2011-11-04 Tom de Vries <tom@codesourcery.com>
[official-gcc.git] / gcc / tree-ssa-pre.c
blob60ae35cef73164e8969d96efc6a48fa6cf721647
1 /* SSA-PRE for trees.
2 Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
5 <stevenb@suse.de>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify
10 it under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
12 any later version.
14 GCC is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 #include "config.h"
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "tree.h"
28 #include "basic-block.h"
29 #include "tree-pretty-print.h"
30 #include "gimple-pretty-print.h"
31 #include "tree-inline.h"
32 #include "tree-flow.h"
33 #include "gimple.h"
34 #include "tree-dump.h"
35 #include "timevar.h"
36 #include "fibheap.h"
37 #include "hashtab.h"
38 #include "tree-iterator.h"
39 #include "alloc-pool.h"
40 #include "obstack.h"
41 #include "tree-pass.h"
42 #include "flags.h"
43 #include "bitmap.h"
44 #include "langhooks.h"
45 #include "cfgloop.h"
46 #include "tree-ssa-sccvn.h"
47 #include "tree-scalar-evolution.h"
48 #include "params.h"
49 #include "dbgcnt.h"
51 /* TODO:
53 1. Avail sets can be shared by making an avail_find_leader that
54 walks up the dominator tree and looks in those avail sets.
55 This might affect code optimality, it's unclear right now.
56 2. Strength reduction can be performed by anticipating expressions
57 we can repair later on.
58 3. We can do back-substitution or smarter value numbering to catch
59 commutative expressions split up over multiple statements.
62 /* For ease of terminology, "expression node" in the below refers to
63 every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
64 represent the actual statement containing the expressions we care about,
65 and we cache the value number by putting it in the expression. */
67 /* Basic algorithm
69 First we walk the statements to generate the AVAIL sets, the
70 EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
71 generation of values/expressions by a given block. We use them
72 when computing the ANTIC sets. The AVAIL sets consist of
73 SSA_NAME's that represent values, so we know what values are
74 available in what blocks. AVAIL is a forward dataflow problem. In
75 SSA, values are never killed, so we don't need a kill set, or a
76 fixpoint iteration, in order to calculate the AVAIL sets. In
77 traditional parlance, AVAIL sets tell us the downsafety of the
78 expressions/values.
80 Next, we generate the ANTIC sets. These sets represent the
81 anticipatable expressions. ANTIC is a backwards dataflow
82 problem. An expression is anticipatable in a given block if it could
83 be generated in that block. This means that if we had to perform
84 an insertion in that block, of the value of that expression, we
85 could. Calculating the ANTIC sets requires phi translation of
86 expressions, because the flow goes backwards through phis. We must
87 iterate to a fixpoint of the ANTIC sets, because we have a kill
88 set. Even in SSA form, values are not live over the entire
89 function, only from their definition point onwards. So we have to
90 remove values from the ANTIC set once we go past the definition
91 point of the leaders that make them up.
92 compute_antic/compute_antic_aux performs this computation.
94 Third, we perform insertions to make partially redundant
95 expressions fully redundant.
97 An expression is partially redundant (excluding partial
98 anticipation) if:
100 1. It is AVAIL in some, but not all, of the predecessors of a
101 given block.
102 2. It is ANTIC in all the predecessors.
104 In order to make it fully redundant, we insert the expression into
105 the predecessors where it is not available, but is ANTIC.
107 For the partial anticipation case, we only perform insertion if it
108 is partially anticipated in some block, and fully available in all
109 of the predecessors.
111 insert/insert_aux/do_regular_insertion/do_partial_partial_insertion
112 performs these steps.
114 Fourth, we eliminate fully redundant expressions.
115 This is a simple statement walk that replaces redundant
116 calculations with the now available values. */
118 /* Representations of value numbers:
120 Value numbers are represented by a representative SSA_NAME. We
121 will create fake SSA_NAME's in situations where we need a
122 representative but do not have one (because it is a complex
123 expression). In order to facilitate storing the value numbers in
124 bitmaps, and keep the number of wasted SSA_NAME's down, we also
125 associate a value_id with each value number, and create full blown
126 ssa_name's only where we actually need them (IE in operands of
127 existing expressions).
129 Theoretically you could replace all the value_id's with
130 SSA_NAME_VERSION, but this would allocate a large number of
131 SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
132 It would also require an additional indirection at each point we
133 use the value id. */
135 /* Representation of expressions on value numbers:
137 Expressions consisting of value numbers are represented the same
138 way as our VN internally represents them, with an additional
139 "pre_expr" wrapping around them in order to facilitate storing all
140 of the expressions in the same sets. */
142 /* Representation of sets:
144 The dataflow sets do not need to be sorted in any particular order
145 for the majority of their lifetime, are simply represented as two
146 bitmaps, one that keeps track of values present in the set, and one
147 that keeps track of expressions present in the set.
149 When we need them in topological order, we produce it on demand by
150 transforming the bitmap into an array and sorting it into topo
151 order. */
153 /* Type of expression, used to know which member of the PRE_EXPR union
154 is valid. */
156 enum pre_expr_kind
158 NAME,
159 NARY,
160 REFERENCE,
161 CONSTANT
164 typedef union pre_expr_union_d
166 tree name;
167 tree constant;
168 vn_nary_op_t nary;
169 vn_reference_t reference;
170 } pre_expr_union;
172 typedef struct pre_expr_d
174 enum pre_expr_kind kind;
175 unsigned int id;
176 pre_expr_union u;
177 } *pre_expr;
179 #define PRE_EXPR_NAME(e) (e)->u.name
180 #define PRE_EXPR_NARY(e) (e)->u.nary
181 #define PRE_EXPR_REFERENCE(e) (e)->u.reference
182 #define PRE_EXPR_CONSTANT(e) (e)->u.constant
184 static int
185 pre_expr_eq (const void *p1, const void *p2)
187 const struct pre_expr_d *e1 = (const struct pre_expr_d *) p1;
188 const struct pre_expr_d *e2 = (const struct pre_expr_d *) p2;
190 if (e1->kind != e2->kind)
191 return false;
193 switch (e1->kind)
195 case CONSTANT:
196 return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
197 PRE_EXPR_CONSTANT (e2));
198 case NAME:
199 return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
200 case NARY:
201 return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
202 case REFERENCE:
203 return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
204 PRE_EXPR_REFERENCE (e2));
205 default:
206 gcc_unreachable ();
210 static hashval_t
211 pre_expr_hash (const void *p1)
213 const struct pre_expr_d *e = (const struct pre_expr_d *) p1;
214 switch (e->kind)
216 case CONSTANT:
217 return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
218 case NAME:
219 return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
220 case NARY:
221 return PRE_EXPR_NARY (e)->hashcode;
222 case REFERENCE:
223 return PRE_EXPR_REFERENCE (e)->hashcode;
224 default:
225 gcc_unreachable ();
230 /* Next global expression id number. */
231 static unsigned int next_expression_id;
233 /* Mapping from expression to id number we can use in bitmap sets. */
234 DEF_VEC_P (pre_expr);
235 DEF_VEC_ALLOC_P (pre_expr, heap);
236 static VEC(pre_expr, heap) *expressions;
237 static htab_t expression_to_id;
238 static VEC(unsigned, heap) *name_to_id;
240 /* Allocate an expression id for EXPR. */
242 static inline unsigned int
243 alloc_expression_id (pre_expr expr)
245 void **slot;
246 /* Make sure we won't overflow. */
247 gcc_assert (next_expression_id + 1 > next_expression_id);
248 expr->id = next_expression_id++;
249 VEC_safe_push (pre_expr, heap, expressions, expr);
250 if (expr->kind == NAME)
252 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
253 /* VEC_safe_grow_cleared allocates no headroom. Avoid frequent
254 re-allocations by using VEC_reserve upfront. There is no
255 VEC_quick_grow_cleared unfortunately. */
256 VEC_reserve (unsigned, heap, name_to_id, num_ssa_names);
257 VEC_safe_grow_cleared (unsigned, heap, name_to_id, num_ssa_names);
258 gcc_assert (VEC_index (unsigned, name_to_id, version) == 0);
259 VEC_replace (unsigned, name_to_id, version, expr->id);
261 else
263 slot = htab_find_slot (expression_to_id, expr, INSERT);
264 gcc_assert (!*slot);
265 *slot = expr;
267 return next_expression_id - 1;
270 /* Return the expression id for tree EXPR. */
272 static inline unsigned int
273 get_expression_id (const pre_expr expr)
275 return expr->id;
278 static inline unsigned int
279 lookup_expression_id (const pre_expr expr)
281 void **slot;
283 if (expr->kind == NAME)
285 unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
286 if (VEC_length (unsigned, name_to_id) <= version)
287 return 0;
288 return VEC_index (unsigned, name_to_id, version);
290 else
292 slot = htab_find_slot (expression_to_id, expr, NO_INSERT);
293 if (!slot)
294 return 0;
295 return ((pre_expr)*slot)->id;
299 /* Return the existing expression id for EXPR, or create one if one
300 does not exist yet. */
302 static inline unsigned int
303 get_or_alloc_expression_id (pre_expr expr)
305 unsigned int id = lookup_expression_id (expr);
306 if (id == 0)
307 return alloc_expression_id (expr);
308 return expr->id = id;
311 /* Return the expression that has expression id ID */
313 static inline pre_expr
314 expression_for_id (unsigned int id)
316 return VEC_index (pre_expr, expressions, id);
319 /* Free the expression id field in all of our expressions,
320 and then destroy the expressions array. */
322 static void
323 clear_expression_ids (void)
325 VEC_free (pre_expr, heap, expressions);
328 static alloc_pool pre_expr_pool;
330 /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
332 static pre_expr
333 get_or_alloc_expr_for_name (tree name)
335 struct pre_expr_d expr;
336 pre_expr result;
337 unsigned int result_id;
339 expr.kind = NAME;
340 expr.id = 0;
341 PRE_EXPR_NAME (&expr) = name;
342 result_id = lookup_expression_id (&expr);
343 if (result_id != 0)
344 return expression_for_id (result_id);
346 result = (pre_expr) pool_alloc (pre_expr_pool);
347 result->kind = NAME;
348 PRE_EXPR_NAME (result) = name;
349 alloc_expression_id (result);
350 return result;
353 static bool in_fre = false;
355 /* An unordered bitmap set. One bitmap tracks values, the other,
356 expressions. */
357 typedef struct bitmap_set
359 bitmap_head expressions;
360 bitmap_head values;
361 } *bitmap_set_t;
363 #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
364 EXECUTE_IF_SET_IN_BITMAP(&(set)->expressions, 0, (id), (bi))
366 #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
367 EXECUTE_IF_SET_IN_BITMAP(&(set)->values, 0, (id), (bi))
369 /* Mapping from value id to expressions with that value_id. */
370 DEF_VEC_P (bitmap_set_t);
371 DEF_VEC_ALLOC_P (bitmap_set_t, heap);
372 static VEC(bitmap_set_t, heap) *value_expressions;
374 /* Sets that we need to keep track of. */
375 typedef struct bb_bitmap_sets
377 /* The EXP_GEN set, which represents expressions/values generated in
378 a basic block. */
379 bitmap_set_t exp_gen;
381 /* The PHI_GEN set, which represents PHI results generated in a
382 basic block. */
383 bitmap_set_t phi_gen;
385 /* The TMP_GEN set, which represents results/temporaries generated
386 in a basic block. IE the LHS of an expression. */
387 bitmap_set_t tmp_gen;
389 /* The AVAIL_OUT set, which represents which values are available in
390 a given basic block. */
391 bitmap_set_t avail_out;
393 /* The ANTIC_IN set, which represents which values are anticipatable
394 in a given basic block. */
395 bitmap_set_t antic_in;
397 /* The PA_IN set, which represents which values are
398 partially anticipatable in a given basic block. */
399 bitmap_set_t pa_in;
401 /* The NEW_SETS set, which is used during insertion to augment the
402 AVAIL_OUT set of blocks with the new insertions performed during
403 the current iteration. */
404 bitmap_set_t new_sets;
406 /* A cache for value_dies_in_block_x. */
407 bitmap expr_dies;
409 /* True if we have visited this block during ANTIC calculation. */
410 unsigned int visited : 1;
412 /* True we have deferred processing this block during ANTIC
413 calculation until its successor is processed. */
414 unsigned int deferred : 1;
416 /* True when the block contains a call that might not return. */
417 unsigned int contains_may_not_return_call : 1;
418 } *bb_value_sets_t;
420 #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
421 #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
422 #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
423 #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
424 #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
425 #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
426 #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
427 #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
428 #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
429 #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred
430 #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
433 /* Basic block list in postorder. */
434 static int *postorder;
436 /* This structure is used to keep track of statistics on what
437 optimization PRE was able to perform. */
438 static struct
440 /* The number of RHS computations eliminated by PRE. */
441 int eliminations;
443 /* The number of new expressions/temporaries generated by PRE. */
444 int insertions;
446 /* The number of inserts found due to partial anticipation */
447 int pa_insert;
449 /* The number of new PHI nodes added by PRE. */
450 int phis;
452 /* The number of values found constant. */
453 int constified;
455 } pre_stats;
457 static bool do_partial_partial;
458 static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int, gimple);
459 static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
460 static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
461 static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
462 static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
463 static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
464 static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr,
465 unsigned int, bool);
466 static bitmap_set_t bitmap_set_new (void);
467 static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
468 gimple, tree);
469 static tree find_or_generate_expression (basic_block, pre_expr, gimple_seq *,
470 gimple);
471 static unsigned int get_expr_value_id (pre_expr);
473 /* We can add and remove elements and entries to and from sets
474 and hash tables, so we use alloc pools for them. */
476 static alloc_pool bitmap_set_pool;
477 static bitmap_obstack grand_bitmap_obstack;
479 /* To avoid adding 300 temporary variables when we only need one, we
480 only create one temporary variable, on demand, and build ssa names
481 off that. We do have to change the variable if the types don't
482 match the current variable's type. */
483 static tree pretemp;
484 static tree storetemp;
485 static tree prephitemp;
487 /* Set of blocks with statements that have had their EH properties changed. */
488 static bitmap need_eh_cleanup;
490 /* Set of blocks with statements that have had their AB properties changed. */
491 static bitmap need_ab_cleanup;
493 /* The phi_translate_table caches phi translations for a given
494 expression and predecessor. */
496 static htab_t phi_translate_table;
498 /* A three tuple {e, pred, v} used to cache phi translations in the
499 phi_translate_table. */
501 typedef struct expr_pred_trans_d
503 /* The expression. */
504 pre_expr e;
506 /* The predecessor block along which we translated the expression. */
507 basic_block pred;
509 /* The value that resulted from the translation. */
510 pre_expr v;
512 /* The hashcode for the expression, pred pair. This is cached for
513 speed reasons. */
514 hashval_t hashcode;
515 } *expr_pred_trans_t;
516 typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
518 /* Return the hash value for a phi translation table entry. */
520 static hashval_t
521 expr_pred_trans_hash (const void *p)
523 const_expr_pred_trans_t const ve = (const_expr_pred_trans_t) p;
524 return ve->hashcode;
527 /* Return true if two phi translation table entries are the same.
528 P1 and P2 should point to the expr_pred_trans_t's to be compared.*/
530 static int
531 expr_pred_trans_eq (const void *p1, const void *p2)
533 const_expr_pred_trans_t const ve1 = (const_expr_pred_trans_t) p1;
534 const_expr_pred_trans_t const ve2 = (const_expr_pred_trans_t) p2;
535 basic_block b1 = ve1->pred;
536 basic_block b2 = ve2->pred;
538 /* If they are not translations for the same basic block, they can't
539 be equal. */
540 if (b1 != b2)
541 return false;
542 return pre_expr_eq (ve1->e, ve2->e);
545 /* Search in the phi translation table for the translation of
546 expression E in basic block PRED.
547 Return the translated value, if found, NULL otherwise. */
549 static inline pre_expr
550 phi_trans_lookup (pre_expr e, basic_block pred)
552 void **slot;
553 struct expr_pred_trans_d ept;
555 ept.e = e;
556 ept.pred = pred;
557 ept.hashcode = iterative_hash_hashval_t (pre_expr_hash (e), pred->index);
558 slot = htab_find_slot_with_hash (phi_translate_table, &ept, ept.hashcode,
559 NO_INSERT);
560 if (!slot)
561 return NULL;
562 else
563 return ((expr_pred_trans_t) *slot)->v;
567 /* Add the tuple mapping from {expression E, basic block PRED} to
568 value V, to the phi translation table. */
570 static inline void
571 phi_trans_add (pre_expr e, pre_expr v, basic_block pred)
573 void **slot;
574 expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d);
575 new_pair->e = e;
576 new_pair->pred = pred;
577 new_pair->v = v;
578 new_pair->hashcode = iterative_hash_hashval_t (pre_expr_hash (e),
579 pred->index);
581 slot = htab_find_slot_with_hash (phi_translate_table, new_pair,
582 new_pair->hashcode, INSERT);
583 free (*slot);
584 *slot = (void *) new_pair;
588 /* Add expression E to the expression set of value id V. */
590 void
591 add_to_value (unsigned int v, pre_expr e)
593 bitmap_set_t set;
595 gcc_assert (get_expr_value_id (e) == v);
597 if (v >= VEC_length (bitmap_set_t, value_expressions))
599 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
600 v + 1);
603 set = VEC_index (bitmap_set_t, value_expressions, v);
604 if (!set)
606 set = bitmap_set_new ();
607 VEC_replace (bitmap_set_t, value_expressions, v, set);
610 bitmap_insert_into_set_1 (set, e, v, true);
613 /* Create a new bitmap set and return it. */
615 static bitmap_set_t
616 bitmap_set_new (void)
618 bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool);
619 bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
620 bitmap_initialize (&ret->values, &grand_bitmap_obstack);
621 return ret;
624 /* Return the value id for a PRE expression EXPR. */
626 static unsigned int
627 get_expr_value_id (pre_expr expr)
629 switch (expr->kind)
631 case CONSTANT:
633 unsigned int id;
634 id = get_constant_value_id (PRE_EXPR_CONSTANT (expr));
635 if (id == 0)
637 id = get_or_alloc_constant_value_id (PRE_EXPR_CONSTANT (expr));
638 add_to_value (id, expr);
640 return id;
642 case NAME:
643 return VN_INFO (PRE_EXPR_NAME (expr))->value_id;
644 case NARY:
645 return PRE_EXPR_NARY (expr)->value_id;
646 case REFERENCE:
647 return PRE_EXPR_REFERENCE (expr)->value_id;
648 default:
649 gcc_unreachable ();
653 /* Remove an expression EXPR from a bitmapped set. */
655 static void
656 bitmap_remove_from_set (bitmap_set_t set, pre_expr expr)
658 unsigned int val = get_expr_value_id (expr);
659 if (!value_id_constant_p (val))
661 bitmap_clear_bit (&set->values, val);
662 bitmap_clear_bit (&set->expressions, get_expression_id (expr));
666 static void
667 bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr,
668 unsigned int val, bool allow_constants)
670 if (allow_constants || !value_id_constant_p (val))
672 /* We specifically expect this and only this function to be able to
673 insert constants into a set. */
674 bitmap_set_bit (&set->values, val);
675 bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr));
679 /* Insert an expression EXPR into a bitmapped set. */
681 static void
682 bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
684 bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false);
687 /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
689 static void
690 bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
692 bitmap_copy (&dest->expressions, &orig->expressions);
693 bitmap_copy (&dest->values, &orig->values);
697 /* Free memory used up by SET. */
698 static void
699 bitmap_set_free (bitmap_set_t set)
701 bitmap_clear (&set->expressions);
702 bitmap_clear (&set->values);
706 /* Generate an topological-ordered array of bitmap set SET. */
708 static VEC(pre_expr, heap) *
709 sorted_array_from_bitmap_set (bitmap_set_t set)
711 unsigned int i, j;
712 bitmap_iterator bi, bj;
713 VEC(pre_expr, heap) *result;
715 /* Pre-allocate roughly enough space for the array. */
716 result = VEC_alloc (pre_expr, heap, bitmap_count_bits (&set->values));
718 FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
720 /* The number of expressions having a given value is usually
721 relatively small. Thus, rather than making a vector of all
722 the expressions and sorting it by value-id, we walk the values
723 and check in the reverse mapping that tells us what expressions
724 have a given value, to filter those in our set. As a result,
725 the expressions are inserted in value-id order, which means
726 topological order.
728 If this is somehow a significant lose for some cases, we can
729 choose which set to walk based on the set size. */
730 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, i);
731 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bj)
733 if (bitmap_bit_p (&set->expressions, j))
734 VEC_safe_push (pre_expr, heap, result, expression_for_id (j));
738 return result;
741 /* Perform bitmapped set operation DEST &= ORIG. */
743 static void
744 bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig)
746 bitmap_iterator bi;
747 unsigned int i;
749 if (dest != orig)
751 bitmap_head temp;
752 bitmap_initialize (&temp, &grand_bitmap_obstack);
754 bitmap_and_into (&dest->values, &orig->values);
755 bitmap_copy (&temp, &dest->expressions);
756 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
758 pre_expr expr = expression_for_id (i);
759 unsigned int value_id = get_expr_value_id (expr);
760 if (!bitmap_bit_p (&dest->values, value_id))
761 bitmap_clear_bit (&dest->expressions, i);
763 bitmap_clear (&temp);
767 /* Subtract all values and expressions contained in ORIG from DEST. */
769 static bitmap_set_t
770 bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig)
772 bitmap_set_t result = bitmap_set_new ();
773 bitmap_iterator bi;
774 unsigned int i;
776 bitmap_and_compl (&result->expressions, &dest->expressions,
777 &orig->expressions);
779 FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
781 pre_expr expr = expression_for_id (i);
782 unsigned int value_id = get_expr_value_id (expr);
783 bitmap_set_bit (&result->values, value_id);
786 return result;
789 /* Subtract all the values in bitmap set B from bitmap set A. */
791 static void
792 bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
794 unsigned int i;
795 bitmap_iterator bi;
796 bitmap_head temp;
798 bitmap_initialize (&temp, &grand_bitmap_obstack);
800 bitmap_copy (&temp, &a->expressions);
801 EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi)
803 pre_expr expr = expression_for_id (i);
804 if (bitmap_set_contains_value (b, get_expr_value_id (expr)))
805 bitmap_remove_from_set (a, expr);
807 bitmap_clear (&temp);
811 /* Return true if bitmapped set SET contains the value VALUE_ID. */
813 static bool
814 bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
816 if (value_id_constant_p (value_id))
817 return true;
819 if (!set || bitmap_empty_p (&set->expressions))
820 return false;
822 return bitmap_bit_p (&set->values, value_id);
825 static inline bool
826 bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr)
828 return bitmap_bit_p (&set->expressions, get_expression_id (expr));
831 /* Replace an instance of value LOOKFOR with expression EXPR in SET. */
833 static void
834 bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor,
835 const pre_expr expr)
837 bitmap_set_t exprset;
838 unsigned int i;
839 bitmap_iterator bi;
841 if (value_id_constant_p (lookfor))
842 return;
844 if (!bitmap_set_contains_value (set, lookfor))
845 return;
847 /* The number of expressions having a given value is usually
848 significantly less than the total number of expressions in SET.
849 Thus, rather than check, for each expression in SET, whether it
850 has the value LOOKFOR, we walk the reverse mapping that tells us
851 what expressions have a given value, and see if any of those
852 expressions are in our set. For large testcases, this is about
853 5-10x faster than walking the bitmap. If this is somehow a
854 significant lose for some cases, we can choose which set to walk
855 based on the set size. */
856 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
857 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
859 if (bitmap_clear_bit (&set->expressions, i))
861 bitmap_set_bit (&set->expressions, get_expression_id (expr));
862 return;
867 /* Return true if two bitmap sets are equal. */
869 static bool
870 bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
872 return bitmap_equal_p (&a->values, &b->values);
875 /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
876 and add it otherwise. */
878 static void
879 bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
881 unsigned int val = get_expr_value_id (expr);
883 if (bitmap_set_contains_value (set, val))
884 bitmap_set_replace_value (set, val, expr);
885 else
886 bitmap_insert_into_set (set, expr);
889 /* Insert EXPR into SET if EXPR's value is not already present in
890 SET. */
892 static void
893 bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
895 unsigned int val = get_expr_value_id (expr);
897 gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr));
899 /* Constant values are always considered to be part of the set. */
900 if (value_id_constant_p (val))
901 return;
903 /* If the value membership changed, add the expression. */
904 if (bitmap_set_bit (&set->values, val))
905 bitmap_set_bit (&set->expressions, expr->id);
908 /* Print out EXPR to outfile. */
910 static void
911 print_pre_expr (FILE *outfile, const pre_expr expr)
913 switch (expr->kind)
915 case CONSTANT:
916 print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0);
917 break;
918 case NAME:
919 print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0);
920 break;
921 case NARY:
923 unsigned int i;
924 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
925 fprintf (outfile, "{%s,", tree_code_name [nary->opcode]);
926 for (i = 0; i < nary->length; i++)
928 print_generic_expr (outfile, nary->op[i], 0);
929 if (i != (unsigned) nary->length - 1)
930 fprintf (outfile, ",");
932 fprintf (outfile, "}");
934 break;
936 case REFERENCE:
938 vn_reference_op_t vro;
939 unsigned int i;
940 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
941 fprintf (outfile, "{");
942 for (i = 0;
943 VEC_iterate (vn_reference_op_s, ref->operands, i, vro);
944 i++)
946 bool closebrace = false;
947 if (vro->opcode != SSA_NAME
948 && TREE_CODE_CLASS (vro->opcode) != tcc_declaration)
950 fprintf (outfile, "%s", tree_code_name [vro->opcode]);
951 if (vro->op0)
953 fprintf (outfile, "<");
954 closebrace = true;
957 if (vro->op0)
959 print_generic_expr (outfile, vro->op0, 0);
960 if (vro->op1)
962 fprintf (outfile, ",");
963 print_generic_expr (outfile, vro->op1, 0);
965 if (vro->op2)
967 fprintf (outfile, ",");
968 print_generic_expr (outfile, vro->op2, 0);
971 if (closebrace)
972 fprintf (outfile, ">");
973 if (i != VEC_length (vn_reference_op_s, ref->operands) - 1)
974 fprintf (outfile, ",");
976 fprintf (outfile, "}");
977 if (ref->vuse)
979 fprintf (outfile, "@");
980 print_generic_expr (outfile, ref->vuse, 0);
983 break;
986 void debug_pre_expr (pre_expr);
988 /* Like print_pre_expr but always prints to stderr. */
989 DEBUG_FUNCTION void
990 debug_pre_expr (pre_expr e)
992 print_pre_expr (stderr, e);
993 fprintf (stderr, "\n");
996 /* Print out SET to OUTFILE. */
998 static void
999 print_bitmap_set (FILE *outfile, bitmap_set_t set,
1000 const char *setname, int blockindex)
1002 fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1003 if (set)
1005 bool first = true;
1006 unsigned i;
1007 bitmap_iterator bi;
1009 FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1011 const pre_expr expr = expression_for_id (i);
1013 if (!first)
1014 fprintf (outfile, ", ");
1015 first = false;
1016 print_pre_expr (outfile, expr);
1018 fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1021 fprintf (outfile, " }\n");
1024 void debug_bitmap_set (bitmap_set_t);
1026 DEBUG_FUNCTION void
1027 debug_bitmap_set (bitmap_set_t set)
1029 print_bitmap_set (stderr, set, "debug", 0);
1032 /* Print out the expressions that have VAL to OUTFILE. */
1034 void
1035 print_value_expressions (FILE *outfile, unsigned int val)
1037 bitmap_set_t set = VEC_index (bitmap_set_t, value_expressions, val);
1038 if (set)
1040 char s[10];
1041 sprintf (s, "%04d", val);
1042 print_bitmap_set (outfile, set, s, 0);
1047 DEBUG_FUNCTION void
1048 debug_value_expressions (unsigned int val)
1050 print_value_expressions (stderr, val);
1053 /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1054 represent it. */
1056 static pre_expr
1057 get_or_alloc_expr_for_constant (tree constant)
1059 unsigned int result_id;
1060 unsigned int value_id;
1061 struct pre_expr_d expr;
1062 pre_expr newexpr;
1064 expr.kind = CONSTANT;
1065 PRE_EXPR_CONSTANT (&expr) = constant;
1066 result_id = lookup_expression_id (&expr);
1067 if (result_id != 0)
1068 return expression_for_id (result_id);
1070 newexpr = (pre_expr) pool_alloc (pre_expr_pool);
1071 newexpr->kind = CONSTANT;
1072 PRE_EXPR_CONSTANT (newexpr) = constant;
1073 alloc_expression_id (newexpr);
1074 value_id = get_or_alloc_constant_value_id (constant);
1075 add_to_value (value_id, newexpr);
1076 return newexpr;
1079 /* Given a value id V, find the actual tree representing the constant
1080 value if there is one, and return it. Return NULL if we can't find
1081 a constant. */
1083 static tree
1084 get_constant_for_value_id (unsigned int v)
1086 if (value_id_constant_p (v))
1088 unsigned int i;
1089 bitmap_iterator bi;
1090 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, v);
1092 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1094 pre_expr expr = expression_for_id (i);
1095 if (expr->kind == CONSTANT)
1096 return PRE_EXPR_CONSTANT (expr);
1099 return NULL;
1102 /* Get or allocate a pre_expr for a piece of GIMPLE, and return it.
1103 Currently only supports constants and SSA_NAMES. */
1104 static pre_expr
1105 get_or_alloc_expr_for (tree t)
1107 if (TREE_CODE (t) == SSA_NAME)
1108 return get_or_alloc_expr_for_name (t);
1109 else if (is_gimple_min_invariant (t))
1110 return get_or_alloc_expr_for_constant (t);
1111 else
1113 /* More complex expressions can result from SCCVN expression
1114 simplification that inserts values for them. As they all
1115 do not have VOPs the get handled by the nary ops struct. */
1116 vn_nary_op_t result;
1117 unsigned int result_id;
1118 vn_nary_op_lookup (t, &result);
1119 if (result != NULL)
1121 pre_expr e = (pre_expr) pool_alloc (pre_expr_pool);
1122 e->kind = NARY;
1123 PRE_EXPR_NARY (e) = result;
1124 result_id = lookup_expression_id (e);
1125 if (result_id != 0)
1127 pool_free (pre_expr_pool, e);
1128 e = expression_for_id (result_id);
1129 return e;
1131 alloc_expression_id (e);
1132 return e;
1135 return NULL;
1138 /* Return the folded version of T if T, when folded, is a gimple
1139 min_invariant. Otherwise, return T. */
1141 static pre_expr
1142 fully_constant_expression (pre_expr e)
1144 switch (e->kind)
1146 case CONSTANT:
1147 return e;
1148 case NARY:
1150 vn_nary_op_t nary = PRE_EXPR_NARY (e);
1151 switch (TREE_CODE_CLASS (nary->opcode))
1153 case tcc_binary:
1154 case tcc_comparison:
1156 /* We have to go from trees to pre exprs to value ids to
1157 constants. */
1158 tree naryop0 = nary->op[0];
1159 tree naryop1 = nary->op[1];
1160 tree result;
1161 if (!is_gimple_min_invariant (naryop0))
1163 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1164 unsigned int vrep0 = get_expr_value_id (rep0);
1165 tree const0 = get_constant_for_value_id (vrep0);
1166 if (const0)
1167 naryop0 = fold_convert (TREE_TYPE (naryop0), const0);
1169 if (!is_gimple_min_invariant (naryop1))
1171 pre_expr rep1 = get_or_alloc_expr_for (naryop1);
1172 unsigned int vrep1 = get_expr_value_id (rep1);
1173 tree const1 = get_constant_for_value_id (vrep1);
1174 if (const1)
1175 naryop1 = fold_convert (TREE_TYPE (naryop1), const1);
1177 result = fold_binary (nary->opcode, nary->type,
1178 naryop0, naryop1);
1179 if (result && is_gimple_min_invariant (result))
1180 return get_or_alloc_expr_for_constant (result);
1181 /* We might have simplified the expression to a
1182 SSA_NAME for example from x_1 * 1. But we cannot
1183 insert a PHI for x_1 unconditionally as x_1 might
1184 not be available readily. */
1185 return e;
1187 case tcc_reference:
1188 if (nary->opcode != REALPART_EXPR
1189 && nary->opcode != IMAGPART_EXPR
1190 && nary->opcode != VIEW_CONVERT_EXPR)
1191 return e;
1192 /* Fallthrough. */
1193 case tcc_unary:
1195 /* We have to go from trees to pre exprs to value ids to
1196 constants. */
1197 tree naryop0 = nary->op[0];
1198 tree const0, result;
1199 if (is_gimple_min_invariant (naryop0))
1200 const0 = naryop0;
1201 else
1203 pre_expr rep0 = get_or_alloc_expr_for (naryop0);
1204 unsigned int vrep0 = get_expr_value_id (rep0);
1205 const0 = get_constant_for_value_id (vrep0);
1207 result = NULL;
1208 if (const0)
1210 tree type1 = TREE_TYPE (nary->op[0]);
1211 const0 = fold_convert (type1, const0);
1212 result = fold_unary (nary->opcode, nary->type, const0);
1214 if (result && is_gimple_min_invariant (result))
1215 return get_or_alloc_expr_for_constant (result);
1216 return e;
1218 default:
1219 return e;
1222 case REFERENCE:
1224 vn_reference_t ref = PRE_EXPR_REFERENCE (e);
1225 tree folded;
1226 if ((folded = fully_constant_vn_reference_p (ref)))
1227 return get_or_alloc_expr_for_constant (folded);
1228 return e;
1230 default:
1231 return e;
1233 return e;
1236 /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that
1237 it has the value it would have in BLOCK. Set *SAME_VALID to true
1238 in case the new vuse doesn't change the value id of the OPERANDS. */
1240 static tree
1241 translate_vuse_through_block (VEC (vn_reference_op_s, heap) *operands,
1242 alias_set_type set, tree type, tree vuse,
1243 basic_block phiblock,
1244 basic_block block, bool *same_valid)
1246 gimple phi = SSA_NAME_DEF_STMT (vuse);
1247 ao_ref ref;
1248 edge e = NULL;
1249 bool use_oracle;
1251 *same_valid = true;
1253 if (gimple_bb (phi) != phiblock)
1254 return vuse;
1256 use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands);
1258 /* Use the alias-oracle to find either the PHI node in this block,
1259 the first VUSE used in this block that is equivalent to vuse or
1260 the first VUSE which definition in this block kills the value. */
1261 if (gimple_code (phi) == GIMPLE_PHI)
1262 e = find_edge (block, phiblock);
1263 else if (use_oracle)
1264 while (!stmt_may_clobber_ref_p_1 (phi, &ref))
1266 vuse = gimple_vuse (phi);
1267 phi = SSA_NAME_DEF_STMT (vuse);
1268 if (gimple_bb (phi) != phiblock)
1269 return vuse;
1270 if (gimple_code (phi) == GIMPLE_PHI)
1272 e = find_edge (block, phiblock);
1273 break;
1276 else
1277 return NULL_TREE;
1279 if (e)
1281 if (use_oracle)
1283 bitmap visited = NULL;
1284 /* Try to find a vuse that dominates this phi node by skipping
1285 non-clobbering statements. */
1286 vuse = get_continuation_for_phi (phi, &ref, &visited);
1287 if (visited)
1288 BITMAP_FREE (visited);
1290 else
1291 vuse = NULL_TREE;
1292 if (!vuse)
1294 /* If we didn't find any, the value ID can't stay the same,
1295 but return the translated vuse. */
1296 *same_valid = false;
1297 vuse = PHI_ARG_DEF (phi, e->dest_idx);
1299 /* ??? We would like to return vuse here as this is the canonical
1300 upmost vdef that this reference is associated with. But during
1301 insertion of the references into the hash tables we only ever
1302 directly insert with their direct gimple_vuse, hence returning
1303 something else would make us not find the other expression. */
1304 return PHI_ARG_DEF (phi, e->dest_idx);
1307 return NULL_TREE;
1310 /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1311 SET2. This is used to avoid making a set consisting of the union
1312 of PA_IN and ANTIC_IN during insert. */
1314 static inline pre_expr
1315 find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2)
1317 pre_expr result;
1319 result = bitmap_find_leader (set1, val, NULL);
1320 if (!result && set2)
1321 result = bitmap_find_leader (set2, val, NULL);
1322 return result;
1325 /* Get the tree type for our PRE expression e. */
1327 static tree
1328 get_expr_type (const pre_expr e)
1330 switch (e->kind)
1332 case NAME:
1333 return TREE_TYPE (PRE_EXPR_NAME (e));
1334 case CONSTANT:
1335 return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1336 case REFERENCE:
1337 return PRE_EXPR_REFERENCE (e)->type;
1338 case NARY:
1339 return PRE_EXPR_NARY (e)->type;
1341 gcc_unreachable();
1344 /* Get a representative SSA_NAME for a given expression.
1345 Since all of our sub-expressions are treated as values, we require
1346 them to be SSA_NAME's for simplicity.
1347 Prior versions of GVNPRE used to use "value handles" here, so that
1348 an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1349 either case, the operands are really values (IE we do not expect
1350 them to be usable without finding leaders). */
1352 static tree
1353 get_representative_for (const pre_expr e)
1355 tree exprtype;
1356 tree name;
1357 unsigned int value_id = get_expr_value_id (e);
1359 switch (e->kind)
1361 case NAME:
1362 return PRE_EXPR_NAME (e);
1363 case CONSTANT:
1364 return PRE_EXPR_CONSTANT (e);
1365 case NARY:
1366 case REFERENCE:
1368 /* Go through all of the expressions representing this value
1369 and pick out an SSA_NAME. */
1370 unsigned int i;
1371 bitmap_iterator bi;
1372 bitmap_set_t exprs = VEC_index (bitmap_set_t, value_expressions,
1373 value_id);
1374 FOR_EACH_EXPR_ID_IN_SET (exprs, i, bi)
1376 pre_expr rep = expression_for_id (i);
1377 if (rep->kind == NAME)
1378 return PRE_EXPR_NAME (rep);
1381 break;
1383 /* If we reached here we couldn't find an SSA_NAME. This can
1384 happen when we've discovered a value that has never appeared in
1385 the program as set to an SSA_NAME, most likely as the result of
1386 phi translation. */
1387 if (dump_file)
1389 fprintf (dump_file,
1390 "Could not find SSA_NAME representative for expression:");
1391 print_pre_expr (dump_file, e);
1392 fprintf (dump_file, "\n");
1395 exprtype = get_expr_type (e);
1397 /* Build and insert the assignment of the end result to the temporary
1398 that we will return. */
1399 if (!pretemp || exprtype != TREE_TYPE (pretemp))
1401 pretemp = create_tmp_reg (exprtype, "pretmp");
1402 add_referenced_var (pretemp);
1405 name = make_ssa_name (pretemp, gimple_build_nop ());
1406 VN_INFO_GET (name)->value_id = value_id;
1407 if (e->kind == CONSTANT)
1408 VN_INFO (name)->valnum = PRE_EXPR_CONSTANT (e);
1409 else
1410 VN_INFO (name)->valnum = name;
1412 add_to_value (value_id, get_or_alloc_expr_for_name (name));
1413 if (dump_file)
1415 fprintf (dump_file, "Created SSA_NAME representative ");
1416 print_generic_expr (dump_file, name, 0);
1417 fprintf (dump_file, " for expression:");
1418 print_pre_expr (dump_file, e);
1419 fprintf (dump_file, "\n");
1422 return name;
1427 static pre_expr
1428 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1429 basic_block pred, basic_block phiblock);
1431 /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1432 the phis in PRED. Return NULL if we can't find a leader for each part
1433 of the translated expression. */
1435 static pre_expr
1436 phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1437 basic_block pred, basic_block phiblock)
1439 switch (expr->kind)
1441 case NARY:
1443 unsigned int i;
1444 bool changed = false;
1445 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1446 vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1447 sizeof_vn_nary_op (nary->length));
1448 memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1450 for (i = 0; i < newnary->length; i++)
1452 if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1453 continue;
1454 else
1456 pre_expr leader, result;
1457 unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1458 leader = find_leader_in_sets (op_val_id, set1, set2);
1459 result = phi_translate (leader, set1, set2, pred, phiblock);
1460 if (result && result != leader)
1462 tree name = get_representative_for (result);
1463 if (!name)
1464 return NULL;
1465 newnary->op[i] = name;
1467 else if (!result)
1468 return NULL;
1470 changed |= newnary->op[i] != nary->op[i];
1473 if (changed)
1475 pre_expr constant;
1476 unsigned int new_val_id;
1478 tree result = vn_nary_op_lookup_pieces (newnary->length,
1479 newnary->opcode,
1480 newnary->type,
1481 &newnary->op[0],
1482 &nary);
1483 if (result && is_gimple_min_invariant (result))
1484 return get_or_alloc_expr_for_constant (result);
1486 expr = (pre_expr) pool_alloc (pre_expr_pool);
1487 expr->kind = NARY;
1488 expr->id = 0;
1489 if (nary)
1491 PRE_EXPR_NARY (expr) = nary;
1492 constant = fully_constant_expression (expr);
1493 if (constant != expr)
1494 return constant;
1496 new_val_id = nary->value_id;
1497 get_or_alloc_expression_id (expr);
1499 else
1501 new_val_id = get_next_value_id ();
1502 VEC_safe_grow_cleared (bitmap_set_t, heap,
1503 value_expressions,
1504 get_max_value_id() + 1);
1505 nary = vn_nary_op_insert_pieces (newnary->length,
1506 newnary->opcode,
1507 newnary->type,
1508 &newnary->op[0],
1509 result, new_val_id);
1510 PRE_EXPR_NARY (expr) = nary;
1511 constant = fully_constant_expression (expr);
1512 if (constant != expr)
1513 return constant;
1514 get_or_alloc_expression_id (expr);
1516 add_to_value (new_val_id, expr);
1518 return expr;
1520 break;
1522 case REFERENCE:
1524 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1525 VEC (vn_reference_op_s, heap) *operands = ref->operands;
1526 tree vuse = ref->vuse;
1527 tree newvuse = vuse;
1528 VEC (vn_reference_op_s, heap) *newoperands = NULL;
1529 bool changed = false, same_valid = true;
1530 unsigned int i, j;
1531 vn_reference_op_t operand;
1532 vn_reference_t newref;
1534 for (i = 0, j = 0;
1535 VEC_iterate (vn_reference_op_s, operands, i, operand); i++, j++)
1537 pre_expr opresult;
1538 pre_expr leader;
1539 tree oldop0 = operand->op0;
1540 tree oldop1 = operand->op1;
1541 tree oldop2 = operand->op2;
1542 tree op0 = oldop0;
1543 tree op1 = oldop1;
1544 tree op2 = oldop2;
1545 tree type = operand->type;
1546 vn_reference_op_s newop = *operand;
1548 if (op0 && TREE_CODE (op0) == SSA_NAME)
1550 unsigned int op_val_id = VN_INFO (op0)->value_id;
1551 leader = find_leader_in_sets (op_val_id, set1, set2);
1552 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1553 if (opresult && opresult != leader)
1555 tree name = get_representative_for (opresult);
1556 if (!name)
1557 break;
1558 op0 = name;
1560 else if (!opresult)
1561 break;
1563 changed |= op0 != oldop0;
1565 if (op1 && TREE_CODE (op1) == SSA_NAME)
1567 unsigned int op_val_id = VN_INFO (op1)->value_id;
1568 leader = find_leader_in_sets (op_val_id, set1, set2);
1569 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1570 if (opresult && opresult != leader)
1572 tree name = get_representative_for (opresult);
1573 if (!name)
1574 break;
1575 op1 = name;
1577 else if (!opresult)
1578 break;
1580 /* We can't possibly insert these. */
1581 else if (op1 && !is_gimple_min_invariant (op1))
1582 break;
1583 changed |= op1 != oldop1;
1584 if (op2 && TREE_CODE (op2) == SSA_NAME)
1586 unsigned int op_val_id = VN_INFO (op2)->value_id;
1587 leader = find_leader_in_sets (op_val_id, set1, set2);
1588 opresult = phi_translate (leader, set1, set2, pred, phiblock);
1589 if (opresult && opresult != leader)
1591 tree name = get_representative_for (opresult);
1592 if (!name)
1593 break;
1594 op2 = name;
1596 else if (!opresult)
1597 break;
1599 /* We can't possibly insert these. */
1600 else if (op2 && !is_gimple_min_invariant (op2))
1601 break;
1602 changed |= op2 != oldop2;
1604 if (!newoperands)
1605 newoperands = VEC_copy (vn_reference_op_s, heap, operands);
1606 /* We may have changed from an SSA_NAME to a constant */
1607 if (newop.opcode == SSA_NAME && TREE_CODE (op0) != SSA_NAME)
1608 newop.opcode = TREE_CODE (op0);
1609 newop.type = type;
1610 newop.op0 = op0;
1611 newop.op1 = op1;
1612 newop.op2 = op2;
1613 /* If it transforms a non-constant ARRAY_REF into a constant
1614 one, adjust the constant offset. */
1615 if (newop.opcode == ARRAY_REF
1616 && newop.off == -1
1617 && TREE_CODE (op0) == INTEGER_CST
1618 && TREE_CODE (op1) == INTEGER_CST
1619 && TREE_CODE (op2) == INTEGER_CST)
1621 double_int off = tree_to_double_int (op0);
1622 off = double_int_add (off,
1623 double_int_neg
1624 (tree_to_double_int (op1)));
1625 off = double_int_mul (off, tree_to_double_int (op2));
1626 if (double_int_fits_in_shwi_p (off))
1627 newop.off = off.low;
1629 VEC_replace (vn_reference_op_s, newoperands, j, &newop);
1630 /* If it transforms from an SSA_NAME to an address, fold with
1631 a preceding indirect reference. */
1632 if (j > 0 && op0 && TREE_CODE (op0) == ADDR_EXPR
1633 && VEC_index (vn_reference_op_s,
1634 newoperands, j - 1)->opcode == MEM_REF)
1635 vn_reference_fold_indirect (&newoperands, &j);
1637 if (i != VEC_length (vn_reference_op_s, operands))
1639 if (newoperands)
1640 VEC_free (vn_reference_op_s, heap, newoperands);
1641 return NULL;
1644 if (vuse)
1646 newvuse = translate_vuse_through_block (newoperands,
1647 ref->set, ref->type,
1648 vuse, phiblock, pred,
1649 &same_valid);
1650 if (newvuse == NULL_TREE)
1652 VEC_free (vn_reference_op_s, heap, newoperands);
1653 return NULL;
1657 if (changed || newvuse != vuse)
1659 unsigned int new_val_id;
1660 pre_expr constant;
1661 bool converted = false;
1663 tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1664 ref->type,
1665 newoperands,
1666 &newref, VN_WALK);
1667 if (result)
1668 VEC_free (vn_reference_op_s, heap, newoperands);
1670 if (result
1671 && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1673 result = fold_build1 (VIEW_CONVERT_EXPR, ref->type, result);
1674 converted = true;
1676 else if (!result && newref
1677 && !useless_type_conversion_p (ref->type, newref->type))
1679 VEC_free (vn_reference_op_s, heap, newoperands);
1680 return NULL;
1683 if (result && is_gimple_min_invariant (result))
1685 gcc_assert (!newoperands);
1686 return get_or_alloc_expr_for_constant (result);
1689 expr = (pre_expr) pool_alloc (pre_expr_pool);
1690 expr->kind = REFERENCE;
1691 expr->id = 0;
1693 if (converted)
1695 vn_nary_op_t nary;
1696 tree nresult;
1698 gcc_assert (CONVERT_EXPR_P (result)
1699 || TREE_CODE (result) == VIEW_CONVERT_EXPR);
1701 nresult = vn_nary_op_lookup_pieces (1, TREE_CODE (result),
1702 TREE_TYPE (result),
1703 &TREE_OPERAND (result, 0),
1704 &nary);
1705 if (nresult && is_gimple_min_invariant (nresult))
1706 return get_or_alloc_expr_for_constant (nresult);
1708 expr->kind = NARY;
1709 if (nary)
1711 PRE_EXPR_NARY (expr) = nary;
1712 constant = fully_constant_expression (expr);
1713 if (constant != expr)
1714 return constant;
1716 new_val_id = nary->value_id;
1717 get_or_alloc_expression_id (expr);
1719 else
1721 new_val_id = get_next_value_id ();
1722 VEC_safe_grow_cleared (bitmap_set_t, heap,
1723 value_expressions,
1724 get_max_value_id() + 1);
1725 nary = vn_nary_op_insert_pieces (1, TREE_CODE (result),
1726 TREE_TYPE (result),
1727 &TREE_OPERAND (result, 0),
1728 NULL_TREE,
1729 new_val_id);
1730 PRE_EXPR_NARY (expr) = nary;
1731 constant = fully_constant_expression (expr);
1732 if (constant != expr)
1733 return constant;
1734 get_or_alloc_expression_id (expr);
1737 else if (newref)
1739 PRE_EXPR_REFERENCE (expr) = newref;
1740 constant = fully_constant_expression (expr);
1741 if (constant != expr)
1742 return constant;
1744 new_val_id = newref->value_id;
1745 get_or_alloc_expression_id (expr);
1747 else
1749 if (changed || !same_valid)
1751 new_val_id = get_next_value_id ();
1752 VEC_safe_grow_cleared (bitmap_set_t, heap,
1753 value_expressions,
1754 get_max_value_id() + 1);
1756 else
1757 new_val_id = ref->value_id;
1758 newref = vn_reference_insert_pieces (newvuse, ref->set,
1759 ref->type,
1760 newoperands,
1761 result, new_val_id);
1762 newoperands = NULL;
1763 PRE_EXPR_REFERENCE (expr) = newref;
1764 constant = fully_constant_expression (expr);
1765 if (constant != expr)
1766 return constant;
1767 get_or_alloc_expression_id (expr);
1769 add_to_value (new_val_id, expr);
1771 VEC_free (vn_reference_op_s, heap, newoperands);
1772 return expr;
1774 break;
1776 case NAME:
1778 gimple phi = NULL;
1779 edge e;
1780 gimple def_stmt;
1781 tree name = PRE_EXPR_NAME (expr);
1783 def_stmt = SSA_NAME_DEF_STMT (name);
1784 if (gimple_code (def_stmt) == GIMPLE_PHI
1785 && gimple_bb (def_stmt) == phiblock)
1786 phi = def_stmt;
1787 else
1788 return expr;
1790 e = find_edge (pred, gimple_bb (phi));
1791 if (e)
1793 tree def = PHI_ARG_DEF (phi, e->dest_idx);
1794 pre_expr newexpr;
1796 if (TREE_CODE (def) == SSA_NAME)
1797 def = VN_INFO (def)->valnum;
1799 /* Handle constant. */
1800 if (is_gimple_min_invariant (def))
1801 return get_or_alloc_expr_for_constant (def);
1803 if (TREE_CODE (def) == SSA_NAME && ssa_undefined_value_p (def))
1804 return NULL;
1806 newexpr = get_or_alloc_expr_for_name (def);
1807 return newexpr;
1810 return expr;
1812 default:
1813 gcc_unreachable ();
1817 /* Wrapper around phi_translate_1 providing caching functionality. */
1819 static pre_expr
1820 phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2,
1821 basic_block pred, basic_block phiblock)
1823 pre_expr phitrans;
1825 if (!expr)
1826 return NULL;
1828 /* Constants contain no values that need translation. */
1829 if (expr->kind == CONSTANT)
1830 return expr;
1832 if (value_id_constant_p (get_expr_value_id (expr)))
1833 return expr;
1835 if (expr->kind != NAME)
1837 phitrans = phi_trans_lookup (expr, pred);
1838 if (phitrans)
1839 return phitrans;
1842 /* Translate. */
1843 phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock);
1845 /* Don't add empty translations to the cache. Neither add
1846 translations of NAMEs as those are cheap to translate. */
1847 if (phitrans
1848 && expr->kind != NAME)
1849 phi_trans_add (expr, phitrans, pred);
1851 return phitrans;
1855 /* For each expression in SET, translate the values through phi nodes
1856 in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1857 expressions in DEST. */
1859 static void
1860 phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred,
1861 basic_block phiblock)
1863 VEC (pre_expr, heap) *exprs;
1864 pre_expr expr;
1865 int i;
1867 if (gimple_seq_empty_p (phi_nodes (phiblock)))
1869 bitmap_set_copy (dest, set);
1870 return;
1873 exprs = sorted_array_from_bitmap_set (set);
1874 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
1876 pre_expr translated;
1877 translated = phi_translate (expr, set, NULL, pred, phiblock);
1878 if (!translated)
1879 continue;
1881 /* We might end up with multiple expressions from SET being
1882 translated to the same value. In this case we do not want
1883 to retain the NARY or REFERENCE expression but prefer a NAME
1884 which would be the leader. */
1885 if (translated->kind == NAME)
1886 bitmap_value_replace_in_set (dest, translated);
1887 else
1888 bitmap_value_insert_into_set (dest, translated);
1890 VEC_free (pre_expr, heap, exprs);
1893 /* Find the leader for a value (i.e., the name representing that
1894 value) in a given set, and return it. If STMT is non-NULL it
1895 makes sure the defining statement for the leader dominates it.
1896 Return NULL if no leader is found. */
1898 static pre_expr
1899 bitmap_find_leader (bitmap_set_t set, unsigned int val, gimple stmt)
1901 if (value_id_constant_p (val))
1903 unsigned int i;
1904 bitmap_iterator bi;
1905 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1907 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
1909 pre_expr expr = expression_for_id (i);
1910 if (expr->kind == CONSTANT)
1911 return expr;
1914 if (bitmap_set_contains_value (set, val))
1916 /* Rather than walk the entire bitmap of expressions, and see
1917 whether any of them has the value we are looking for, we look
1918 at the reverse mapping, which tells us the set of expressions
1919 that have a given value (IE value->expressions with that
1920 value) and see if any of those expressions are in our set.
1921 The number of expressions per value is usually significantly
1922 less than the number of expressions in the set. In fact, for
1923 large testcases, doing it this way is roughly 5-10x faster
1924 than walking the bitmap.
1925 If this is somehow a significant lose for some cases, we can
1926 choose which set to walk based on which set is smaller. */
1927 unsigned int i;
1928 bitmap_iterator bi;
1929 bitmap_set_t exprset = VEC_index (bitmap_set_t, value_expressions, val);
1931 EXECUTE_IF_AND_IN_BITMAP (&exprset->expressions,
1932 &set->expressions, 0, i, bi)
1934 pre_expr val = expression_for_id (i);
1935 /* At the point where stmt is not null, there should always
1936 be an SSA_NAME first in the list of expressions. */
1937 if (stmt)
1939 gimple def_stmt = SSA_NAME_DEF_STMT (PRE_EXPR_NAME (val));
1940 if (gimple_code (def_stmt) != GIMPLE_PHI
1941 && gimple_bb (def_stmt) == gimple_bb (stmt)
1942 /* PRE insertions are at the end of the basic-block
1943 and have UID 0. */
1944 && (gimple_uid (def_stmt) == 0
1945 || gimple_uid (def_stmt) >= gimple_uid (stmt)))
1946 continue;
1948 return val;
1951 return NULL;
1954 /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1955 BLOCK by seeing if it is not killed in the block. Note that we are
1956 only determining whether there is a store that kills it. Because
1957 of the order in which clean iterates over values, we are guaranteed
1958 that altered operands will have caused us to be eliminated from the
1959 ANTIC_IN set already. */
1961 static bool
1962 value_dies_in_block_x (pre_expr expr, basic_block block)
1964 tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1965 vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1966 gimple def;
1967 gimple_stmt_iterator gsi;
1968 unsigned id = get_expression_id (expr);
1969 bool res = false;
1970 ao_ref ref;
1972 if (!vuse)
1973 return false;
1975 /* Lookup a previously calculated result. */
1976 if (EXPR_DIES (block)
1977 && bitmap_bit_p (EXPR_DIES (block), id * 2))
1978 return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1980 /* A memory expression {e, VUSE} dies in the block if there is a
1981 statement that may clobber e. If, starting statement walk from the
1982 top of the basic block, a statement uses VUSE there can be no kill
1983 inbetween that use and the original statement that loaded {e, VUSE},
1984 so we can stop walking. */
1985 ref.base = NULL_TREE;
1986 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1988 tree def_vuse, def_vdef;
1989 def = gsi_stmt (gsi);
1990 def_vuse = gimple_vuse (def);
1991 def_vdef = gimple_vdef (def);
1993 /* Not a memory statement. */
1994 if (!def_vuse)
1995 continue;
1997 /* Not a may-def. */
1998 if (!def_vdef)
2000 /* A load with the same VUSE, we're done. */
2001 if (def_vuse == vuse)
2002 break;
2004 continue;
2007 /* Init ref only if we really need it. */
2008 if (ref.base == NULL_TREE
2009 && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type,
2010 refx->operands))
2012 res = true;
2013 break;
2015 /* If the statement may clobber expr, it dies. */
2016 if (stmt_may_clobber_ref_p_1 (def, &ref))
2018 res = true;
2019 break;
2023 /* Remember the result. */
2024 if (!EXPR_DIES (block))
2025 EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
2026 bitmap_set_bit (EXPR_DIES (block), id * 2);
2027 if (res)
2028 bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
2030 return res;
2034 #define union_contains_value(SET1, SET2, VAL) \
2035 (bitmap_set_contains_value ((SET1), (VAL)) \
2036 || ((SET2) && bitmap_set_contains_value ((SET2), (VAL))))
2038 /* Determine if vn_reference_op_t VRO is legal in SET1 U SET2.
2040 static bool
2041 vro_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2,
2042 vn_reference_op_t vro)
2044 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
2046 struct pre_expr_d temp;
2047 temp.kind = NAME;
2048 temp.id = 0;
2049 PRE_EXPR_NAME (&temp) = vro->op0;
2050 temp.id = lookup_expression_id (&temp);
2051 if (temp.id == 0)
2052 return false;
2053 if (!union_contains_value (set1, set2,
2054 get_expr_value_id (&temp)))
2055 return false;
2057 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
2059 struct pre_expr_d temp;
2060 temp.kind = NAME;
2061 temp.id = 0;
2062 PRE_EXPR_NAME (&temp) = vro->op1;
2063 temp.id = lookup_expression_id (&temp);
2064 if (temp.id == 0)
2065 return false;
2066 if (!union_contains_value (set1, set2,
2067 get_expr_value_id (&temp)))
2068 return false;
2071 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
2073 struct pre_expr_d temp;
2074 temp.kind = NAME;
2075 temp.id = 0;
2076 PRE_EXPR_NAME (&temp) = vro->op2;
2077 temp.id = lookup_expression_id (&temp);
2078 if (temp.id == 0)
2079 return false;
2080 if (!union_contains_value (set1, set2,
2081 get_expr_value_id (&temp)))
2082 return false;
2085 return true;
2088 /* Determine if the expression EXPR is valid in SET1 U SET2.
2089 ONLY SET2 CAN BE NULL.
2090 This means that we have a leader for each part of the expression
2091 (if it consists of values), or the expression is an SSA_NAME.
2092 For loads/calls, we also see if the vuse is killed in this block. */
2094 static bool
2095 valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr,
2096 basic_block block)
2098 switch (expr->kind)
2100 case NAME:
2101 return bitmap_set_contains_expr (AVAIL_OUT (block), expr);
2102 case NARY:
2104 unsigned int i;
2105 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2106 for (i = 0; i < nary->length; i++)
2108 if (TREE_CODE (nary->op[i]) == SSA_NAME)
2110 struct pre_expr_d temp;
2111 temp.kind = NAME;
2112 temp.id = 0;
2113 PRE_EXPR_NAME (&temp) = nary->op[i];
2114 temp.id = lookup_expression_id (&temp);
2115 if (temp.id == 0)
2116 return false;
2117 if (!union_contains_value (set1, set2,
2118 get_expr_value_id (&temp)))
2119 return false;
2122 /* If the NARY may trap make sure the block does not contain
2123 a possible exit point.
2124 ??? This is overly conservative if we translate AVAIL_OUT
2125 as the available expression might be after the exit point. */
2126 if (BB_MAY_NOTRETURN (block)
2127 && vn_nary_may_trap (nary))
2128 return false;
2129 return true;
2131 break;
2132 case REFERENCE:
2134 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2135 vn_reference_op_t vro;
2136 unsigned int i;
2138 FOR_EACH_VEC_ELT (vn_reference_op_s, ref->operands, i, vro)
2140 if (!vro_valid_in_sets (set1, set2, vro))
2141 return false;
2143 if (ref->vuse)
2145 gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2146 if (!gimple_nop_p (def_stmt)
2147 && gimple_bb (def_stmt) != block
2148 && !dominated_by_p (CDI_DOMINATORS,
2149 block, gimple_bb (def_stmt)))
2150 return false;
2152 return !value_dies_in_block_x (expr, block);
2154 default:
2155 gcc_unreachable ();
2159 /* Clean the set of expressions that are no longer valid in SET1 or
2160 SET2. This means expressions that are made up of values we have no
2161 leaders for in SET1 or SET2. This version is used for partial
2162 anticipation, which means it is not valid in either ANTIC_IN or
2163 PA_IN. */
2165 static void
2166 dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block)
2168 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set1);
2169 pre_expr expr;
2170 int i;
2172 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2174 if (!valid_in_sets (set1, set2, expr, block))
2175 bitmap_remove_from_set (set1, expr);
2177 VEC_free (pre_expr, heap, exprs);
2180 /* Clean the set of expressions that are no longer valid in SET. This
2181 means expressions that are made up of values we have no leaders for
2182 in SET. */
2184 static void
2185 clean (bitmap_set_t set, basic_block block)
2187 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (set);
2188 pre_expr expr;
2189 int i;
2191 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
2193 if (!valid_in_sets (set, NULL, expr, block))
2194 bitmap_remove_from_set (set, expr);
2196 VEC_free (pre_expr, heap, exprs);
2199 static sbitmap has_abnormal_preds;
2201 /* List of blocks that may have changed during ANTIC computation and
2202 thus need to be iterated over. */
2204 static sbitmap changed_blocks;
2206 /* Decide whether to defer a block for a later iteration, or PHI
2207 translate SOURCE to DEST using phis in PHIBLOCK. Return false if we
2208 should defer the block, and true if we processed it. */
2210 static bool
2211 defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source,
2212 basic_block block, basic_block phiblock)
2214 if (!BB_VISITED (phiblock))
2216 SET_BIT (changed_blocks, block->index);
2217 BB_VISITED (block) = 0;
2218 BB_DEFERRED (block) = 1;
2219 return false;
2221 else
2222 phi_translate_set (dest, source, block, phiblock);
2223 return true;
2226 /* Compute the ANTIC set for BLOCK.
2228 If succs(BLOCK) > 1 then
2229 ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2230 else if succs(BLOCK) == 1 then
2231 ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2233 ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2236 static bool
2237 compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2239 bool changed = false;
2240 bitmap_set_t S, old, ANTIC_OUT;
2241 bitmap_iterator bi;
2242 unsigned int bii;
2243 edge e;
2244 edge_iterator ei;
2246 old = ANTIC_OUT = S = NULL;
2247 BB_VISITED (block) = 1;
2249 /* If any edges from predecessors are abnormal, antic_in is empty,
2250 so do nothing. */
2251 if (block_has_abnormal_pred_edge)
2252 goto maybe_dump_sets;
2254 old = ANTIC_IN (block);
2255 ANTIC_OUT = bitmap_set_new ();
2257 /* If the block has no successors, ANTIC_OUT is empty. */
2258 if (EDGE_COUNT (block->succs) == 0)
2260 /* If we have one successor, we could have some phi nodes to
2261 translate through. */
2262 else if (single_succ_p (block))
2264 basic_block succ_bb = single_succ (block);
2266 /* We trade iterations of the dataflow equations for having to
2267 phi translate the maximal set, which is incredibly slow
2268 (since the maximal set often has 300+ members, even when you
2269 have a small number of blocks).
2270 Basically, we defer the computation of ANTIC for this block
2271 until we have processed it's successor, which will inevitably
2272 have a *much* smaller set of values to phi translate once
2273 clean has been run on it.
2274 The cost of doing this is that we technically perform more
2275 iterations, however, they are lower cost iterations.
2277 Timings for PRE on tramp3d-v4:
2278 without maximal set fix: 11 seconds
2279 with maximal set fix/without deferring: 26 seconds
2280 with maximal set fix/with deferring: 11 seconds
2283 if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb),
2284 block, succ_bb))
2286 changed = true;
2287 goto maybe_dump_sets;
2290 /* If we have multiple successors, we take the intersection of all of
2291 them. Note that in the case of loop exit phi nodes, we may have
2292 phis to translate through. */
2293 else
2295 VEC(basic_block, heap) * worklist;
2296 size_t i;
2297 basic_block bprime, first = NULL;
2299 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2300 FOR_EACH_EDGE (e, ei, block->succs)
2302 if (!first
2303 && BB_VISITED (e->dest))
2304 first = e->dest;
2305 else if (BB_VISITED (e->dest))
2306 VEC_quick_push (basic_block, worklist, e->dest);
2309 /* Of multiple successors we have to have visited one already. */
2310 if (!first)
2312 SET_BIT (changed_blocks, block->index);
2313 BB_VISITED (block) = 0;
2314 BB_DEFERRED (block) = 1;
2315 changed = true;
2316 VEC_free (basic_block, heap, worklist);
2317 goto maybe_dump_sets;
2320 if (!gimple_seq_empty_p (phi_nodes (first)))
2321 phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first);
2322 else
2323 bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first));
2325 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2327 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2329 bitmap_set_t tmp = bitmap_set_new ();
2330 phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime);
2331 bitmap_set_and (ANTIC_OUT, tmp);
2332 bitmap_set_free (tmp);
2334 else
2335 bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime));
2337 VEC_free (basic_block, heap, worklist);
2340 /* Generate ANTIC_OUT - TMP_GEN. */
2341 S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block));
2343 /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2344 ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block),
2345 TMP_GEN (block));
2347 /* Then union in the ANTIC_OUT - TMP_GEN values,
2348 to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2349 FOR_EACH_EXPR_ID_IN_SET (S, bii, bi)
2350 bitmap_value_insert_into_set (ANTIC_IN (block),
2351 expression_for_id (bii));
2353 clean (ANTIC_IN (block), block);
2355 if (!bitmap_set_equal (old, ANTIC_IN (block)))
2357 changed = true;
2358 SET_BIT (changed_blocks, block->index);
2359 FOR_EACH_EDGE (e, ei, block->preds)
2360 SET_BIT (changed_blocks, e->src->index);
2362 else
2363 RESET_BIT (changed_blocks, block->index);
2365 maybe_dump_sets:
2366 if (dump_file && (dump_flags & TDF_DETAILS))
2368 if (!BB_DEFERRED (block) || BB_VISITED (block))
2370 if (ANTIC_OUT)
2371 print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2373 print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2374 block->index);
2376 if (S)
2377 print_bitmap_set (dump_file, S, "S", block->index);
2379 else
2381 fprintf (dump_file,
2382 "Block %d was deferred for a future iteration.\n",
2383 block->index);
2386 if (old)
2387 bitmap_set_free (old);
2388 if (S)
2389 bitmap_set_free (S);
2390 if (ANTIC_OUT)
2391 bitmap_set_free (ANTIC_OUT);
2392 return changed;
2395 /* Compute PARTIAL_ANTIC for BLOCK.
2397 If succs(BLOCK) > 1 then
2398 PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2399 in ANTIC_OUT for all succ(BLOCK)
2400 else if succs(BLOCK) == 1 then
2401 PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2403 PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK]
2404 - ANTIC_IN[BLOCK])
2407 static bool
2408 compute_partial_antic_aux (basic_block block,
2409 bool block_has_abnormal_pred_edge)
2411 bool changed = false;
2412 bitmap_set_t old_PA_IN;
2413 bitmap_set_t PA_OUT;
2414 edge e;
2415 edge_iterator ei;
2416 unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH);
2418 old_PA_IN = PA_OUT = NULL;
2420 /* If any edges from predecessors are abnormal, antic_in is empty,
2421 so do nothing. */
2422 if (block_has_abnormal_pred_edge)
2423 goto maybe_dump_sets;
2425 /* If there are too many partially anticipatable values in the
2426 block, phi_translate_set can take an exponential time: stop
2427 before the translation starts. */
2428 if (max_pa
2429 && single_succ_p (block)
2430 && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2431 goto maybe_dump_sets;
2433 old_PA_IN = PA_IN (block);
2434 PA_OUT = bitmap_set_new ();
2436 /* If the block has no successors, ANTIC_OUT is empty. */
2437 if (EDGE_COUNT (block->succs) == 0)
2439 /* If we have one successor, we could have some phi nodes to
2440 translate through. Note that we can't phi translate across DFS
2441 back edges in partial antic, because it uses a union operation on
2442 the successors. For recurrences like IV's, we will end up
2443 generating a new value in the set on each go around (i + 3 (VH.1)
2444 VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2445 else if (single_succ_p (block))
2447 basic_block succ = single_succ (block);
2448 if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK))
2449 phi_translate_set (PA_OUT, PA_IN (succ), block, succ);
2451 /* If we have multiple successors, we take the union of all of
2452 them. */
2453 else
2455 VEC(basic_block, heap) * worklist;
2456 size_t i;
2457 basic_block bprime;
2459 worklist = VEC_alloc (basic_block, heap, EDGE_COUNT (block->succs));
2460 FOR_EACH_EDGE (e, ei, block->succs)
2462 if (e->flags & EDGE_DFS_BACK)
2463 continue;
2464 VEC_quick_push (basic_block, worklist, e->dest);
2466 if (VEC_length (basic_block, worklist) > 0)
2468 FOR_EACH_VEC_ELT (basic_block, worklist, i, bprime)
2470 unsigned int i;
2471 bitmap_iterator bi;
2473 FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi)
2474 bitmap_value_insert_into_set (PA_OUT,
2475 expression_for_id (i));
2476 if (!gimple_seq_empty_p (phi_nodes (bprime)))
2478 bitmap_set_t pa_in = bitmap_set_new ();
2479 phi_translate_set (pa_in, PA_IN (bprime), block, bprime);
2480 FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2481 bitmap_value_insert_into_set (PA_OUT,
2482 expression_for_id (i));
2483 bitmap_set_free (pa_in);
2485 else
2486 FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi)
2487 bitmap_value_insert_into_set (PA_OUT,
2488 expression_for_id (i));
2491 VEC_free (basic_block, heap, worklist);
2494 /* PA_IN starts with PA_OUT - TMP_GEN.
2495 Then we subtract things from ANTIC_IN. */
2496 PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block));
2498 /* For partial antic, we want to put back in the phi results, since
2499 we will properly avoid making them partially antic over backedges. */
2500 bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2501 bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2503 /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2504 bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2506 dependent_clean (PA_IN (block), ANTIC_IN (block), block);
2508 if (!bitmap_set_equal (old_PA_IN, PA_IN (block)))
2510 changed = true;
2511 SET_BIT (changed_blocks, block->index);
2512 FOR_EACH_EDGE (e, ei, block->preds)
2513 SET_BIT (changed_blocks, e->src->index);
2515 else
2516 RESET_BIT (changed_blocks, block->index);
2518 maybe_dump_sets:
2519 if (dump_file && (dump_flags & TDF_DETAILS))
2521 if (PA_OUT)
2522 print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2524 print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2526 if (old_PA_IN)
2527 bitmap_set_free (old_PA_IN);
2528 if (PA_OUT)
2529 bitmap_set_free (PA_OUT);
2530 return changed;
2533 /* Compute ANTIC and partial ANTIC sets. */
2535 static void
2536 compute_antic (void)
2538 bool changed = true;
2539 int num_iterations = 0;
2540 basic_block block;
2541 int i;
2543 /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2544 We pre-build the map of blocks with incoming abnormal edges here. */
2545 has_abnormal_preds = sbitmap_alloc (last_basic_block);
2546 sbitmap_zero (has_abnormal_preds);
2548 FOR_EACH_BB (block)
2550 edge_iterator ei;
2551 edge e;
2553 FOR_EACH_EDGE (e, ei, block->preds)
2555 e->flags &= ~EDGE_DFS_BACK;
2556 if (e->flags & EDGE_ABNORMAL)
2558 SET_BIT (has_abnormal_preds, block->index);
2559 break;
2563 BB_VISITED (block) = 0;
2564 BB_DEFERRED (block) = 0;
2566 /* While we are here, give empty ANTIC_IN sets to each block. */
2567 ANTIC_IN (block) = bitmap_set_new ();
2568 PA_IN (block) = bitmap_set_new ();
2571 /* At the exit block we anticipate nothing. */
2572 ANTIC_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2573 BB_VISITED (EXIT_BLOCK_PTR) = 1;
2574 PA_IN (EXIT_BLOCK_PTR) = bitmap_set_new ();
2576 changed_blocks = sbitmap_alloc (last_basic_block + 1);
2577 sbitmap_ones (changed_blocks);
2578 while (changed)
2580 if (dump_file && (dump_flags & TDF_DETAILS))
2581 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2582 /* ??? We need to clear our PHI translation cache here as the
2583 ANTIC sets shrink and we restrict valid translations to
2584 those having operands with leaders in ANTIC. Same below
2585 for PA ANTIC computation. */
2586 num_iterations++;
2587 changed = false;
2588 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1; i >= 0; i--)
2590 if (TEST_BIT (changed_blocks, postorder[i]))
2592 basic_block block = BASIC_BLOCK (postorder[i]);
2593 changed |= compute_antic_aux (block,
2594 TEST_BIT (has_abnormal_preds,
2595 block->index));
2598 /* Theoretically possible, but *highly* unlikely. */
2599 gcc_checking_assert (num_iterations < 500);
2602 statistics_histogram_event (cfun, "compute_antic iterations",
2603 num_iterations);
2605 if (do_partial_partial)
2607 sbitmap_ones (changed_blocks);
2608 mark_dfs_back_edges ();
2609 num_iterations = 0;
2610 changed = true;
2611 while (changed)
2613 if (dump_file && (dump_flags & TDF_DETAILS))
2614 fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2615 num_iterations++;
2616 changed = false;
2617 for (i = n_basic_blocks - NUM_FIXED_BLOCKS - 1 ; i >= 0; i--)
2619 if (TEST_BIT (changed_blocks, postorder[i]))
2621 basic_block block = BASIC_BLOCK (postorder[i]);
2622 changed
2623 |= compute_partial_antic_aux (block,
2624 TEST_BIT (has_abnormal_preds,
2625 block->index));
2628 /* Theoretically possible, but *highly* unlikely. */
2629 gcc_checking_assert (num_iterations < 500);
2631 statistics_histogram_event (cfun, "compute_partial_antic iterations",
2632 num_iterations);
2634 sbitmap_free (has_abnormal_preds);
2635 sbitmap_free (changed_blocks);
2638 /* Return true if we can value number the call in STMT. This is true
2639 if we have a pure or constant call to a real function. */
2641 static bool
2642 can_value_number_call (gimple stmt)
2644 if (gimple_call_internal_p (stmt))
2645 return false;
2646 if (gimple_call_flags (stmt) & (ECF_PURE | ECF_CONST))
2647 return true;
2648 return false;
2651 /* Return true if OP is a tree which we can perform PRE on.
2652 This may not match the operations we can value number, but in
2653 a perfect world would. */
2655 static bool
2656 can_PRE_operation (tree op)
2658 return UNARY_CLASS_P (op)
2659 || BINARY_CLASS_P (op)
2660 || COMPARISON_CLASS_P (op)
2661 || TREE_CODE (op) == MEM_REF
2662 || TREE_CODE (op) == COMPONENT_REF
2663 || TREE_CODE (op) == VIEW_CONVERT_EXPR
2664 || TREE_CODE (op) == CALL_EXPR
2665 || TREE_CODE (op) == ARRAY_REF;
2669 /* Inserted expressions are placed onto this worklist, which is used
2670 for performing quick dead code elimination of insertions we made
2671 that didn't turn out to be necessary. */
2672 static bitmap inserted_exprs;
2674 /* Pool allocated fake store expressions are placed onto this
2675 worklist, which, after performing dead code elimination, is walked
2676 to see which expressions need to be put into GC'able memory */
2677 static VEC(gimple, heap) *need_creation;
2679 /* The actual worker for create_component_ref_by_pieces. */
2681 static tree
2682 create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2683 unsigned int *operand, gimple_seq *stmts,
2684 gimple domstmt)
2686 vn_reference_op_t currop = VEC_index (vn_reference_op_s, ref->operands,
2687 *operand);
2688 tree genop;
2689 ++*operand;
2690 switch (currop->opcode)
2692 case CALL_EXPR:
2694 tree folded, sc = NULL_TREE;
2695 unsigned int nargs = 0;
2696 tree fn, *args;
2697 if (TREE_CODE (currop->op0) == FUNCTION_DECL)
2698 fn = currop->op0;
2699 else
2701 pre_expr op0 = get_or_alloc_expr_for (currop->op0);
2702 fn = find_or_generate_expression (block, op0, stmts, domstmt);
2703 if (!fn)
2704 return NULL_TREE;
2706 if (currop->op1)
2708 pre_expr scexpr = get_or_alloc_expr_for (currop->op1);
2709 sc = find_or_generate_expression (block, scexpr, stmts, domstmt);
2710 if (!sc)
2711 return NULL_TREE;
2713 args = XNEWVEC (tree, VEC_length (vn_reference_op_s,
2714 ref->operands) - 1);
2715 while (*operand < VEC_length (vn_reference_op_s, ref->operands))
2717 args[nargs] = create_component_ref_by_pieces_1 (block, ref,
2718 operand, stmts,
2719 domstmt);
2720 if (!args[nargs])
2722 free (args);
2723 return NULL_TREE;
2725 nargs++;
2727 folded = build_call_array (currop->type,
2728 (TREE_CODE (fn) == FUNCTION_DECL
2729 ? build_fold_addr_expr (fn) : fn),
2730 nargs, args);
2731 free (args);
2732 if (sc)
2733 CALL_EXPR_STATIC_CHAIN (folded) = sc;
2734 return folded;
2736 break;
2737 case MEM_REF:
2739 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2740 stmts, domstmt);
2741 tree offset = currop->op0;
2742 if (!baseop)
2743 return NULL_TREE;
2744 if (TREE_CODE (baseop) == ADDR_EXPR
2745 && handled_component_p (TREE_OPERAND (baseop, 0)))
2747 HOST_WIDE_INT off;
2748 tree base;
2749 base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2750 &off);
2751 gcc_assert (base);
2752 offset = int_const_binop (PLUS_EXPR, offset,
2753 build_int_cst (TREE_TYPE (offset),
2754 off));
2755 baseop = build_fold_addr_expr (base);
2757 return fold_build2 (MEM_REF, currop->type, baseop, offset);
2759 break;
2760 case TARGET_MEM_REF:
2762 pre_expr op0expr, op1expr;
2763 tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2764 vn_reference_op_t nextop = VEC_index (vn_reference_op_s, ref->operands,
2765 ++*operand);
2766 tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2767 stmts, domstmt);
2768 if (!baseop)
2769 return NULL_TREE;
2770 if (currop->op0)
2772 op0expr = get_or_alloc_expr_for (currop->op0);
2773 genop0 = find_or_generate_expression (block, op0expr,
2774 stmts, domstmt);
2775 if (!genop0)
2776 return NULL_TREE;
2778 if (nextop->op0)
2780 op1expr = get_or_alloc_expr_for (nextop->op0);
2781 genop1 = find_or_generate_expression (block, op1expr,
2782 stmts, domstmt);
2783 if (!genop1)
2784 return NULL_TREE;
2786 return build5 (TARGET_MEM_REF, currop->type,
2787 baseop, currop->op2, genop0, currop->op1, genop1);
2789 break;
2790 case ADDR_EXPR:
2791 if (currop->op0)
2793 gcc_assert (is_gimple_min_invariant (currop->op0));
2794 return currop->op0;
2796 /* Fallthrough. */
2797 case REALPART_EXPR:
2798 case IMAGPART_EXPR:
2799 case VIEW_CONVERT_EXPR:
2801 tree folded;
2802 tree genop0 = create_component_ref_by_pieces_1 (block, ref,
2803 operand,
2804 stmts, domstmt);
2805 if (!genop0)
2806 return NULL_TREE;
2807 folded = fold_build1 (currop->opcode, currop->type,
2808 genop0);
2809 return folded;
2811 break;
2812 case BIT_FIELD_REF:
2814 tree folded;
2815 tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2816 stmts, domstmt);
2817 pre_expr op1expr = get_or_alloc_expr_for (currop->op0);
2818 pre_expr op2expr = get_or_alloc_expr_for (currop->op1);
2819 tree genop1;
2820 tree genop2;
2822 if (!genop0)
2823 return NULL_TREE;
2824 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2825 if (!genop1)
2826 return NULL_TREE;
2827 genop2 = find_or_generate_expression (block, op2expr, stmts, domstmt);
2828 if (!genop2)
2829 return NULL_TREE;
2830 folded = fold_build3 (BIT_FIELD_REF, currop->type, genop0, genop1,
2831 genop2);
2832 return folded;
2835 /* For array ref vn_reference_op's, operand 1 of the array ref
2836 is op0 of the reference op and operand 3 of the array ref is
2837 op1. */
2838 case ARRAY_RANGE_REF:
2839 case ARRAY_REF:
2841 tree genop0;
2842 tree genop1 = currop->op0;
2843 pre_expr op1expr;
2844 tree genop2 = currop->op1;
2845 pre_expr op2expr;
2846 tree genop3 = currop->op2;
2847 pre_expr op3expr;
2848 genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2849 stmts, domstmt);
2850 if (!genop0)
2851 return NULL_TREE;
2852 op1expr = get_or_alloc_expr_for (genop1);
2853 genop1 = find_or_generate_expression (block, op1expr, stmts, domstmt);
2854 if (!genop1)
2855 return NULL_TREE;
2856 if (genop2)
2858 tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2859 /* Drop zero minimum index if redundant. */
2860 if (integer_zerop (genop2)
2861 && (!domain_type
2862 || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2863 genop2 = NULL_TREE;
2864 else
2866 op2expr = get_or_alloc_expr_for (genop2);
2867 genop2 = find_or_generate_expression (block, op2expr, stmts,
2868 domstmt);
2869 if (!genop2)
2870 return NULL_TREE;
2873 if (genop3)
2875 tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2876 /* We can't always put a size in units of the element alignment
2877 here as the element alignment may be not visible. See
2878 PR43783. Simply drop the element size for constant
2879 sizes. */
2880 if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type)))
2881 genop3 = NULL_TREE;
2882 else
2884 genop3 = size_binop (EXACT_DIV_EXPR, genop3,
2885 size_int (TYPE_ALIGN_UNIT (elmt_type)));
2886 op3expr = get_or_alloc_expr_for (genop3);
2887 genop3 = find_or_generate_expression (block, op3expr, stmts,
2888 domstmt);
2889 if (!genop3)
2890 return NULL_TREE;
2893 return build4 (currop->opcode, currop->type, genop0, genop1,
2894 genop2, genop3);
2896 case COMPONENT_REF:
2898 tree op0;
2899 tree op1;
2900 tree genop2 = currop->op1;
2901 pre_expr op2expr;
2902 op0 = create_component_ref_by_pieces_1 (block, ref, operand,
2903 stmts, domstmt);
2904 if (!op0)
2905 return NULL_TREE;
2906 /* op1 should be a FIELD_DECL, which are represented by
2907 themselves. */
2908 op1 = currop->op0;
2909 if (genop2)
2911 op2expr = get_or_alloc_expr_for (genop2);
2912 genop2 = find_or_generate_expression (block, op2expr, stmts,
2913 domstmt);
2914 if (!genop2)
2915 return NULL_TREE;
2918 return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1,
2919 genop2);
2921 break;
2922 case SSA_NAME:
2924 pre_expr op0expr = get_or_alloc_expr_for (currop->op0);
2925 genop = find_or_generate_expression (block, op0expr, stmts, domstmt);
2926 return genop;
2928 case STRING_CST:
2929 case INTEGER_CST:
2930 case COMPLEX_CST:
2931 case VECTOR_CST:
2932 case REAL_CST:
2933 case CONSTRUCTOR:
2934 case VAR_DECL:
2935 case PARM_DECL:
2936 case CONST_DECL:
2937 case RESULT_DECL:
2938 case FUNCTION_DECL:
2939 return currop->op0;
2941 default:
2942 gcc_unreachable ();
2946 /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2947 COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2948 trying to rename aggregates into ssa form directly, which is a no no.
2950 Thus, this routine doesn't create temporaries, it just builds a
2951 single access expression for the array, calling
2952 find_or_generate_expression to build the innermost pieces.
2954 This function is a subroutine of create_expression_by_pieces, and
2955 should not be called on it's own unless you really know what you
2956 are doing. */
2958 static tree
2959 create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2960 gimple_seq *stmts, gimple domstmt)
2962 unsigned int op = 0;
2963 return create_component_ref_by_pieces_1 (block, ref, &op, stmts, domstmt);
2966 /* Find a leader for an expression, or generate one using
2967 create_expression_by_pieces if it's ANTIC but
2968 complex.
2969 BLOCK is the basic_block we are looking for leaders in.
2970 EXPR is the expression to find a leader or generate for.
2971 STMTS is the statement list to put the inserted expressions on.
2972 Returns the SSA_NAME of the LHS of the generated expression or the
2973 leader.
2974 DOMSTMT if non-NULL is a statement that should be dominated by
2975 all uses in the generated expression. If DOMSTMT is non-NULL this
2976 routine can fail and return NULL_TREE. Otherwise it will assert
2977 on failure. */
2979 static tree
2980 find_or_generate_expression (basic_block block, pre_expr expr,
2981 gimple_seq *stmts, gimple domstmt)
2983 pre_expr leader = bitmap_find_leader (AVAIL_OUT (block),
2984 get_expr_value_id (expr), domstmt);
2985 tree genop = NULL;
2986 if (leader)
2988 if (leader->kind == NAME)
2989 genop = PRE_EXPR_NAME (leader);
2990 else if (leader->kind == CONSTANT)
2991 genop = PRE_EXPR_CONSTANT (leader);
2994 /* If it's still NULL, it must be a complex expression, so generate
2995 it recursively. Not so if inserting expressions for values generated
2996 by SCCVN. */
2997 if (genop == NULL
2998 && !domstmt)
3000 bitmap_set_t exprset;
3001 unsigned int lookfor = get_expr_value_id (expr);
3002 bool handled = false;
3003 bitmap_iterator bi;
3004 unsigned int i;
3006 exprset = VEC_index (bitmap_set_t, value_expressions, lookfor);
3007 FOR_EACH_EXPR_ID_IN_SET (exprset, i, bi)
3009 pre_expr temp = expression_for_id (i);
3010 if (temp->kind != NAME)
3012 handled = true;
3013 genop = create_expression_by_pieces (block, temp, stmts,
3014 domstmt,
3015 get_expr_type (expr));
3016 break;
3019 if (!handled && domstmt)
3020 return NULL_TREE;
3022 gcc_assert (handled);
3024 return genop;
3027 #define NECESSARY GF_PLF_1
3029 /* Create an expression in pieces, so that we can handle very complex
3030 expressions that may be ANTIC, but not necessary GIMPLE.
3031 BLOCK is the basic block the expression will be inserted into,
3032 EXPR is the expression to insert (in value form)
3033 STMTS is a statement list to append the necessary insertions into.
3035 This function will die if we hit some value that shouldn't be
3036 ANTIC but is (IE there is no leader for it, or its components).
3037 This function may also generate expressions that are themselves
3038 partially or fully redundant. Those that are will be either made
3039 fully redundant during the next iteration of insert (for partially
3040 redundant ones), or eliminated by eliminate (for fully redundant
3041 ones).
3043 If DOMSTMT is non-NULL then we make sure that all uses in the
3044 expressions dominate that statement. In this case the function
3045 can return NULL_TREE to signal failure. */
3047 static tree
3048 create_expression_by_pieces (basic_block block, pre_expr expr,
3049 gimple_seq *stmts, gimple domstmt, tree type)
3051 tree temp, name;
3052 tree folded;
3053 gimple_seq forced_stmts = NULL;
3054 unsigned int value_id;
3055 gimple_stmt_iterator gsi;
3056 tree exprtype = type ? type : get_expr_type (expr);
3057 pre_expr nameexpr;
3058 gimple newstmt;
3060 switch (expr->kind)
3062 /* We may hit the NAME/CONSTANT case if we have to convert types
3063 that value numbering saw through. */
3064 case NAME:
3065 folded = PRE_EXPR_NAME (expr);
3066 break;
3067 case CONSTANT:
3068 folded = PRE_EXPR_CONSTANT (expr);
3069 break;
3070 case REFERENCE:
3072 vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
3073 folded = create_component_ref_by_pieces (block, ref, stmts, domstmt);
3075 break;
3076 case NARY:
3078 vn_nary_op_t nary = PRE_EXPR_NARY (expr);
3079 tree genop[4];
3080 unsigned i;
3081 for (i = 0; i < nary->length; ++i)
3083 pre_expr op = get_or_alloc_expr_for (nary->op[i]);
3084 genop[i] = find_or_generate_expression (block, op,
3085 stmts, domstmt);
3086 if (!genop[i])
3087 return NULL_TREE;
3088 /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
3089 may have conversions stripped. */
3090 if (nary->opcode == POINTER_PLUS_EXPR)
3092 if (i == 0)
3093 genop[i] = fold_convert (nary->type, genop[i]);
3094 else if (i == 1)
3095 genop[i] = convert_to_ptrofftype (genop[i]);
3097 else
3098 genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]);
3100 if (nary->opcode == CONSTRUCTOR)
3102 VEC(constructor_elt,gc) *elts = NULL;
3103 for (i = 0; i < nary->length; ++i)
3104 CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
3105 folded = build_constructor (nary->type, elts);
3107 else
3109 switch (nary->length)
3111 case 1:
3112 folded = fold_build1 (nary->opcode, nary->type,
3113 genop[0]);
3114 break;
3115 case 2:
3116 folded = fold_build2 (nary->opcode, nary->type,
3117 genop[0], genop[1]);
3118 break;
3119 case 3:
3120 folded = fold_build3 (nary->opcode, nary->type,
3121 genop[0], genop[1], genop[3]);
3122 break;
3123 default:
3124 gcc_unreachable ();
3128 break;
3129 default:
3130 return NULL_TREE;
3133 if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
3134 folded = fold_convert (exprtype, folded);
3136 /* Force the generated expression to be a sequence of GIMPLE
3137 statements.
3138 We have to call unshare_expr because force_gimple_operand may
3139 modify the tree we pass to it. */
3140 folded = force_gimple_operand (unshare_expr (folded), &forced_stmts,
3141 false, NULL);
3143 /* If we have any intermediate expressions to the value sets, add them
3144 to the value sets and chain them in the instruction stream. */
3145 if (forced_stmts)
3147 gsi = gsi_start (forced_stmts);
3148 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3150 gimple stmt = gsi_stmt (gsi);
3151 tree forcedname = gimple_get_lhs (stmt);
3152 pre_expr nameexpr;
3154 if (TREE_CODE (forcedname) == SSA_NAME)
3156 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3157 VN_INFO_GET (forcedname)->valnum = forcedname;
3158 VN_INFO (forcedname)->value_id = get_next_value_id ();
3159 nameexpr = get_or_alloc_expr_for_name (forcedname);
3160 add_to_value (VN_INFO (forcedname)->value_id, nameexpr);
3161 if (!in_fre)
3162 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3163 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3165 mark_symbols_for_renaming (stmt);
3167 gimple_seq_add_seq (stmts, forced_stmts);
3170 /* Build and insert the assignment of the end result to the temporary
3171 that we will return. */
3172 if (!pretemp || exprtype != TREE_TYPE (pretemp))
3173 pretemp = create_tmp_reg (exprtype, "pretmp");
3175 temp = pretemp;
3176 add_referenced_var (temp);
3178 newstmt = gimple_build_assign (temp, folded);
3179 name = make_ssa_name (temp, newstmt);
3180 gimple_assign_set_lhs (newstmt, name);
3181 gimple_set_plf (newstmt, NECESSARY, false);
3183 gimple_seq_add_stmt (stmts, newstmt);
3184 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
3186 /* All the symbols in NEWEXPR should be put into SSA form. */
3187 mark_symbols_for_renaming (newstmt);
3189 /* Fold the last statement. */
3190 gsi = gsi_last (*stmts);
3191 if (fold_stmt_inplace (&gsi))
3192 update_stmt (gsi_stmt (gsi));
3194 /* Add a value number to the temporary.
3195 The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3196 we are creating the expression by pieces, and this particular piece of
3197 the expression may have been represented. There is no harm in replacing
3198 here. */
3199 VN_INFO_GET (name)->valnum = name;
3200 value_id = get_expr_value_id (expr);
3201 VN_INFO (name)->value_id = value_id;
3202 nameexpr = get_or_alloc_expr_for_name (name);
3203 add_to_value (value_id, nameexpr);
3204 if (NEW_SETS (block))
3205 bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3206 bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3208 pre_stats.insertions++;
3209 if (dump_file && (dump_flags & TDF_DETAILS))
3211 fprintf (dump_file, "Inserted ");
3212 print_gimple_stmt (dump_file, newstmt, 0, 0);
3213 fprintf (dump_file, " in predecessor %d\n", block->index);
3216 return name;
3220 /* Returns true if we want to inhibit the insertions of PHI nodes
3221 for the given EXPR for basic block BB (a member of a loop).
3222 We want to do this, when we fear that the induction variable we
3223 create might inhibit vectorization. */
3225 static bool
3226 inhibit_phi_insertion (basic_block bb, pre_expr expr)
3228 vn_reference_t vr = PRE_EXPR_REFERENCE (expr);
3229 VEC (vn_reference_op_s, heap) *ops = vr->operands;
3230 vn_reference_op_t op;
3231 unsigned i;
3233 /* If we aren't going to vectorize we don't inhibit anything. */
3234 if (!flag_tree_vectorize)
3235 return false;
3237 /* Otherwise we inhibit the insertion when the address of the
3238 memory reference is a simple induction variable. In other
3239 cases the vectorizer won't do anything anyway (either it's
3240 loop invariant or a complicated expression). */
3241 FOR_EACH_VEC_ELT (vn_reference_op_s, ops, i, op)
3243 switch (op->opcode)
3245 case ARRAY_REF:
3246 case ARRAY_RANGE_REF:
3247 if (TREE_CODE (op->op0) != SSA_NAME)
3248 break;
3249 /* Fallthru. */
3250 case SSA_NAME:
3252 basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0));
3253 affine_iv iv;
3254 /* Default defs are loop invariant. */
3255 if (!defbb)
3256 break;
3257 /* Defined outside this loop, also loop invariant. */
3258 if (!flow_bb_inside_loop_p (bb->loop_father, defbb))
3259 break;
3260 /* If it's a simple induction variable inhibit insertion,
3261 the vectorizer might be interested in this one. */
3262 if (simple_iv (bb->loop_father, bb->loop_father,
3263 op->op0, &iv, true))
3264 return true;
3265 /* No simple IV, vectorizer can't do anything, hence no
3266 reason to inhibit the transformation for this operand. */
3267 break;
3269 default:
3270 break;
3273 return false;
3276 /* Insert the to-be-made-available values of expression EXPRNUM for each
3277 predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3278 merge the result with a phi node, given the same value number as
3279 NODE. Return true if we have inserted new stuff. */
3281 static bool
3282 insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3283 pre_expr *avail)
3285 pre_expr expr = expression_for_id (exprnum);
3286 pre_expr newphi;
3287 unsigned int val = get_expr_value_id (expr);
3288 edge pred;
3289 bool insertions = false;
3290 bool nophi = false;
3291 basic_block bprime;
3292 pre_expr eprime;
3293 edge_iterator ei;
3294 tree type = get_expr_type (expr);
3295 tree temp;
3296 gimple phi;
3298 if (dump_file && (dump_flags & TDF_DETAILS))
3300 fprintf (dump_file, "Found partial redundancy for expression ");
3301 print_pre_expr (dump_file, expr);
3302 fprintf (dump_file, " (%04d)\n", val);
3305 /* Make sure we aren't creating an induction variable. */
3306 if (block->loop_depth > 0 && EDGE_COUNT (block->preds) == 2)
3308 bool firstinsideloop = false;
3309 bool secondinsideloop = false;
3310 firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3311 EDGE_PRED (block, 0)->src);
3312 secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3313 EDGE_PRED (block, 1)->src);
3314 /* Induction variables only have one edge inside the loop. */
3315 if ((firstinsideloop ^ secondinsideloop)
3316 && (expr->kind != REFERENCE
3317 || inhibit_phi_insertion (block, expr)))
3319 if (dump_file && (dump_flags & TDF_DETAILS))
3320 fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n");
3321 nophi = true;
3325 /* Make the necessary insertions. */
3326 FOR_EACH_EDGE (pred, ei, block->preds)
3328 gimple_seq stmts = NULL;
3329 tree builtexpr;
3330 bprime = pred->src;
3331 eprime = avail[bprime->index];
3333 if (eprime->kind != NAME && eprime->kind != CONSTANT)
3335 builtexpr = create_expression_by_pieces (bprime,
3336 eprime,
3337 &stmts, NULL,
3338 type);
3339 gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3340 gsi_insert_seq_on_edge (pred, stmts);
3341 avail[bprime->index] = get_or_alloc_expr_for_name (builtexpr);
3342 insertions = true;
3344 else if (eprime->kind == CONSTANT)
3346 /* Constants may not have the right type, fold_convert
3347 should give us back a constant with the right type.
3349 tree constant = PRE_EXPR_CONSTANT (eprime);
3350 if (!useless_type_conversion_p (type, TREE_TYPE (constant)))
3352 tree builtexpr = fold_convert (type, constant);
3353 if (!is_gimple_min_invariant (builtexpr))
3355 tree forcedexpr = force_gimple_operand (builtexpr,
3356 &stmts, true,
3357 NULL);
3358 if (!is_gimple_min_invariant (forcedexpr))
3360 if (forcedexpr != builtexpr)
3362 VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime);
3363 VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime);
3365 if (stmts)
3367 gimple_stmt_iterator gsi;
3368 gsi = gsi_start (stmts);
3369 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3371 gimple stmt = gsi_stmt (gsi);
3372 tree lhs = gimple_get_lhs (stmt);
3373 if (TREE_CODE (lhs) == SSA_NAME)
3374 bitmap_set_bit (inserted_exprs,
3375 SSA_NAME_VERSION (lhs));
3376 gimple_set_plf (stmt, NECESSARY, false);
3378 gsi_insert_seq_on_edge (pred, stmts);
3380 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3383 else
3384 avail[bprime->index] = get_or_alloc_expr_for_constant (builtexpr);
3387 else if (eprime->kind == NAME)
3389 /* We may have to do a conversion because our value
3390 numbering can look through types in certain cases, but
3391 our IL requires all operands of a phi node have the same
3392 type. */
3393 tree name = PRE_EXPR_NAME (eprime);
3394 if (!useless_type_conversion_p (type, TREE_TYPE (name)))
3396 tree builtexpr;
3397 tree forcedexpr;
3398 builtexpr = fold_convert (type, name);
3399 forcedexpr = force_gimple_operand (builtexpr,
3400 &stmts, true,
3401 NULL);
3403 if (forcedexpr != name)
3405 VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum;
3406 VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id;
3409 if (stmts)
3411 gimple_stmt_iterator gsi;
3412 gsi = gsi_start (stmts);
3413 for (; !gsi_end_p (gsi); gsi_next (&gsi))
3415 gimple stmt = gsi_stmt (gsi);
3416 tree lhs = gimple_get_lhs (stmt);
3417 if (TREE_CODE (lhs) == SSA_NAME)
3418 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
3419 gimple_set_plf (stmt, NECESSARY, false);
3421 gsi_insert_seq_on_edge (pred, stmts);
3423 avail[bprime->index] = get_or_alloc_expr_for_name (forcedexpr);
3427 /* If we didn't want a phi node, and we made insertions, we still have
3428 inserted new stuff, and thus return true. If we didn't want a phi node,
3429 and didn't make insertions, we haven't added anything new, so return
3430 false. */
3431 if (nophi && insertions)
3432 return true;
3433 else if (nophi && !insertions)
3434 return false;
3436 /* Now build a phi for the new variable. */
3437 if (!prephitemp || TREE_TYPE (prephitemp) != type)
3438 prephitemp = create_tmp_var (type, "prephitmp");
3440 temp = prephitemp;
3441 add_referenced_var (temp);
3443 if (TREE_CODE (type) == COMPLEX_TYPE
3444 || TREE_CODE (type) == VECTOR_TYPE)
3445 DECL_GIMPLE_REG_P (temp) = 1;
3446 phi = create_phi_node (temp, block);
3448 gimple_set_plf (phi, NECESSARY, false);
3449 VN_INFO_GET (gimple_phi_result (phi))->valnum = gimple_phi_result (phi);
3450 VN_INFO (gimple_phi_result (phi))->value_id = val;
3451 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (gimple_phi_result (phi)));
3452 FOR_EACH_EDGE (pred, ei, block->preds)
3454 pre_expr ae = avail[pred->src->index];
3455 gcc_assert (get_expr_type (ae) == type
3456 || useless_type_conversion_p (type, get_expr_type (ae)));
3457 if (ae->kind == CONSTANT)
3458 add_phi_arg (phi, PRE_EXPR_CONSTANT (ae), pred, UNKNOWN_LOCATION);
3459 else
3460 add_phi_arg (phi, PRE_EXPR_NAME (avail[pred->src->index]), pred,
3461 UNKNOWN_LOCATION);
3464 newphi = get_or_alloc_expr_for_name (gimple_phi_result (phi));
3465 add_to_value (val, newphi);
3467 /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3468 this insertion, since we test for the existence of this value in PHI_GEN
3469 before proceeding with the partial redundancy checks in insert_aux.
3471 The value may exist in AVAIL_OUT, in particular, it could be represented
3472 by the expression we are trying to eliminate, in which case we want the
3473 replacement to occur. If it's not existing in AVAIL_OUT, we want it
3474 inserted there.
3476 Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3477 this block, because if it did, it would have existed in our dominator's
3478 AVAIL_OUT, and would have been skipped due to the full redundancy check.
3481 bitmap_insert_into_set (PHI_GEN (block), newphi);
3482 bitmap_value_replace_in_set (AVAIL_OUT (block),
3483 newphi);
3484 bitmap_insert_into_set (NEW_SETS (block),
3485 newphi);
3487 if (dump_file && (dump_flags & TDF_DETAILS))
3489 fprintf (dump_file, "Created phi ");
3490 print_gimple_stmt (dump_file, phi, 0, 0);
3491 fprintf (dump_file, " in block %d\n", block->index);
3493 pre_stats.phis++;
3494 return true;
3499 /* Perform insertion of partially redundant values.
3500 For BLOCK, do the following:
3501 1. Propagate the NEW_SETS of the dominator into the current block.
3502 If the block has multiple predecessors,
3503 2a. Iterate over the ANTIC expressions for the block to see if
3504 any of them are partially redundant.
3505 2b. If so, insert them into the necessary predecessors to make
3506 the expression fully redundant.
3507 2c. Insert a new PHI merging the values of the predecessors.
3508 2d. Insert the new PHI, and the new expressions, into the
3509 NEW_SETS set.
3510 3. Recursively call ourselves on the dominator children of BLOCK.
3512 Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by
3513 do_regular_insertion and do_partial_insertion.
3517 static bool
3518 do_regular_insertion (basic_block block, basic_block dom)
3520 bool new_stuff = false;
3521 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (ANTIC_IN (block));
3522 pre_expr expr;
3523 int i;
3525 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3527 if (expr->kind != NAME)
3529 pre_expr *avail;
3530 unsigned int val;
3531 bool by_some = false;
3532 bool cant_insert = false;
3533 bool all_same = true;
3534 pre_expr first_s = NULL;
3535 edge pred;
3536 basic_block bprime;
3537 pre_expr eprime = NULL;
3538 edge_iterator ei;
3539 pre_expr edoubleprime = NULL;
3540 bool do_insertion = false;
3542 val = get_expr_value_id (expr);
3543 if (bitmap_set_contains_value (PHI_GEN (block), val))
3544 continue;
3545 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3547 if (dump_file && (dump_flags & TDF_DETAILS))
3548 fprintf (dump_file, "Found fully redundant value\n");
3549 continue;
3552 avail = XCNEWVEC (pre_expr, last_basic_block);
3553 FOR_EACH_EDGE (pred, ei, block->preds)
3555 unsigned int vprime;
3557 /* We should never run insertion for the exit block
3558 and so not come across fake pred edges. */
3559 gcc_assert (!(pred->flags & EDGE_FAKE));
3560 bprime = pred->src;
3561 eprime = phi_translate (expr, ANTIC_IN (block), NULL,
3562 bprime, block);
3564 /* eprime will generally only be NULL if the
3565 value of the expression, translated
3566 through the PHI for this predecessor, is
3567 undefined. If that is the case, we can't
3568 make the expression fully redundant,
3569 because its value is undefined along a
3570 predecessor path. We can thus break out
3571 early because it doesn't matter what the
3572 rest of the results are. */
3573 if (eprime == NULL)
3575 cant_insert = true;
3576 break;
3579 eprime = fully_constant_expression (eprime);
3580 vprime = get_expr_value_id (eprime);
3581 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3582 vprime, NULL);
3583 if (edoubleprime == NULL)
3585 avail[bprime->index] = eprime;
3586 all_same = false;
3588 else
3590 avail[bprime->index] = edoubleprime;
3591 by_some = true;
3592 /* We want to perform insertions to remove a redundancy on
3593 a path in the CFG we want to optimize for speed. */
3594 if (optimize_edge_for_speed_p (pred))
3595 do_insertion = true;
3596 if (first_s == NULL)
3597 first_s = edoubleprime;
3598 else if (!pre_expr_eq (first_s, edoubleprime))
3599 all_same = false;
3602 /* If we can insert it, it's not the same value
3603 already existing along every predecessor, and
3604 it's defined by some predecessor, it is
3605 partially redundant. */
3606 if (!cant_insert && !all_same && by_some)
3608 if (!do_insertion)
3610 if (dump_file && (dump_flags & TDF_DETAILS))
3612 fprintf (dump_file, "Skipping partial redundancy for "
3613 "expression ");
3614 print_pre_expr (dump_file, expr);
3615 fprintf (dump_file, " (%04d), no redundancy on to be "
3616 "optimized for speed edge\n", val);
3619 else if (dbg_cnt (treepre_insert)
3620 && insert_into_preds_of_block (block,
3621 get_expression_id (expr),
3622 avail))
3623 new_stuff = true;
3625 /* If all edges produce the same value and that value is
3626 an invariant, then the PHI has the same value on all
3627 edges. Note this. */
3628 else if (!cant_insert && all_same && eprime
3629 && (edoubleprime->kind == CONSTANT
3630 || edoubleprime->kind == NAME)
3631 && !value_id_constant_p (val))
3633 unsigned int j;
3634 bitmap_iterator bi;
3635 bitmap_set_t exprset = VEC_index (bitmap_set_t,
3636 value_expressions, val);
3638 unsigned int new_val = get_expr_value_id (edoubleprime);
3639 FOR_EACH_EXPR_ID_IN_SET (exprset, j, bi)
3641 pre_expr expr = expression_for_id (j);
3643 if (expr->kind == NAME)
3645 vn_ssa_aux_t info = VN_INFO (PRE_EXPR_NAME (expr));
3646 /* Just reset the value id and valnum so it is
3647 the same as the constant we have discovered. */
3648 if (edoubleprime->kind == CONSTANT)
3650 info->valnum = PRE_EXPR_CONSTANT (edoubleprime);
3651 pre_stats.constified++;
3653 else
3654 info->valnum = VN_INFO (PRE_EXPR_NAME (edoubleprime))->valnum;
3655 info->value_id = new_val;
3659 free (avail);
3663 VEC_free (pre_expr, heap, exprs);
3664 return new_stuff;
3668 /* Perform insertion for partially anticipatable expressions. There
3669 is only one case we will perform insertion for these. This case is
3670 if the expression is partially anticipatable, and fully available.
3671 In this case, we know that putting it earlier will enable us to
3672 remove the later computation. */
3675 static bool
3676 do_partial_partial_insertion (basic_block block, basic_block dom)
3678 bool new_stuff = false;
3679 VEC (pre_expr, heap) *exprs = sorted_array_from_bitmap_set (PA_IN (block));
3680 pre_expr expr;
3681 int i;
3683 FOR_EACH_VEC_ELT (pre_expr, exprs, i, expr)
3685 if (expr->kind != NAME)
3687 pre_expr *avail;
3688 unsigned int val;
3689 bool by_all = true;
3690 bool cant_insert = false;
3691 edge pred;
3692 basic_block bprime;
3693 pre_expr eprime = NULL;
3694 edge_iterator ei;
3696 val = get_expr_value_id (expr);
3697 if (bitmap_set_contains_value (PHI_GEN (block), val))
3698 continue;
3699 if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3700 continue;
3702 avail = XCNEWVEC (pre_expr, last_basic_block);
3703 FOR_EACH_EDGE (pred, ei, block->preds)
3705 unsigned int vprime;
3706 pre_expr edoubleprime;
3708 /* We should never run insertion for the exit block
3709 and so not come across fake pred edges. */
3710 gcc_assert (!(pred->flags & EDGE_FAKE));
3711 bprime = pred->src;
3712 eprime = phi_translate (expr, ANTIC_IN (block),
3713 PA_IN (block),
3714 bprime, block);
3716 /* eprime will generally only be NULL if the
3717 value of the expression, translated
3718 through the PHI for this predecessor, is
3719 undefined. If that is the case, we can't
3720 make the expression fully redundant,
3721 because its value is undefined along a
3722 predecessor path. We can thus break out
3723 early because it doesn't matter what the
3724 rest of the results are. */
3725 if (eprime == NULL)
3727 cant_insert = true;
3728 break;
3731 eprime = fully_constant_expression (eprime);
3732 vprime = get_expr_value_id (eprime);
3733 edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3734 vprime, NULL);
3735 if (edoubleprime == NULL)
3737 by_all = false;
3738 break;
3740 else
3741 avail[bprime->index] = edoubleprime;
3745 /* If we can insert it, it's not the same value
3746 already existing along every predecessor, and
3747 it's defined by some predecessor, it is
3748 partially redundant. */
3749 if (!cant_insert && by_all && dbg_cnt (treepre_insert))
3751 pre_stats.pa_insert++;
3752 if (insert_into_preds_of_block (block, get_expression_id (expr),
3753 avail))
3754 new_stuff = true;
3756 free (avail);
3760 VEC_free (pre_expr, heap, exprs);
3761 return new_stuff;
3764 static bool
3765 insert_aux (basic_block block)
3767 basic_block son;
3768 bool new_stuff = false;
3770 if (block)
3772 basic_block dom;
3773 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3774 if (dom)
3776 unsigned i;
3777 bitmap_iterator bi;
3778 bitmap_set_t newset = NEW_SETS (dom);
3779 if (newset)
3781 /* Note that we need to value_replace both NEW_SETS, and
3782 AVAIL_OUT. For both the case of NEW_SETS, the value may be
3783 represented by some non-simple expression here that we want
3784 to replace it with. */
3785 FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3787 pre_expr expr = expression_for_id (i);
3788 bitmap_value_replace_in_set (NEW_SETS (block), expr);
3789 bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3792 if (!single_pred_p (block))
3794 new_stuff |= do_regular_insertion (block, dom);
3795 if (do_partial_partial)
3796 new_stuff |= do_partial_partial_insertion (block, dom);
3800 for (son = first_dom_son (CDI_DOMINATORS, block);
3801 son;
3802 son = next_dom_son (CDI_DOMINATORS, son))
3804 new_stuff |= insert_aux (son);
3807 return new_stuff;
3810 /* Perform insertion of partially redundant values. */
3812 static void
3813 insert (void)
3815 bool new_stuff = true;
3816 basic_block bb;
3817 int num_iterations = 0;
3819 FOR_ALL_BB (bb)
3820 NEW_SETS (bb) = bitmap_set_new ();
3822 while (new_stuff)
3824 num_iterations++;
3825 new_stuff = insert_aux (ENTRY_BLOCK_PTR);
3827 statistics_histogram_event (cfun, "insert iterations", num_iterations);
3831 /* Add OP to EXP_GEN (block), and possibly to the maximal set. */
3833 static void
3834 add_to_exp_gen (basic_block block, tree op)
3836 if (!in_fre)
3838 pre_expr result;
3839 if (TREE_CODE (op) == SSA_NAME && ssa_undefined_value_p (op))
3840 return;
3841 result = get_or_alloc_expr_for_name (op);
3842 bitmap_value_insert_into_set (EXP_GEN (block), result);
3846 /* Create value ids for PHI in BLOCK. */
3848 static void
3849 make_values_for_phi (gimple phi, basic_block block)
3851 tree result = gimple_phi_result (phi);
3853 /* We have no need for virtual phis, as they don't represent
3854 actual computations. */
3855 if (is_gimple_reg (result))
3857 pre_expr e = get_or_alloc_expr_for_name (result);
3858 add_to_value (get_expr_value_id (e), e);
3859 bitmap_insert_into_set (PHI_GEN (block), e);
3860 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3861 if (!in_fre)
3863 unsigned i;
3864 for (i = 0; i < gimple_phi_num_args (phi); ++i)
3866 tree arg = gimple_phi_arg_def (phi, i);
3867 if (TREE_CODE (arg) == SSA_NAME)
3869 e = get_or_alloc_expr_for_name (arg);
3870 add_to_value (get_expr_value_id (e), e);
3877 /* Compute the AVAIL set for all basic blocks.
3879 This function performs value numbering of the statements in each basic
3880 block. The AVAIL sets are built from information we glean while doing
3881 this value numbering, since the AVAIL sets contain only one entry per
3882 value.
3884 AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3885 AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3887 static void
3888 compute_avail (void)
3891 basic_block block, son;
3892 basic_block *worklist;
3893 size_t sp = 0;
3894 unsigned i;
3896 /* We pretend that default definitions are defined in the entry block.
3897 This includes function arguments and the static chain decl. */
3898 for (i = 1; i < num_ssa_names; ++i)
3900 tree name = ssa_name (i);
3901 pre_expr e;
3902 if (!name
3903 || !SSA_NAME_IS_DEFAULT_DEF (name)
3904 || has_zero_uses (name)
3905 || !is_gimple_reg (name))
3906 continue;
3908 e = get_or_alloc_expr_for_name (name);
3909 add_to_value (get_expr_value_id (e), e);
3910 if (!in_fre)
3911 bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e);
3912 bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e);
3915 /* Allocate the worklist. */
3916 worklist = XNEWVEC (basic_block, n_basic_blocks);
3918 /* Seed the algorithm by putting the dominator children of the entry
3919 block on the worklist. */
3920 for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR);
3921 son;
3922 son = next_dom_son (CDI_DOMINATORS, son))
3923 worklist[sp++] = son;
3925 /* Loop until the worklist is empty. */
3926 while (sp)
3928 gimple_stmt_iterator gsi;
3929 gimple stmt;
3930 basic_block dom;
3931 unsigned int stmt_uid = 1;
3933 /* Pick a block from the worklist. */
3934 block = worklist[--sp];
3936 /* Initially, the set of available values in BLOCK is that of
3937 its immediate dominator. */
3938 dom = get_immediate_dominator (CDI_DOMINATORS, block);
3939 if (dom)
3940 bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3942 /* Generate values for PHI nodes. */
3943 for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi))
3944 make_values_for_phi (gsi_stmt (gsi), block);
3946 BB_MAY_NOTRETURN (block) = 0;
3948 /* Now compute value numbers and populate value sets with all
3949 the expressions computed in BLOCK. */
3950 for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
3952 ssa_op_iter iter;
3953 tree op;
3955 stmt = gsi_stmt (gsi);
3956 gimple_set_uid (stmt, stmt_uid++);
3958 /* Cache whether the basic-block has any non-visible side-effect
3959 or control flow.
3960 If this isn't a call or it is the last stmt in the
3961 basic-block then the CFG represents things correctly. */
3962 if (is_gimple_call (stmt)
3963 && !stmt_ends_bb_p (stmt))
3965 /* Non-looping const functions always return normally.
3966 Otherwise the call might not return or have side-effects
3967 that forbids hoisting possibly trapping expressions
3968 before it. */
3969 int flags = gimple_call_flags (stmt);
3970 if (!(flags & ECF_CONST)
3971 || (flags & ECF_LOOPING_CONST_OR_PURE))
3972 BB_MAY_NOTRETURN (block) = 1;
3975 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
3977 pre_expr e = get_or_alloc_expr_for_name (op);
3979 add_to_value (get_expr_value_id (e), e);
3980 if (!in_fre)
3981 bitmap_insert_into_set (TMP_GEN (block), e);
3982 bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3985 if (gimple_has_volatile_ops (stmt)
3986 || stmt_could_throw_p (stmt))
3987 continue;
3989 switch (gimple_code (stmt))
3991 case GIMPLE_RETURN:
3992 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
3993 add_to_exp_gen (block, op);
3994 continue;
3996 case GIMPLE_CALL:
3998 vn_reference_t ref;
3999 unsigned int i;
4000 vn_reference_op_t vro;
4001 pre_expr result = NULL;
4002 VEC(vn_reference_op_s, heap) *ops = NULL;
4004 if (!can_value_number_call (stmt))
4005 continue;
4007 copy_reference_ops_from_call (stmt, &ops);
4008 vn_reference_lookup_pieces (gimple_vuse (stmt), 0,
4009 gimple_expr_type (stmt),
4010 ops, &ref, VN_NOWALK);
4011 VEC_free (vn_reference_op_s, heap, ops);
4012 if (!ref)
4013 continue;
4015 for (i = 0; VEC_iterate (vn_reference_op_s,
4016 ref->operands, i,
4017 vro); i++)
4019 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4020 add_to_exp_gen (block, vro->op0);
4021 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4022 add_to_exp_gen (block, vro->op1);
4023 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4024 add_to_exp_gen (block, vro->op2);
4026 result = (pre_expr) pool_alloc (pre_expr_pool);
4027 result->kind = REFERENCE;
4028 result->id = 0;
4029 PRE_EXPR_REFERENCE (result) = ref;
4031 get_or_alloc_expression_id (result);
4032 add_to_value (get_expr_value_id (result), result);
4033 if (!in_fre)
4034 bitmap_value_insert_into_set (EXP_GEN (block), result);
4035 continue;
4038 case GIMPLE_ASSIGN:
4040 pre_expr result = NULL;
4041 switch (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)))
4043 case tcc_unary:
4044 case tcc_binary:
4045 case tcc_comparison:
4047 vn_nary_op_t nary;
4048 unsigned int i;
4050 vn_nary_op_lookup_pieces (gimple_num_ops (stmt) - 1,
4051 gimple_assign_rhs_code (stmt),
4052 gimple_expr_type (stmt),
4053 gimple_assign_rhs1_ptr (stmt),
4054 &nary);
4056 if (!nary)
4057 continue;
4059 for (i = 0; i < nary->length; i++)
4060 if (TREE_CODE (nary->op[i]) == SSA_NAME)
4061 add_to_exp_gen (block, nary->op[i]);
4063 result = (pre_expr) pool_alloc (pre_expr_pool);
4064 result->kind = NARY;
4065 result->id = 0;
4066 PRE_EXPR_NARY (result) = nary;
4067 break;
4070 case tcc_declaration:
4071 case tcc_reference:
4073 vn_reference_t ref;
4074 unsigned int i;
4075 vn_reference_op_t vro;
4077 vn_reference_lookup (gimple_assign_rhs1 (stmt),
4078 gimple_vuse (stmt),
4079 VN_WALK, &ref);
4080 if (!ref)
4081 continue;
4083 for (i = 0; VEC_iterate (vn_reference_op_s,
4084 ref->operands, i,
4085 vro); i++)
4087 if (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)
4088 add_to_exp_gen (block, vro->op0);
4089 if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME)
4090 add_to_exp_gen (block, vro->op1);
4091 if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME)
4092 add_to_exp_gen (block, vro->op2);
4094 result = (pre_expr) pool_alloc (pre_expr_pool);
4095 result->kind = REFERENCE;
4096 result->id = 0;
4097 PRE_EXPR_REFERENCE (result) = ref;
4098 break;
4101 default:
4102 /* For any other statement that we don't
4103 recognize, simply add all referenced
4104 SSA_NAMEs to EXP_GEN. */
4105 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4106 add_to_exp_gen (block, op);
4107 continue;
4110 get_or_alloc_expression_id (result);
4111 add_to_value (get_expr_value_id (result), result);
4112 if (!in_fre)
4113 bitmap_value_insert_into_set (EXP_GEN (block), result);
4115 continue;
4117 default:
4118 break;
4122 /* Put the dominator children of BLOCK on the worklist of blocks
4123 to compute available sets for. */
4124 for (son = first_dom_son (CDI_DOMINATORS, block);
4125 son;
4126 son = next_dom_son (CDI_DOMINATORS, son))
4127 worklist[sp++] = son;
4130 free (worklist);
4133 /* Insert the expression for SSA_VN that SCCVN thought would be simpler
4134 than the available expressions for it. The insertion point is
4135 right before the first use in STMT. Returns the SSA_NAME that should
4136 be used for replacement. */
4138 static tree
4139 do_SCCVN_insertion (gimple stmt, tree ssa_vn)
4141 basic_block bb = gimple_bb (stmt);
4142 gimple_stmt_iterator gsi;
4143 gimple_seq stmts = NULL;
4144 tree expr;
4145 pre_expr e;
4147 /* First create a value expression from the expression we want
4148 to insert and associate it with the value handle for SSA_VN. */
4149 e = get_or_alloc_expr_for (vn_get_expr_for (ssa_vn));
4150 if (e == NULL)
4151 return NULL_TREE;
4153 /* Then use create_expression_by_pieces to generate a valid
4154 expression to insert at this point of the IL stream. */
4155 expr = create_expression_by_pieces (bb, e, &stmts, stmt, NULL);
4156 if (expr == NULL_TREE)
4157 return NULL_TREE;
4158 gsi = gsi_for_stmt (stmt);
4159 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
4161 return expr;
4164 /* Eliminate fully redundant computations. */
4166 static unsigned int
4167 eliminate (void)
4169 VEC (gimple, heap) *to_remove = NULL;
4170 VEC (gimple, heap) *to_update = NULL;
4171 basic_block b;
4172 unsigned int todo = 0;
4173 gimple_stmt_iterator gsi;
4174 gimple stmt;
4175 unsigned i;
4177 FOR_EACH_BB (b)
4179 for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi))
4181 stmt = gsi_stmt (gsi);
4183 /* Lookup the RHS of the expression, see if we have an
4184 available computation for it. If so, replace the RHS with
4185 the available computation. */
4186 if (gimple_has_lhs (stmt)
4187 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
4188 && !gimple_assign_ssa_name_copy_p (stmt)
4189 && (!gimple_assign_single_p (stmt)
4190 || !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
4191 && !gimple_has_volatile_ops (stmt)
4192 && !has_zero_uses (gimple_get_lhs (stmt)))
4194 tree lhs = gimple_get_lhs (stmt);
4195 tree rhs = NULL_TREE;
4196 tree sprime = NULL;
4197 pre_expr lhsexpr = get_or_alloc_expr_for_name (lhs);
4198 pre_expr sprimeexpr;
4200 if (gimple_assign_single_p (stmt))
4201 rhs = gimple_assign_rhs1 (stmt);
4203 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4204 get_expr_value_id (lhsexpr),
4205 NULL);
4207 if (sprimeexpr)
4209 if (sprimeexpr->kind == CONSTANT)
4210 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4211 else if (sprimeexpr->kind == NAME)
4212 sprime = PRE_EXPR_NAME (sprimeexpr);
4213 else
4214 gcc_unreachable ();
4217 /* If there is no existing leader but SCCVN knows this
4218 value is constant, use that constant. */
4219 if (!sprime && is_gimple_min_invariant (VN_INFO (lhs)->valnum))
4221 sprime = VN_INFO (lhs)->valnum;
4222 if (!useless_type_conversion_p (TREE_TYPE (lhs),
4223 TREE_TYPE (sprime)))
4224 sprime = fold_convert (TREE_TYPE (lhs), sprime);
4226 if (dump_file && (dump_flags & TDF_DETAILS))
4228 fprintf (dump_file, "Replaced ");
4229 print_gimple_expr (dump_file, stmt, 0, 0);
4230 fprintf (dump_file, " with ");
4231 print_generic_expr (dump_file, sprime, 0);
4232 fprintf (dump_file, " in ");
4233 print_gimple_stmt (dump_file, stmt, 0, 0);
4235 pre_stats.eliminations++;
4236 propagate_tree_value_into_stmt (&gsi, sprime);
4237 stmt = gsi_stmt (gsi);
4238 update_stmt (stmt);
4239 continue;
4242 /* If there is no existing usable leader but SCCVN thinks
4243 it has an expression it wants to use as replacement,
4244 insert that. */
4245 if (!sprime || sprime == lhs)
4247 tree val = VN_INFO (lhs)->valnum;
4248 if (val != VN_TOP
4249 && TREE_CODE (val) == SSA_NAME
4250 && VN_INFO (val)->needs_insertion
4251 && can_PRE_operation (vn_get_expr_for (val)))
4252 sprime = do_SCCVN_insertion (stmt, val);
4254 if (sprime
4255 && sprime != lhs
4256 && (rhs == NULL_TREE
4257 || TREE_CODE (rhs) != SSA_NAME
4258 || may_propagate_copy (rhs, sprime)))
4260 bool can_make_abnormal_goto
4261 = is_gimple_call (stmt)
4262 && stmt_can_make_abnormal_goto (stmt);
4264 gcc_assert (sprime != rhs);
4266 if (dump_file && (dump_flags & TDF_DETAILS))
4268 fprintf (dump_file, "Replaced ");
4269 print_gimple_expr (dump_file, stmt, 0, 0);
4270 fprintf (dump_file, " with ");
4271 print_generic_expr (dump_file, sprime, 0);
4272 fprintf (dump_file, " in ");
4273 print_gimple_stmt (dump_file, stmt, 0, 0);
4276 if (TREE_CODE (sprime) == SSA_NAME)
4277 gimple_set_plf (SSA_NAME_DEF_STMT (sprime),
4278 NECESSARY, true);
4279 /* We need to make sure the new and old types actually match,
4280 which may require adding a simple cast, which fold_convert
4281 will do for us. */
4282 if ((!rhs || TREE_CODE (rhs) != SSA_NAME)
4283 && !useless_type_conversion_p (gimple_expr_type (stmt),
4284 TREE_TYPE (sprime)))
4285 sprime = fold_convert (gimple_expr_type (stmt), sprime);
4287 pre_stats.eliminations++;
4288 propagate_tree_value_into_stmt (&gsi, sprime);
4289 stmt = gsi_stmt (gsi);
4290 update_stmt (stmt);
4292 /* If we removed EH side-effects from the statement, clean
4293 its EH information. */
4294 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4296 bitmap_set_bit (need_eh_cleanup,
4297 gimple_bb (stmt)->index);
4298 if (dump_file && (dump_flags & TDF_DETAILS))
4299 fprintf (dump_file, " Removed EH side-effects.\n");
4302 /* Likewise for AB side-effects. */
4303 if (can_make_abnormal_goto
4304 && !stmt_can_make_abnormal_goto (stmt))
4306 bitmap_set_bit (need_ab_cleanup,
4307 gimple_bb (stmt)->index);
4308 if (dump_file && (dump_flags & TDF_DETAILS))
4309 fprintf (dump_file, " Removed AB side-effects.\n");
4313 /* If the statement is a scalar store, see if the expression
4314 has the same value number as its rhs. If so, the store is
4315 dead. */
4316 else if (gimple_assign_single_p (stmt)
4317 && !is_gimple_reg (gimple_assign_lhs (stmt))
4318 && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
4319 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
4321 tree rhs = gimple_assign_rhs1 (stmt);
4322 tree val;
4323 val = vn_reference_lookup (gimple_assign_lhs (stmt),
4324 gimple_vuse (stmt), VN_WALK, NULL);
4325 if (TREE_CODE (rhs) == SSA_NAME)
4326 rhs = VN_INFO (rhs)->valnum;
4327 if (val
4328 && operand_equal_p (val, rhs, 0))
4330 if (dump_file && (dump_flags & TDF_DETAILS))
4332 fprintf (dump_file, "Deleted redundant store ");
4333 print_gimple_stmt (dump_file, stmt, 0, 0);
4336 /* Queue stmt for removal. */
4337 VEC_safe_push (gimple, heap, to_remove, stmt);
4340 /* Visit COND_EXPRs and fold the comparison with the
4341 available value-numbers. */
4342 else if (gimple_code (stmt) == GIMPLE_COND)
4344 tree op0 = gimple_cond_lhs (stmt);
4345 tree op1 = gimple_cond_rhs (stmt);
4346 tree result;
4348 if (TREE_CODE (op0) == SSA_NAME)
4349 op0 = VN_INFO (op0)->valnum;
4350 if (TREE_CODE (op1) == SSA_NAME)
4351 op1 = VN_INFO (op1)->valnum;
4352 result = fold_binary (gimple_cond_code (stmt), boolean_type_node,
4353 op0, op1);
4354 if (result && TREE_CODE (result) == INTEGER_CST)
4356 if (integer_zerop (result))
4357 gimple_cond_make_false (stmt);
4358 else
4359 gimple_cond_make_true (stmt);
4360 update_stmt (stmt);
4361 todo = TODO_cleanup_cfg;
4364 /* Visit indirect calls and turn them into direct calls if
4365 possible. */
4366 if (is_gimple_call (stmt))
4368 tree orig_fn = gimple_call_fn (stmt);
4369 tree fn;
4370 if (!orig_fn)
4371 continue;
4372 if (TREE_CODE (orig_fn) == SSA_NAME)
4373 fn = VN_INFO (orig_fn)->valnum;
4374 else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF
4375 && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME)
4376 fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum;
4377 else
4378 continue;
4379 if (gimple_call_addr_fndecl (fn) != NULL_TREE
4380 && useless_type_conversion_p (TREE_TYPE (orig_fn),
4381 TREE_TYPE (fn)))
4383 bool can_make_abnormal_goto
4384 = stmt_can_make_abnormal_goto (stmt);
4385 bool was_noreturn = gimple_call_noreturn_p (stmt);
4387 if (dump_file && (dump_flags & TDF_DETAILS))
4389 fprintf (dump_file, "Replacing call target with ");
4390 print_generic_expr (dump_file, fn, 0);
4391 fprintf (dump_file, " in ");
4392 print_gimple_stmt (dump_file, stmt, 0, 0);
4395 gimple_call_set_fn (stmt, fn);
4396 VEC_safe_push (gimple, heap, to_update, stmt);
4398 /* When changing a call into a noreturn call, cfg cleanup
4399 is needed to fix up the noreturn call. */
4400 if (!was_noreturn && gimple_call_noreturn_p (stmt))
4401 todo |= TODO_cleanup_cfg;
4403 /* If we removed EH side-effects from the statement, clean
4404 its EH information. */
4405 if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
4407 bitmap_set_bit (need_eh_cleanup,
4408 gimple_bb (stmt)->index);
4409 if (dump_file && (dump_flags & TDF_DETAILS))
4410 fprintf (dump_file, " Removed EH side-effects.\n");
4413 /* Likewise for AB side-effects. */
4414 if (can_make_abnormal_goto
4415 && !stmt_can_make_abnormal_goto (stmt))
4417 bitmap_set_bit (need_ab_cleanup,
4418 gimple_bb (stmt)->index);
4419 if (dump_file && (dump_flags & TDF_DETAILS))
4420 fprintf (dump_file, " Removed AB side-effects.\n");
4423 /* Changing an indirect call to a direct call may
4424 have exposed different semantics. This may
4425 require an SSA update. */
4426 todo |= TODO_update_ssa_only_virtuals;
4431 for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
4433 gimple stmt, phi = gsi_stmt (gsi);
4434 tree sprime = NULL_TREE, res = PHI_RESULT (phi);
4435 pre_expr sprimeexpr, resexpr;
4436 gimple_stmt_iterator gsi2;
4438 /* We want to perform redundant PHI elimination. Do so by
4439 replacing the PHI with a single copy if possible.
4440 Do not touch inserted, single-argument or virtual PHIs. */
4441 if (gimple_phi_num_args (phi) == 1
4442 || !is_gimple_reg (res))
4444 gsi_next (&gsi);
4445 continue;
4448 resexpr = get_or_alloc_expr_for_name (res);
4449 sprimeexpr = bitmap_find_leader (AVAIL_OUT (b),
4450 get_expr_value_id (resexpr), NULL);
4451 if (sprimeexpr)
4453 if (sprimeexpr->kind == CONSTANT)
4454 sprime = PRE_EXPR_CONSTANT (sprimeexpr);
4455 else if (sprimeexpr->kind == NAME)
4456 sprime = PRE_EXPR_NAME (sprimeexpr);
4457 else
4458 gcc_unreachable ();
4460 if (!sprime && is_gimple_min_invariant (VN_INFO (res)->valnum))
4462 sprime = VN_INFO (res)->valnum;
4463 if (!useless_type_conversion_p (TREE_TYPE (res),
4464 TREE_TYPE (sprime)))
4465 sprime = fold_convert (TREE_TYPE (res), sprime);
4467 if (!sprime
4468 || sprime == res)
4470 gsi_next (&gsi);
4471 continue;
4474 if (dump_file && (dump_flags & TDF_DETAILS))
4476 fprintf (dump_file, "Replaced redundant PHI node defining ");
4477 print_generic_expr (dump_file, res, 0);
4478 fprintf (dump_file, " with ");
4479 print_generic_expr (dump_file, sprime, 0);
4480 fprintf (dump_file, "\n");
4483 remove_phi_node (&gsi, false);
4485 if (!bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))
4486 && TREE_CODE (sprime) == SSA_NAME)
4487 gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true);
4489 if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime)))
4490 sprime = fold_convert (TREE_TYPE (res), sprime);
4491 stmt = gimple_build_assign (res, sprime);
4492 SSA_NAME_DEF_STMT (res) = stmt;
4493 gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY));
4495 gsi2 = gsi_after_labels (b);
4496 gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT);
4497 /* Queue the copy for eventual removal. */
4498 VEC_safe_push (gimple, heap, to_remove, stmt);
4499 /* If we inserted this PHI node ourself, it's not an elimination. */
4500 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)))
4501 pre_stats.phis--;
4502 else
4503 pre_stats.eliminations++;
4507 /* We cannot remove stmts during BB walk, especially not release SSA
4508 names there as this confuses the VN machinery. The stmts ending
4509 up in to_remove are either stores or simple copies. */
4510 FOR_EACH_VEC_ELT (gimple, to_remove, i, stmt)
4512 tree lhs = gimple_assign_lhs (stmt);
4513 tree rhs = gimple_assign_rhs1 (stmt);
4514 use_operand_p use_p;
4515 gimple use_stmt;
4517 /* If there is a single use only, propagate the equivalency
4518 instead of keeping the copy. */
4519 if (TREE_CODE (lhs) == SSA_NAME
4520 && TREE_CODE (rhs) == SSA_NAME
4521 && single_imm_use (lhs, &use_p, &use_stmt)
4522 && may_propagate_copy (USE_FROM_PTR (use_p), rhs))
4524 SET_USE (use_p, rhs);
4525 update_stmt (use_stmt);
4526 if (bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs))
4527 && TREE_CODE (rhs) == SSA_NAME)
4528 gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true);
4531 /* If this is a store or a now unused copy, remove it. */
4532 if (TREE_CODE (lhs) != SSA_NAME
4533 || has_zero_uses (lhs))
4535 basic_block bb = gimple_bb (stmt);
4536 gsi = gsi_for_stmt (stmt);
4537 unlink_stmt_vdef (stmt);
4538 gsi_remove (&gsi, true);
4539 if (gimple_purge_dead_eh_edges (bb))
4540 todo |= TODO_cleanup_cfg;
4541 if (TREE_CODE (lhs) == SSA_NAME)
4542 bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs));
4543 release_defs (stmt);
4546 VEC_free (gimple, heap, to_remove);
4548 /* We cannot update call statements with virtual operands during
4549 SSA walk. This might remove them which in turn makes our
4550 VN lattice invalid. */
4551 FOR_EACH_VEC_ELT (gimple, to_update, i, stmt)
4552 update_stmt (stmt);
4553 VEC_free (gimple, heap, to_update);
4555 return todo;
4558 /* Borrow a bit of tree-ssa-dce.c for the moment.
4559 XXX: In 4.1, we should be able to just run a DCE pass after PRE, though
4560 this may be a bit faster, and we may want critical edges kept split. */
4562 /* If OP's defining statement has not already been determined to be necessary,
4563 mark that statement necessary. Return the stmt, if it is newly
4564 necessary. */
4566 static inline gimple
4567 mark_operand_necessary (tree op)
4569 gimple stmt;
4571 gcc_assert (op);
4573 if (TREE_CODE (op) != SSA_NAME)
4574 return NULL;
4576 stmt = SSA_NAME_DEF_STMT (op);
4577 gcc_assert (stmt);
4579 if (gimple_plf (stmt, NECESSARY)
4580 || gimple_nop_p (stmt))
4581 return NULL;
4583 gimple_set_plf (stmt, NECESSARY, true);
4584 return stmt;
4587 /* Because we don't follow exactly the standard PRE algorithm, and decide not
4588 to insert PHI nodes sometimes, and because value numbering of casts isn't
4589 perfect, we sometimes end up inserting dead code. This simple DCE-like
4590 pass removes any insertions we made that weren't actually used. */
4592 static void
4593 remove_dead_inserted_code (void)
4595 bitmap worklist;
4596 unsigned i;
4597 bitmap_iterator bi;
4598 gimple t;
4600 worklist = BITMAP_ALLOC (NULL);
4601 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4603 t = SSA_NAME_DEF_STMT (ssa_name (i));
4604 if (gimple_plf (t, NECESSARY))
4605 bitmap_set_bit (worklist, i);
4607 while (!bitmap_empty_p (worklist))
4609 i = bitmap_first_set_bit (worklist);
4610 bitmap_clear_bit (worklist, i);
4611 t = SSA_NAME_DEF_STMT (ssa_name (i));
4613 /* PHI nodes are somewhat special in that each PHI alternative has
4614 data and control dependencies. All the statements feeding the
4615 PHI node's arguments are always necessary. */
4616 if (gimple_code (t) == GIMPLE_PHI)
4618 unsigned k;
4620 for (k = 0; k < gimple_phi_num_args (t); k++)
4622 tree arg = PHI_ARG_DEF (t, k);
4623 if (TREE_CODE (arg) == SSA_NAME)
4625 gimple n = mark_operand_necessary (arg);
4626 if (n)
4627 bitmap_set_bit (worklist, SSA_NAME_VERSION (arg));
4631 else
4633 /* Propagate through the operands. Examine all the USE, VUSE and
4634 VDEF operands in this statement. Mark all the statements
4635 which feed this statement's uses as necessary. */
4636 ssa_op_iter iter;
4637 tree use;
4639 /* The operands of VDEF expressions are also needed as they
4640 represent potential definitions that may reach this
4641 statement (VDEF operands allow us to follow def-def
4642 links). */
4644 FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES)
4646 gimple n = mark_operand_necessary (use);
4647 if (n)
4648 bitmap_set_bit (worklist, SSA_NAME_VERSION (use));
4653 EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi)
4655 t = SSA_NAME_DEF_STMT (ssa_name (i));
4656 if (!gimple_plf (t, NECESSARY))
4658 gimple_stmt_iterator gsi;
4660 if (dump_file && (dump_flags & TDF_DETAILS))
4662 fprintf (dump_file, "Removing unnecessary insertion:");
4663 print_gimple_stmt (dump_file, t, 0, 0);
4666 gsi = gsi_for_stmt (t);
4667 if (gimple_code (t) == GIMPLE_PHI)
4668 remove_phi_node (&gsi, true);
4669 else
4671 gsi_remove (&gsi, true);
4672 release_defs (t);
4676 BITMAP_FREE (worklist);
4679 /* Compute a reverse post-order in *POST_ORDER. If INCLUDE_ENTRY_EXIT is
4680 true, then then ENTRY_BLOCK and EXIT_BLOCK are included. Returns
4681 the number of visited blocks. */
4683 static int
4684 my_rev_post_order_compute (int *post_order, bool include_entry_exit)
4686 edge_iterator *stack;
4687 int sp;
4688 int post_order_num = 0;
4689 sbitmap visited;
4691 if (include_entry_exit)
4692 post_order[post_order_num++] = EXIT_BLOCK;
4694 /* Allocate stack for back-tracking up CFG. */
4695 stack = XNEWVEC (edge_iterator, n_basic_blocks + 1);
4696 sp = 0;
4698 /* Allocate bitmap to track nodes that have been visited. */
4699 visited = sbitmap_alloc (last_basic_block);
4701 /* None of the nodes in the CFG have been visited yet. */
4702 sbitmap_zero (visited);
4704 /* Push the last edge on to the stack. */
4705 stack[sp++] = ei_start (EXIT_BLOCK_PTR->preds);
4707 while (sp)
4709 edge_iterator ei;
4710 basic_block src;
4711 basic_block dest;
4713 /* Look at the edge on the top of the stack. */
4714 ei = stack[sp - 1];
4715 src = ei_edge (ei)->src;
4716 dest = ei_edge (ei)->dest;
4718 /* Check if the edge destination has been visited yet. */
4719 if (src != ENTRY_BLOCK_PTR && ! TEST_BIT (visited, src->index))
4721 /* Mark that we have visited the destination. */
4722 SET_BIT (visited, src->index);
4724 if (EDGE_COUNT (src->preds) > 0)
4725 /* Since the DEST node has been visited for the first
4726 time, check its successors. */
4727 stack[sp++] = ei_start (src->preds);
4728 else
4729 post_order[post_order_num++] = src->index;
4731 else
4733 if (ei_one_before_end_p (ei) && dest != EXIT_BLOCK_PTR)
4734 post_order[post_order_num++] = dest->index;
4736 if (!ei_one_before_end_p (ei))
4737 ei_next (&stack[sp - 1]);
4738 else
4739 sp--;
4743 if (include_entry_exit)
4744 post_order[post_order_num++] = ENTRY_BLOCK;
4746 free (stack);
4747 sbitmap_free (visited);
4748 return post_order_num;
4752 /* Initialize data structures used by PRE. */
4754 static void
4755 init_pre (bool do_fre)
4757 basic_block bb;
4759 next_expression_id = 1;
4760 expressions = NULL;
4761 VEC_safe_push (pre_expr, heap, expressions, NULL);
4762 value_expressions = VEC_alloc (bitmap_set_t, heap, get_max_value_id () + 1);
4763 VEC_safe_grow_cleared (bitmap_set_t, heap, value_expressions,
4764 get_max_value_id() + 1);
4765 name_to_id = NULL;
4767 in_fre = do_fre;
4769 inserted_exprs = BITMAP_ALLOC (NULL);
4770 need_creation = NULL;
4771 pretemp = NULL_TREE;
4772 storetemp = NULL_TREE;
4773 prephitemp = NULL_TREE;
4775 connect_infinite_loops_to_exit ();
4776 memset (&pre_stats, 0, sizeof (pre_stats));
4779 postorder = XNEWVEC (int, n_basic_blocks - NUM_FIXED_BLOCKS);
4780 my_rev_post_order_compute (postorder, false);
4782 alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4784 calculate_dominance_info (CDI_POST_DOMINATORS);
4785 calculate_dominance_info (CDI_DOMINATORS);
4787 bitmap_obstack_initialize (&grand_bitmap_obstack);
4788 phi_translate_table = htab_create (5110, expr_pred_trans_hash,
4789 expr_pred_trans_eq, free);
4790 expression_to_id = htab_create (num_ssa_names * 3,
4791 pre_expr_hash,
4792 pre_expr_eq, NULL);
4793 bitmap_set_pool = create_alloc_pool ("Bitmap sets",
4794 sizeof (struct bitmap_set), 30);
4795 pre_expr_pool = create_alloc_pool ("pre_expr nodes",
4796 sizeof (struct pre_expr_d), 30);
4797 FOR_ALL_BB (bb)
4799 EXP_GEN (bb) = bitmap_set_new ();
4800 PHI_GEN (bb) = bitmap_set_new ();
4801 TMP_GEN (bb) = bitmap_set_new ();
4802 AVAIL_OUT (bb) = bitmap_set_new ();
4805 need_eh_cleanup = BITMAP_ALLOC (NULL);
4806 need_ab_cleanup = BITMAP_ALLOC (NULL);
4810 /* Deallocate data structures used by PRE. */
4812 static void
4813 fini_pre (bool do_fre)
4815 free (postorder);
4816 VEC_free (bitmap_set_t, heap, value_expressions);
4817 BITMAP_FREE (inserted_exprs);
4818 VEC_free (gimple, heap, need_creation);
4819 bitmap_obstack_release (&grand_bitmap_obstack);
4820 free_alloc_pool (bitmap_set_pool);
4821 free_alloc_pool (pre_expr_pool);
4822 htab_delete (phi_translate_table);
4823 htab_delete (expression_to_id);
4824 VEC_free (unsigned, heap, name_to_id);
4826 free_aux_for_blocks ();
4828 free_dominance_info (CDI_POST_DOMINATORS);
4830 if (!bitmap_empty_p (need_eh_cleanup))
4832 gimple_purge_all_dead_eh_edges (need_eh_cleanup);
4833 cleanup_tree_cfg ();
4836 BITMAP_FREE (need_eh_cleanup);
4838 if (!bitmap_empty_p (need_ab_cleanup))
4840 gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
4841 cleanup_tree_cfg ();
4844 BITMAP_FREE (need_ab_cleanup);
4846 if (!do_fre)
4847 loop_optimizer_finalize ();
4850 /* Main entry point to the SSA-PRE pass. DO_FRE is true if the caller
4851 only wants to do full redundancy elimination. */
4853 static unsigned int
4854 execute_pre (bool do_fre)
4856 unsigned int todo = 0;
4858 do_partial_partial = optimize > 2 && optimize_function_for_speed_p (cfun);
4860 /* This has to happen before SCCVN runs because
4861 loop_optimizer_init may create new phis, etc. */
4862 if (!do_fre)
4863 loop_optimizer_init (LOOPS_NORMAL);
4865 if (!run_scc_vn (do_fre ? VN_WALKREWRITE : VN_WALK))
4867 if (!do_fre)
4868 loop_optimizer_finalize ();
4870 return 0;
4873 init_pre (do_fre);
4874 scev_initialize ();
4876 /* Collect and value number expressions computed in each basic block. */
4877 compute_avail ();
4879 if (dump_file && (dump_flags & TDF_DETAILS))
4881 basic_block bb;
4883 FOR_ALL_BB (bb)
4885 print_bitmap_set (dump_file, EXP_GEN (bb), "exp_gen", bb->index);
4886 print_bitmap_set (dump_file, PHI_GEN (bb), "phi_gen", bb->index);
4887 print_bitmap_set (dump_file, TMP_GEN (bb), "tmp_gen", bb->index);
4888 print_bitmap_set (dump_file, AVAIL_OUT (bb), "avail_out", bb->index);
4892 /* Insert can get quite slow on an incredibly large number of basic
4893 blocks due to some quadratic behavior. Until this behavior is
4894 fixed, don't run it when he have an incredibly large number of
4895 bb's. If we aren't going to run insert, there is no point in
4896 computing ANTIC, either, even though it's plenty fast. */
4897 if (!do_fre && n_basic_blocks < 4000)
4899 compute_antic ();
4900 insert ();
4903 /* Make sure to remove fake edges before committing our inserts.
4904 This makes sure we don't end up with extra critical edges that
4905 we would need to split. */
4906 remove_fake_exit_edges ();
4907 gsi_commit_edge_inserts ();
4909 /* Remove all the redundant expressions. */
4910 todo |= eliminate ();
4912 statistics_counter_event (cfun, "Insertions", pre_stats.insertions);
4913 statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert);
4914 statistics_counter_event (cfun, "New PHIs", pre_stats.phis);
4915 statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations);
4916 statistics_counter_event (cfun, "Constified", pre_stats.constified);
4918 clear_expression_ids ();
4919 if (!do_fre)
4921 remove_dead_inserted_code ();
4922 todo |= TODO_verify_flow;
4925 scev_finalize ();
4926 fini_pre (do_fre);
4928 if (!do_fre)
4929 /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4930 case we can merge the block with the remaining predecessor of the block.
4931 It should either:
4932 - call merge_blocks after each tail merge iteration
4933 - call merge_blocks after all tail merge iterations
4934 - mark TODO_cleanup_cfg when necessary
4935 - share the cfg cleanup with fini_pre. */
4936 todo |= tail_merge_optimize (todo);
4937 free_scc_vn ();
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 */