2011-08-19 Andrew Stubbs <ams@codesourcery.com>
[official-gcc.git] / gcc / tree-ssa-pre.c
blobd784bac6818dceab657912cfe8cc695f1ad1ee6c
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 get_var_ann (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 struct vn_nary_op_s newnary;
1447 /* The NARY structure is only guaranteed to have been
1448 allocated to the nary->length operands. */
1449 memcpy (&newnary, nary, (sizeof (struct vn_nary_op_s)
1450 - sizeof (tree) * (4 - nary->length)));
1452 for (i = 0; i < newnary.length; i++)
1454 if (TREE_CODE (newnary.op[i]) != SSA_NAME)
1455 continue;
1456 else
1458 pre_expr leader, result;
1459 unsigned int op_val_id = VN_INFO (newnary.op[i])->value_id;
1460 leader = find_leader_in_sets (op_val_id, set1, set2);
1461 result = phi_translate (leader, set1, set2, pred, phiblock);
1462 if (result && result != leader)
1464 tree name = get_representative_for (result);
1465 if (!name)
1466 return NULL;
1467 newnary.op[i] = name;
1469 else if (!result)
1470 return NULL;
1472 changed |= newnary.op[i] != nary->op[i];
1475 if (changed)
1477 pre_expr constant;
1478 unsigned int new_val_id;
1480 tree result = vn_nary_op_lookup_pieces (newnary.length,
1481 newnary.opcode,
1482 newnary.type,
1483 newnary.op[0],
1484 newnary.op[1],
1485 newnary.op[2],
1486 newnary.op[3],
1487 &nary);
1488 if (result && is_gimple_min_invariant (result))
1489 return get_or_alloc_expr_for_constant (result);
1491 expr = (pre_expr) pool_alloc (pre_expr_pool);
1492 expr->kind = NARY;
1493 expr->id = 0;
1494 if (nary)
1496 PRE_EXPR_NARY (expr) = nary;
1497 constant = fully_constant_expression (expr);
1498 if (constant != expr)
1499 return constant;
1501 new_val_id = nary->value_id;
1502 get_or_alloc_expression_id (expr);
1504 else
1506 new_val_id = get_next_value_id ();
1507 VEC_safe_grow_cleared (bitmap_set_t, heap,
1508 value_expressions,
1509 get_max_value_id() + 1);
1510 nary = vn_nary_op_insert_pieces (newnary.length,
1511 newnary.opcode,
1512 newnary.type,
1513 newnary.op[0],
1514 newnary.op[1],
1515 newnary.op[2],
1516 newnary.op[3],
1517 result, new_val_id);
1518 PRE_EXPR_NARY (expr) = nary;
1519 constant = fully_constant_expression (expr);
1520 if (constant != expr)
1521 return constant;
1522 get_or_alloc_expression_id (expr);
1524 add_to_value (new_val_id, expr);
1526 return expr;
1528 break;
1530 case REFERENCE:
1532 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1533 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1534 tree vuse = ref->vuse;
1535 tree newvuse = vuse;
1536 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1537 bool changed = false, same_valid = true;
1538 unsigned int i, j;
1539 vn_reference_op_t operand;
1540 vn_reference_t newref;
1542 for (i = 0, j = 0;
1543 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1545 pre_expr opresult;
1546 pre_expr leader;
1547 tree oldop0 = operand->op0;
1548 tree oldop1 = operand->op1;
1549 tree oldop2 = operand->op2;
1550 tree op0 = oldop0;
1551 tree op1 = oldop1;
1552 tree op2 = oldop2;
1553 tree type = operand->type;
1554 vn_reference_op_s newop = *operand;
1556 if (op0 && TREE_CODE (op0) == SSA_NAME)
1558 unsigned int op_val_id = VN_INFO (op0)->value_id;
1559 leader = find_leader_in_sets (op_val_id, set1, set2);
1560 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1561 if (opresult && opresult != leader)
1563 tree name = get_representative_for (opresult);
1564 if (!name)
1565 break;
1566 op0 = name;
1568 else if (!opresult)
1569 break;
1571 changed |= op0 != oldop0;
1573 if (op1 && TREE_CODE (op1) == SSA_NAME)
1575 unsigned int op_val_id = VN_INFO (op1)->value_id;
1576 leader = find_leader_in_sets (op_val_id, set1, set2);
1577 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1578 if (opresult && opresult != leader)
1580 tree name = get_representative_for (opresult);
1581 if (!name)
1582 break;
1583 op1 = name;
1585 else if (!opresult)
1586 break;
1588 /* We can't possibly insert these. */
1589 else if (op1 && !is_gimple_min_invariant (op1))
1590 break;
1591 changed |= op1 != oldop1;
1592 if (op2 && TREE_CODE (op2) == SSA_NAME)
1594 unsigned int op_val_id = VN_INFO (op2)->value_id;
1595 leader = find_leader_in_sets (op_val_id, set1, set2);
1596 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1597 if (opresult && opresult != leader)
1599 tree name = get_representative_for (opresult);
1600 if (!name)
1601 break;
1602 op2 = name;
1604 else if (!opresult)
1605 break;
1607 /* We can't possibly insert these. */
1608 else if (op2 && !is_gimple_min_invariant (op2))
1609 break;
1610 changed |= op2 != oldop2;
1612 if (!newoperands)
1613 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1614 /* We may have changed from an SSA_NAME to a constant */
1615 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1616 newop.opcode = TREE_CODE (op0);
1617 newop.type = type;
1618 newop.op0 = op0;
1619 newop.op1 = op1;
1620 newop.op2 = op2;
1621 /* If it transforms a non-constant ARRAY_REF into a constant
1622 one, adjust the constant offset. */
1623 if (newop.opcode == ARRAY_REF
1624 && newop.off == -1
1625 && TREE_CODE (op0) == INTEGER_CST
1626 && TREE_CODE (op1) == INTEGER_CST
1627 && TREE_CODE (op2) == INTEGER_CST)
1629 double_int off = tree_to_double_int (op0);
1630 off = double_int_add (off,
1631 double_int_neg
1632 (tree_to_double_int (op1)));
1633 off = double_int_mul (off, tree_to_double_int (op2));
1634 if (double_int_fits_in_shwi_p (off))
1635 newop.off = off.low;
1637 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1638 /* If it transforms from an SSA_NAME to an address, fold with
1639 a preceding indirect reference. */
1640 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1641 && VEC_index (vn_reference_op_s,
1642 newoperands, j - 1)->opcode == MEM_REF)
1643 vn_reference_fold_indirect (&newoperands, &j);
1645 if (i != VEC_length (vn_reference_op_s, operands))
1647 if (newoperands)
1648 VEC_free (vn_reference_op_s, heap, newoperands);
1649 return NULL;
1652 if (vuse)
1654 newvuse = translate_vuse_through_block (newoperands,
1655 ref->set, ref->type,
1656 vuse, phiblock, pred,
1657 &same_valid);
1658 if (newvuse == NULL_TREE)
1660 VEC_free (vn_reference_op_s, heap, newoperands);
1661 return NULL;
1665 if (changed || newvuse != vuse)
1667 unsigned int new_val_id;
1668 pre_expr constant;
1669 bool converted = false;
1671 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1672 ref->type,
1673 newoperands,
1674 &newref, VN_WALK);
1675 if (result)
1676 VEC_free (vn_reference_op_s, heap, newoperands);
1678 if (result
1679 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1681 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1682 converted = true;
1684 else if (!result && newref
1685 && !useless_type_conversion_p (ref->type, newref->type))
1687 VEC_free (vn_reference_op_s, heap, newoperands);
1688 return NULL;
1691 if (result && is_gimple_min_invariant (result))
1693 gcc_assert (!newoperands);
1694 return get_or_alloc_expr_for_constant (result);
1697 expr = (pre_expr) pool_alloc (pre_expr_pool);
1698 expr->kind = REFERENCE;
1699 expr->id = 0;
1701 if (converted)
1703 vn_nary_op_t nary;
1704 tree nresult;
1706 gcc_assert (CONVERT_EXPR_P (result)
1707 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1709 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1710 TREE_TYPE (result),
1711 TREE_OPERAND (result, 0),
1712 NULL_TREE, NULL_TREE,
1713 NULL_TREE,
1714 &nary);
1715 if (nresult && is_gimple_min_invariant (nresult))
1716 return get_or_alloc_expr_for_constant (nresult);
1718 expr->kind = NARY;
1719 if (nary)
1721 PRE_EXPR_NARY (expr) = nary;
1722 constant = fully_constant_expression (expr);
1723 if (constant != expr)
1724 return constant;
1726 new_val_id = nary->value_id;
1727 get_or_alloc_expression_id (expr);
1729 else
1731 new_val_id = get_next_value_id ();
1732 VEC_safe_grow_cleared (bitmap_set_t, heap,
1733 value_expressions,
1734 get_max_value_id() + 1);
1735 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1736 TREE_TYPE (result),
1737 TREE_OPERAND (result, 0),
1738 NULL_TREE, NULL_TREE,
1739 NULL_TREE, NULL_TREE,
1740 new_val_id);
1741 PRE_EXPR_NARY (expr) = nary;
1742 constant = fully_constant_expression (expr);
1743 if (constant != expr)
1744 return constant;
1745 get_or_alloc_expression_id (expr);
1748 else if (newref)
1750 PRE_EXPR_REFERENCE (expr) = newref;
1751 constant = fully_constant_expression (expr);
1752 if (constant != expr)
1753 return constant;
1755 new_val_id = newref->value_id;
1756 get_or_alloc_expression_id (expr);
1758 else
1760 if (changed || !same_valid)
1762 new_val_id = get_next_value_id ();
1763 VEC_safe_grow_cleared (bitmap_set_t, heap,
1764 value_expressions,
1765 get_max_value_id() + 1);
1767 else
1768 new_val_id = ref->value_id;
1769 newref = vn_reference_insert_pieces (newvuse, ref->set,
1770 ref->type,
1771 newoperands,
1772 result, new_val_id);
1773 newoperands = NULL;
1774 PRE_EXPR_REFERENCE (expr) = newref;
1775 constant = fully_constant_expression (expr);
1776 if (constant != expr)
1777 return constant;
1778 get_or_alloc_expression_id (expr);
1780 add_to_value (new_val_id, expr);
1782 VEC_free (vn_reference_op_s, heap, newoperands);
1783 return expr;
1785 break;
1787 case NAME:
1789 gimple phi = NULL;
1790 edge e;
1791 gimple def_stmt;
1792 tree name = PRE_EXPR_NAME (expr);
1794 def_stmt = SSA_NAME_DEF_STMT (name);
1795 if (gimple_code (def_stmt) == GIMPLE_PHI
1796 && gimple_bb (def_stmt) == phiblock)
1797 phi = def_stmt;
1798 else
1799 return expr;
1801 e = find_edge (pred, gimple_bb (phi));
1802 if (e)
1804 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1805 pre_expr newexpr;
1807 if (TREE_CODE (def) == SSA_NAME)
1808 def = VN_INFO (def)->valnum;
1810 /* Handle constant. */
1811 if (is_gimple_min_invariant (def))
1812 return get_or_alloc_expr_for_constant (def);
1814 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1815 return NULL;
1817 newexpr = get_or_alloc_expr_for_name (def);
1818 return newexpr;
1821 return expr;
1823 default:
1824 gcc_unreachable ();
1828 /* Wrapper around phi_translate_1 providing caching functionality. */
1830 static pre_expr
1831 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1832 basic_block pred, basic_block phiblock)
1834 pre_expr phitrans;
1836 if (!expr)
1837 return NULL;
1839 /* Constants contain no values that need translation. */
1840 if (expr->kind == CONSTANT)
1841 return expr;
1843 if (value_id_constant_p (get_expr_value_id (expr)))
1844 return expr;
1846 if (expr->kind != NAME)
1848 phitrans = phi_trans_lookup (expr, pred);
1849 if (phitrans)
1850 return phitrans;
1853 /* Translate. */
1854 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1856 /* Don't add empty translations to the cache. Neither add
1857 translations of NAMEs as those are cheap to translate. */
1858 if (phitrans
1859 && expr->kind != NAME)
1860 phi_trans_add (expr, phitrans, pred);
1862 return phitrans;
1866 /* For each expression in SET, translate the values through phi nodes
1867 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1868 expressions in DEST. */
1870 static void
1871 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1872 basic_block phiblock)
1874 VEC (pre_expr, heap) *exprs;
1875 pre_expr expr;
1876 int i;
1878 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1880 bitmap_set_copy (dest, set);
1881 return;
1884 exprs = sorted_array_from_bitmap_set (set);
1885 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1887 pre_expr translated;
1888 translated = phi_translate (expr, set, NULL, pred, phiblock);
1889 if (!translated)
1890 continue;
1892 /* We might end up with multiple expressions from SET being
1893 translated to the same value. In this case we do not want
1894 to retain the NARY or REFERENCE expression but prefer a NAME
1895 which would be the leader. */
1896 if (translated->kind == NAME)
1897 bitmap_value_replace_in_set (dest, translated);
1898 else
1899 bitmap_value_insert_into_set (dest, translated);
1901 VEC_free (pre_expr, heap, exprs);
1904 /* Find the leader for a value (i.e., the name representing that
1905 value) in a given set, and return it. If STMT is non-NULL it
1906 makes sure the defining statement for the leader dominates it.
1907 Return NULL if no leader is found. */
1909 static pre_expr
1910 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1912 if (value_id_constant_p (val))
1914 unsigned int i;
1915 bitmap_iterator bi;
1916 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1918 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1920 pre_expr expr = expression_for_id (i);
1921 if (expr->kind == CONSTANT)
1922 return expr;
1925 if (bitmap_set_contains_value (set, val))
1927 /* Rather than walk the entire bitmap of expressions, and see
1928 whether any of them has the value we are looking for, we look
1929 at the reverse mapping, which tells us the set of expressions
1930 that have a given value (IE value->expressions with that
1931 value) and see if any of those expressions are in our set.
1932 The number of expressions per value is usually significantly
1933 less than the number of expressions in the set. In fact, for
1934 large testcases, doing it this way is roughly 5-10x faster
1935 than walking the bitmap.
1936 If this is somehow a significant lose for some cases, we can
1937 choose which set to walk based on which set is smaller. */
1938 unsigned int i;
1939 bitmap_iterator bi;
1940 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1942 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1943 &set->expressions, 0, i, bi)
1945 pre_expr val = expression_for_id (i);
1946 /* At the point where stmt is not null, there should always
1947 be an SSA_NAME first in the list of expressions. */
1948 if (stmt)
1950 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1951 if (gimple_code (def_stmt) != GIMPLE_PHI
1952 && gimple_bb (def_stmt) == gimple_bb (stmt)
1953 /* PRE insertions are at the end of the basic-block
1954 and have UID 0. */
1955 && (gimple_uid (def_stmt) == 0
1956 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1957 continue;
1959 return val;
1962 return NULL;
1965 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1966 BLOCK by seeing if it is not killed in the block. Note that we are
1967 only determining whether there is a store that kills it. Because
1968 of the order in which clean iterates over values, we are guaranteed
1969 that altered operands will have caused us to be eliminated from the
1970 ANTIC_IN set already. */
1972 static bool
1973 value_dies_in_block_x (pre_expr expr, basic_block block)
1975 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1976 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1977 gimple def;
1978 gimple_stmt_iterator gsi;
1979 unsigned id = get_expression_id (expr);
1980 bool res = false;
1981 ao_ref ref;
1983 if (!vuse)
1984 return false;
1986 /* Lookup a previously calculated result. */
1987 if (EXPR_DIES (block)
1988 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1989 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1991 /* A memory expression {e, VUSE} dies in the block if there is a
1992 statement that may clobber e. If, starting statement walk from the
1993 top of the basic block, a statement uses VUSE there can be no kill
1994 inbetween that use and the original statement that loaded {e, VUSE},
1995 so we can stop walking. */
1996 ref.base = NULL_TREE;
1997 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1999 tree def_vuse, def_vdef;
2000 def = gsi_stmt (gsi);
2001 def_vuse = gimple_vuse (def);
2002 def_vdef = gimple_vdef (def);
2004 /* Not a memory statement. */
2005 if (!def_vuse)
2006 continue;
2008 /* Not a may-def. */
2009 if (!def_vdef)
2011 /* A load with the same VUSE, we're done. */
2012 if (def_vuse == vuse)
2013 break;
2015 continue;
2018 /* Init ref only if we really need it. */
2019 if (ref.base == NULL_TREE
2020 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
2021 refx->operands))
2023 res = true;
2024 break;
2026 /* If the statement may clobber expr, it dies. */
2027 if (stmt_may_clobber_ref_p_1 (def, &ref))
2029 res = true;
2030 break;
2034 /* Remember the result. */
2035 if (!EXPR_DIES (block))
2036 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2037 bitmap_set_bit (EXPR_DIES (block), id * 2);
2038 if (res)
2039 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2041 return res;
2045 #define union_contains_value(SET1, SET2, VAL) \
2046 (bitmap_set_contains_value ((SET1), (VAL)) \
2047 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
2049 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2051 static bool
2052 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2053 vn_reference_op_t vro)
2055 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2057 struct pre_expr_d temp;
2058 temp.kind = NAME;
2059 temp.id = 0;
2060 PRE_EXPR_NAME (&temp) = vro->op0;
2061 temp.id = lookup_expression_id (&temp);
2062 if (temp.id == 0)
2063 return false;
2064 if (!union_contains_value (set1, set2,
2065 get_expr_value_id (&temp)))
2066 return false;
2068 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2070 struct pre_expr_d temp;
2071 temp.kind = NAME;
2072 temp.id = 0;
2073 PRE_EXPR_NAME (&temp) = vro->op1;
2074 temp.id = lookup_expression_id (&temp);
2075 if (temp.id == 0)
2076 return false;
2077 if (!union_contains_value (set1, set2,
2078 get_expr_value_id (&temp)))
2079 return false;
2082 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2084 struct pre_expr_d temp;
2085 temp.kind = NAME;
2086 temp.id = 0;
2087 PRE_EXPR_NAME (&temp) = vro->op2;
2088 temp.id = lookup_expression_id (&temp);
2089 if (temp.id == 0)
2090 return false;
2091 if (!union_contains_value (set1, set2,
2092 get_expr_value_id (&temp)))
2093 return false;
2096 return true;
2099 /* Determine if the expression EXPR is valid in SET1 U SET2.
2100 ONLY SET2 CAN BE NULL.
2101 This means that we have a leader for each part of the expression
2102 (if it consists of values), or the expression is an SSA_NAME.
2103 For loads/calls, we also see if the vuse is killed in this block. */
2105 static bool
2106 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2107 basic_block block)
2109 switch (expr->kind)
2111 case NAME:
2112 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2113 case NARY:
2115 unsigned int i;
2116 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2117 for (i = 0; i < nary->length; i++)
2119 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2121 struct pre_expr_d temp;
2122 temp.kind = NAME;
2123 temp.id = 0;
2124 PRE_EXPR_NAME (&temp) = nary->op[i];
2125 temp.id = lookup_expression_id (&temp);
2126 if (temp.id == 0)
2127 return false;
2128 if (!union_contains_value (set1, set2,
2129 get_expr_value_id (&temp)))
2130 return false;
2133 /* If the NARY may trap make sure the block does not contain
2134 a possible exit point.
2135 ??? This is overly conservative if we translate AVAIL_OUT
2136 as the available expression might be after the exit point. */
2137 if (BB_MAY_NOTRETURN (block)
2138 && vn_nary_may_trap (nary))
2139 return false;
2140 return true;
2142 break;
2143 case REFERENCE:
2145 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2146 vn_reference_op_t vro;
2147 unsigned int i;
2149 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2151 if (!vro_valid_in_sets (set1, set2, vro))
2152 return false;
2154 if (ref->vuse)
2156 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2157 if (!gimple_nop_p (def_stmt)
2158 && gimple_bb (def_stmt) != block
2159 && !dominated_by_p (CDI_DOMINATORS,
2160 block, gimple_bb (def_stmt)))
2161 return false;
2163 return !value_dies_in_block_x (expr, block);
2165 default:
2166 gcc_unreachable ();
2170 /* Clean the set of expressions that are no longer valid in SET1 or
2171 SET2. This means expressions that are made up of values we have no
2172 leaders for in SET1 or SET2. This version is used for partial
2173 anticipation, which means it is not valid in either ANTIC_IN or
2174 PA_IN. */
2176 static void
2177 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2179 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2180 pre_expr expr;
2181 int i;
2183 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2185 if (!valid_in_sets (set1, set2, expr, block))
2186 bitmap_remove_from_set (set1, expr);
2188 VEC_free (pre_expr, heap, exprs);
2191 /* Clean the set of expressions that are no longer valid in SET. This
2192 means expressions that are made up of values we have no leaders for
2193 in SET. */
2195 static void
2196 clean (bitmap_set_t set, basic_block block)
2198 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2199 pre_expr expr;
2200 int i;
2202 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2204 if (!valid_in_sets (set, NULL, expr, block))
2205 bitmap_remove_from_set (set, expr);
2207 VEC_free (pre_expr, heap, exprs);
2210 static sbitmap has_abnormal_preds;
2212 /* List of blocks that may have changed during ANTIC computation and
2213 thus need to be iterated over. */
2215 static sbitmap changed_blocks;
2217 /* Decide whether to defer a block for a later iteration, or PHI
2218 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2219 should defer the block, and true if we processed it. */
2221 static bool
2222 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2223 basic_block block, basic_block phiblock)
2225 if (!BB_VISITED (phiblock))
2227 SET_BIT (changed_blocks, block->index);
2228 BB_VISITED (block) = 0;
2229 BB_DEFERRED (block) = 1;
2230 return false;
2232 else
2233 phi_translate_set (dest, source, block, phiblock);
2234 return true;
2237 /* Compute the ANTIC set for BLOCK.
2239 If succs(BLOCK) > 1 then
2240 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2241 else if succs(BLOCK) == 1 then
2242 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2244 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2247 static bool
2248 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2250 bool changed = false;
2251 bitmap_set_t S, old, ANTIC_OUT;
2252 bitmap_iterator bi;
2253 unsigned int bii;
2254 edge e;
2255 edge_iterator ei;
2257 old = ANTIC_OUT = S = NULL;
2258 BB_VISITED (block) = 1;
2260 /* If any edges from predecessors are abnormal, antic_in is empty,
2261 so do nothing. */
2262 if (block_has_abnormal_pred_edge)
2263 goto maybe_dump_sets;
2265 old = ANTIC_IN (block);
2266 ANTIC_OUT = bitmap_set_new ();
2268 /* If the block has no successors, ANTIC_OUT is empty. */
2269 if (EDGE_COUNT (block->succs) == 0)
2271 /* If we have one successor, we could have some phi nodes to
2272 translate through. */
2273 else if (single_succ_p (block))
2275 basic_block succ_bb = single_succ (block);
2277 /* We trade iterations of the dataflow equations for having to
2278 phi translate the maximal set, which is incredibly slow
2279 (since the maximal set often has 300+ members, even when you
2280 have a small number of blocks).
2281 Basically, we defer the computation of ANTIC for this block
2282 until we have processed it's successor, which will inevitably
2283 have a *much* smaller set of values to phi translate once
2284 clean has been run on it.
2285 The cost of doing this is that we technically perform more
2286 iterations, however, they are lower cost iterations.
2288 Timings for PRE on tramp3d-v4:
2289 without maximal set fix: 11 seconds
2290 with maximal set fix/without deferring: 26 seconds
2291 with maximal set fix/with deferring: 11 seconds
2294 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2295 block, succ_bb))
2297 changed = true;
2298 goto maybe_dump_sets;
2301 /* If we have multiple successors, we take the intersection of all of
2302 them. Note that in the case of loop exit phi nodes, we may have
2303 phis to translate through. */
2304 else
2306 VEC(basic_block, heap) * worklist;
2307 size_t i;
2308 basic_block bprime, first = NULL;
2310 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2311 FOR_EACH_EDGE (e, ei, block->succs)
2313 if (!first
2314 && BB_VISITED (e->dest))
2315 first = e->dest;
2316 else if (BB_VISITED (e->dest))
2317 VEC_quick_push (basic_block, worklist, e->dest);
2320 /* Of multiple successors we have to have visited one already. */
2321 if (!first)
2323 SET_BIT (changed_blocks, block->index);
2324 BB_VISITED (block) = 0;
2325 BB_DEFERRED (block) = 1;
2326 changed = true;
2327 VEC_free (basic_block, heap, worklist);
2328 goto maybe_dump_sets;
2331 if (!gimple_seq_empty_p (phi_nodes (first)))
2332 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2333 else
2334 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2336 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2338 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2340 bitmap_set_t tmp = bitmap_set_new ();
2341 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2342 bitmap_set_and (ANTIC_OUT, tmp);
2343 bitmap_set_free (tmp);
2345 else
2346 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2348 VEC_free (basic_block, heap, worklist);
2351 /* Generate ANTIC_OUT - TMP_GEN. */
2352 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2354 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2355 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2356 TMP_GEN (block));
2358 /* Then union in the ANTIC_OUT - TMP_GEN values,
2359 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2360 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2361 bitmap_value_insert_into_set (ANTIC_IN (block),
2362 expression_for_id (bii));
2364 clean (ANTIC_IN (block), block);
2366 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2368 changed = true;
2369 SET_BIT (changed_blocks, block->index);
2370 FOR_EACH_EDGE (e, ei, block->preds)
2371 SET_BIT (changed_blocks, e->src->index);
2373 else
2374 RESET_BIT (changed_blocks, block->index);
2376 maybe_dump_sets:
2377 if (dump_file && (dump_flags & TDF_DETAILS))
2379 if (!BB_DEFERRED (block) || BB_VISITED (block))
2381 if (ANTIC_OUT)
2382 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2384 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2385 block->index);
2387 if (S)
2388 print_bitmap_set (dump_file, S, "S", block->index);
2390 else
2392 fprintf (dump_file,
2393 "Block %d was deferred for a future iteration.\n",
2394 block->index);
2397 if (old)
2398 bitmap_set_free (old);
2399 if (S)
2400 bitmap_set_free (S);
2401 if (ANTIC_OUT)
2402 bitmap_set_free (ANTIC_OUT);
2403 return changed;
2406 /* Compute PARTIAL_ANTIC for BLOCK.
2408 If succs(BLOCK) > 1 then
2409 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2410 in ANTIC_OUT for all succ(BLOCK)
2411 else if succs(BLOCK) == 1 then
2412 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2414 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2415 - ANTIC_IN[BLOCK])
2418 static bool
2419 compute_partial_antic_aux (basic_block block,
2420 bool block_has_abnormal_pred_edge)
2422 bool changed = false;
2423 bitmap_set_t old_PA_IN;
2424 bitmap_set_t PA_OUT;
2425 edge e;
2426 edge_iterator ei;
2427 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2429 old_PA_IN = PA_OUT = NULL;
2431 /* If any edges from predecessors are abnormal, antic_in is empty,
2432 so do nothing. */
2433 if (block_has_abnormal_pred_edge)
2434 goto maybe_dump_sets;
2436 /* If there are too many partially anticipatable values in the
2437 block, phi_translate_set can take an exponential time: stop
2438 before the translation starts. */
2439 if (max_pa
2440 && single_succ_p (block)
2441 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2442 goto maybe_dump_sets;
2444 old_PA_IN = PA_IN (block);
2445 PA_OUT = bitmap_set_new ();
2447 /* If the block has no successors, ANTIC_OUT is empty. */
2448 if (EDGE_COUNT (block->succs) == 0)
2450 /* If we have one successor, we could have some phi nodes to
2451 translate through. Note that we can't phi translate across DFS
2452 back edges in partial antic, because it uses a union operation on
2453 the successors. For recurrences like IV's, we will end up
2454 generating a new value in the set on each go around (i + 3 (VH.1)
2455 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2456 else if (single_succ_p (block))
2458 basic_block succ = single_succ (block);
2459 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2460 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2462 /* If we have multiple successors, we take the union of all of
2463 them. */
2464 else
2466 VEC(basic_block, heap) * worklist;
2467 size_t i;
2468 basic_block bprime;
2470 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2471 FOR_EACH_EDGE (e, ei, block->succs)
2473 if (e->flags & EDGE_DFS_BACK)
2474 continue;
2475 VEC_quick_push (basic_block, worklist, e->dest);
2477 if (VEC_length (basic_block, worklist) > 0)
2479 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2481 unsigned int i;
2482 bitmap_iterator bi;
2484 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2485 bitmap_value_insert_into_set (PA_OUT,
2486 expression_for_id (i));
2487 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2489 bitmap_set_t pa_in = bitmap_set_new ();
2490 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2491 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2492 bitmap_value_insert_into_set (PA_OUT,
2493 expression_for_id (i));
2494 bitmap_set_free (pa_in);
2496 else
2497 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2498 bitmap_value_insert_into_set (PA_OUT,
2499 expression_for_id (i));
2502 VEC_free (basic_block, heap, worklist);
2505 /* PA_IN starts with PA_OUT - TMP_GEN.
2506 Then we subtract things from ANTIC_IN. */
2507 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2509 /* For partial antic, we want to put back in the phi results, since
2510 we will properly avoid making them partially antic over backedges. */
2511 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2512 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2514 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2515 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2517 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2519 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2521 changed = true;
2522 SET_BIT (changed_blocks, block->index);
2523 FOR_EACH_EDGE (e, ei, block->preds)
2524 SET_BIT (changed_blocks, e->src->index);
2526 else
2527 RESET_BIT (changed_blocks, block->index);
2529 maybe_dump_sets:
2530 if (dump_file && (dump_flags & TDF_DETAILS))
2532 if (PA_OUT)
2533 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2535 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2537 if (old_PA_IN)
2538 bitmap_set_free (old_PA_IN);
2539 if (PA_OUT)
2540 bitmap_set_free (PA_OUT);
2541 return changed;
2544 /* Compute ANTIC and partial ANTIC sets. */
2546 static void
2547 compute_antic (void)
2549 bool changed = true;
2550 int num_iterations = 0;
2551 basic_block block;
2552 int i;
2554 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2555 We pre-build the map of blocks with incoming abnormal edges here. */
2556 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2557 sbitmap_zero (has_abnormal_preds);
2559 FOR_EACH_BB (block)
2561 edge_iterator ei;
2562 edge e;
2564 FOR_EACH_EDGE (e, ei, block->preds)
2566 e->flags &= ~EDGE_DFS_BACK;
2567 if (e->flags & EDGE_ABNORMAL)
2569 SET_BIT (has_abnormal_preds, block->index);
2570 break;
2574 BB_VISITED (block) = 0;
2575 BB_DEFERRED (block) = 0;
2577 /* While we are here, give empty ANTIC_IN sets to each block. */
2578 ANTIC_IN (block) = bitmap_set_new ();
2579 PA_IN (block) = bitmap_set_new ();
2582 /* At the exit block we anticipate nothing. */
2583 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2584 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2585 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2587 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2588 sbitmap_ones (changed_blocks);
2589 while (changed)
2591 if (dump_file && (dump_flags & TDF_DETAILS))
2592 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2593 /* ??? We need to clear our PHI translation cache here as the
2594 ANTIC sets shrink and we restrict valid translations to
2595 those having operands with leaders in ANTIC. Same below
2596 for PA ANTIC computation. */
2597 num_iterations++;
2598 changed = false;
2599 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2601 if (TEST_BIT (changed_blocks, postorder[i]))
2603 basic_block block = BASIC_BLOCK (postorder[i]);
2604 changed |= compute_antic_aux (block,
2605 TEST_BIT (has_abnormal_preds,
2606 block->index));
2609 /* Theoretically possible, but *highly* unlikely. */
2610 gcc_checking_assert (num_iterations < 500);
2613 statistics_histogram_event (cfun, "compute_antic iterations",
2614 num_iterations);
2616 if (do_partial_partial)
2618 sbitmap_ones (changed_blocks);
2619 mark_dfs_back_edges ();
2620 num_iterations = 0;
2621 changed = true;
2622 while (changed)
2624 if (dump_file && (dump_flags & TDF_DETAILS))
2625 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2626 num_iterations++;
2627 changed = false;
2628 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2630 if (TEST_BIT (changed_blocks, postorder[i]))
2632 basic_block block = BASIC_BLOCK (postorder[i]);
2633 changed
2634 |= compute_partial_antic_aux (block,
2635 TEST_BIT (has_abnormal_preds,
2636 block->index));
2639 /* Theoretically possible, but *highly* unlikely. */
2640 gcc_checking_assert (num_iterations < 500);
2642 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2643 num_iterations);
2645 sbitmap_free (has_abnormal_preds);
2646 sbitmap_free (changed_blocks);
2649 /* Return true if we can value number the call in STMT. This is true
2650 if we have a pure or constant call to a real function. */
2652 static bool
2653 can_value_number_call (gimple stmt)
2655 if (gimple_call_internal_p (stmt))
2656 return false;
2657 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2658 return true;
2659 return false;
2662 /* Return true if OP is a tree which we can perform PRE on.
2663 This may not match the operations we can value number, but in
2664 a perfect world would. */
2666 static bool
2667 can_PRE_operation (tree op)
2669 return UNARY_CLASS_P (op)
2670 || BINARY_CLASS_P (op)
2671 || COMPARISON_CLASS_P (op)
2672 || TREE_CODE (op) == MEM_REF
2673 || TREE_CODE (op) == COMPONENT_REF
2674 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2675 || TREE_CODE (op) == CALL_EXPR
2676 || TREE_CODE (op) == ARRAY_REF;
2680 /* Inserted expressions are placed onto this worklist, which is used
2681 for performing quick dead code elimination of insertions we made
2682 that didn't turn out to be necessary. */
2683 static bitmap inserted_exprs;
2685 /* Pool allocated fake store expressions are placed onto this
2686 worklist, which, after performing dead code elimination, is walked
2687 to see which expressions need to be put into GC'able memory */
2688 static VEC(gimple, heap) *need_creation;
2690 /* The actual worker for create_component_ref_by_pieces. */
2692 static tree
2693 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2694 unsigned int *operand, gimple_seq *stmts,
2695 gimple domstmt)
2697 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2698 *operand);
2699 tree genop;
2700 ++*operand;
2701 switch (currop->opcode)
2703 case CALL_EXPR:
2705 tree folded, sc = NULL_TREE;
2706 unsigned int nargs = 0;
2707 tree fn, *args;
2708 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2709 fn = currop->op0;
2710 else
2712 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2713 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2714 if (!fn)
2715 return NULL_TREE;
2717 if (currop->op1)
2719 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2720 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2721 if (!sc)
2722 return NULL_TREE;
2724 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2725 ref->operands) - 1);
2726 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2728 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2729 operand, stmts,
2730 domstmt);
2731 if (!args[nargs])
2733 free (args);
2734 return NULL_TREE;
2736 nargs++;
2738 folded = build_call_array (currop->type,
2739 (TREE_CODE (fn) == FUNCTION_DECL
2740 ? build_fold_addr_expr (fn) : fn),
2741 nargs, args);
2742 free (args);
2743 if (sc)
2744 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2745 return folded;
2747 break;
2748 case MEM_REF:
2750 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2751 stmts, domstmt);
2752 tree offset = currop->op0;
2753 if (!baseop)
2754 return NULL_TREE;
2755 if (TREE_CODE (baseop) == ADDR_EXPR
2756 && handled_component_p (TREE_OPERAND (baseop, 0)))
2758 HOST_WIDE_INT off;
2759 tree base;
2760 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2761 &off);
2762 gcc_assert (base);
2763 offset = int_const_binop (PLUS_EXPR, offset,
2764 build_int_cst (TREE_TYPE (offset),
2765 off));
2766 baseop = build_fold_addr_expr (base);
2768 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2770 break;
2771 case TARGET_MEM_REF:
2773 pre_expr op0expr, op1expr;
2774 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2775 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2776 ++*operand);
2777 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2778 stmts, domstmt);
2779 if (!baseop)
2780 return NULL_TREE;
2781 if (currop->op0)
2783 op0expr = get_or_alloc_expr_for (currop->op0);
2784 genop0 = find_or_generate_expression (block, op0expr,
2785 stmts, domstmt);
2786 if (!genop0)
2787 return NULL_TREE;
2789 if (nextop->op0)
2791 op1expr = get_or_alloc_expr_for (nextop->op0);
2792 genop1 = find_or_generate_expression (block, op1expr,
2793 stmts, domstmt);
2794 if (!genop1)
2795 return NULL_TREE;
2797 return build5 (TARGET_MEM_REF, currop->type,
2798 baseop, currop->op2, genop0, currop->op1, genop1);
2800 break;
2801 case ADDR_EXPR:
2802 if (currop->op0)
2804 gcc_assert (is_gimple_min_invariant (currop->op0));
2805 return currop->op0;
2807 /* Fallthrough. */
2808 case REALPART_EXPR:
2809 case IMAGPART_EXPR:
2810 case VIEW_CONVERT_EXPR:
2812 tree folded;
2813 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2814 operand,
2815 stmts, domstmt);
2816 if (!genop0)
2817 return NULL_TREE;
2818 folded = fold_build1 (currop->opcode, currop->type,
2819 genop0);
2820 return folded;
2822 break;
2823 case BIT_FIELD_REF:
2825 tree folded;
2826 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2827 stmts, domstmt);
2828 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2829 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2830 tree genop1;
2831 tree genop2;
2833 if (!genop0)
2834 return NULL_TREE;
2835 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2836 if (!genop1)
2837 return NULL_TREE;
2838 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2839 if (!genop2)
2840 return NULL_TREE;
2841 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2842 genop2);
2843 return folded;
2846 /* For array ref vn_reference_op's, operand 1 of the array ref
2847 is op0 of the reference op and operand 3 of the array ref is
2848 op1. */
2849 case ARRAY_RANGE_REF:
2850 case ARRAY_REF:
2852 tree genop0;
2853 tree genop1 = currop->op0;
2854 pre_expr op1expr;
2855 tree genop2 = currop->op1;
2856 pre_expr op2expr;
2857 tree genop3 = currop->op2;
2858 pre_expr op3expr;
2859 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2860 stmts, domstmt);
2861 if (!genop0)
2862 return NULL_TREE;
2863 op1expr = get_or_alloc_expr_for (genop1);
2864 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2865 if (!genop1)
2866 return NULL_TREE;
2867 if (genop2)
2869 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2870 /* Drop zero minimum index if redundant. */
2871 if (integer_zerop (genop2)
2872 && (!domain_type
2873 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2874 genop2 = NULL_TREE;
2875 else
2877 op2expr = get_or_alloc_expr_for (genop2);
2878 genop2 = find_or_generate_expression (block, op2expr, stmts,
2879 domstmt);
2880 if (!genop2)
2881 return NULL_TREE;
2884 if (genop3)
2886 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2887 /* We can't always put a size in units of the element alignment
2888 here as the element alignment may be not visible. See
2889 PR43783. Simply drop the element size for constant
2890 sizes. */
2891 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2892 genop3 = NULL_TREE;
2893 else
2895 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2896 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2897 op3expr = get_or_alloc_expr_for (genop3);
2898 genop3 = find_or_generate_expression (block, op3expr, stmts,
2899 domstmt);
2900 if (!genop3)
2901 return NULL_TREE;
2904 return build4 (currop->opcode, currop->type, genop0, genop1,
2905 genop2, genop3);
2907 case COMPONENT_REF:
2909 tree op0;
2910 tree op1;
2911 tree genop2 = currop->op1;
2912 pre_expr op2expr;
2913 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2914 stmts, domstmt);
2915 if (!op0)
2916 return NULL_TREE;
2917 /* op1 should be a FIELD_DECL, which are represented by
2918 themselves. */
2919 op1 = currop->op0;
2920 if (genop2)
2922 op2expr = get_or_alloc_expr_for (genop2);
2923 genop2 = find_or_generate_expression (block, op2expr, stmts,
2924 domstmt);
2925 if (!genop2)
2926 return NULL_TREE;
2929 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2930 genop2);
2932 break;
2933 case SSA_NAME:
2935 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2936 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2937 return genop;
2939 case STRING_CST:
2940 case INTEGER_CST:
2941 case COMPLEX_CST:
2942 case VECTOR_CST:
2943 case REAL_CST:
2944 case CONSTRUCTOR:
2945 case VAR_DECL:
2946 case PARM_DECL:
2947 case CONST_DECL:
2948 case RESULT_DECL:
2949 case FUNCTION_DECL:
2950 return currop->op0;
2952 default:
2953 gcc_unreachable ();
2957 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2958 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2959 trying to rename aggregates into ssa form directly, which is a no no.
2961 Thus, this routine doesn't create temporaries, it just builds a
2962 single access expression for the array, calling
2963 find_or_generate_expression to build the innermost pieces.
2965 This function is a subroutine of create_expression_by_pieces, and
2966 should not be called on it's own unless you really know what you
2967 are doing. */
2969 static tree
2970 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2971 gimple_seq *stmts, gimple domstmt)
2973 unsigned int op = 0;
2974 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2977 /* Find a leader for an expression, or generate one using
2978 create_expression_by_pieces if it's ANTIC but
2979 complex.
2980 BLOCK is the basic_block we are looking for leaders in.
2981 EXPR is the expression to find a leader or generate for.
2982 STMTS is the statement list to put the inserted expressions on.
2983 Returns the SSA_NAME of the LHS of the generated expression or the
2984 leader.
2985 DOMSTMT if non-NULL is a statement that should be dominated by
2986 all uses in the generated expression. If DOMSTMT is non-NULL this
2987 routine can fail and return NULL_TREE. Otherwise it will assert
2988 on failure. */
2990 static tree
2991 find_or_generate_expression (basic_block block, pre_expr expr,
2992 gimple_seq *stmts, gimple domstmt)
2994 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2995 get_expr_value_id (expr), domstmt);
2996 tree genop = NULL;
2997 if (leader)
2999 if (leader->kind == NAME)
3000 genop = PRE_EXPR_NAME (leader);
3001 else if (leader->kind == CONSTANT)
3002 genop = PRE_EXPR_CONSTANT (leader);
3005 /* If it's still NULL, it must be a complex expression, so generate
3006 it recursively. Not so if inserting expressions for values generated
3007 by SCCVN. */
3008 if (genop == NULL
3009 && !domstmt)
3011 bitmap_set_t exprset;
3012 unsigned int lookfor = get_expr_value_id (expr);
3013 bool handled = false;
3014 bitmap_iterator bi;
3015 unsigned int i;
3017 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3018 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3020 pre_expr temp = expression_for_id (i);
3021 if (temp->kind != NAME)
3023 handled = true;
3024 genop = create_expression_by_pieces (block, temp, stmts,
3025 domstmt,
3026 get_expr_type (expr));
3027 break;
3030 if (!handled && domstmt)
3031 return NULL_TREE;
3033 gcc_assert (handled);
3035 return genop;
3038 #define NECESSARY GF_PLF_1
3040 /* Create an expression in pieces, so that we can handle very complex
3041 expressions that may be ANTIC, but not necessary GIMPLE.
3042 BLOCK is the basic block the expression will be inserted into,
3043 EXPR is the expression to insert (in value form)
3044 STMTS is a statement list to append the necessary insertions into.
3046 This function will die if we hit some value that shouldn't be
3047 ANTIC but is (IE there is no leader for it, or its components).
3048 This function may also generate expressions that are themselves
3049 partially or fully redundant. Those that are will be either made
3050 fully redundant during the next iteration of insert (for partially
3051 redundant ones), or eliminated by eliminate (for fully redundant
3052 ones).
3054 If DOMSTMT is non-NULL then we make sure that all uses in the
3055 expressions dominate that statement. In this case the function
3056 can return NULL_TREE to signal failure. */
3058 static tree
3059 create_expression_by_pieces (basic_block block, pre_expr expr,
3060 gimple_seq *stmts, gimple domstmt, tree type)
3062 tree temp, name;
3063 tree folded;
3064 gimple_seq forced_stmts = NULL;
3065 unsigned int value_id;
3066 gimple_stmt_iterator gsi;
3067 tree exprtype = type ? type : get_expr_type (expr);
3068 pre_expr nameexpr;
3069 gimple newstmt;
3071 switch (expr->kind)
3073 /* We may hit the NAME/CONSTANT case if we have to convert types
3074 that value numbering saw through. */
3075 case NAME:
3076 folded = PRE_EXPR_NAME (expr);
3077 break;
3078 case CONSTANT:
3079 folded = PRE_EXPR_CONSTANT (expr);
3080 break;
3081 case REFERENCE:
3083 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3084 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3086 break;
3087 case NARY:
3089 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3090 switch (nary->length)
3092 case 2:
3094 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3095 pre_expr op2 = get_or_alloc_expr_for (nary->op[1]);
3096 tree genop1 = find_or_generate_expression (block, op1,
3097 stmts, domstmt);
3098 tree genop2 = find_or_generate_expression (block, op2,
3099 stmts, domstmt);
3100 if (!genop1 || !genop2)
3101 return NULL_TREE;
3102 /* Ensure op2 is a ptrofftype for POINTER_PLUS_EXPR. It
3103 may be a constant with the wrong type. */
3104 if (nary->opcode == POINTER_PLUS_EXPR)
3106 genop1 = fold_convert (nary->type, genop1);
3107 genop2 = convert_to_ptrofftype (genop2);
3109 else
3111 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3112 genop2 = fold_convert (TREE_TYPE (nary->op[1]), genop2);
3115 folded = fold_build2 (nary->opcode, nary->type,
3116 genop1, genop2);
3118 break;
3119 case 1:
3121 pre_expr op1 = get_or_alloc_expr_for (nary->op[0]);
3122 tree genop1 = find_or_generate_expression (block, op1,
3123 stmts, domstmt);
3124 if (!genop1)
3125 return NULL_TREE;
3126 genop1 = fold_convert (TREE_TYPE (nary->op[0]), genop1);
3128 folded = fold_build1 (nary->opcode, nary->type,
3129 genop1);
3131 break;
3132 default:
3133 return NULL_TREE;
3136 break;
3137 default:
3138 return NULL_TREE;
3141 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3142 folded = fold_convert (exprtype, folded);
3144 /* Force the generated expression to be a sequence of GIMPLE
3145 statements.
3146 We have to call unshare_expr because force_gimple_operand may
3147 modify the tree we pass to it. */
3148 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3149 false, NULL);
3151 /* If we have any intermediate expressions to the value sets, add them
3152 to the value sets and chain them in the instruction stream. */
3153 if (forced_stmts)
3155 gsi = gsi_start (forced_stmts);
3156 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3158 gimple stmt = gsi_stmt (gsi);
3159 tree forcedname = gimple_get_lhs (stmt);
3160 pre_expr nameexpr;
3162 if (TREE_CODE (forcedname) == SSA_NAME)
3164 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3165 VN_INFO_GET (forcedname)->valnum = forcedname;
3166 VN_INFO (forcedname)->value_id = get_next_value_id ();
3167 nameexpr = get_or_alloc_expr_for_name (forcedname);
3168 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3169 if (!in_fre)
3170 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3171 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3173 mark_symbols_for_renaming (stmt);
3175 gimple_seq_add_seq (stmts, forced_stmts);
3178 /* Build and insert the assignment of the end result to the temporary
3179 that we will return. */
3180 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3182 pretemp = create_tmp_reg (exprtype, "pretmp");
3183 get_var_ann (pretemp);
3186 temp = pretemp;
3187 add_referenced_var (temp);
3189 newstmt = gimple_build_assign (temp, folded);
3190 name = make_ssa_name (temp, newstmt);
3191 gimple_assign_set_lhs (newstmt, name);
3192 gimple_set_plf (newstmt, NECESSARY, false);
3194 gimple_seq_add_stmt (stmts, newstmt);
3195 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3197 /* All the symbols in NEWEXPR should be put into SSA form. */
3198 mark_symbols_for_renaming (newstmt);
3200 /* Add a value number to the temporary.
3201 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3202 we are creating the expression by pieces, and this particular piece of
3203 the expression may have been represented. There is no harm in replacing
3204 here. */
3205 VN_INFO_GET (name)->valnum = name;
3206 value_id = get_expr_value_id (expr);
3207 VN_INFO (name)->value_id = value_id;
3208 nameexpr = get_or_alloc_expr_for_name (name);
3209 add_to_value (value_id, nameexpr);
3210 if (NEW_SETS (block))
3211 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3212 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3214 pre_stats.insertions++;
3215 if (dump_file && (dump_flags & TDF_DETAILS))
3217 fprintf (dump_file, "Inserted ");
3218 print_gimple_stmt (dump_file, newstmt, 0, 0);
3219 fprintf (dump_file, " in predecessor %d\n", block->index);
3222 return name;
3226 /* Returns true if we want to inhibit the insertions of PHI nodes
3227 for the given EXPR for basic block BB (a member of a loop).
3228 We want to do this, when we fear that the induction variable we
3229 create might inhibit vectorization. */
3231 static bool
3232 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3234 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3235 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3236 vn_reference_op_t op;
3237 unsigned i;
3239 /* If we aren't going to vectorize we don't inhibit anything. */
3240 if (!flag_tree_vectorize)
3241 return false;
3243 /* Otherwise we inhibit the insertion when the address of the
3244 memory reference is a simple induction variable. In other
3245 cases the vectorizer won't do anything anyway (either it's
3246 loop invariant or a complicated expression). */
3247 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3249 switch (op->opcode)
3251 case ARRAY_REF:
3252 case ARRAY_RANGE_REF:
3253 if (TREE_CODE (op->op0) != SSA_NAME)
3254 break;
3255 /* Fallthru. */
3256 case SSA_NAME:
3258 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3259 affine_iv iv;
3260 /* Default defs are loop invariant. */
3261 if (!defbb)
3262 break;
3263 /* Defined outside this loop, also loop invariant. */
3264 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3265 break;
3266 /* If it's a simple induction variable inhibit insertion,
3267 the vectorizer might be interested in this one. */
3268 if (simple_iv (bb->loop_father, bb->loop_father,
3269 op->op0, &iv, true))
3270 return true;
3271 /* No simple IV, vectorizer can't do anything, hence no
3272 reason to inhibit the transformation for this operand. */
3273 break;
3275 default:
3276 break;
3279 return false;
3282 /* Insert the to-be-made-available values of expression EXPRNUM for each
3283 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3284 merge the result with a phi node, given the same value number as
3285 NODE. Return true if we have inserted new stuff. */
3287 static bool
3288 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3289 pre_expr *avail)
3291 pre_expr expr = expression_for_id (exprnum);
3292 pre_expr newphi;
3293 unsigned int val = get_expr_value_id (expr);
3294 edge pred;
3295 bool insertions = false;
3296 bool nophi = false;
3297 basic_block bprime;
3298 pre_expr eprime;
3299 edge_iterator ei;
3300 tree type = get_expr_type (expr);
3301 tree temp;
3302 gimple phi;
3304 if (dump_file && (dump_flags & TDF_DETAILS))
3306 fprintf (dump_file, "Found partial redundancy for expression ");
3307 print_pre_expr (dump_file, expr);
3308 fprintf (dump_file, " (%04d)\n", val);
3311 /* Make sure we aren't creating an induction variable. */
3312 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3314 bool firstinsideloop = false;
3315 bool secondinsideloop = false;
3316 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3317 EDGE_PRED (block, 0)->src);
3318 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3319 EDGE_PRED (block, 1)->src);
3320 /* Induction variables only have one edge inside the loop. */
3321 if ((firstinsideloop ^ secondinsideloop)
3322 && (expr->kind != REFERENCE
3323 || inhibit_phi_insertion (block, expr)))
3325 if (dump_file && (dump_flags & TDF_DETAILS))
3326 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3327 nophi = true;
3331 /* Make the necessary insertions. */
3332 FOR_EACH_EDGE (pred, ei, block->preds)
3334 gimple_seq stmts = NULL;
3335 tree builtexpr;
3336 bprime = pred->src;
3337 eprime = avail[bprime->index];
3339 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3341 builtexpr = create_expression_by_pieces (bprime,
3342 eprime,
3343 &stmts, NULL,
3344 type);
3345 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3346 gsi_insert_seq_on_edge (pred, stmts);
3347 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3348 insertions = true;
3350 else if (eprime->kind == CONSTANT)
3352 /* Constants may not have the right type, fold_convert
3353 should give us back a constant with the right type.
3355 tree constant = PRE_EXPR_CONSTANT (eprime);
3356 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3358 tree builtexpr = fold_convert (type, constant);
3359 if (!is_gimple_min_invariant (builtexpr))
3361 tree forcedexpr = force_gimple_operand (builtexpr,
3362 &stmts, true,
3363 NULL);
3364 if (!is_gimple_min_invariant (forcedexpr))
3366 if (forcedexpr != builtexpr)
3368 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3369 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3371 if (stmts)
3373 gimple_stmt_iterator gsi;
3374 gsi = gsi_start (stmts);
3375 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3377 gimple stmt = gsi_stmt (gsi);
3378 tree lhs = gimple_get_lhs (stmt);
3379 if (TREE_CODE (lhs) == SSA_NAME)
3380 bitmap_set_bit (inserted_exprs,
3381 SSA_NAME_VERSION (lhs));
3382 gimple_set_plf (stmt, NECESSARY, false);
3384 gsi_insert_seq_on_edge (pred, stmts);
3386 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3389 else
3390 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3393 else if (eprime->kind == NAME)
3395 /* We may have to do a conversion because our value
3396 numbering can look through types in certain cases, but
3397 our IL requires all operands of a phi node have the same
3398 type. */
3399 tree name = PRE_EXPR_NAME (eprime);
3400 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3402 tree builtexpr;
3403 tree forcedexpr;
3404 builtexpr = fold_convert (type, name);
3405 forcedexpr = force_gimple_operand (builtexpr,
3406 &stmts, true,
3407 NULL);
3409 if (forcedexpr != name)
3411 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3412 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3415 if (stmts)
3417 gimple_stmt_iterator gsi;
3418 gsi = gsi_start (stmts);
3419 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3421 gimple stmt = gsi_stmt (gsi);
3422 tree lhs = gimple_get_lhs (stmt);
3423 if (TREE_CODE (lhs) == SSA_NAME)
3424 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3425 gimple_set_plf (stmt, NECESSARY, false);
3427 gsi_insert_seq_on_edge (pred, stmts);
3429 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3433 /* If we didn't want a phi node, and we made insertions, we still have
3434 inserted new stuff, and thus return true. If we didn't want a phi node,
3435 and didn't make insertions, we haven't added anything new, so return
3436 false. */
3437 if (nophi && insertions)
3438 return true;
3439 else if (nophi && !insertions)
3440 return false;
3442 /* Now build a phi for the new variable. */
3443 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3445 prephitemp = create_tmp_var (type, "prephitmp");
3446 get_var_ann (prephitemp);
3449 temp = prephitemp;
3450 add_referenced_var (temp);
3452 if (TREE_CODE (type) == COMPLEX_TYPE
3453 || TREE_CODE (type) == VECTOR_TYPE)
3454 DECL_GIMPLE_REG_P (temp) = 1;
3455 phi = create_phi_node (temp, block);
3457 gimple_set_plf (phi, NECESSARY, false);
3458 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3459 VN_INFO (gimple_phi_result (phi))->value_id = val;
3460 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3461 FOR_EACH_EDGE (pred, ei, block->preds)
3463 pre_expr ae = avail[pred->src->index];
3464 gcc_assert (get_expr_type (ae) == type
3465 || useless_type_conversion_p (type, get_expr_type (ae)));
3466 if (ae->kind == CONSTANT)
3467 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3468 else
3469 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3470 UNKNOWN_LOCATION);
3473 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3474 add_to_value (val, newphi);
3476 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3477 this insertion, since we test for the existence of this value in PHI_GEN
3478 before proceeding with the partial redundancy checks in insert_aux.
3480 The value may exist in AVAIL_OUT, in particular, it could be represented
3481 by the expression we are trying to eliminate, in which case we want the
3482 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3483 inserted there.
3485 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3486 this block, because if it did, it would have existed in our dominator's
3487 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3490 bitmap_insert_into_set (PHI_GEN (block), newphi);
3491 bitmap_value_replace_in_set (AVAIL_OUT (block),
3492 newphi);
3493 bitmap_insert_into_set (NEW_SETS (block),
3494 newphi);
3496 if (dump_file && (dump_flags & TDF_DETAILS))
3498 fprintf (dump_file, "Created phi ");
3499 print_gimple_stmt (dump_file, phi, 0, 0);
3500 fprintf (dump_file, " in block %d\n", block->index);
3502 pre_stats.phis++;
3503 return true;
3508 /* Perform insertion of partially redundant values.
3509 For BLOCK, do the following:
3510 1. Propagate the NEW_SETS of the dominator into the current block.
3511 If the block has multiple predecessors,
3512 2a. Iterate over the ANTIC expressions for the block to see if
3513 any of them are partially redundant.
3514 2b. If so, insert them into the necessary predecessors to make
3515 the expression fully redundant.
3516 2c. Insert a new PHI merging the values of the predecessors.
3517 2d. Insert the new PHI, and the new expressions, into the
3518 NEW_SETS set.
3519 3. Recursively call ourselves on the dominator children of BLOCK.
3521 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3522 do_regular_insertion and do_partial_insertion.
3526 static bool
3527 do_regular_insertion (basic_block block, basic_block dom)
3529 bool new_stuff = false;
3530 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3531 pre_expr expr;
3532 int i;
3534 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3536 if (expr->kind != NAME)
3538 pre_expr *avail;
3539 unsigned int val;
3540 bool by_some = false;
3541 bool cant_insert = false;
3542 bool all_same = true;
3543 pre_expr first_s = NULL;
3544 edge pred;
3545 basic_block bprime;
3546 pre_expr eprime = NULL;
3547 edge_iterator ei;
3548 pre_expr edoubleprime = NULL;
3549 bool do_insertion = false;
3551 val = get_expr_value_id (expr);
3552 if (bitmap_set_contains_value (PHI_GEN (block), val))
3553 continue;
3554 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3556 if (dump_file && (dump_flags & TDF_DETAILS))
3557 fprintf (dump_file, "Found fully redundant value\n");
3558 continue;
3561 avail = XCNEWVEC (pre_expr, last_basic_block);
3562 FOR_EACH_EDGE (pred, ei, block->preds)
3564 unsigned int vprime;
3566 /* We should never run insertion for the exit block
3567 and so not come across fake pred edges. */
3568 gcc_assert (!(pred->flags & EDGE_FAKE));
3569 bprime = pred->src;
3570 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3571 bprime, block);
3573 /* eprime will generally only be NULL if the
3574 value of the expression, translated
3575 through the PHI for this predecessor, is
3576 undefined. If that is the case, we can't
3577 make the expression fully redundant,
3578 because its value is undefined along a
3579 predecessor path. We can thus break out
3580 early because it doesn't matter what the
3581 rest of the results are. */
3582 if (eprime == NULL)
3584 cant_insert = true;
3585 break;
3588 eprime = fully_constant_expression (eprime);
3589 vprime = get_expr_value_id (eprime);
3590 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3591 vprime, NULL);
3592 if (edoubleprime == NULL)
3594 avail[bprime->index] = eprime;
3595 all_same = false;
3597 else
3599 avail[bprime->index] = edoubleprime;
3600 by_some = true;
3601 /* We want to perform insertions to remove a redundancy on
3602 a path in the CFG we want to optimize for speed. */
3603 if (optimize_edge_for_speed_p (pred))
3604 do_insertion = true;
3605 if (first_s == NULL)
3606 first_s = edoubleprime;
3607 else if (!pre_expr_eq (first_s, edoubleprime))
3608 all_same = false;
3611 /* If we can insert it, it's not the same value
3612 already existing along every predecessor, and
3613 it's defined by some predecessor, it is
3614 partially redundant. */
3615 if (!cant_insert && !all_same && by_some)
3617 if (!do_insertion)
3619 if (dump_file && (dump_flags & TDF_DETAILS))
3621 fprintf (dump_file, "Skipping partial redundancy for "
3622 "expression ");
3623 print_pre_expr (dump_file, expr);
3624 fprintf (dump_file, " (%04d), no redundancy on to be "
3625 "optimized for speed edge\n", val);
3628 else if (dbg_cnt (treepre_insert)
3629 && insert_into_preds_of_block (block,
3630 get_expression_id (expr),
3631 avail))
3632 new_stuff = true;
3634 /* If all edges produce the same value and that value is
3635 an invariant, then the PHI has the same value on all
3636 edges. Note this. */
3637 else if (!cant_insert && all_same && eprime
3638 && (edoubleprime->kind == CONSTANT
3639 || edoubleprime->kind == NAME)
3640 && !value_id_constant_p (val))
3642 unsigned int j;
3643 bitmap_iterator bi;
3644 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3645 value_expressions, val);
3647 unsigned int new_val = get_expr_value_id (edoubleprime);
3648 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3650 pre_expr expr = expression_for_id (j);
3652 if (expr->kind == NAME)
3654 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3655 /* Just reset the value id and valnum so it is
3656 the same as the constant we have discovered. */
3657 if (edoubleprime->kind == CONSTANT)
3659 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3660 pre_stats.constified++;
3662 else
3663 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3664 info->value_id = new_val;
3668 free (avail);
3672 VEC_free (pre_expr, heap, exprs);
3673 return new_stuff;
3677 /* Perform insertion for partially anticipatable expressions. There
3678 is only one case we will perform insertion for these. This case is
3679 if the expression is partially anticipatable, and fully available.
3680 In this case, we know that putting it earlier will enable us to
3681 remove the later computation. */
3684 static bool
3685 do_partial_partial_insertion (basic_block block, basic_block dom)
3687 bool new_stuff = false;
3688 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3689 pre_expr expr;
3690 int i;
3692 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3694 if (expr->kind != NAME)
3696 pre_expr *avail;
3697 unsigned int val;
3698 bool by_all = true;
3699 bool cant_insert = false;
3700 edge pred;
3701 basic_block bprime;
3702 pre_expr eprime = NULL;
3703 edge_iterator ei;
3705 val = get_expr_value_id (expr);
3706 if (bitmap_set_contains_value (PHI_GEN (block), val))
3707 continue;
3708 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3709 continue;
3711 avail = XCNEWVEC (pre_expr, last_basic_block);
3712 FOR_EACH_EDGE (pred, ei, block->preds)
3714 unsigned int vprime;
3715 pre_expr edoubleprime;
3717 /* We should never run insertion for the exit block
3718 and so not come across fake pred edges. */
3719 gcc_assert (!(pred->flags & EDGE_FAKE));
3720 bprime = pred->src;
3721 eprime = phi_translate (expr, ANTIC_IN (block),
3722 PA_IN (block),
3723 bprime, block);
3725 /* eprime will generally only be NULL if the
3726 value of the expression, translated
3727 through the PHI for this predecessor, is
3728 undefined. If that is the case, we can't
3729 make the expression fully redundant,
3730 because its value is undefined along a
3731 predecessor path. We can thus break out
3732 early because it doesn't matter what the
3733 rest of the results are. */
3734 if (eprime == NULL)
3736 cant_insert = true;
3737 break;
3740 eprime = fully_constant_expression (eprime);
3741 vprime = get_expr_value_id (eprime);
3742 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3743 vprime, NULL);
3744 if (edoubleprime == NULL)
3746 by_all = false;
3747 break;
3749 else
3750 avail[bprime->index] = edoubleprime;
3754 /* If we can insert it, it's not the same value
3755 already existing along every predecessor, and
3756 it's defined by some predecessor, it is
3757 partially redundant. */
3758 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3760 pre_stats.pa_insert++;
3761 if (insert_into_preds_of_block (block, get_expression_id (expr),
3762 avail))
3763 new_stuff = true;
3765 free (avail);
3769 VEC_free (pre_expr, heap, exprs);
3770 return new_stuff;
3773 static bool
3774 insert_aux (basic_block block)
3776 basic_block son;
3777 bool new_stuff = false;
3779 if (block)
3781 basic_block dom;
3782 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3783 if (dom)
3785 unsigned i;
3786 bitmap_iterator bi;
3787 bitmap_set_t newset = NEW_SETS (dom);
3788 if (newset)
3790 /* Note that we need to value_replace both NEW_SETS, and
3791 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3792 represented by some non-simple expression here that we want
3793 to replace it with. */
3794 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3796 pre_expr expr = expression_for_id (i);
3797 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3798 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3801 if (!single_pred_p (block))
3803 new_stuff |= do_regular_insertion (block, dom);
3804 if (do_partial_partial)
3805 new_stuff |= do_partial_partial_insertion (block, dom);
3809 for (son = first_dom_son (CDI_DOMINATORS, block);
3810 son;
3811 son = next_dom_son (CDI_DOMINATORS, son))
3813 new_stuff |= insert_aux (son);
3816 return new_stuff;
3819 /* Perform insertion of partially redundant values. */
3821 static void
3822 insert (void)
3824 bool new_stuff = true;
3825 basic_block bb;
3826 int num_iterations = 0;
3828 FOR_ALL_BB (bb)
3829 NEW_SETS (bb) = bitmap_set_new ();
3831 while (new_stuff)
3833 num_iterations++;
3834 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3836 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3840 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3842 static void
3843 add_to_exp_gen (basic_block block, tree op)
3845 if (!in_fre)
3847 pre_expr result;
3848 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3849 return;
3850 result = get_or_alloc_expr_for_name (op);
3851 bitmap_value_insert_into_set (EXP_GEN (block), result);
3855 /* Create value ids for PHI in BLOCK. */
3857 static void
3858 make_values_for_phi (gimple phi, basic_block block)
3860 tree result = gimple_phi_result (phi);
3862 /* We have no need for virtual phis, as they don't represent
3863 actual computations. */
3864 if (is_gimple_reg (result))
3866 pre_expr e = get_or_alloc_expr_for_name (result);
3867 add_to_value (get_expr_value_id (e), e);
3868 bitmap_insert_into_set (PHI_GEN (block), e);
3869 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3870 if (!in_fre)
3872 unsigned i;
3873 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3875 tree arg = gimple_phi_arg_def (phi, i);
3876 if (TREE_CODE (arg) == SSA_NAME)
3878 e = get_or_alloc_expr_for_name (arg);
3879 add_to_value (get_expr_value_id (e), e);
3886 /* Compute the AVAIL set for all basic blocks.
3888 This function performs value numbering of the statements in each basic
3889 block. The AVAIL sets are built from information we glean while doing
3890 this value numbering, since the AVAIL sets contain only one entry per
3891 value.
3893 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3894 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3896 static void
3897 compute_avail (void)
3900 basic_block block, son;
3901 basic_block *worklist;
3902 size_t sp = 0;
3903 unsigned i;
3905 /* We pretend that default definitions are defined in the entry block.
3906 This includes function arguments and the static chain decl. */
3907 for (i = 1; i < num_ssa_names; ++i)
3909 tree name = ssa_name (i);
3910 pre_expr e;
3911 if (!name
3912 || !SSA_NAME_IS_DEFAULT_DEF (name)
3913 || has_zero_uses (name)
3914 || !is_gimple_reg (name))
3915 continue;
3917 e = get_or_alloc_expr_for_name (name);
3918 add_to_value (get_expr_value_id (e), e);
3919 if (!in_fre)
3920 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3921 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3924 /* Allocate the worklist. */
3925 worklist = XNEWVEC (basic_block, n_basic_blocks);
3927 /* Seed the algorithm by putting the dominator children of the entry
3928 block on the worklist. */
3929 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3930 son;
3931 son = next_dom_son (CDI_DOMINATORS, son))
3932 worklist[sp++] = son;
3934 /* Loop until the worklist is empty. */
3935 while (sp)
3937 gimple_stmt_iterator gsi;
3938 gimple stmt;
3939 basic_block dom;
3940 unsigned int stmt_uid = 1;
3942 /* Pick a block from the worklist. */
3943 block = worklist[--sp];
3945 /* Initially, the set of available values in BLOCK is that of
3946 its immediate dominator. */
3947 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3948 if (dom)
3949 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3951 /* Generate values for PHI nodes. */
3952 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3953 make_values_for_phi (gsi_stmt (gsi), block);
3955 BB_MAY_NOTRETURN (block) = 0;
3957 /* Now compute value numbers and populate value sets with all
3958 the expressions computed in BLOCK. */
3959 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3961 ssa_op_iter iter;
3962 tree op;
3964 stmt = gsi_stmt (gsi);
3965 gimple_set_uid (stmt, stmt_uid++);
3967 /* Cache whether the basic-block has any non-visible side-effect
3968 or control flow.
3969 If this isn't a call or it is the last stmt in the
3970 basic-block then the CFG represents things correctly. */
3971 if (is_gimple_call (stmt)
3972 && !stmt_ends_bb_p (stmt))
3974 /* Non-looping const functions always return normally.
3975 Otherwise the call might not return or have side-effects
3976 that forbids hoisting possibly trapping expressions
3977 before it. */
3978 int flags = gimple_call_flags (stmt);
3979 if (!(flags & ECF_CONST)
3980 || (flags & ECF_LOOPING_CONST_OR_PURE))
3981 BB_MAY_NOTRETURN (block) = 1;
3984 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3986 pre_expr e = get_or_alloc_expr_for_name (op);
3988 add_to_value (get_expr_value_id (e), e);
3989 if (!in_fre)
3990 bitmap_insert_into_set (TMP_GEN (block), e);
3991 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3994 if (gimple_has_volatile_ops (stmt)
3995 || stmt_could_throw_p (stmt))
3996 continue;
3998 switch (gimple_code (stmt))
4000 case GIMPLE_RETURN:
4001 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4002 add_to_exp_gen (block, op);
4003 continue;
4005 case GIMPLE_CALL:
4007 vn_reference_t ref;
4008 unsigned int i;
4009 vn_reference_op_t vro;
4010 pre_expr result = NULL;
4011 VEC(vn_reference_op_s, heap) *ops = NULL;
4013 if (!can_value_number_call (stmt))
4014 continue;
4016 copy_reference_ops_from_call (stmt, &ops);
4017 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4018 gimple_expr_type (stmt),
4019 ops, &ref, VN_NOWALK);
4020 VEC_free (vn_reference_op_s, heap, ops);
4021 if (!ref)
4022 continue;
4024 for (i = 0; VEC_iterate (vn_reference_op_s,
4025 ref->operands, i,
4026 vro); i++)
4028 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4029 add_to_exp_gen (block, vro->op0);
4030 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4031 add_to_exp_gen (block, vro->op1);
4032 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4033 add_to_exp_gen (block, vro->op2);
4035 result = (pre_expr) pool_alloc (pre_expr_pool);
4036 result->kind = REFERENCE;
4037 result->id = 0;
4038 PRE_EXPR_REFERENCE (result) = ref;
4040 get_or_alloc_expression_id (result);
4041 add_to_value (get_expr_value_id (result), result);
4042 if (!in_fre)
4043 bitmap_value_insert_into_set (EXP_GEN (block), result);
4044 continue;
4047 case GIMPLE_ASSIGN:
4049 pre_expr result = NULL;
4050 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4052 case tcc_unary:
4053 case tcc_binary:
4054 case tcc_comparison:
4056 vn_nary_op_t nary;
4057 unsigned int i;
4059 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4060 gimple_assign_rhs_code (stmt),
4061 gimple_expr_type (stmt),
4062 gimple_assign_rhs1 (stmt),
4063 gimple_assign_rhs2 (stmt),
4064 NULL_TREE, NULL_TREE, &nary);
4066 if (!nary)
4067 continue;
4069 for (i = 0; i < nary->length; i++)
4070 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4071 add_to_exp_gen (block, nary->op[i]);
4073 result = (pre_expr) pool_alloc (pre_expr_pool);
4074 result->kind = NARY;
4075 result->id = 0;
4076 PRE_EXPR_NARY (result) = nary;
4077 break;
4080 case tcc_declaration:
4081 case tcc_reference:
4083 vn_reference_t ref;
4084 unsigned int i;
4085 vn_reference_op_t vro;
4087 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4088 gimple_vuse (stmt),
4089 VN_WALK, &ref);
4090 if (!ref)
4091 continue;
4093 for (i = 0; VEC_iterate (vn_reference_op_s,
4094 ref->operands, i,
4095 vro); i++)
4097 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4098 add_to_exp_gen (block, vro->op0);
4099 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4100 add_to_exp_gen (block, vro->op1);
4101 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4102 add_to_exp_gen (block, vro->op2);
4104 result = (pre_expr) pool_alloc (pre_expr_pool);
4105 result->kind = REFERENCE;
4106 result->id = 0;
4107 PRE_EXPR_REFERENCE (result) = ref;
4108 break;
4111 default:
4112 /* For any other statement that we don't
4113 recognize, simply add all referenced
4114 SSA_NAMEs to EXP_GEN. */
4115 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4116 add_to_exp_gen (block, op);
4117 continue;
4120 get_or_alloc_expression_id (result);
4121 add_to_value (get_expr_value_id (result), result);
4122 if (!in_fre)
4123 bitmap_value_insert_into_set (EXP_GEN (block), result);
4125 continue;
4127 default:
4128 break;
4132 /* Put the dominator children of BLOCK on the worklist of blocks
4133 to compute available sets for. */
4134 for (son = first_dom_son (CDI_DOMINATORS, block);
4135 son;
4136 son = next_dom_son (CDI_DOMINATORS, son))
4137 worklist[sp++] = son;
4140 free (worklist);
4143 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4144 than the available expressions for it. The insertion point is
4145 right before the first use in STMT. Returns the SSA_NAME that should
4146 be used for replacement. */
4148 static tree
4149 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4151 basic_block bb = gimple_bb (stmt);
4152 gimple_stmt_iterator gsi;
4153 gimple_seq stmts = NULL;
4154 tree expr;
4155 pre_expr e;
4157 /* First create a value expression from the expression we want
4158 to insert and associate it with the value handle for SSA_VN. */
4159 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4160 if (e == NULL)
4161 return NULL_TREE;
4163 /* Then use create_expression_by_pieces to generate a valid
4164 expression to insert at this point of the IL stream. */
4165 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4166 if (expr == NULL_TREE)
4167 return NULL_TREE;
4168 gsi = gsi_for_stmt (stmt);
4169 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4171 return expr;
4174 /* Eliminate fully redundant computations. */
4176 static unsigned int
4177 eliminate (void)
4179 VEC (gimple, heap) *to_remove = NULL;
4180 VEC (gimple, heap) *to_update = NULL;
4181 basic_block b;
4182 unsigned int todo = 0;
4183 gimple_stmt_iterator gsi;
4184 gimple stmt;
4185 unsigned i;
4187 FOR_EACH_BB (b)
4189 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4191 stmt = gsi_stmt (gsi);
4193 /* Lookup the RHS of the expression, see if we have an
4194 available computation for it. If so, replace the RHS with
4195 the available computation. */
4196 if (gimple_has_lhs (stmt)
4197 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4198 && !gimple_assign_ssa_name_copy_p (stmt)
4199 && (!gimple_assign_single_p (stmt)
4200 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4201 && !gimple_has_volatile_ops (stmt)
4202 && !has_zero_uses (gimple_get_lhs (stmt)))
4204 tree lhs = gimple_get_lhs (stmt);
4205 tree rhs = NULL_TREE;
4206 tree sprime = NULL;
4207 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4208 pre_expr sprimeexpr;
4210 if (gimple_assign_single_p (stmt))
4211 rhs = gimple_assign_rhs1 (stmt);
4213 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4214 get_expr_value_id (lhsexpr),
4215 NULL);
4217 if (sprimeexpr)
4219 if (sprimeexpr->kind == CONSTANT)
4220 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4221 else if (sprimeexpr->kind == NAME)
4222 sprime = PRE_EXPR_NAME (sprimeexpr);
4223 else
4224 gcc_unreachable ();
4227 /* If there is no existing leader but SCCVN knows this
4228 value is constant, use that constant. */
4229 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4231 sprime = VN_INFO (lhs)->valnum;
4232 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4233 TREE_TYPE (sprime)))
4234 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4236 if (dump_file && (dump_flags & TDF_DETAILS))
4238 fprintf (dump_file, "Replaced ");
4239 print_gimple_expr (dump_file, stmt, 0, 0);
4240 fprintf (dump_file, " with ");
4241 print_generic_expr (dump_file, sprime, 0);
4242 fprintf (dump_file, " in ");
4243 print_gimple_stmt (dump_file, stmt, 0, 0);
4245 pre_stats.eliminations++;
4246 propagate_tree_value_into_stmt (&gsi, sprime);
4247 stmt = gsi_stmt (gsi);
4248 update_stmt (stmt);
4249 continue;
4252 /* If there is no existing usable leader but SCCVN thinks
4253 it has an expression it wants to use as replacement,
4254 insert that. */
4255 if (!sprime || sprime == lhs)
4257 tree val = VN_INFO (lhs)->valnum;
4258 if (val != VN_TOP
4259 && TREE_CODE (val) == SSA_NAME
4260 && VN_INFO (val)->needs_insertion
4261 && can_PRE_operation (vn_get_expr_for (val)))
4262 sprime = do_SCCVN_insertion (stmt, val);
4264 if (sprime
4265 && sprime != lhs
4266 && (rhs == NULL_TREE
4267 || TREE_CODE (rhs) != SSA_NAME
4268 || may_propagate_copy (rhs, sprime)))
4270 bool can_make_abnormal_goto
4271 = is_gimple_call (stmt)
4272 && stmt_can_make_abnormal_goto (stmt);
4274 gcc_assert (sprime != rhs);
4276 if (dump_file && (dump_flags & TDF_DETAILS))
4278 fprintf (dump_file, "Replaced ");
4279 print_gimple_expr (dump_file, stmt, 0, 0);
4280 fprintf (dump_file, " with ");
4281 print_generic_expr (dump_file, sprime, 0);
4282 fprintf (dump_file, " in ");
4283 print_gimple_stmt (dump_file, stmt, 0, 0);
4286 if (TREE_CODE (sprime) == SSA_NAME)
4287 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4288 NECESSARY, true);
4289 /* We need to make sure the new and old types actually match,
4290 which may require adding a simple cast, which fold_convert
4291 will do for us. */
4292 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4293 && !useless_type_conversion_p (gimple_expr_type (stmt),
4294 TREE_TYPE (sprime)))
4295 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4297 pre_stats.eliminations++;
4298 propagate_tree_value_into_stmt (&gsi, sprime);
4299 stmt = gsi_stmt (gsi);
4300 update_stmt (stmt);
4302 /* If we removed EH side-effects from the statement, clean
4303 its EH information. */
4304 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4306 bitmap_set_bit (need_eh_cleanup,
4307 gimple_bb (stmt)->index);
4308 if (dump_file && (dump_flags & TDF_DETAILS))
4309 fprintf (dump_file, " Removed EH side-effects.\n");
4312 /* Likewise for AB side-effects. */
4313 if (can_make_abnormal_goto
4314 && !stmt_can_make_abnormal_goto (stmt))
4316 bitmap_set_bit (need_ab_cleanup,
4317 gimple_bb (stmt)->index);
4318 if (dump_file && (dump_flags & TDF_DETAILS))
4319 fprintf (dump_file, " Removed AB side-effects.\n");
4323 /* If the statement is a scalar store, see if the expression
4324 has the same value number as its rhs. If so, the store is
4325 dead. */
4326 else if (gimple_assign_single_p (stmt)
4327 && !is_gimple_reg (gimple_assign_lhs (stmt))
4328 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4329 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4331 tree rhs = gimple_assign_rhs1 (stmt);
4332 tree val;
4333 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4334 gimple_vuse (stmt), VN_WALK, NULL);
4335 if (TREE_CODE (rhs) == SSA_NAME)
4336 rhs = VN_INFO (rhs)->valnum;
4337 if (val
4338 && operand_equal_p (val, rhs, 0))
4340 if (dump_file && (dump_flags & TDF_DETAILS))
4342 fprintf (dump_file, "Deleted redundant store ");
4343 print_gimple_stmt (dump_file, stmt, 0, 0);
4346 /* Queue stmt for removal. */
4347 VEC_safe_push (gimple, heap, to_remove, stmt);
4350 /* Visit COND_EXPRs and fold the comparison with the
4351 available value-numbers. */
4352 else if (gimple_code (stmt) == GIMPLE_COND)
4354 tree op0 = gimple_cond_lhs (stmt);
4355 tree op1 = gimple_cond_rhs (stmt);
4356 tree result;
4358 if (TREE_CODE (op0) == SSA_NAME)
4359 op0 = VN_INFO (op0)->valnum;
4360 if (TREE_CODE (op1) == SSA_NAME)
4361 op1 = VN_INFO (op1)->valnum;
4362 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4363 op0, op1);
4364 if (result && TREE_CODE (result) == INTEGER_CST)
4366 if (integer_zerop (result))
4367 gimple_cond_make_false (stmt);
4368 else
4369 gimple_cond_make_true (stmt);
4370 update_stmt (stmt);
4371 todo = TODO_cleanup_cfg;
4374 /* Visit indirect calls and turn them into direct calls if
4375 possible. */
4376 if (is_gimple_call (stmt))
4378 tree orig_fn = gimple_call_fn (stmt);
4379 tree fn;
4380 if (!orig_fn)
4381 continue;
4382 if (TREE_CODE (orig_fn) == SSA_NAME)
4383 fn = VN_INFO (orig_fn)->valnum;
4384 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4385 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4386 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4387 else
4388 continue;
4389 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4390 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4391 TREE_TYPE (fn)))
4393 bool can_make_abnormal_goto
4394 = stmt_can_make_abnormal_goto (stmt);
4395 bool was_noreturn = gimple_call_noreturn_p (stmt);
4397 if (dump_file && (dump_flags & TDF_DETAILS))
4399 fprintf (dump_file, "Replacing call target with ");
4400 print_generic_expr (dump_file, fn, 0);
4401 fprintf (dump_file, " in ");
4402 print_gimple_stmt (dump_file, stmt, 0, 0);
4405 gimple_call_set_fn (stmt, fn);
4406 VEC_safe_push (gimple, heap, to_update, stmt);
4408 /* When changing a call into a noreturn call, cfg cleanup
4409 is needed to fix up the noreturn call. */
4410 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4411 todo |= TODO_cleanup_cfg;
4413 /* If we removed EH side-effects from the statement, clean
4414 its EH information. */
4415 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4417 bitmap_set_bit (need_eh_cleanup,
4418 gimple_bb (stmt)->index);
4419 if (dump_file && (dump_flags & TDF_DETAILS))
4420 fprintf (dump_file, " Removed EH side-effects.\n");
4423 /* Likewise for AB side-effects. */
4424 if (can_make_abnormal_goto
4425 && !stmt_can_make_abnormal_goto (stmt))
4427 bitmap_set_bit (need_ab_cleanup,
4428 gimple_bb (stmt)->index);
4429 if (dump_file && (dump_flags & TDF_DETAILS))
4430 fprintf (dump_file, " Removed AB side-effects.\n");
4433 /* Changing an indirect call to a direct call may
4434 have exposed different semantics. This may
4435 require an SSA update. */
4436 todo |= TODO_update_ssa_only_virtuals;
4441 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4443 gimple stmt, phi = gsi_stmt (gsi);
4444 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4445 pre_expr sprimeexpr, resexpr;
4446 gimple_stmt_iterator gsi2;
4448 /* We want to perform redundant PHI elimination. Do so by
4449 replacing the PHI with a single copy if possible.
4450 Do not touch inserted, single-argument or virtual PHIs. */
4451 if (gimple_phi_num_args (phi) == 1
4452 || !is_gimple_reg (res))
4454 gsi_next (&gsi);
4455 continue;
4458 resexpr = get_or_alloc_expr_for_name (res);
4459 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4460 get_expr_value_id (resexpr), NULL);
4461 if (sprimeexpr)
4463 if (sprimeexpr->kind == CONSTANT)
4464 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4465 else if (sprimeexpr->kind == NAME)
4466 sprime = PRE_EXPR_NAME (sprimeexpr);
4467 else
4468 gcc_unreachable ();
4470 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4472 sprime = VN_INFO (res)->valnum;
4473 if (!useless_type_conversion_p (TREE_TYPE (res),
4474 TREE_TYPE (sprime)))
4475 sprime = fold_convert (TREE_TYPE (res), sprime);
4477 if (!sprime
4478 || sprime == res)
4480 gsi_next (&gsi);
4481 continue;
4484 if (dump_file && (dump_flags & TDF_DETAILS))
4486 fprintf (dump_file, "Replaced redundant PHI node defining ");
4487 print_generic_expr (dump_file, res, 0);
4488 fprintf (dump_file, " with ");
4489 print_generic_expr (dump_file, sprime, 0);
4490 fprintf (dump_file, "\n");
4493 remove_phi_node (&gsi, false);
4495 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4496 && TREE_CODE (sprime) == SSA_NAME)
4497 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4499 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4500 sprime = fold_convert (TREE_TYPE (res), sprime);
4501 stmt = gimple_build_assign (res, sprime);
4502 SSA_NAME_DEF_STMT (res) = stmt;
4503 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4505 gsi2 = gsi_after_labels (b);
4506 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4507 /* Queue the copy for eventual removal. */
4508 VEC_safe_push (gimple, heap, to_remove, stmt);
4509 /* If we inserted this PHI node ourself, it's not an elimination. */
4510 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4511 pre_stats.phis--;
4512 else
4513 pre_stats.eliminations++;
4517 /* We cannot remove stmts during BB walk, especially not release SSA
4518 names there as this confuses the VN machinery. The stmts ending
4519 up in to_remove are either stores or simple copies. */
4520 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4522 tree lhs = gimple_assign_lhs (stmt);
4523 tree rhs = gimple_assign_rhs1 (stmt);
4524 use_operand_p use_p;
4525 gimple use_stmt;
4527 /* If there is a single use only, propagate the equivalency
4528 instead of keeping the copy. */
4529 if (TREE_CODE (lhs) == SSA_NAME
4530 && TREE_CODE (rhs) == SSA_NAME
4531 && single_imm_use (lhs, &use_p, &use_stmt)
4532 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4534 SET_USE (use_p, rhs);
4535 update_stmt (use_stmt);
4536 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4537 && TREE_CODE (rhs) == SSA_NAME)
4538 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4541 /* If this is a store or a now unused copy, remove it. */
4542 if (TREE_CODE (lhs) != SSA_NAME
4543 || has_zero_uses (lhs))
4545 basic_block bb = gimple_bb (stmt);
4546 gsi = gsi_for_stmt (stmt);
4547 unlink_stmt_vdef (stmt);
4548 gsi_remove (&gsi, true);
4549 if (gimple_purge_dead_eh_edges (bb))
4550 todo |= TODO_cleanup_cfg;
4551 if (TREE_CODE (lhs) == SSA_NAME)
4552 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4553 release_defs (stmt);
4556 VEC_free (gimple, heap, to_remove);
4558 /* We cannot update call statements with virtual operands during
4559 SSA walk. This might remove them which in turn makes our
4560 VN lattice invalid. */
4561 FOR_EACH_VEC_ELT (gimple, to_update, i, stmt)
4562 update_stmt (stmt);
4563 VEC_free (gimple, heap, to_update);
4565 return todo;
4568 /* Borrow a bit of tree-ssa-dce.c for the moment.
4569 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4570 this may be a bit faster, and we may want critical edges kept split. */
4572 /* If OP's defining statement has not already been determined to be necessary,
4573 mark that statement necessary. Return the stmt, if it is newly
4574 necessary. */
4576 static inline gimple
4577 mark_operand_necessary (tree op)
4579 gimple stmt;
4581 gcc_assert (op);
4583 if (TREE_CODE (op) != SSA_NAME)
4584 return NULL;
4586 stmt = SSA_NAME_DEF_STMT (op);
4587 gcc_assert (stmt);
4589 if (gimple_plf (stmt, NECESSARY)
4590 || gimple_nop_p (stmt))
4591 return NULL;
4593 gimple_set_plf (stmt, NECESSARY, true);
4594 return stmt;
4597 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4598 to insert PHI nodes sometimes, and because value numbering of casts isn't
4599 perfect, we sometimes end up inserting dead code. This simple DCE-like
4600 pass removes any insertions we made that weren't actually used. */
4602 static void
4603 remove_dead_inserted_code (void)
4605 bitmap worklist;
4606 unsigned i;
4607 bitmap_iterator bi;
4608 gimple t;
4610 worklist = BITMAP_ALLOC (NULL);
4611 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4613 t = SSA_NAME_DEF_STMT (ssa_name (i));
4614 if (gimple_plf (t, NECESSARY))
4615 bitmap_set_bit (worklist, i);
4617 while (!bitmap_empty_p (worklist))
4619 i = bitmap_first_set_bit (worklist);
4620 bitmap_clear_bit (worklist, i);
4621 t = SSA_NAME_DEF_STMT (ssa_name (i));
4623 /* PHI nodes are somewhat special in that each PHI alternative has
4624 data and control dependencies. All the statements feeding the
4625 PHI node's arguments are always necessary. */
4626 if (gimple_code (t) == GIMPLE_PHI)
4628 unsigned k;
4630 for (k = 0; k < gimple_phi_num_args (t); k++)
4632 tree arg = PHI_ARG_DEF (t, k);
4633 if (TREE_CODE (arg) == SSA_NAME)
4635 gimple n = mark_operand_necessary (arg);
4636 if (n)
4637 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4641 else
4643 /* Propagate through the operands. Examine all the USE, VUSE and
4644 VDEF operands in this statement. Mark all the statements
4645 which feed this statement's uses as necessary. */
4646 ssa_op_iter iter;
4647 tree use;
4649 /* The operands of VDEF expressions are also needed as they
4650 represent potential definitions that may reach this
4651 statement (VDEF operands allow us to follow def-def
4652 links). */
4654 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4656 gimple n = mark_operand_necessary (use);
4657 if (n)
4658 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4663 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4665 t = SSA_NAME_DEF_STMT (ssa_name (i));
4666 if (!gimple_plf (t, NECESSARY))
4668 gimple_stmt_iterator gsi;
4670 if (dump_file && (dump_flags & TDF_DETAILS))
4672 fprintf (dump_file, "Removing unnecessary insertion:");
4673 print_gimple_stmt (dump_file, t, 0, 0);
4676 gsi = gsi_for_stmt (t);
4677 if (gimple_code (t) == GIMPLE_PHI)
4678 remove_phi_node (&gsi, true);
4679 else
4681 gsi_remove (&gsi, true);
4682 release_defs (t);
4686 BITMAP_FREE (worklist);
4689 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4690 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4691 the number of visited blocks. */
4693 static int
4694 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4696 edge_iterator *stack;
4697 int sp;
4698 int post_order_num = 0;
4699 sbitmap visited;
4701 if (include_entry_exit)
4702 post_order[post_order_num++] = EXIT_BLOCK;
4704 /* Allocate stack for back-tracking up CFG. */
4705 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4706 sp = 0;
4708 /* Allocate bitmap to track nodes that have been visited. */
4709 visited = sbitmap_alloc (last_basic_block);
4711 /* None of the nodes in the CFG have been visited yet. */
4712 sbitmap_zero (visited);
4714 /* Push the last edge on to the stack. */
4715 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4717 while (sp)
4719 edge_iterator ei;
4720 basic_block src;
4721 basic_block dest;
4723 /* Look at the edge on the top of the stack. */
4724 ei = stack[sp - 1];
4725 src = ei_edge (ei)->src;
4726 dest = ei_edge (ei)->dest;
4728 /* Check if the edge destination has been visited yet. */
4729 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4731 /* Mark that we have visited the destination. */
4732 SET_BIT (visited, src->index);
4734 if (EDGE_COUNT (src->preds) > 0)
4735 /* Since the DEST node has been visited for the first
4736 time, check its successors. */
4737 stack[sp++] = ei_start (src->preds);
4738 else
4739 post_order[post_order_num++] = src->index;
4741 else
4743 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4744 post_order[post_order_num++] = dest->index;
4746 if (!ei_one_before_end_p (ei))
4747 ei_next (&stack[sp - 1]);
4748 else
4749 sp--;
4753 if (include_entry_exit)
4754 post_order[post_order_num++] = ENTRY_BLOCK;
4756 free (stack);
4757 sbitmap_free (visited);
4758 return post_order_num;
4762 /* Initialize data structures used by PRE. */
4764 static void
4765 init_pre (bool do_fre)
4767 basic_block bb;
4769 next_expression_id = 1;
4770 expressions = NULL;
4771 VEC_safe_push (pre_expr, heap, expressions, NULL);
4772 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4773 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4774 get_max_value_id() + 1);
4775 name_to_id = NULL;
4777 in_fre = do_fre;
4779 inserted_exprs = BITMAP_ALLOC (NULL);
4780 need_creation = NULL;
4781 pretemp = NULL_TREE;
4782 storetemp = NULL_TREE;
4783 prephitemp = NULL_TREE;
4785 connect_infinite_loops_to_exit ();
4786 memset (&pre_stats, 0, sizeof (pre_stats));
4789 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4790 my_rev_post_order_compute (postorder, false);
4792 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4794 calculate_dominance_info (CDI_POST_DOMINATORS);
4795 calculate_dominance_info (CDI_DOMINATORS);
4797 bitmap_obstack_initialize (&grand_bitmap_obstack);
4798 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4799 expr_pred_trans_eq, free);
4800 expression_to_id = htab_create (num_ssa_names * 3,
4801 pre_expr_hash,
4802 pre_expr_eq, NULL);
4803 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4804 sizeof (struct bitmap_set), 30);
4805 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4806 sizeof (struct pre_expr_d), 30);
4807 FOR_ALL_BB (bb)
4809 EXP_GEN (bb) = bitmap_set_new ();
4810 PHI_GEN (bb) = bitmap_set_new ();
4811 TMP_GEN (bb) = bitmap_set_new ();
4812 AVAIL_OUT (bb) = bitmap_set_new ();
4815 need_eh_cleanup = BITMAP_ALLOC (NULL);
4816 need_ab_cleanup = BITMAP_ALLOC (NULL);
4820 /* Deallocate data structures used by PRE. */
4822 static void
4823 fini_pre (bool do_fre)
4825 free (postorder);
4826 VEC_free (bitmap_set_t, heap, value_expressions);
4827 BITMAP_FREE (inserted_exprs);
4828 VEC_free (gimple, heap, need_creation);
4829 bitmap_obstack_release (&grand_bitmap_obstack);
4830 free_alloc_pool (bitmap_set_pool);
4831 free_alloc_pool (pre_expr_pool);
4832 htab_delete (phi_translate_table);
4833 htab_delete (expression_to_id);
4834 VEC_free (unsigned, heap, name_to_id);
4836 free_aux_for_blocks ();
4838 free_dominance_info (CDI_POST_DOMINATORS);
4840 if (!bitmap_empty_p (need_eh_cleanup))
4842 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4843 cleanup_tree_cfg ();
4846 BITMAP_FREE (need_eh_cleanup);
4848 if (!bitmap_empty_p (need_ab_cleanup))
4850 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4851 cleanup_tree_cfg ();
4854 BITMAP_FREE (need_ab_cleanup);
4856 if (!do_fre)
4857 loop_optimizer_finalize ();
4860 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4861 only wants to do full redundancy elimination. */
4863 static unsigned int
4864 execute_pre (bool do_fre)
4866 unsigned int todo = 0;
4868 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4870 /* This has to happen before SCCVN runs because
4871 loop_optimizer_init may create new phis, etc. */
4872 if (!do_fre)
4873 loop_optimizer_init (LOOPS_NORMAL);
4875 if (!run_scc_vn (do_fre ? VN_WALKREWRITE : VN_WALK))
4877 if (!do_fre)
4878 loop_optimizer_finalize ();
4880 return 0;
4883 init_pre (do_fre);
4884 scev_initialize ();
4886 /* Collect and value number expressions computed in each basic block. */
4887 compute_avail ();
4889 if (dump_file && (dump_flags & TDF_DETAILS))
4891 basic_block bb;
4893 FOR_ALL_BB (bb)
4895 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4896 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4897 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4898 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4902 /* Insert can get quite slow on an incredibly large number of basic
4903 blocks due to some quadratic behavior. Until this behavior is
4904 fixed, don't run it when he have an incredibly large number of
4905 bb's. If we aren't going to run insert, there is no point in
4906 computing ANTIC, either, even though it's plenty fast. */
4907 if (!do_fre && n_basic_blocks < 4000)
4909 compute_antic ();
4910 insert ();
4913 /* Make sure to remove fake edges before committing our inserts.
4914 This makes sure we don't end up with extra critical edges that
4915 we would need to split. */
4916 remove_fake_exit_edges ();
4917 gsi_commit_edge_inserts ();
4919 /* Remove all the redundant expressions. */
4920 todo |= eliminate ();
4922 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4923 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4924 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4925 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4926 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4928 clear_expression_ids ();
4929 free_scc_vn ();
4930 if (!do_fre)
4932 remove_dead_inserted_code ();
4933 todo |= TODO_verify_flow;
4936 scev_finalize ();
4937 fini_pre (do_fre);
4939 return todo;
4942 /* Gate and execute functions for PRE. */
4944 static unsigned int
4945 do_pre (void)
4947 return execute_pre (false);
4950 static bool
4951 gate_pre (void)
4953 return flag_tree_pre != 0;
4956 struct gimple_opt_pass pass_pre =
4959 GIMPLE_PASS,
4960 "pre", /* name */
4961 gate_pre, /* gate */
4962 do_pre, /* execute */
4963 NULL, /* sub */
4964 NULL, /* next */
4965 0, /* static_pass_number */
4966 TV_TREE_PRE, /* tv_id */
4967 PROP_no_crit_edges | PROP_cfg
4968 | PROP_ssa, /* properties_required */
4969 0, /* properties_provided */
4970 0, /* properties_destroyed */
4971 TODO_rebuild_alias, /* todo_flags_start */
4972 TODO_update_ssa_only_virtuals | TODO_ggc_collect
4973 | TODO_verify_ssa /* todo_flags_finish */
4978 /* Gate and execute functions for FRE. */
4980 static unsigned int
4981 execute_fre (void)
4983 return execute_pre (true);
4986 static bool
4987 gate_fre (void)
4989 return flag_tree_fre != 0;
4992 struct gimple_opt_pass pass_fre =
4995 GIMPLE_PASS,
4996 "fre", /* name */
4997 gate_fre, /* gate */
4998 execute_fre, /* execute */
4999 NULL, /* sub */
5000 NULL, /* next */
5001 0, /* static_pass_number */
5002 TV_TREE_FRE, /* tv_id */
5003 PROP_cfg | PROP_ssa, /* properties_required */
5004 0, /* properties_provided */
5005 0, /* properties_destroyed */
5006 0, /* todo_flags_start */
5007 TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */