or_alloc_expr_for_name (name);
3904 add_to_value (get_expr_value_id (e), e);
3905 if (!in_fre)
3906 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3907 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3910 /* Allocate the worklist. */
3911 worklist = XNEWVEC (basic_block, n_basic_blocks);
3913 /* Seed the algorithm by putting the dominator children of the entry
3914 block on the worklist. */
3915 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3916 son;
3917 son = next_dom_son (CDI_DOMINATORS, son))
3918 worklist[sp++] = son;
3920 /* Loop until the worklist is empty. */
3921 while (sp)
3923 gimple_stmt_iterator gsi;
3924 gimple stmt;
3925 basic_block dom;
3926 unsigned int stmt_uid = 1;
3928 /* Pick a block from the worklist. */
3929 block = worklist[--sp];
3931 /* Initially, the set of available values in BLOCK is that of
3932 its immediate dominator. */
3933 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3934 if (dom)
3935 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3937 /* Generate values for PHI nodes. */
3938 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3939 make_values_for_phi (gsi_stmt (gsi), block);
3941 BB_MAY_NOTRETURN (block) = 0;
3943 /* Now compute value numbers and populate value sets with all
3944 the expressions computed in BLOCK. */
3945 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3947 ssa_op_iter iter;
3948 tree op;
3950 stmt = gsi_stmt (gsi);
3951 gimple_set_uid (stmt, stmt_uid++);
3953 /* Cache whether the basic-block has any non-visible side-effect
3954 or control flow.
3955 If this isn't a call or it is the last stmt in the
3956 basic-block then the CFG represents things correctly. */
3957 if (is_gimple_call (stmt)
3958 && !stmt_ends_bb_p (stmt))
3960 /* Non-looping const functions always return normally.
3961 Otherwise the call might not return or have side-effects
3962 that forbids hoisting possibly trapping expressions
3963 before it. */
3964 int flags = gimple_call_flags (stmt);
3965 if (!(flags & ECF_CONST)
3966 || (flags & ECF_LOOPING_CONST_OR_PURE))
3967 BB_MAY_NOTRETURN (block) = 1;
3970 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3972 pre_expr e = get_or_alloc_expr_for_name (op);
3974 add_to_value (get_expr_value_id (e), e);
3975 if (!in_fre)
3976 bitmap_insert_into_set (TMP_GEN (block), e);
3977 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3980 if (gimple_has_volatile_ops (stmt)
3981 || stmt_could_throw_p (stmt))
3982 continue;
3984 switch (gimple_code (stmt))
3986 case GIMPLE_RETURN:
3987 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3988 add_to_exp_gen (block, op);
3989 continue;
3991 case GIMPLE_CALL:
3993 vn_reference_t ref;
3994 unsigned int i;
3995 vn_reference_op_t vro;
3996 pre_expr result = NULL;
3997 VEC(vn_reference_op_s, heap) *ops = NULL;
3999 if (!can_value_number_call (stmt))
4000 continue;
4002 copy_reference_ops_from_call (stmt, &ops);
4003 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4004 gimple_expr_type (stmt),
4005 ops, &ref, VN_NOWALK);
4006 VEC_free (vn_reference_op_s, heap, ops);
4007 if (!ref)
4008 continue;
4010 for (i = 0; VEC_iterate (vn_reference_op_s,
4011 ref->operands, i,
4012 vro); i++)
4014 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4015 add_to_exp_gen (block, vro->op0);
4016 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4017 add_to_exp_gen (block, vro->op1);
4018 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4019 add_to_exp_gen (block, vro->op2);
4021 result = (pre_expr) pool_alloc (pre_expr_pool);
4022 result->kind = REFERENCE;
4023 result->id = 0;
4024 PRE_EXPR_REFERENCE (result) = ref;
4026 get_or_alloc_expression_id (result);
4027 add_to_value (get_expr_value_id (result), result);
4028 if (!in_fre)
4029 bitmap_value_insert_into_set (EXP_GEN (block), result);
4030 continue;
4033 case GIMPLE_ASSIGN:
4035 pre_expr result = NULL;
4036 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4038 case tcc_unary:
4039 case tcc_binary:
4040 case tcc_comparison:
4042 vn_nary_op_t nary;
4043 unsigned int i;
4045 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4046 gimple_assign_rhs_code (stmt),
4047 gimple_expr_type (stmt),
4048 gimple_assign_rhs1_ptr (stmt),
4049 &nary);
4051 if (!nary)
4052 continue;
4054 for (i = 0; i < nary->length; i++)
4055 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4056 add_to_exp_gen (block, nary->op[i]);
4058 result = (pre_expr) pool_alloc (pre_expr_pool);
4059 result->kind = NARY;
4060 result->id = 0;
4061 PRE_EXPR_NARY (result) = nary;
4062 break;
4065 case tcc_declaration:
4066 case tcc_reference:
4068 vn_reference_t ref;
4069 unsigned int i;
4070 vn_reference_op_t vro;
4072 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4073 gimple_vuse (stmt),
4074 VN_WALK, &ref);
4075 if (!ref)
4076 continue;
4078 for (i = 0; VEC_iterate (vn_reference_op_s,
4079 ref->operands, i,
4080 vro); i++)
4082 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4083 add_to_exp_gen (block, vro->op0);
4084 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4085 add_to_exp_gen (block, vro->op1);
4086 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4087 add_to_exp_gen (block, vro->op2);
4089 result = (pre_expr) pool_alloc (pre_expr_pool);
4090 result->kind = REFERENCE;
4091 result->id = 0;
4092 PRE_EXPR_REFERENCE (result) = ref;
4093 break;
4096 default:
4097 /* For any other statement that we don't
4098 recognize, simply add all referenced
4099 SSA_NAMEs to EXP_GEN. */
4100 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4101 add_to_exp_gen (block, op);
4102 continue;
4105 get_or_alloc_expression_id (result);
4106 add_to_value (get_expr_value_id (result), result);
4107 if (!in_fre)
4108 bitmap_value_insert_into_set (EXP_GEN (block), result);
4110 continue;
4112 default:
4113 break;
4117 /* Put the dominator children of BLOCK on the worklist of blocks
4118 to compute available sets for. */
4119 for (son = first_dom_son (CDI_DOMINATORS, block);
4120 son;
4121 son = next_dom_son (CDI_DOMINATORS, son))
4122 worklist[sp++] = son;
4125 free (worklist);
4128 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4129 than the available expressions for it. The insertion point is
4130 right before the first use in STMT. Returns the SSA_NAME that should
4131 be used for replacement. */
4133 static tree
4134 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4136 basic_block bb = gimple_bb (stmt);
4137 gimple_stmt_iterator gsi;
4138 gimple_seq stmts = NULL;
4139 tree expr;
4140 pre_expr e;
4142 /* First create a value expression from the expression we want
4143 to insert and associate it with the value handle for SSA_VN. */
4144 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4145 if (e == NULL)
4146 return NULL_TREE;
4148 /* Then use create_expression_by_pieces to generate a valid
4149 expression to insert at this point of the IL stream. */
4150 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4151 if (expr == NULL_TREE)
4152 return NULL_TREE;
4153 gsi = gsi_for_stmt (stmt);
4154 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4156 return expr;
4159 /* Eliminate fully redundant computations. */
4161 static unsigned int
4162 eliminate (void)
4164 VEC (gimple, heap) *to_remove = NULL;
4165 VEC (gimple, heap) *to_update = NULL;
4166 basic_block b;
4167 unsigned int todo = 0;
4168 gimple_stmt_iterator gsi;
4169 gimple stmt;
4170 unsigned i;
4172 FOR_EACH_BB (b)
4174 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4176 stmt = gsi_stmt (gsi);
4178 /* Lookup the RHS of the expression, see if we have an
4179 available computation for it. If so, replace the RHS with
4180 the available computation. */
4181 if (gimple_has_lhs (stmt)
4182 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4183 && !gimple_assign_ssa_name_copy_p (stmt)
4184 && (!gimple_assign_single_p (stmt)
4185 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4186 && !gimple_has_volatile_ops (stmt)
4187 && !has_zero_uses (gimple_get_lhs (stmt)))
4189 tree lhs = gimple_get_lhs (stmt);
4190 tree rhs = NULL_TREE;
4191 tree sprime = NULL;
4192 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4193 pre_expr sprimeexpr;
4195 if (gimple_assign_single_p (stmt))
4196 rhs = gimple_assign_rhs1 (stmt);
4198 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4199 get_expr_value_id (lhsexpr),
4200 NULL);
4202 if (sprimeexpr)
4204 if (sprimeexpr->kind == CONSTANT)
4205 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4206 else if (sprimeexpr->kind == NAME)
4207 sprime = PRE_EXPR_NAME (sprimeexpr);
4208 else
4209 gcc_unreachable ();
4212 /* If there is no existing leader but SCCVN knows this
4213 value is constant, use that constant. */
4214 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4216 sprime = VN_INFO (lhs)->valnum;
4217 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4218 TREE_TYPE (sprime)))
4219 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4221 if (dump_file && (dump_flags & TDF_DETAILS))
4223 fprintf (dump_file, "Replaced ");
4224 print_gimple_expr (dump_file, stmt, 0, 0);
4225 fprintf (dump_file, " with ");
4226 print_generic_expr (dump_file, sprime, 0);
4227 fprintf (dump_file, " in ");
4228 print_gimple_stmt (dump_file, stmt, 0, 0);
4230 pre_stats.eliminations++;
4231 propagate_tree_value_into_stmt (&gsi, sprime);
4232 stmt = gsi_stmt (gsi);
4233 update_stmt (stmt);
4234 continue;
4237 /* If there is no existing usable leader but SCCVN thinks
4238 it has an expression it wants to use as replacement,
4239 insert that. */
4240 if (!sprime || sprime == lhs)
4242 tree val = VN_INFO (lhs)->valnum;
4243 if (val != VN_TOP
4244 && TREE_CODE (val) == SSA_NAME
4245 && VN_INFO (val)->needs_insertion
4246 && can_PRE_operation (vn_get_expr_for (val)))
4247 sprime = do_SCCVN_insertion (stmt, val);
4249 if (sprime
4250 && sprime != lhs
4251 && (rhs == NULL_TREE
4252 || TREE_CODE (rhs) != SSA_NAME
4253 || may_propagate_copy (rhs, sprime)))
4255 bool can_make_abnormal_goto
4256 = is_gimple_call (stmt)
4257 && stmt_can_make_abnormal_goto (stmt);
4259 gcc_assert (sprime != rhs);
4261 if (dump_file && (dump_flags & TDF_DETAILS))
4263 fprintf (dump_file, "Replaced ");
4264 print_gimple_expr (dump_file, stmt, 0, 0);
4265 fprintf (dump_file, " with ");
4266 print_generic_expr (dump_file, sprime, 0);
4267 fprintf (dump_file, " in ");
4268 print_gimple_stmt (dump_file, stmt, 0, 0);
4271 if (TREE_CODE (sprime) == SSA_NAME)
4272 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4273 NECESSARY, true);
4274 /* We need to make sure the new and old types actually match,
4275 which may require adding a simple cast, which fold_convert
4276 will do for us. */
4277 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4278 && !useless_type_conversion_p (gimple_expr_type (stmt),
4279 TREE_TYPE (sprime)))
4280 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4282 pre_stats.eliminations++;
4283 propagate_tree_value_into_stmt (&gsi, sprime);
4284 stmt = gsi_stmt (gsi);
4285 update_stmt (stmt);
4287 /* If we removed EH side-effects from the statement, clean
4288 its EH information. */
4289 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4291 bitmap_set_bit (need_eh_cleanup,
4292 gimple_bb (stmt)->index);
4293 if (dump_file && (dump_flags & TDF_DETAILS))
4294 fprintf (dump_file, " Removed EH side-effects.\n");
4297 /* Likewise for AB side-effects. */
4298 if (can_make_abnormal_goto
4299 && !stmt_can_make_abnormal_goto (stmt))
4301 bitmap_set_bit (need_ab_cleanup,
4302 gimple_bb (stmt)->index);
4303 if (dump_file && (dump_flags & TDF_DETAILS))
4304 fprintf (dump_file, " Removed AB side-effects.\n");
4308 /* If the statement is a scalar store, see if the expression
4309 has the same value number as its rhs. If so, the store is
4310 dead. */
4311 else if (gimple_assign_single_p (stmt)
4312 && !is_gimple_reg (gimple_assign_lhs (stmt))
4313 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4314 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4316 tree rhs = gimple_assign_rhs1 (stmt);
4317 tree val;
4318 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4319 gimple_vuse (stmt), VN_WALK, NULL);
4320 if (TREE_CODE (rhs) == SSA_NAME)
4321 rhs = VN_INFO (rhs)->valnum;
4322 if (val
4323 && operand_equal_p (val, rhs, 0))
4325 if (dump_file && (dump_flags & TDF_DETAILS))
4327 fprintf (dump_file, "Deleted redundant store ");
4328 print_gimple_stmt (dump_file, stmt, 0, 0);
4331 /* Queue stmt for removal. */
4332 VEC_safe_push (gimple, heap, to_remove, stmt);
4335 /* Visit COND_EXPRs and fold the comparison with the
4336 available value-numbers. */
4337 else if (gimple_code (stmt) == GIMPLE_COND)
4339 tree op0 = gimple_cond_lhs (stmt);
4340 tree op1 = gimple_cond_rhs (stmt);
4341 tree result;
4343 if (TREE_CODE (op0) == SSA_NAME)
4344 op0 = VN_INFO (op0)->valnum;
4345 if (TREE_CODE (op1) == SSA_NAME)
4346 op1 = VN_INFO (op1)->valnum;
4347 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4348 op0, op1);
4349 if (result && TREE_CODE (result) == INTEGER_CST)
4351 if (integer_zerop (result))
4352 gimple_cond_make_false (stmt);
4353 else
4354 gimple_cond_make_true (stmt);
4355 update_stmt (stmt);
4356 todo = TODO_cleanup_cfg;
4359 /* Visit indirect calls and turn them into direct calls if
4360 possible. */
4361 if (is_gimple_call (stmt))
4363 tree orig_fn = gimple_call_fn (stmt);
4364 tree fn;
4365 if (!orig_fn)
4366 continue;
4367 if (TREE_CODE (orig_fn) == SSA_NAME)
4368 fn = VN_INFO (orig_fn)->valnum;
4369 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4370 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4371 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4372 else
4373 continue;
4374 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4375 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4376 TREE_TYPE (fn)))
4378 bool can_make_abnormal_goto
4379 = stmt_can_make_abnormal_goto (stmt);
4380 bool was_noreturn = gimple_call_noreturn_p (stmt);
4382 if (dump_file && (dump_flags & TDF_DETAILS))
4384 fprintf (dump_file, "Replacing call target with ");
4385 print_generic_expr (dump_file, fn, 0);
4386 fprintf (dump_file, " in ");
4387 print_gimple_stmt (dump_file, stmt, 0, 0);
4390 gimple_call_set_fn (stmt, fn);
4391 VEC_safe_push (gimple, heap, to_update, stmt);
4393 /* When changing a call into a noreturn call, cfg cleanup
4394 is needed to fix up the noreturn call. */
4395 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4396 todo |= TODO_cleanup_cfg;
4398 /* If we removed EH side-effects from the statement, clean
4399 its EH information. */
4400 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4402 bitmap_set_bit (need_eh_cleanup,
4403 gimple_bb (stmt)->index);
4404 if (dump_file && (dump_flags & TDF_DETAILS))
4405 fprintf (dump_file, " Removed EH side-effects.\n");
4408 /* Likewise for AB side-effects. */
4409 if (can_make_abnormal_goto
4410 && !stmt_can_make_abnormal_goto (stmt))
4412 bitmap_set_bit (need_ab_cleanup,
4413 gimple_bb (stmt)->index);
4414 if (dump_file && (dump_flags & TDF_DETAILS))
4415 fprintf (dump_file, " Removed AB side-effects.\n");
4418 /* Changing an indirect call to a direct call may
4419 have exposed different semantics. This may
4420 require an SSA update. */
4421 todo |= TODO_update_ssa_only_virtuals;
4426 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4428 gimple stmt, phi = gsi_stmt (gsi);
4429 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4430 pre_expr sprimeexpr, resexpr;
4431 gimple_stmt_iterator gsi2;
4433 /* We want to perform redundant PHI elimination. Do so by
4434 replacing the PHI with a single copy if possible.
4435 Do not touch inserted, single-argument or virtual PHIs. */
4436 if (gimple_phi_num_args (phi) == 1
4437 || !is_gimple_reg (res))
4439 gsi_next (&gsi);
4440 continue;
4443 resexpr = get_or_alloc_expr_for_name (res);
4444 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4445 get_expr_value_id (resexpr), NULL);
4446 if (sprimeexpr)
4448 if (sprimeexpr->kind == CONSTANT)
4449 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4450 else if (sprimeexpr->kind == NAME)
4451 sprime = PRE_EXPR_NAME (sprimeexpr);
4452 else
4453 gcc_unreachable ();
4455 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4457 sprime = VN_INFO (res)->valnum;
4458 if (!useless_type_conversion_p (TREE_TYPE (res),
4459 TREE_TYPE (sprime)))
4460 sprime = fold_convert (TREE_TYPE (res), sprime);
4462 if (!sprime
4463 || sprime == res)
4465 gsi_next (&gsi);
4466 continue;
4469 if (dump_file && (dump_flags & TDF_DETAILS))
4471 fprintf (dump_file, "Replaced redundant PHI node defining ");
4472 print_generic_expr (dump_file, res, 0);
4473 fprintf (dump_file, " with ");
4474 print_generic_expr (dump_file, sprime, 0);
4475 fprintf (dump_file, "\n");
4478 remove_phi_node (&gsi, false);
4480 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4481 && TREE_CODE (sprime) == SSA_NAME)
4482 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4484 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4485 sprime = fold_convert (TREE_TYPE (res), sprime);
4486 stmt = gimple_build_assign (res, sprime);
4487 SSA_NAME_DEF_STMT (res) = stmt;
4488 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4490 gsi2 = gsi_after_labels (b);
4491 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4492 /* Queue the copy for eventual removal. */
4493 VEC_safe_push (gimple, heap, to_remove, stmt);
4494 /* If we inserted this PHI node ourself, it's not an elimination. */
4495 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4496 pre_stats.phis--;
4497 else
4498 pre_stats.eliminations++;
4502 /* We cannot remove stmts during BB walk, especially not release SSA
4503 names there as this confuses the VN machinery. The stmts ending
4504 up in to_remove are either stores or simple copies. */
4505 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4507 tree lhs = gimple_assign_lhs (stmt);
4508 tree rhs = gimple_assign_rhs1 (stmt);
4509 use_operand_p use_p;
4510 gimple use_stmt;
4512 /* If there is a single use only, propagate the equivalency
4513 instead of keeping the copy. */
4514 if (TREE_CODE (lhs) == SSA_NAME
4515 && TREE_CODE (rhs) == SSA_NAME
4516 && single_imm_use (lhs, &use_p, &use_stmt)
4517 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4519 SET_USE (use_p, rhs);
4520 update_stmt (use_stmt);
4521 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4522 && TREE_CODE (rhs) == SSA_NAME)
4523 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4526 /* If this is a store or a now unused copy, remove it. */
4527 if (TREE_CODE (lhs) != SSA_NAME
4528 || has_zero_uses (lhs))
4530 basic_block bb = gimple_bb (stmt);
4531 gsi = gsi_for_stmt (stmt);
4532 unlink_stmt_vdef (stmt);
4533 gsi_remove (&gsi, true);
4534 if (gimple_purge_dead_eh_edges (bb))
4535 todo |= TODO_cleanup_cfg;
4536 if (TREE_CODE (lhs) == SSA_NAME)
4537 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4538 release_defs (stmt);
4541 VEC_free (gimple, heap, to_remove);
4543 /* We cannot update call statements with virtual operands during
4544 SSA walk. This might remove them which in turn makes our
4545 VN lattice invalid. */
4546 FOR_EACH_VEC_ELT (gimple, to_update, i, stmt)
4547 update_stmt (stmt);
4548 VEC_free (gimple, heap, to_update);
4550 return todo;
4553 /* Borrow a bit of tree-ssa-dce.c for the moment.
4554 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4555 this may be a bit faster, and we may want critical edges kept split. */
4557 /* If OP's defining statement has not already been determined to be necessary,
4558 mark that statement necessary. Return the stmt, if it is newly
4559 necessary. */
4561 static inline gimple
4562 mark_operand_necessary (tree op)
4564 gimple stmt;
4566 gcc_assert (op);
4568 if (TREE_CODE (op) != SSA_NAME)
4569 return NULL;
4571 stmt = SSA_NAME_DEF_STMT (op);
4572 gcc_assert (stmt);
4574 if (gimple_plf (stmt, NECESSARY)
4575 || gimple_nop_p (stmt))
4576 return NULL;
4578 gimple_set_plf (stmt, NECESSARY, true);
4579 return stmt;
4582 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4583 to insert PHI nodes sometimes, and because value numbering of casts isn't
4584 perfect, we sometimes end up inserting dead code. This simple DCE-like
4585 pass removes any insertions we made that weren't actually used. */
4587 static void
4588 remove_dead_inserted_code (void)
4590 bitmap worklist;
4591 unsigned i;
4592 bitmap_iterator bi;
4593 gimple t;
4595 worklist = BITMAP_ALLOC (NULL);
4596 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4598 t = SSA_NAME_DEF_STMT (ssa_name (i));
4599 if (gimple_plf (t, NECESSARY))
4600 bitmap_set_bit (worklist, i);
4602 while (!bitmap_empty_p (worklist))
4604 i = bitmap_first_set_bit (worklist);
4605 bitmap_clear_bit (worklist, i);
4606 t = SSA_NAME_DEF_STMT (ssa_name (i));
4608 /* PHI nodes are somewhat special in that each PHI alternative has
4609 data and control dependencies. All the statements feeding the
4610 PHI node's arguments are always necessary. */
4611 if (gimple_code (t) == GIMPLE_PHI)
4613 unsigned k;
4615 for (k = 0; k < gimple_phi_num_args (t); k++)
4617 tree arg = PHI_ARG_DEF (t, k);
4618 if (TREE_CODE (arg) == SSA_NAME)
4620 gimple n = mark_operand_necessary (arg);
4621 if (n)
4622 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4626 else
4628 /* Propagate through the operands. Examine all the USE, VUSE and
4629 VDEF operands in this statement. Mark all the statements
4630 which feed this statement's uses as necessary. */
4631 ssa_op_iter iter;
4632 tree use;
4634 /* The operands of VDEF expressions are also needed as they
4635 represent potential definitions that may reach this
4636 statement (VDEF operands allow us to follow def-def
4637 links). */
4639 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4641 gimple n = mark_operand_necessary (use);
4642 if (n)
4643 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4648 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4650 t = SSA_NAME_DEF_STMT (ssa_name (i));
4651 if (!gimple_plf (t, NECESSARY))
4653 gimple_stmt_iterator gsi;
4655 if (dump_file && (dump_flags & TDF_DETAILS))
4657 fprintf (dump_file, "Removing unnecessary insertion:");
4658 print_gimple_stmt (dump_file, t, 0, 0);
4661 gsi = gsi_for_stmt (t);
4662 if (gimple_code (t) == GIMPLE_PHI)
4663 remove_phi_node (&gsi, true);
4664 else
4666 gsi_remove (&gsi, true);
4667 release_defs (t);
4671 BITMAP_FREE (worklist);
4674 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4675 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4676 the number of visited blocks. */
4678 static int
4679 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4681 edge_iterator *stack;
4682 int sp;
4683 int post_order_num = 0;
4684 sbitmap visited;
4686 if (include_entry_exit)
4687 post_order[post_order_num++] = EXIT_BLOCK;
4689 /* Allocate stack for back-tracking up CFG. */
4690 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4691 sp = 0;
4693 /* Allocate bitmap to track nodes that have been visited. */
4694 visited = sbitmap_alloc (last_basic_block);
4696 /* None of the nodes in the CFG have been visited yet. */
4697 sbitmap_zero (visited);
4699 /* Push the last edge on to the stack. */
4700 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4702 while (sp)
4704 edge_iterator ei;
4705 basic_block src;
4706 basic_block dest;
4708 /* Look at the edge on the top of the stack. */
4709 ei = stack[sp - 1];
4710 src = ei_edge (ei)->src;
4711 dest = ei_edge (ei)->dest;
4713 /* Check if the edge destination has been visited yet. */
4714 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4716 /* Mark that we have visited the destination. */
4717 SET_BIT (visited, src->index);
4719 if (EDGE_COUNT (src->preds) > 0)
4720 /* Since the DEST node has been visited for the first
4721 time, check its successors. */
4722 stack[sp++] = ei_start (src->preds);
4723 else
4724 post_order[post_order_num++] = src->index;
4726 else
4728 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4729 post_order[post_order_num++] = dest->index;
4731 if (!ei_one_before_end_p (ei))
4732 ei_next (&stack[sp - 1]);
4733 else
4734 sp--;
4738 if (include_entry_exit)
4739 post_order[post_order_num++] = ENTRY_BLOCK;
4741 free (stack);
4742 sbitmap_free (visited);
4743 return post_order_num;
4747 /* Initialize data structures used by PRE. */
4749 static void
4750 init_pre (bool do_fre)
4752 basic_block bb;
4754 next_expression_id = 1;
4755 expressions = NULL;
4756 VEC_safe_push (pre_expr, heap, expressions, NULL);
4757 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4758 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4759 get_max_value_id() + 1);
4760 name_to_id = NULL;
4762 in_fre = do_fre;
4764 inserted_exprs = BITMAP_ALLOC (NULL);
4765 need_creation = NULL;
4766 pretemp = NULL_TREE;
4767 storetemp = NULL_TREE;
4768 prephitemp = NULL_TREE;
4770 connect_infinite_loops_to_exit ();
4771 memset (&pre_stats, 0, sizeof (pre_stats));
4774 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4775 my_rev_post_order_compute (postorder, false);
4777 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4779 calculate_dominance_info (CDI_POST_DOMINATORS);
4780 calculate_dominance_info (CDI_DOMINATORS);
4782 bitmap_obstack_initialize (&grand_bitmap_obstack);
4783 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4784 expr_pred_trans_eq, free);
4785 expression_to_id = htab_create (num_ssa_names * 3,
4786 pre_expr_hash,
4787 pre_expr_eq, NULL);
4788 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4789 sizeof (struct bitmap_set), 30);
4790 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4791 sizeof (struct pre_expr_d), 30);
4792 FOR_ALL_BB (bb)
4794 EXP_GEN (bb) = bitmap_set_new ();
4795 PHI_GEN (bb) = bitmap_set_new ();
4796 TMP_GEN (bb) = bitmap_set_new ();
4797 AVAIL_OUT (bb) = bitmap_set_new ();
4800 need_eh_cleanup = BITMAP_ALLOC (NULL);
4801 need_ab_cleanup = BITMAP_ALLOC (NULL);
4805 /* Deallocate data structures used by PRE. */
4807 static void
4808 fini_pre (bool do_fre)
4810 free (postorder);
4811 VEC_free (bitmap_set_t, heap, value_expressions);
4812 BITMAP_FREE (inserted_exprs);
4813 VEC_free (gimple, heap, need_creation);
4814 bitmap_obstack_release (&grand_bitmap_obstack);
4815 free_alloc_pool (bitmap_set_pool);
4816 free_alloc_pool (pre_expr_pool);
4817 htab_delete (phi_translate_table);
4818 htab_delete (expression_to_id);
4819 VEC_free (unsigned, heap, name_to_id);
4821 free_aux_for_blocks ();
4823 free_dominance_info (CDI_POST_DOMINATORS);
4825 if (!bitmap_empty_p (need_eh_cleanup))
4827 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4828 cleanup_tree_cfg ();
4831 BITMAP_FREE (need_eh_cleanup);
4833 if (!bitmap_empty_p (need_ab_cleanup))
4835 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4836 cleanup_tree_cfg ();
4839 BITMAP_FREE (need_ab_cleanup);
4841 if (!do_fre)
4842 loop_optimizer_finalize ();
4845 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4846 only wants to do full redundancy elimination. */
4848 static unsigned int
4849 execute_pre (bool do_fre)
4851 unsigned int todo = 0;
4853 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4855 /* This has to happen before SCCVN runs because
4856 loop_optimizer_init may create new phis, etc. */
4857 if (!do_fre)
4858 loop_optimizer_init (LOOPS_NORMAL);
4860 if (!run_scc_vn (do_fre ? VN_WALKREWRITE : VN_WALK))
4862 if (!do_fre)
4863 loop_optimizer_finalize ();
4865 return 0;
4868 init_pre (do_fre);
4869 scev_initialize ();
4871 /* Collect and value number expressions computed in each basic block. */
4872 compute_avail ();
4874 if (dump_file && (dump_flags & TDF_DETAILS))
4876 basic_block bb;
4878 FOR_ALL_BB (bb)
4880 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4881 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4882 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4883 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4887 /* Insert can get quite slow on an incredibly large number of basic
4888 blocks due to some quadratic behavior. Until this behavior is
4889 fixed, don't run it when he have an incredibly large number of
4890 bb's. If we aren't going to run insert, there is no point in
4891 computing ANTIC, either, even though it's plenty fast. */
4892 if (!do_fre && n_basic_blocks < 4000)
4894 compute_antic ();
4895 insert ();
4898 /* Make sure to remove fake edges before committing our inserts.
4899 This makes sure we don't end up with extra critical edges that
4900 we would need to split. */
4901 remove_fake_exit_edges ();
4902 gsi_commit_edge_inserts ();
4904 /* Remove all the redundant expressions. */
4905 todo |= eliminate ();
4907 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4908 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4909 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4910 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4911 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4913 clear_expression_ids ();
4914 free_scc_vn ();
4915 if (!do_fre)
4917 remove_dead_inserted_code ();
4918 todo |= TODO_verify_flow;
4921 scev_finalize ();
4922 fini_pre (do_fre);
4924 return todo;
4927 /* Gate and execute functions for PRE. */
4929 static unsigned int
4930 do_pre (void)
4932 return execute_pre (false);
4935 static bool
4936 gate_pre (void)
4938 return flag_tree_pre != 0;
4941 struct gimple_opt_pass pass_pre =
4944 GIMPLE_PASS,
4945 "pre", /* name */
4946 gate_pre, /* gate */
4947 do_pre, /* execute */
4948 NULL, /* sub */
4949 NULL, /* next */
4950 0, /* static_pass_number */
4951 TV_TREE_PRE, /* tv_id */
4952 PROP_no_crit_edges | PROP_cfg
4953 | PROP_ssa, /* properties_required */
4954 0, /* properties_provided */
4955 0, /* properties_destroyed */
4956 TODO_rebuild_alias, /* todo_flags_start */
4957 TODO_update_ssa_only_virtuals | TODO_ggc_collect
4958 | TODO_verify_ssa /* todo_flags_finish */
4963 /* Gate and execute functions for FRE. */
4965 static unsigned int
4966 execute_fre (void)
4968 return execute_pre (true);
4971 static bool
4972 gate_fre (void)
4974 return flag_tree_fre != 0;
4977 struct gimple_opt_pass pass_fre =
4980 GIMPLE_PASS,
4981 "fre", /* name */
4982 gate_fre, /* gate */
4983 execute_fre, /* execute */
4984 NULL, /* sub */
4985 NULL, /* next */
4986 0, /* static_pass_number */
4987 TV_TREE_FRE, /* tv_id */
4988 PROP_cfg | PROP_ssa, /* properties_required */
4989 0, /* properties_provided */
4990 0, /* properties_destroyed */
4991 0, /* todo_flags_start */
4992 TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */