1 /* Induction variable optimizations.
2 Copyright (C) 2003-2016 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 /* This pass tries to find the optimal set of induction variables for the loop.
21 It optimizes just the basic linear induction variables (although adding
22 support for other types should not be too hard). It includes the
23 optimizations commonly known as strength reduction, induction variable
24 coalescing and induction variable elimination. It does it in the
27 1) The interesting uses of induction variables are found. This includes
29 -- uses of induction variables in non-linear expressions
30 -- addresses of arrays
31 -- comparisons of induction variables
33 Note the interesting uses are categorized and handled in group.
34 Generally, address type uses are grouped together if their iv bases
35 are different in constant offset.
37 2) Candidates for the induction variables are found. This includes
39 -- old induction variables
40 -- the variables defined by expressions derived from the "interesting
43 3) The optimal (w.r. to a cost function) set of variables is chosen. The
44 cost function assigns a cost to sets of induction variables and consists
47 -- The group/use costs. Each of the interesting groups/uses chooses
48 the best induction variable in the set and adds its cost to the sum.
49 The cost reflects the time spent on modifying the induction variables
50 value to be usable for the given purpose (adding base and offset for
52 -- The variable costs. Each of the variables has a cost assigned that
53 reflects the costs associated with incrementing the value of the
54 variable. The original variables are somewhat preferred.
55 -- The set cost. Depending on the size of the set, extra cost may be
56 added to reflect register pressure.
58 All the costs are defined in a machine-specific way, using the target
59 hooks and machine descriptions to determine them.
61 4) The trees are transformed to use the new variables, the dead code is
64 All of this is done loop by loop. Doing it globally is theoretically
65 possible, it might give a better performance and it might enable us
66 to decide costs more precisely, but getting all the interactions right
67 would be complicated. */
71 #include "coretypes.h"
77 #include "tree-pass.h"
81 #include "insn-config.h"
85 #include "gimple-pretty-print.h"
87 #include "fold-const.h"
88 #include "stor-layout.h"
91 #include "gimple-iterator.h"
92 #include "gimplify-me.h"
94 #include "tree-ssa-loop-ivopts.h"
95 #include "tree-ssa-loop-manip.h"
96 #include "tree-ssa-loop-niter.h"
97 #include "tree-ssa-loop.h"
100 #include "tree-dfa.h"
101 #include "tree-ssa.h"
103 #include "tree-scalar-evolution.h"
105 #include "tree-affine.h"
106 #include "tree-ssa-propagate.h"
107 #include "tree-ssa-address.h"
108 #include "builtins.h"
109 #include "tree-vectorizer.h"
111 /* FIXME: Expressions are expanded to RTL in this pass to determine the
112 cost of different addressing modes. This should be moved to a TBD
113 interface between the GIMPLE and RTL worlds. */
115 /* The infinite cost. */
116 #define INFTY 10000000
118 /* Returns the expected number of loop iterations for LOOP.
119 The average trip count is computed from profile data if it
122 static inline HOST_WIDE_INT
123 avg_loop_niter (struct loop
*loop
)
125 HOST_WIDE_INT niter
= estimated_stmt_executions_int (loop
);
128 niter
= likely_max_stmt_executions_int (loop
);
130 if (niter
== -1 || niter
> PARAM_VALUE (PARAM_AVG_LOOP_NITER
))
131 return PARAM_VALUE (PARAM_AVG_LOOP_NITER
);
139 /* Representation of the induction variable. */
142 tree base
; /* Initial value of the iv. */
143 tree base_object
; /* A memory object to that the induction variable points. */
144 tree step
; /* Step of the iv (constant only). */
145 tree ssa_name
; /* The ssa name with the value. */
146 struct iv_use
*nonlin_use
; /* The identifier in the use if it is the case. */
147 bool biv_p
; /* Is it a biv? */
148 bool no_overflow
; /* True if the iv doesn't overflow. */
149 bool have_address_use
;/* For biv, indicate if it's used in any address
153 /* Per-ssa version information (induction variable descriptions, etc.). */
156 tree name
; /* The ssa name. */
157 struct iv
*iv
; /* Induction variable description. */
158 bool has_nonlin_use
; /* For a loop-level invariant, whether it is used in
159 an expression that is not an induction variable. */
160 bool preserve_biv
; /* For the original biv, whether to preserve it. */
161 unsigned inv_id
; /* Id of an invariant. */
167 USE_NONLINEAR_EXPR
, /* Use in a nonlinear expression. */
168 USE_ADDRESS
, /* Use in an address. */
169 USE_COMPARE
/* Use is a compare. */
172 /* Cost of a computation. */
175 comp_cost (): cost (0), complexity (0), scratch (0)
178 comp_cost (int cost
, unsigned complexity
, int scratch
= 0)
179 : cost (cost
), complexity (complexity
), scratch (scratch
)
182 /* Returns true if COST is infinite. */
183 bool infinite_cost_p ();
185 /* Adds costs COST1 and COST2. */
186 friend comp_cost
operator+ (comp_cost cost1
, comp_cost cost2
);
188 /* Adds COST to the comp_cost. */
189 comp_cost
operator+= (comp_cost cost
);
191 /* Adds constant C to this comp_cost. */
192 comp_cost
operator+= (HOST_WIDE_INT c
);
194 /* Subtracts constant C to this comp_cost. */
195 comp_cost
operator-= (HOST_WIDE_INT c
);
197 /* Divide the comp_cost by constant C. */
198 comp_cost
operator/= (HOST_WIDE_INT c
);
200 /* Multiply the comp_cost by constant C. */
201 comp_cost
operator*= (HOST_WIDE_INT c
);
203 /* Subtracts costs COST1 and COST2. */
204 friend comp_cost
operator- (comp_cost cost1
, comp_cost cost2
);
206 /* Subtracts COST from this comp_cost. */
207 comp_cost
operator-= (comp_cost cost
);
209 /* Returns true if COST1 is smaller than COST2. */
210 friend bool operator< (comp_cost cost1
, comp_cost cost2
);
212 /* Returns true if COST1 and COST2 are equal. */
213 friend bool operator== (comp_cost cost1
, comp_cost cost2
);
215 /* Returns true if COST1 is smaller or equal than COST2. */
216 friend bool operator<= (comp_cost cost1
, comp_cost cost2
);
218 int cost
; /* The runtime cost. */
219 unsigned complexity
; /* The estimate of the complexity of the code for
220 the computation (in no concrete units --
221 complexity field should be larger for more
222 complex expressions and addressing modes). */
223 int scratch
; /* Scratch used during cost computation. */
226 static const comp_cost no_cost
;
227 static const comp_cost
infinite_cost (INFTY
, INFTY
, INFTY
);
230 comp_cost::infinite_cost_p ()
232 return cost
== INFTY
;
236 operator+ (comp_cost cost1
, comp_cost cost2
)
238 if (cost1
.infinite_cost_p () || cost2
.infinite_cost_p ())
239 return infinite_cost
;
241 cost1
.cost
+= cost2
.cost
;
242 cost1
.complexity
+= cost2
.complexity
;
248 operator- (comp_cost cost1
, comp_cost cost2
)
250 if (cost1
.infinite_cost_p ())
251 return infinite_cost
;
253 gcc_assert (!cost2
.infinite_cost_p ());
255 cost1
.cost
-= cost2
.cost
;
256 cost1
.complexity
-= cost2
.complexity
;
262 comp_cost::operator+= (comp_cost cost
)
264 *this = *this + cost
;
269 comp_cost::operator+= (HOST_WIDE_INT c
)
271 if (infinite_cost_p ())
280 comp_cost::operator-= (HOST_WIDE_INT c
)
282 if (infinite_cost_p ())
291 comp_cost::operator/= (HOST_WIDE_INT c
)
293 if (infinite_cost_p ())
302 comp_cost::operator*= (HOST_WIDE_INT c
)
304 if (infinite_cost_p ())
313 comp_cost::operator-= (comp_cost cost
)
315 *this = *this - cost
;
320 operator< (comp_cost cost1
, comp_cost cost2
)
322 if (cost1
.cost
== cost2
.cost
)
323 return cost1
.complexity
< cost2
.complexity
;
325 return cost1
.cost
< cost2
.cost
;
329 operator== (comp_cost cost1
, comp_cost cost2
)
331 return cost1
.cost
== cost2
.cost
332 && cost1
.complexity
== cost2
.complexity
;
336 operator<= (comp_cost cost1
, comp_cost cost2
)
338 return cost1
< cost2
|| cost1
== cost2
;
341 struct iv_inv_expr_ent
;
343 /* The candidate - cost pair. */
346 struct iv_cand
*cand
; /* The candidate. */
347 comp_cost cost
; /* The cost. */
348 bitmap depends_on
; /* The list of invariants that have to be
350 tree value
; /* For final value elimination, the expression for
351 the final value of the iv. For iv elimination,
352 the new bound to compare with. */
353 enum tree_code comp
; /* For iv elimination, the comparison. */
354 iv_inv_expr_ent
*inv_expr
; /* Loop invariant expression. */
360 unsigned id
; /* The id of the use. */
361 unsigned group_id
; /* The group id the use belongs to. */
362 enum use_type type
; /* Type of the use. */
363 struct iv
*iv
; /* The induction variable it is based on. */
364 gimple
*stmt
; /* Statement in that it occurs. */
365 tree
*op_p
; /* The place where it occurs. */
367 tree addr_base
; /* Base address with const offset stripped. */
368 unsigned HOST_WIDE_INT addr_offset
;
369 /* Const offset stripped from base address. */
375 /* The id of the group. */
377 /* Uses of the group are of the same type. */
379 /* The set of "related" IV candidates, plus the important ones. */
380 bitmap related_cands
;
381 /* Number of IV candidates in the cost_map. */
382 unsigned n_map_members
;
383 /* The costs wrto the iv candidates. */
384 struct cost_pair
*cost_map
;
385 /* The selected candidate for the group. */
386 struct iv_cand
*selected
;
387 /* Uses in the group. */
388 vec
<struct iv_use
*> vuses
;
391 /* The position where the iv is computed. */
394 IP_NORMAL
, /* At the end, just before the exit condition. */
395 IP_END
, /* At the end of the latch block. */
396 IP_BEFORE_USE
, /* Immediately before a specific use. */
397 IP_AFTER_USE
, /* Immediately after a specific use. */
398 IP_ORIGINAL
/* The original biv. */
401 /* The induction variable candidate. */
404 unsigned id
; /* The number of the candidate. */
405 bool important
; /* Whether this is an "important" candidate, i.e. such
406 that it should be considered by all uses. */
407 ENUM_BITFIELD(iv_position
) pos
: 8; /* Where it is computed. */
408 gimple
*incremented_at
;/* For original biv, the statement where it is
410 tree var_before
; /* The variable used for it before increment. */
411 tree var_after
; /* The variable used for it after increment. */
412 struct iv
*iv
; /* The value of the candidate. NULL for
413 "pseudocandidate" used to indicate the possibility
414 to replace the final value of an iv by direct
415 computation of the value. */
416 unsigned cost
; /* Cost of the candidate. */
417 unsigned cost_step
; /* Cost of the candidate's increment operation. */
418 struct iv_use
*ainc_use
; /* For IP_{BEFORE,AFTER}_USE candidates, the place
419 where it is incremented. */
420 bitmap depends_on
; /* The list of invariants that are used in step of the
422 struct iv
*orig_iv
; /* The original iv if this cand is added from biv with
426 /* Hashtable entry for common candidate derived from iv uses. */
427 struct iv_common_cand
431 /* IV uses from which this common candidate is derived. */
432 auto_vec
<struct iv_use
*> uses
;
436 /* Hashtable helpers. */
438 struct iv_common_cand_hasher
: delete_ptr_hash
<iv_common_cand
>
440 static inline hashval_t
hash (const iv_common_cand
*);
441 static inline bool equal (const iv_common_cand
*, const iv_common_cand
*);
444 /* Hash function for possible common candidates. */
447 iv_common_cand_hasher::hash (const iv_common_cand
*ccand
)
452 /* Hash table equality function for common candidates. */
455 iv_common_cand_hasher::equal (const iv_common_cand
*ccand1
,
456 const iv_common_cand
*ccand2
)
458 return (ccand1
->hash
== ccand2
->hash
459 && operand_equal_p (ccand1
->base
, ccand2
->base
, 0)
460 && operand_equal_p (ccand1
->step
, ccand2
->step
, 0)
461 && (TYPE_PRECISION (TREE_TYPE (ccand1
->base
))
462 == TYPE_PRECISION (TREE_TYPE (ccand2
->base
))));
465 /* Loop invariant expression hashtable entry. */
467 struct iv_inv_expr_ent
469 /* Tree expression of the entry. */
471 /* Unique indentifier. */
477 /* Sort iv_inv_expr_ent pair A and B by id field. */
480 sort_iv_inv_expr_ent (const void *a
, const void *b
)
482 const iv_inv_expr_ent
* const *e1
= (const iv_inv_expr_ent
* const *) (a
);
483 const iv_inv_expr_ent
* const *e2
= (const iv_inv_expr_ent
* const *) (b
);
485 unsigned id1
= (*e1
)->id
;
486 unsigned id2
= (*e2
)->id
;
496 /* Hashtable helpers. */
498 struct iv_inv_expr_hasher
: free_ptr_hash
<iv_inv_expr_ent
>
500 static inline hashval_t
hash (const iv_inv_expr_ent
*);
501 static inline bool equal (const iv_inv_expr_ent
*, const iv_inv_expr_ent
*);
504 /* Hash function for loop invariant expressions. */
507 iv_inv_expr_hasher::hash (const iv_inv_expr_ent
*expr
)
512 /* Hash table equality function for expressions. */
515 iv_inv_expr_hasher::equal (const iv_inv_expr_ent
*expr1
,
516 const iv_inv_expr_ent
*expr2
)
518 return expr1
->hash
== expr2
->hash
519 && operand_equal_p (expr1
->expr
, expr2
->expr
, 0);
524 /* The currently optimized loop. */
525 struct loop
*current_loop
;
526 source_location loop_loc
;
528 /* Numbers of iterations for all exits of the current loop. */
529 hash_map
<edge
, tree_niter_desc
*> *niters
;
531 /* Number of registers used in it. */
534 /* The size of version_info array allocated. */
535 unsigned version_info_size
;
537 /* The array of information for the ssa names. */
538 struct version_info
*version_info
;
540 /* The hashtable of loop invariant expressions created
542 hash_table
<iv_inv_expr_hasher
> *inv_expr_tab
;
544 /* Loop invariant expression id. */
547 /* The bitmap of indices in version_info whose value was changed. */
550 /* The uses of induction variables. */
551 vec
<iv_group
*> vgroups
;
553 /* The candidates. */
554 vec
<iv_cand
*> vcands
;
556 /* A bitmap of important candidates. */
557 bitmap important_candidates
;
559 /* Cache used by tree_to_aff_combination_expand. */
560 hash_map
<tree
, name_expansion
*> *name_expansion_cache
;
562 /* The hashtable of common candidates derived from iv uses. */
563 hash_table
<iv_common_cand_hasher
> *iv_common_cand_tab
;
565 /* The common candidates. */
566 vec
<iv_common_cand
*> iv_common_cands
;
568 /* The maximum invariant id. */
571 /* Number of no_overflow BIVs which are not used in memory address. */
572 unsigned bivs_not_used_in_addr
;
574 /* Obstack for iv structure. */
575 struct obstack iv_obstack
;
577 /* Whether to consider just related and important candidates when replacing a
579 bool consider_all_candidates
;
581 /* Are we optimizing for speed? */
584 /* Whether the loop body includes any function calls. */
585 bool body_includes_call
;
587 /* Whether the loop body can only be exited via single exit. */
588 bool loop_single_exit_p
;
591 /* An assignment of iv candidates to uses. */
595 /* The number of uses covered by the assignment. */
598 /* Number of uses that cannot be expressed by the candidates in the set. */
601 /* Candidate assigned to a use, together with the related costs. */
602 struct cost_pair
**cand_for_group
;
604 /* Number of times each candidate is used. */
605 unsigned *n_cand_uses
;
607 /* The candidates used. */
610 /* The number of candidates in the set. */
613 /* Total number of registers needed. */
616 /* Total cost of expressing uses. */
617 comp_cost cand_use_cost
;
619 /* Total cost of candidates. */
622 /* Number of times each invariant is used. */
623 unsigned *n_invariant_uses
;
625 /* Hash set with used invariant expression. */
626 hash_map
<iv_inv_expr_ent
*, unsigned> *used_inv_exprs
;
628 /* Total cost of the assignment. */
632 /* Difference of two iv candidate assignments. */
637 struct iv_group
*group
;
639 /* An old assignment (for rollback purposes). */
640 struct cost_pair
*old_cp
;
642 /* A new assignment. */
643 struct cost_pair
*new_cp
;
645 /* Next change in the list. */
646 struct iv_ca_delta
*next
;
649 /* Bound on number of candidates below that all candidates are considered. */
651 #define CONSIDER_ALL_CANDIDATES_BOUND \
652 ((unsigned) PARAM_VALUE (PARAM_IV_CONSIDER_ALL_CANDIDATES_BOUND))
654 /* If there are more iv occurrences, we just give up (it is quite unlikely that
655 optimizing such a loop would help, and it would take ages). */
657 #define MAX_CONSIDERED_GROUPS \
658 ((unsigned) PARAM_VALUE (PARAM_IV_MAX_CONSIDERED_USES))
660 /* If there are at most this number of ivs in the set, try removing unnecessary
661 ivs from the set always. */
663 #define ALWAYS_PRUNE_CAND_SET_BOUND \
664 ((unsigned) PARAM_VALUE (PARAM_IV_ALWAYS_PRUNE_CAND_SET_BOUND))
666 /* The list of trees for that the decl_rtl field must be reset is stored
669 static vec
<tree
> decl_rtl_to_reset
;
671 static comp_cost
force_expr_to_var_cost (tree
, bool);
673 /* The single loop exit if it dominates the latch, NULL otherwise. */
676 single_dom_exit (struct loop
*loop
)
678 edge exit
= single_exit (loop
);
683 if (!just_once_each_iteration_p (loop
, exit
->src
))
689 /* Dumps information about the induction variable IV to FILE. Don't dump
690 variable's name if DUMP_NAME is FALSE. The information is dumped with
691 preceding spaces indicated by INDENT_LEVEL. */
694 dump_iv (FILE *file
, struct iv
*iv
, bool dump_name
, unsigned indent_level
)
697 const char spaces
[9] = {' ', ' ', ' ', ' ', ' ', ' ', ' ', ' ', '\0'};
699 if (indent_level
> 4)
701 p
= spaces
+ 8 - (indent_level
<< 1);
703 fprintf (file
, "%sIV struct:\n", p
);
704 if (iv
->ssa_name
&& dump_name
)
706 fprintf (file
, "%s SSA_NAME:\t", p
);
707 print_generic_expr (file
, iv
->ssa_name
, TDF_SLIM
);
708 fprintf (file
, "\n");
711 fprintf (file
, "%s Type:\t", p
);
712 print_generic_expr (file
, TREE_TYPE (iv
->base
), TDF_SLIM
);
713 fprintf (file
, "\n");
715 fprintf (file
, "%s Base:\t", p
);
716 print_generic_expr (file
, iv
->base
, TDF_SLIM
);
717 fprintf (file
, "\n");
719 fprintf (file
, "%s Step:\t", p
);
720 print_generic_expr (file
, iv
->step
, TDF_SLIM
);
721 fprintf (file
, "\n");
725 fprintf (file
, "%s Object:\t", p
);
726 print_generic_expr (file
, iv
->base_object
, TDF_SLIM
);
727 fprintf (file
, "\n");
730 fprintf (file
, "%s Biv:\t%c\n", p
, iv
->biv_p
? 'Y' : 'N');
732 fprintf (file
, "%s Overflowness wrto loop niter:\t%s\n",
733 p
, iv
->no_overflow
? "No-overflow" : "Overflow");
736 /* Dumps information about the USE to FILE. */
739 dump_use (FILE *file
, struct iv_use
*use
)
741 fprintf (file
, " Use %d.%d:\n", use
->group_id
, use
->id
);
742 fprintf (file
, " At stmt:\t");
743 print_gimple_stmt (file
, use
->stmt
, 0, 0);
744 fprintf (file
, " At pos:\t");
746 print_generic_expr (file
, *use
->op_p
, TDF_SLIM
);
747 fprintf (file
, "\n");
748 dump_iv (file
, use
->iv
, false, 2);
751 /* Dumps information about the uses to FILE. */
754 dump_groups (FILE *file
, struct ivopts_data
*data
)
757 struct iv_group
*group
;
759 for (i
= 0; i
< data
->vgroups
.length (); i
++)
761 group
= data
->vgroups
[i
];
762 fprintf (file
, "Group %d:\n", group
->id
);
763 if (group
->type
== USE_NONLINEAR_EXPR
)
764 fprintf (file
, " Type:\tGENERIC\n");
765 else if (group
->type
== USE_ADDRESS
)
766 fprintf (file
, " Type:\tADDRESS\n");
769 gcc_assert (group
->type
== USE_COMPARE
);
770 fprintf (file
, " Type:\tCOMPARE\n");
772 for (j
= 0; j
< group
->vuses
.length (); j
++)
773 dump_use (file
, group
->vuses
[j
]);
777 /* Dumps information about induction variable candidate CAND to FILE. */
780 dump_cand (FILE *file
, struct iv_cand
*cand
)
782 struct iv
*iv
= cand
->iv
;
784 fprintf (file
, "Candidate %d:\n", cand
->id
);
785 if (cand
->depends_on
)
787 fprintf (file
, " Depend on: ");
788 dump_bitmap (file
, cand
->depends_on
);
791 if (cand
->var_before
)
793 fprintf (file
, " Var befor: ");
794 print_generic_expr (file
, cand
->var_before
, TDF_SLIM
);
795 fprintf (file
, "\n");
799 fprintf (file
, " Var after: ");
800 print_generic_expr (file
, cand
->var_after
, TDF_SLIM
);
801 fprintf (file
, "\n");
807 fprintf (file
, " Incr POS: before exit test\n");
811 fprintf (file
, " Incr POS: before use %d\n", cand
->ainc_use
->id
);
815 fprintf (file
, " Incr POS: after use %d\n", cand
->ainc_use
->id
);
819 fprintf (file
, " Incr POS: at end\n");
823 fprintf (file
, " Incr POS: orig biv\n");
827 dump_iv (file
, iv
, false, 1);
830 /* Returns the info for ssa version VER. */
832 static inline struct version_info
*
833 ver_info (struct ivopts_data
*data
, unsigned ver
)
835 return data
->version_info
+ ver
;
838 /* Returns the info for ssa name NAME. */
840 static inline struct version_info
*
841 name_info (struct ivopts_data
*data
, tree name
)
843 return ver_info (data
, SSA_NAME_VERSION (name
));
846 /* Returns true if STMT is after the place where the IP_NORMAL ivs will be
850 stmt_after_ip_normal_pos (struct loop
*loop
, gimple
*stmt
)
852 basic_block bb
= ip_normal_pos (loop
), sbb
= gimple_bb (stmt
);
856 if (sbb
== loop
->latch
)
862 return stmt
== last_stmt (bb
);
865 /* Returns true if STMT if after the place where the original induction
866 variable CAND is incremented. If TRUE_IF_EQUAL is set, we return true
867 if the positions are identical. */
870 stmt_after_inc_pos (struct iv_cand
*cand
, gimple
*stmt
, bool true_if_equal
)
872 basic_block cand_bb
= gimple_bb (cand
->incremented_at
);
873 basic_block stmt_bb
= gimple_bb (stmt
);
875 if (!dominated_by_p (CDI_DOMINATORS
, stmt_bb
, cand_bb
))
878 if (stmt_bb
!= cand_bb
)
882 && gimple_uid (stmt
) == gimple_uid (cand
->incremented_at
))
884 return gimple_uid (stmt
) > gimple_uid (cand
->incremented_at
);
887 /* Returns true if STMT if after the place where the induction variable
888 CAND is incremented in LOOP. */
891 stmt_after_increment (struct loop
*loop
, struct iv_cand
*cand
, gimple
*stmt
)
899 return stmt_after_ip_normal_pos (loop
, stmt
);
903 return stmt_after_inc_pos (cand
, stmt
, false);
906 return stmt_after_inc_pos (cand
, stmt
, true);
913 /* Returns true if EXP is a ssa name that occurs in an abnormal phi node. */
916 abnormal_ssa_name_p (tree exp
)
921 if (TREE_CODE (exp
) != SSA_NAME
)
924 return SSA_NAME_OCCURS_IN_ABNORMAL_PHI (exp
) != 0;
927 /* Returns false if BASE or INDEX contains a ssa name that occurs in an
928 abnormal phi node. Callback for for_each_index. */
931 idx_contains_abnormal_ssa_name_p (tree base
, tree
*index
,
932 void *data ATTRIBUTE_UNUSED
)
934 if (TREE_CODE (base
) == ARRAY_REF
|| TREE_CODE (base
) == ARRAY_RANGE_REF
)
936 if (abnormal_ssa_name_p (TREE_OPERAND (base
, 2)))
938 if (abnormal_ssa_name_p (TREE_OPERAND (base
, 3)))
942 return !abnormal_ssa_name_p (*index
);
945 /* Returns true if EXPR contains a ssa name that occurs in an
946 abnormal phi node. */
949 contains_abnormal_ssa_name_p (tree expr
)
952 enum tree_code_class codeclass
;
957 code
= TREE_CODE (expr
);
958 codeclass
= TREE_CODE_CLASS (code
);
960 if (code
== SSA_NAME
)
961 return SSA_NAME_OCCURS_IN_ABNORMAL_PHI (expr
) != 0;
963 if (code
== INTEGER_CST
964 || is_gimple_min_invariant (expr
))
967 if (code
== ADDR_EXPR
)
968 return !for_each_index (&TREE_OPERAND (expr
, 0),
969 idx_contains_abnormal_ssa_name_p
,
972 if (code
== COND_EXPR
)
973 return contains_abnormal_ssa_name_p (TREE_OPERAND (expr
, 0))
974 || contains_abnormal_ssa_name_p (TREE_OPERAND (expr
, 1))
975 || contains_abnormal_ssa_name_p (TREE_OPERAND (expr
, 2));
981 if (contains_abnormal_ssa_name_p (TREE_OPERAND (expr
, 1)))
986 if (contains_abnormal_ssa_name_p (TREE_OPERAND (expr
, 0)))
998 /* Returns the structure describing number of iterations determined from
999 EXIT of DATA->current_loop, or NULL if something goes wrong. */
1001 static struct tree_niter_desc
*
1002 niter_for_exit (struct ivopts_data
*data
, edge exit
)
1004 struct tree_niter_desc
*desc
;
1005 tree_niter_desc
**slot
;
1009 data
->niters
= new hash_map
<edge
, tree_niter_desc
*>;
1013 slot
= data
->niters
->get (exit
);
1017 /* Try to determine number of iterations. We cannot safely work with ssa
1018 names that appear in phi nodes on abnormal edges, so that we do not
1019 create overlapping life ranges for them (PR 27283). */
1020 desc
= XNEW (struct tree_niter_desc
);
1021 if (!number_of_iterations_exit (data
->current_loop
,
1023 || contains_abnormal_ssa_name_p (desc
->niter
))
1028 data
->niters
->put (exit
, desc
);
1036 /* Returns the structure describing number of iterations determined from
1037 single dominating exit of DATA->current_loop, or NULL if something
1040 static struct tree_niter_desc
*
1041 niter_for_single_dom_exit (struct ivopts_data
*data
)
1043 edge exit
= single_dom_exit (data
->current_loop
);
1048 return niter_for_exit (data
, exit
);
1051 /* Initializes data structures used by the iv optimization pass, stored
1055 tree_ssa_iv_optimize_init (struct ivopts_data
*data
)
1057 data
->version_info_size
= 2 * num_ssa_names
;
1058 data
->version_info
= XCNEWVEC (struct version_info
, data
->version_info_size
);
1059 data
->relevant
= BITMAP_ALLOC (NULL
);
1060 data
->important_candidates
= BITMAP_ALLOC (NULL
);
1061 data
->max_inv_id
= 0;
1062 data
->niters
= NULL
;
1063 data
->vgroups
.create (20);
1064 data
->vcands
.create (20);
1065 data
->inv_expr_tab
= new hash_table
<iv_inv_expr_hasher
> (10);
1066 data
->max_inv_expr_id
= 0;
1067 data
->name_expansion_cache
= NULL
;
1068 data
->iv_common_cand_tab
= new hash_table
<iv_common_cand_hasher
> (10);
1069 data
->iv_common_cands
.create (20);
1070 decl_rtl_to_reset
.create (20);
1071 gcc_obstack_init (&data
->iv_obstack
);
1074 /* Returns a memory object to that EXPR points. In case we are able to
1075 determine that it does not point to any such object, NULL is returned. */
1078 determine_base_object (tree expr
)
1080 enum tree_code code
= TREE_CODE (expr
);
1083 /* If this is a pointer casted to any type, we need to determine
1084 the base object for the pointer; so handle conversions before
1085 throwing away non-pointer expressions. */
1086 if (CONVERT_EXPR_P (expr
))
1087 return determine_base_object (TREE_OPERAND (expr
, 0));
1089 if (!POINTER_TYPE_P (TREE_TYPE (expr
)))
1098 obj
= TREE_OPERAND (expr
, 0);
1099 base
= get_base_address (obj
);
1104 if (TREE_CODE (base
) == MEM_REF
)
1105 return determine_base_object (TREE_OPERAND (base
, 0));
1107 return fold_convert (ptr_type_node
,
1108 build_fold_addr_expr (base
));
1110 case POINTER_PLUS_EXPR
:
1111 return determine_base_object (TREE_OPERAND (expr
, 0));
1115 /* Pointer addition is done solely using POINTER_PLUS_EXPR. */
1119 return fold_convert (ptr_type_node
, expr
);
1123 /* Return true if address expression with non-DECL_P operand appears
1127 contain_complex_addr_expr (tree expr
)
1132 switch (TREE_CODE (expr
))
1134 case POINTER_PLUS_EXPR
:
1137 res
|= contain_complex_addr_expr (TREE_OPERAND (expr
, 0));
1138 res
|= contain_complex_addr_expr (TREE_OPERAND (expr
, 1));
1142 return (!DECL_P (TREE_OPERAND (expr
, 0)));
1151 /* Allocates an induction variable with given initial value BASE and step STEP
1152 for loop LOOP. NO_OVERFLOW implies the iv doesn't overflow. */
1155 alloc_iv (struct ivopts_data
*data
, tree base
, tree step
,
1156 bool no_overflow
= false)
1159 struct iv
*iv
= (struct iv
*) obstack_alloc (&data
->iv_obstack
,
1160 sizeof (struct iv
));
1161 gcc_assert (step
!= NULL_TREE
);
1163 /* Lower address expression in base except ones with DECL_P as operand.
1165 1) More accurate cost can be computed for address expressions;
1166 2) Duplicate candidates won't be created for bases in different
1167 forms, like &a[0] and &a. */
1169 if ((TREE_CODE (expr
) == ADDR_EXPR
&& !DECL_P (TREE_OPERAND (expr
, 0)))
1170 || contain_complex_addr_expr (expr
))
1173 tree_to_aff_combination (expr
, TREE_TYPE (base
), &comb
);
1174 base
= fold_convert (TREE_TYPE (base
), aff_combination_to_tree (&comb
));
1178 iv
->base_object
= determine_base_object (base
);
1181 iv
->nonlin_use
= NULL
;
1182 iv
->ssa_name
= NULL_TREE
;
1184 && !iv_can_overflow_p (data
->current_loop
, TREE_TYPE (base
),
1187 iv
->no_overflow
= no_overflow
;
1188 iv
->have_address_use
= false;
1193 /* Sets STEP and BASE for induction variable IV. NO_OVERFLOW implies the IV
1194 doesn't overflow. */
1197 set_iv (struct ivopts_data
*data
, tree iv
, tree base
, tree step
,
1200 struct version_info
*info
= name_info (data
, iv
);
1202 gcc_assert (!info
->iv
);
1204 bitmap_set_bit (data
->relevant
, SSA_NAME_VERSION (iv
));
1205 info
->iv
= alloc_iv (data
, base
, step
, no_overflow
);
1206 info
->iv
->ssa_name
= iv
;
1209 /* Finds induction variable declaration for VAR. */
1212 get_iv (struct ivopts_data
*data
, tree var
)
1215 tree type
= TREE_TYPE (var
);
1217 if (!POINTER_TYPE_P (type
)
1218 && !INTEGRAL_TYPE_P (type
))
1221 if (!name_info (data
, var
)->iv
)
1223 bb
= gimple_bb (SSA_NAME_DEF_STMT (var
));
1226 || !flow_bb_inside_loop_p (data
->current_loop
, bb
))
1227 set_iv (data
, var
, var
, build_int_cst (type
, 0), true);
1230 return name_info (data
, var
)->iv
;
1233 /* Return the first non-invariant ssa var found in EXPR. */
1236 extract_single_var_from_expr (tree expr
)
1240 enum tree_code code
;
1242 if (!expr
|| is_gimple_min_invariant (expr
))
1245 code
= TREE_CODE (expr
);
1246 if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code
)))
1248 n
= TREE_OPERAND_LENGTH (expr
);
1249 for (i
= 0; i
< n
; i
++)
1251 tmp
= extract_single_var_from_expr (TREE_OPERAND (expr
, i
));
1257 return (TREE_CODE (expr
) == SSA_NAME
) ? expr
: NULL
;
1260 /* Finds basic ivs. */
1263 find_bivs (struct ivopts_data
*data
)
1267 tree step
, type
, base
, stop
;
1269 struct loop
*loop
= data
->current_loop
;
1272 for (psi
= gsi_start_phis (loop
->header
); !gsi_end_p (psi
); gsi_next (&psi
))
1276 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (PHI_RESULT (phi
)))
1279 if (virtual_operand_p (PHI_RESULT (phi
)))
1282 if (!simple_iv (loop
, loop
, PHI_RESULT (phi
), &iv
, true))
1285 if (integer_zerop (iv
.step
))
1289 base
= PHI_ARG_DEF_FROM_EDGE (phi
, loop_preheader_edge (loop
));
1290 /* Stop expanding iv base at the first ssa var referred by iv step.
1291 Ideally we should stop at any ssa var, because that's expensive
1292 and unusual to happen, we just do it on the first one.
1294 See PR64705 for the rationale. */
1295 stop
= extract_single_var_from_expr (step
);
1296 base
= expand_simple_operations (base
, stop
);
1297 if (contains_abnormal_ssa_name_p (base
)
1298 || contains_abnormal_ssa_name_p (step
))
1301 type
= TREE_TYPE (PHI_RESULT (phi
));
1302 base
= fold_convert (type
, base
);
1305 if (POINTER_TYPE_P (type
))
1306 step
= convert_to_ptrofftype (step
);
1308 step
= fold_convert (type
, step
);
1311 set_iv (data
, PHI_RESULT (phi
), base
, step
, iv
.no_overflow
);
1318 /* Marks basic ivs. */
1321 mark_bivs (struct ivopts_data
*data
)
1326 struct iv
*iv
, *incr_iv
;
1327 struct loop
*loop
= data
->current_loop
;
1328 basic_block incr_bb
;
1331 data
->bivs_not_used_in_addr
= 0;
1332 for (psi
= gsi_start_phis (loop
->header
); !gsi_end_p (psi
); gsi_next (&psi
))
1336 iv
= get_iv (data
, PHI_RESULT (phi
));
1340 var
= PHI_ARG_DEF_FROM_EDGE (phi
, loop_latch_edge (loop
));
1341 def
= SSA_NAME_DEF_STMT (var
);
1342 /* Don't mark iv peeled from other one as biv. */
1344 && gimple_code (def
) == GIMPLE_PHI
1345 && gimple_bb (def
) == loop
->header
)
1348 incr_iv
= get_iv (data
, var
);
1352 /* If the increment is in the subloop, ignore it. */
1353 incr_bb
= gimple_bb (SSA_NAME_DEF_STMT (var
));
1354 if (incr_bb
->loop_father
!= data
->current_loop
1355 || (incr_bb
->flags
& BB_IRREDUCIBLE_LOOP
))
1359 incr_iv
->biv_p
= true;
1360 if (iv
->no_overflow
)
1361 data
->bivs_not_used_in_addr
++;
1362 if (incr_iv
->no_overflow
)
1363 data
->bivs_not_used_in_addr
++;
1367 /* Checks whether STMT defines a linear induction variable and stores its
1368 parameters to IV. */
1371 find_givs_in_stmt_scev (struct ivopts_data
*data
, gimple
*stmt
, affine_iv
*iv
)
1374 struct loop
*loop
= data
->current_loop
;
1376 iv
->base
= NULL_TREE
;
1377 iv
->step
= NULL_TREE
;
1379 if (gimple_code (stmt
) != GIMPLE_ASSIGN
)
1382 lhs
= gimple_assign_lhs (stmt
);
1383 if (TREE_CODE (lhs
) != SSA_NAME
)
1386 if (!simple_iv (loop
, loop_containing_stmt (stmt
), lhs
, iv
, true))
1389 /* Stop expanding iv base at the first ssa var referred by iv step.
1390 Ideally we should stop at any ssa var, because that's expensive
1391 and unusual to happen, we just do it on the first one.
1393 See PR64705 for the rationale. */
1394 stop
= extract_single_var_from_expr (iv
->step
);
1395 iv
->base
= expand_simple_operations (iv
->base
, stop
);
1396 if (contains_abnormal_ssa_name_p (iv
->base
)
1397 || contains_abnormal_ssa_name_p (iv
->step
))
1400 /* If STMT could throw, then do not consider STMT as defining a GIV.
1401 While this will suppress optimizations, we can not safely delete this
1402 GIV and associated statements, even if it appears it is not used. */
1403 if (stmt_could_throw_p (stmt
))
1409 /* Finds general ivs in statement STMT. */
1412 find_givs_in_stmt (struct ivopts_data
*data
, gimple
*stmt
)
1416 if (!find_givs_in_stmt_scev (data
, stmt
, &iv
))
1419 set_iv (data
, gimple_assign_lhs (stmt
), iv
.base
, iv
.step
, iv
.no_overflow
);
1422 /* Finds general ivs in basic block BB. */
1425 find_givs_in_bb (struct ivopts_data
*data
, basic_block bb
)
1427 gimple_stmt_iterator bsi
;
1429 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
1430 find_givs_in_stmt (data
, gsi_stmt (bsi
));
1433 /* Finds general ivs. */
1436 find_givs (struct ivopts_data
*data
)
1438 struct loop
*loop
= data
->current_loop
;
1439 basic_block
*body
= get_loop_body_in_dom_order (loop
);
1442 for (i
= 0; i
< loop
->num_nodes
; i
++)
1443 find_givs_in_bb (data
, body
[i
]);
1447 /* For each ssa name defined in LOOP determines whether it is an induction
1448 variable and if so, its initial value and step. */
1451 find_induction_variables (struct ivopts_data
*data
)
1456 if (!find_bivs (data
))
1462 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1464 struct tree_niter_desc
*niter
= niter_for_single_dom_exit (data
);
1468 fprintf (dump_file
, " number of iterations ");
1469 print_generic_expr (dump_file
, niter
->niter
, TDF_SLIM
);
1470 if (!integer_zerop (niter
->may_be_zero
))
1472 fprintf (dump_file
, "; zero if ");
1473 print_generic_expr (dump_file
, niter
->may_be_zero
, TDF_SLIM
);
1475 fprintf (dump_file
, "\n");
1478 fprintf (dump_file
, "\n<Induction Vars>:\n");
1479 EXECUTE_IF_SET_IN_BITMAP (data
->relevant
, 0, i
, bi
)
1481 struct version_info
*info
= ver_info (data
, i
);
1482 if (info
->iv
&& info
->iv
->step
&& !integer_zerop (info
->iv
->step
))
1483 dump_iv (dump_file
, ver_info (data
, i
)->iv
, true, 0);
1490 /* Records a use of TYPE at *USE_P in STMT whose value is IV in GROUP.
1491 For address type use, ADDR_BASE is the stripped IV base, ADDR_OFFSET
1492 is the const offset stripped from IV base; for other types use, both
1493 are zero by default. */
1495 static struct iv_use
*
1496 record_use (struct iv_group
*group
, tree
*use_p
, struct iv
*iv
,
1497 gimple
*stmt
, enum use_type type
, tree addr_base
,
1498 unsigned HOST_WIDE_INT addr_offset
)
1500 struct iv_use
*use
= XCNEW (struct iv_use
);
1502 use
->id
= group
->vuses
.length ();
1503 use
->group_id
= group
->id
;
1508 use
->addr_base
= addr_base
;
1509 use
->addr_offset
= addr_offset
;
1511 group
->vuses
.safe_push (use
);
1515 /* Checks whether OP is a loop-level invariant and if so, records it.
1516 NONLINEAR_USE is true if the invariant is used in a way we do not
1517 handle specially. */
1520 record_invariant (struct ivopts_data
*data
, tree op
, bool nonlinear_use
)
1523 struct version_info
*info
;
1525 if (TREE_CODE (op
) != SSA_NAME
1526 || virtual_operand_p (op
))
1529 bb
= gimple_bb (SSA_NAME_DEF_STMT (op
));
1531 && flow_bb_inside_loop_p (data
->current_loop
, bb
))
1534 info
= name_info (data
, op
);
1536 info
->has_nonlin_use
|= nonlinear_use
;
1538 info
->inv_id
= ++data
->max_inv_id
;
1539 bitmap_set_bit (data
->relevant
, SSA_NAME_VERSION (op
));
1543 strip_offset (tree expr
, unsigned HOST_WIDE_INT
*offset
);
1545 /* Record a group of TYPE. */
1547 static struct iv_group
*
1548 record_group (struct ivopts_data
*data
, enum use_type type
)
1550 struct iv_group
*group
= XCNEW (struct iv_group
);
1552 group
->id
= data
->vgroups
.length ();
1554 group
->related_cands
= BITMAP_ALLOC (NULL
);
1555 group
->vuses
.create (1);
1557 data
->vgroups
.safe_push (group
);
1561 /* Record a use of TYPE at *USE_P in STMT whose value is IV in a group.
1562 New group will be created if there is no existing group for the use. */
1564 static struct iv_use
*
1565 record_group_use (struct ivopts_data
*data
, tree
*use_p
,
1566 struct iv
*iv
, gimple
*stmt
, enum use_type type
)
1568 tree addr_base
= NULL
;
1569 struct iv_group
*group
= NULL
;
1570 unsigned HOST_WIDE_INT addr_offset
= 0;
1572 /* Record non address type use in a new group. */
1573 if (type
== USE_ADDRESS
&& iv
->base_object
)
1577 addr_base
= strip_offset (iv
->base
, &addr_offset
);
1578 for (i
= 0; i
< data
->vgroups
.length (); i
++)
1582 group
= data
->vgroups
[i
];
1583 use
= group
->vuses
[0];
1584 if (use
->type
!= USE_ADDRESS
|| !use
->iv
->base_object
)
1587 /* Check if it has the same stripped base and step. */
1588 if (operand_equal_p (iv
->base_object
, use
->iv
->base_object
, 0)
1589 && operand_equal_p (iv
->step
, use
->iv
->step
, 0)
1590 && operand_equal_p (addr_base
, use
->addr_base
, 0))
1593 if (i
== data
->vgroups
.length ())
1598 group
= record_group (data
, type
);
1600 return record_use (group
, use_p
, iv
, stmt
, type
, addr_base
, addr_offset
);
1603 /* Checks whether the use OP is interesting and if so, records it. */
1605 static struct iv_use
*
1606 find_interesting_uses_op (struct ivopts_data
*data
, tree op
)
1612 if (TREE_CODE (op
) != SSA_NAME
)
1615 iv
= get_iv (data
, op
);
1621 gcc_assert (iv
->nonlin_use
->type
== USE_NONLINEAR_EXPR
);
1622 return iv
->nonlin_use
;
1625 if (integer_zerop (iv
->step
))
1627 record_invariant (data
, op
, true);
1631 stmt
= SSA_NAME_DEF_STMT (op
);
1632 gcc_assert (gimple_code (stmt
) == GIMPLE_PHI
|| is_gimple_assign (stmt
));
1634 use
= record_group_use (data
, NULL
, iv
, stmt
, USE_NONLINEAR_EXPR
);
1635 iv
->nonlin_use
= use
;
1639 /* Given a condition in statement STMT, checks whether it is a compare
1640 of an induction variable and an invariant. If this is the case,
1641 CONTROL_VAR is set to location of the iv, BOUND to the location of
1642 the invariant, IV_VAR and IV_BOUND are set to the corresponding
1643 induction variable descriptions, and true is returned. If this is not
1644 the case, CONTROL_VAR and BOUND are set to the arguments of the
1645 condition and false is returned. */
1648 extract_cond_operands (struct ivopts_data
*data
, gimple
*stmt
,
1649 tree
**control_var
, tree
**bound
,
1650 struct iv
**iv_var
, struct iv
**iv_bound
)
1652 /* The objects returned when COND has constant operands. */
1653 static struct iv const_iv
;
1655 tree
*op0
= &zero
, *op1
= &zero
;
1656 struct iv
*iv0
= &const_iv
, *iv1
= &const_iv
;
1659 if (gimple_code (stmt
) == GIMPLE_COND
)
1661 gcond
*cond_stmt
= as_a
<gcond
*> (stmt
);
1662 op0
= gimple_cond_lhs_ptr (cond_stmt
);
1663 op1
= gimple_cond_rhs_ptr (cond_stmt
);
1667 op0
= gimple_assign_rhs1_ptr (stmt
);
1668 op1
= gimple_assign_rhs2_ptr (stmt
);
1671 zero
= integer_zero_node
;
1672 const_iv
.step
= integer_zero_node
;
1674 if (TREE_CODE (*op0
) == SSA_NAME
)
1675 iv0
= get_iv (data
, *op0
);
1676 if (TREE_CODE (*op1
) == SSA_NAME
)
1677 iv1
= get_iv (data
, *op1
);
1679 /* Exactly one of the compared values must be an iv, and the other one must
1684 if (integer_zerop (iv0
->step
))
1686 /* Control variable may be on the other side. */
1687 std::swap (op0
, op1
);
1688 std::swap (iv0
, iv1
);
1690 ret
= !integer_zerop (iv0
->step
) && integer_zerop (iv1
->step
);
1705 /* Checks whether the condition in STMT is interesting and if so,
1709 find_interesting_uses_cond (struct ivopts_data
*data
, gimple
*stmt
)
1711 tree
*var_p
, *bound_p
;
1714 if (!extract_cond_operands (data
, stmt
, &var_p
, &bound_p
, &var_iv
, NULL
))
1716 find_interesting_uses_op (data
, *var_p
);
1717 find_interesting_uses_op (data
, *bound_p
);
1721 record_group_use (data
, NULL
, var_iv
, stmt
, USE_COMPARE
);
1724 /* Returns the outermost loop EXPR is obviously invariant in
1725 relative to the loop LOOP, i.e. if all its operands are defined
1726 outside of the returned loop. Returns NULL if EXPR is not
1727 even obviously invariant in LOOP. */
1730 outermost_invariant_loop_for_expr (struct loop
*loop
, tree expr
)
1735 if (is_gimple_min_invariant (expr
))
1736 return current_loops
->tree_root
;
1738 if (TREE_CODE (expr
) == SSA_NAME
)
1740 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (expr
));
1743 if (flow_bb_inside_loop_p (loop
, def_bb
))
1745 return superloop_at_depth (loop
,
1746 loop_depth (def_bb
->loop_father
) + 1);
1749 return current_loops
->tree_root
;
1755 unsigned maxdepth
= 0;
1756 len
= TREE_OPERAND_LENGTH (expr
);
1757 for (i
= 0; i
< len
; i
++)
1759 struct loop
*ivloop
;
1760 if (!TREE_OPERAND (expr
, i
))
1763 ivloop
= outermost_invariant_loop_for_expr (loop
, TREE_OPERAND (expr
, i
));
1766 maxdepth
= MAX (maxdepth
, loop_depth (ivloop
));
1769 return superloop_at_depth (loop
, maxdepth
);
1772 /* Returns true if expression EXPR is obviously invariant in LOOP,
1773 i.e. if all its operands are defined outside of the LOOP. LOOP
1774 should not be the function body. */
1777 expr_invariant_in_loop_p (struct loop
*loop
, tree expr
)
1782 gcc_assert (loop_depth (loop
) > 0);
1784 if (is_gimple_min_invariant (expr
))
1787 if (TREE_CODE (expr
) == SSA_NAME
)
1789 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (expr
));
1791 && flow_bb_inside_loop_p (loop
, def_bb
))
1800 len
= TREE_OPERAND_LENGTH (expr
);
1801 for (i
= 0; i
< len
; i
++)
1802 if (TREE_OPERAND (expr
, i
)
1803 && !expr_invariant_in_loop_p (loop
, TREE_OPERAND (expr
, i
)))
1809 /* Given expression EXPR which computes inductive values with respect
1810 to loop recorded in DATA, this function returns biv from which EXPR
1811 is derived by tracing definition chains of ssa variables in EXPR. */
1814 find_deriving_biv_for_expr (struct ivopts_data
*data
, tree expr
)
1819 enum tree_code code
;
1822 if (expr
== NULL_TREE
)
1825 if (is_gimple_min_invariant (expr
))
1828 code
= TREE_CODE (expr
);
1829 if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code
)))
1831 n
= TREE_OPERAND_LENGTH (expr
);
1832 for (i
= 0; i
< n
; i
++)
1834 iv
= find_deriving_biv_for_expr (data
, TREE_OPERAND (expr
, i
));
1840 /* Stop if it's not ssa name. */
1841 if (code
!= SSA_NAME
)
1844 iv
= get_iv (data
, expr
);
1845 if (!iv
|| integer_zerop (iv
->step
))
1850 stmt
= SSA_NAME_DEF_STMT (expr
);
1851 if (gphi
*phi
= dyn_cast
<gphi
*> (stmt
))
1854 use_operand_p use_p
;
1856 if (virtual_operand_p (gimple_phi_result (phi
)))
1859 FOR_EACH_PHI_ARG (use_p
, phi
, iter
, SSA_OP_USE
)
1861 tree use
= USE_FROM_PTR (use_p
);
1862 iv
= find_deriving_biv_for_expr (data
, use
);
1868 if (gimple_code (stmt
) != GIMPLE_ASSIGN
)
1871 e1
= gimple_assign_rhs1 (stmt
);
1872 code
= gimple_assign_rhs_code (stmt
);
1873 if (get_gimple_rhs_class (code
) == GIMPLE_SINGLE_RHS
)
1874 return find_deriving_biv_for_expr (data
, e1
);
1881 case POINTER_PLUS_EXPR
:
1882 /* Increments, decrements and multiplications by a constant
1884 e2
= gimple_assign_rhs2 (stmt
);
1885 iv
= find_deriving_biv_for_expr (data
, e2
);
1891 /* Casts are simple. */
1892 return find_deriving_biv_for_expr (data
, e1
);
1901 /* Record BIV, its predecessor and successor that they are used in
1902 address type uses. */
1905 record_biv_for_address_use (struct ivopts_data
*data
, struct iv
*biv
)
1908 tree type
, base_1
, base_2
;
1911 if (!biv
|| !biv
->biv_p
|| integer_zerop (biv
->step
)
1912 || biv
->have_address_use
|| !biv
->no_overflow
)
1915 type
= TREE_TYPE (biv
->base
);
1916 if (!INTEGRAL_TYPE_P (type
))
1919 biv
->have_address_use
= true;
1920 data
->bivs_not_used_in_addr
--;
1921 base_1
= fold_build2 (PLUS_EXPR
, type
, biv
->base
, biv
->step
);
1922 EXECUTE_IF_SET_IN_BITMAP (data
->relevant
, 0, i
, bi
)
1924 struct iv
*iv
= ver_info (data
, i
)->iv
;
1926 if (!iv
|| !iv
->biv_p
|| integer_zerop (iv
->step
)
1927 || iv
->have_address_use
|| !iv
->no_overflow
)
1930 if (type
!= TREE_TYPE (iv
->base
)
1931 || !INTEGRAL_TYPE_P (TREE_TYPE (iv
->base
)))
1934 if (!operand_equal_p (biv
->step
, iv
->step
, 0))
1937 base_2
= fold_build2 (PLUS_EXPR
, type
, iv
->base
, iv
->step
);
1938 if (operand_equal_p (base_1
, iv
->base
, 0)
1939 || operand_equal_p (base_2
, biv
->base
, 0))
1941 iv
->have_address_use
= true;
1942 data
->bivs_not_used_in_addr
--;
1947 /* Cumulates the steps of indices into DATA and replaces their values with the
1948 initial ones. Returns false when the value of the index cannot be determined.
1949 Callback for for_each_index. */
1951 struct ifs_ivopts_data
1953 struct ivopts_data
*ivopts_data
;
1959 idx_find_step (tree base
, tree
*idx
, void *data
)
1961 struct ifs_ivopts_data
*dta
= (struct ifs_ivopts_data
*) data
;
1963 bool use_overflow_semantics
= false;
1964 tree step
, iv_base
, iv_step
, lbound
, off
;
1965 struct loop
*loop
= dta
->ivopts_data
->current_loop
;
1967 /* If base is a component ref, require that the offset of the reference
1969 if (TREE_CODE (base
) == COMPONENT_REF
)
1971 off
= component_ref_field_offset (base
);
1972 return expr_invariant_in_loop_p (loop
, off
);
1975 /* If base is array, first check whether we will be able to move the
1976 reference out of the loop (in order to take its address in strength
1977 reduction). In order for this to work we need both lower bound
1978 and step to be loop invariants. */
1979 if (TREE_CODE (base
) == ARRAY_REF
|| TREE_CODE (base
) == ARRAY_RANGE_REF
)
1981 /* Moreover, for a range, the size needs to be invariant as well. */
1982 if (TREE_CODE (base
) == ARRAY_RANGE_REF
1983 && !expr_invariant_in_loop_p (loop
, TYPE_SIZE (TREE_TYPE (base
))))
1986 step
= array_ref_element_size (base
);
1987 lbound
= array_ref_low_bound (base
);
1989 if (!expr_invariant_in_loop_p (loop
, step
)
1990 || !expr_invariant_in_loop_p (loop
, lbound
))
1994 if (TREE_CODE (*idx
) != SSA_NAME
)
1997 iv
= get_iv (dta
->ivopts_data
, *idx
);
2001 /* XXX We produce for a base of *D42 with iv->base being &x[0]
2002 *&x[0], which is not folded and does not trigger the
2003 ARRAY_REF path below. */
2006 if (integer_zerop (iv
->step
))
2009 if (TREE_CODE (base
) == ARRAY_REF
|| TREE_CODE (base
) == ARRAY_RANGE_REF
)
2011 step
= array_ref_element_size (base
);
2013 /* We only handle addresses whose step is an integer constant. */
2014 if (TREE_CODE (step
) != INTEGER_CST
)
2018 /* The step for pointer arithmetics already is 1 byte. */
2019 step
= size_one_node
;
2023 if (iv
->no_overflow
&& nowrap_type_p (TREE_TYPE (iv_step
)))
2024 use_overflow_semantics
= true;
2026 if (!convert_affine_scev (dta
->ivopts_data
->current_loop
,
2027 sizetype
, &iv_base
, &iv_step
, dta
->stmt
,
2028 use_overflow_semantics
))
2030 /* The index might wrap. */
2034 step
= fold_build2 (MULT_EXPR
, sizetype
, step
, iv_step
);
2035 dta
->step
= fold_build2 (PLUS_EXPR
, sizetype
, dta
->step
, step
);
2037 if (dta
->ivopts_data
->bivs_not_used_in_addr
)
2040 iv
= find_deriving_biv_for_expr (dta
->ivopts_data
, iv
->ssa_name
);
2042 record_biv_for_address_use (dta
->ivopts_data
, iv
);
2047 /* Records use in index IDX. Callback for for_each_index. Ivopts data
2048 object is passed to it in DATA. */
2051 idx_record_use (tree base
, tree
*idx
,
2054 struct ivopts_data
*data
= (struct ivopts_data
*) vdata
;
2055 find_interesting_uses_op (data
, *idx
);
2056 if (TREE_CODE (base
) == ARRAY_REF
|| TREE_CODE (base
) == ARRAY_RANGE_REF
)
2058 find_interesting_uses_op (data
, array_ref_element_size (base
));
2059 find_interesting_uses_op (data
, array_ref_low_bound (base
));
2064 /* If we can prove that TOP = cst * BOT for some constant cst,
2065 store cst to MUL and return true. Otherwise return false.
2066 The returned value is always sign-extended, regardless of the
2067 signedness of TOP and BOT. */
2070 constant_multiple_of (tree top
, tree bot
, widest_int
*mul
)
2073 enum tree_code code
;
2074 unsigned precision
= TYPE_PRECISION (TREE_TYPE (top
));
2075 widest_int res
, p0
, p1
;
2080 if (operand_equal_p (top
, bot
, 0))
2086 code
= TREE_CODE (top
);
2090 mby
= TREE_OPERAND (top
, 1);
2091 if (TREE_CODE (mby
) != INTEGER_CST
)
2094 if (!constant_multiple_of (TREE_OPERAND (top
, 0), bot
, &res
))
2097 *mul
= wi::sext (res
* wi::to_widest (mby
), precision
);
2102 if (!constant_multiple_of (TREE_OPERAND (top
, 0), bot
, &p0
)
2103 || !constant_multiple_of (TREE_OPERAND (top
, 1), bot
, &p1
))
2106 if (code
== MINUS_EXPR
)
2108 *mul
= wi::sext (p0
+ p1
, precision
);
2112 if (TREE_CODE (bot
) != INTEGER_CST
)
2115 p0
= widest_int::from (top
, SIGNED
);
2116 p1
= widest_int::from (bot
, SIGNED
);
2119 *mul
= wi::sext (wi::divmod_trunc (p0
, p1
, SIGNED
, &res
), precision
);
2127 /* Return true if memory reference REF with step STEP may be unaligned. */
2130 may_be_unaligned_p (tree ref
, tree step
)
2132 /* TARGET_MEM_REFs are translated directly to valid MEMs on the target,
2133 thus they are not misaligned. */
2134 if (TREE_CODE (ref
) == TARGET_MEM_REF
)
2137 unsigned int align
= TYPE_ALIGN (TREE_TYPE (ref
));
2138 if (GET_MODE_ALIGNMENT (TYPE_MODE (TREE_TYPE (ref
))) > align
)
2139 align
= GET_MODE_ALIGNMENT (TYPE_MODE (TREE_TYPE (ref
)));
2141 unsigned HOST_WIDE_INT bitpos
;
2142 unsigned int ref_align
;
2143 get_object_alignment_1 (ref
, &ref_align
, &bitpos
);
2144 if (ref_align
< align
2145 || (bitpos
% align
) != 0
2146 || (bitpos
% BITS_PER_UNIT
) != 0)
2149 unsigned int trailing_zeros
= tree_ctz (step
);
2150 if (trailing_zeros
< HOST_BITS_PER_INT
2151 && (1U << trailing_zeros
) * BITS_PER_UNIT
< align
)
2157 /* Return true if EXPR may be non-addressable. */
2160 may_be_nonaddressable_p (tree expr
)
2162 switch (TREE_CODE (expr
))
2164 case TARGET_MEM_REF
:
2165 /* TARGET_MEM_REFs are translated directly to valid MEMs on the
2166 target, thus they are always addressable. */
2170 /* Likewise for MEM_REFs, modulo the storage order. */
2171 return REF_REVERSE_STORAGE_ORDER (expr
);
2174 if (REF_REVERSE_STORAGE_ORDER (expr
))
2176 return may_be_nonaddressable_p (TREE_OPERAND (expr
, 0));
2179 if (TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_OPERAND (expr
, 0))))
2181 return DECL_NONADDRESSABLE_P (TREE_OPERAND (expr
, 1))
2182 || may_be_nonaddressable_p (TREE_OPERAND (expr
, 0));
2185 case ARRAY_RANGE_REF
:
2186 if (TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_OPERAND (expr
, 0))))
2188 return may_be_nonaddressable_p (TREE_OPERAND (expr
, 0));
2190 case VIEW_CONVERT_EXPR
:
2191 /* This kind of view-conversions may wrap non-addressable objects
2192 and make them look addressable. After some processing the
2193 non-addressability may be uncovered again, causing ADDR_EXPRs
2194 of inappropriate objects to be built. */
2195 if (is_gimple_reg (TREE_OPERAND (expr
, 0))
2196 || !is_gimple_addressable (TREE_OPERAND (expr
, 0)))
2198 return may_be_nonaddressable_p (TREE_OPERAND (expr
, 0));
2210 /* Finds addresses in *OP_P inside STMT. */
2213 find_interesting_uses_address (struct ivopts_data
*data
, gimple
*stmt
,
2216 tree base
= *op_p
, step
= size_zero_node
;
2218 struct ifs_ivopts_data ifs_ivopts_data
;
2220 /* Do not play with volatile memory references. A bit too conservative,
2221 perhaps, but safe. */
2222 if (gimple_has_volatile_ops (stmt
))
2225 /* Ignore bitfields for now. Not really something terribly complicated
2227 if (TREE_CODE (base
) == BIT_FIELD_REF
)
2230 base
= unshare_expr (base
);
2232 if (TREE_CODE (base
) == TARGET_MEM_REF
)
2234 tree type
= build_pointer_type (TREE_TYPE (base
));
2238 && TREE_CODE (TMR_BASE (base
)) == SSA_NAME
)
2240 civ
= get_iv (data
, TMR_BASE (base
));
2244 TMR_BASE (base
) = civ
->base
;
2247 if (TMR_INDEX2 (base
)
2248 && TREE_CODE (TMR_INDEX2 (base
)) == SSA_NAME
)
2250 civ
= get_iv (data
, TMR_INDEX2 (base
));
2254 TMR_INDEX2 (base
) = civ
->base
;
2257 if (TMR_INDEX (base
)
2258 && TREE_CODE (TMR_INDEX (base
)) == SSA_NAME
)
2260 civ
= get_iv (data
, TMR_INDEX (base
));
2264 TMR_INDEX (base
) = civ
->base
;
2269 if (TMR_STEP (base
))
2270 astep
= fold_build2 (MULT_EXPR
, type
, TMR_STEP (base
), astep
);
2272 step
= fold_build2 (PLUS_EXPR
, type
, step
, astep
);
2276 if (integer_zerop (step
))
2278 base
= tree_mem_ref_addr (type
, base
);
2282 ifs_ivopts_data
.ivopts_data
= data
;
2283 ifs_ivopts_data
.stmt
= stmt
;
2284 ifs_ivopts_data
.step
= size_zero_node
;
2285 if (!for_each_index (&base
, idx_find_step
, &ifs_ivopts_data
)
2286 || integer_zerop (ifs_ivopts_data
.step
))
2288 step
= ifs_ivopts_data
.step
;
2290 /* Check that the base expression is addressable. This needs
2291 to be done after substituting bases of IVs into it. */
2292 if (may_be_nonaddressable_p (base
))
2295 /* Moreover, on strict alignment platforms, check that it is
2296 sufficiently aligned. */
2297 if (STRICT_ALIGNMENT
&& may_be_unaligned_p (base
, step
))
2300 base
= build_fold_addr_expr (base
);
2302 /* Substituting bases of IVs into the base expression might
2303 have caused folding opportunities. */
2304 if (TREE_CODE (base
) == ADDR_EXPR
)
2306 tree
*ref
= &TREE_OPERAND (base
, 0);
2307 while (handled_component_p (*ref
))
2308 ref
= &TREE_OPERAND (*ref
, 0);
2309 if (TREE_CODE (*ref
) == MEM_REF
)
2311 tree tem
= fold_binary (MEM_REF
, TREE_TYPE (*ref
),
2312 TREE_OPERAND (*ref
, 0),
2313 TREE_OPERAND (*ref
, 1));
2320 civ
= alloc_iv (data
, base
, step
);
2321 record_group_use (data
, op_p
, civ
, stmt
, USE_ADDRESS
);
2325 for_each_index (op_p
, idx_record_use
, data
);
2328 /* Finds and records invariants used in STMT. */
2331 find_invariants_stmt (struct ivopts_data
*data
, gimple
*stmt
)
2334 use_operand_p use_p
;
2337 FOR_EACH_PHI_OR_STMT_USE (use_p
, stmt
, iter
, SSA_OP_USE
)
2339 op
= USE_FROM_PTR (use_p
);
2340 record_invariant (data
, op
, false);
2344 /* Finds interesting uses of induction variables in the statement STMT. */
2347 find_interesting_uses_stmt (struct ivopts_data
*data
, gimple
*stmt
)
2350 tree op
, *lhs
, *rhs
;
2352 use_operand_p use_p
;
2353 enum tree_code code
;
2355 find_invariants_stmt (data
, stmt
);
2357 if (gimple_code (stmt
) == GIMPLE_COND
)
2359 find_interesting_uses_cond (data
, stmt
);
2363 if (is_gimple_assign (stmt
))
2365 lhs
= gimple_assign_lhs_ptr (stmt
);
2366 rhs
= gimple_assign_rhs1_ptr (stmt
);
2368 if (TREE_CODE (*lhs
) == SSA_NAME
)
2370 /* If the statement defines an induction variable, the uses are not
2371 interesting by themselves. */
2373 iv
= get_iv (data
, *lhs
);
2375 if (iv
&& !integer_zerop (iv
->step
))
2379 code
= gimple_assign_rhs_code (stmt
);
2380 if (get_gimple_rhs_class (code
) == GIMPLE_SINGLE_RHS
2381 && (REFERENCE_CLASS_P (*rhs
)
2382 || is_gimple_val (*rhs
)))
2384 if (REFERENCE_CLASS_P (*rhs
))
2385 find_interesting_uses_address (data
, stmt
, rhs
);
2387 find_interesting_uses_op (data
, *rhs
);
2389 if (REFERENCE_CLASS_P (*lhs
))
2390 find_interesting_uses_address (data
, stmt
, lhs
);
2393 else if (TREE_CODE_CLASS (code
) == tcc_comparison
)
2395 find_interesting_uses_cond (data
, stmt
);
2399 /* TODO -- we should also handle address uses of type
2401 memory = call (whatever);
2408 if (gimple_code (stmt
) == GIMPLE_PHI
2409 && gimple_bb (stmt
) == data
->current_loop
->header
)
2411 iv
= get_iv (data
, PHI_RESULT (stmt
));
2413 if (iv
&& !integer_zerop (iv
->step
))
2417 FOR_EACH_PHI_OR_STMT_USE (use_p
, stmt
, iter
, SSA_OP_USE
)
2419 op
= USE_FROM_PTR (use_p
);
2421 if (TREE_CODE (op
) != SSA_NAME
)
2424 iv
= get_iv (data
, op
);
2428 find_interesting_uses_op (data
, op
);
2432 /* Finds interesting uses of induction variables outside of loops
2433 on loop exit edge EXIT. */
2436 find_interesting_uses_outside (struct ivopts_data
*data
, edge exit
)
2442 for (psi
= gsi_start_phis (exit
->dest
); !gsi_end_p (psi
); gsi_next (&psi
))
2445 def
= PHI_ARG_DEF_FROM_EDGE (phi
, exit
);
2446 if (!virtual_operand_p (def
))
2447 find_interesting_uses_op (data
, def
);
2451 /* Compute maximum offset of [base + offset] addressing mode
2452 for memory reference represented by USE. */
2454 static HOST_WIDE_INT
2455 compute_max_addr_offset (struct iv_use
*use
)
2459 HOST_WIDE_INT i
, off
;
2460 unsigned list_index
, num
;
2462 machine_mode mem_mode
, addr_mode
;
2463 static vec
<HOST_WIDE_INT
> max_offset_list
;
2465 as
= TYPE_ADDR_SPACE (TREE_TYPE (use
->iv
->base
));
2466 mem_mode
= TYPE_MODE (TREE_TYPE (*use
->op_p
));
2468 num
= max_offset_list
.length ();
2469 list_index
= (unsigned) as
* MAX_MACHINE_MODE
+ (unsigned) mem_mode
;
2470 if (list_index
>= num
)
2472 max_offset_list
.safe_grow (list_index
+ MAX_MACHINE_MODE
);
2473 for (; num
< max_offset_list
.length (); num
++)
2474 max_offset_list
[num
] = -1;
2477 off
= max_offset_list
[list_index
];
2481 addr_mode
= targetm
.addr_space
.address_mode (as
);
2482 reg
= gen_raw_REG (addr_mode
, LAST_VIRTUAL_REGISTER
+ 1);
2483 addr
= gen_rtx_fmt_ee (PLUS
, addr_mode
, reg
, NULL_RTX
);
2485 width
= GET_MODE_BITSIZE (addr_mode
) - 1;
2486 if (width
> (HOST_BITS_PER_WIDE_INT
- 1))
2487 width
= HOST_BITS_PER_WIDE_INT
- 1;
2489 for (i
= width
; i
> 0; i
--)
2491 off
= (HOST_WIDE_INT_1U
<< i
) - 1;
2492 XEXP (addr
, 1) = gen_int_mode (off
, addr_mode
);
2493 if (memory_address_addr_space_p (mem_mode
, addr
, as
))
2496 /* For some strict-alignment targets, the offset must be naturally
2497 aligned. Try an aligned offset if mem_mode is not QImode. */
2498 off
= (HOST_WIDE_INT_1U
<< i
);
2499 if (off
> GET_MODE_SIZE (mem_mode
) && mem_mode
!= QImode
)
2501 off
-= GET_MODE_SIZE (mem_mode
);
2502 XEXP (addr
, 1) = gen_int_mode (off
, addr_mode
);
2503 if (memory_address_addr_space_p (mem_mode
, addr
, as
))
2510 max_offset_list
[list_index
] = off
;
2514 /* Comparison function to sort group in ascending order of addr_offset. */
2517 group_compare_offset (const void *a
, const void *b
)
2519 const struct iv_use
*const *u1
= (const struct iv_use
*const *) a
;
2520 const struct iv_use
*const *u2
= (const struct iv_use
*const *) b
;
2522 if ((*u1
)->addr_offset
!= (*u2
)->addr_offset
)
2523 return (*u1
)->addr_offset
< (*u2
)->addr_offset
? -1 : 1;
2528 /* Check if small groups should be split. Return true if no group
2529 contains more than two uses with distinct addr_offsets. Return
2530 false otherwise. We want to split such groups because:
2532 1) Small groups don't have much benefit and may interfer with
2533 general candidate selection.
2534 2) Size for problem with only small groups is usually small and
2535 general algorithm can handle it well.
2537 TODO -- Above claim may not hold when we want to merge memory
2538 accesses with conseuctive addresses. */
2541 split_small_address_groups_p (struct ivopts_data
*data
)
2543 unsigned int i
, j
, distinct
= 1;
2545 struct iv_group
*group
;
2547 for (i
= 0; i
< data
->vgroups
.length (); i
++)
2549 group
= data
->vgroups
[i
];
2550 if (group
->vuses
.length () == 1)
2553 gcc_assert (group
->type
== USE_ADDRESS
);
2554 if (group
->vuses
.length () == 2)
2556 if (group
->vuses
[0]->addr_offset
> group
->vuses
[1]->addr_offset
)
2557 std::swap (group
->vuses
[0], group
->vuses
[1]);
2560 group
->vuses
.qsort (group_compare_offset
);
2566 for (pre
= group
->vuses
[0], j
= 1; j
< group
->vuses
.length (); j
++)
2568 if (group
->vuses
[j
]->addr_offset
!= pre
->addr_offset
)
2570 pre
= group
->vuses
[j
];
2579 return (distinct
<= 2);
2582 /* For each group of address type uses, this function further groups
2583 these uses according to the maximum offset supported by target's
2584 [base + offset] addressing mode. */
2587 split_address_groups (struct ivopts_data
*data
)
2590 HOST_WIDE_INT max_offset
= -1;
2592 /* Reset max offset to split all small groups. */
2593 if (split_small_address_groups_p (data
))
2596 for (i
= 0; i
< data
->vgroups
.length (); i
++)
2598 struct iv_group
*group
= data
->vgroups
[i
];
2599 struct iv_use
*use
= group
->vuses
[0];
2602 use
->group_id
= group
->id
;
2603 if (group
->vuses
.length () == 1)
2606 if (max_offset
!= 0)
2607 max_offset
= compute_max_addr_offset (use
);
2609 for (j
= 1; j
< group
->vuses
.length (); j
++)
2611 struct iv_use
*next
= group
->vuses
[j
];
2613 /* Only uses with offset that can fit in offset part against
2614 the first use can be grouped together. */
2615 if (next
->addr_offset
- use
->addr_offset
2616 > (unsigned HOST_WIDE_INT
) max_offset
)
2620 next
->group_id
= group
->id
;
2623 if (j
< group
->vuses
.length ())
2625 struct iv_group
*new_group
= record_group (data
, group
->type
);
2626 new_group
->vuses
.safe_splice (group
->vuses
);
2627 new_group
->vuses
.block_remove (0, j
);
2628 group
->vuses
.truncate (j
);
2633 /* Finds uses of the induction variables that are interesting. */
2636 find_interesting_uses (struct ivopts_data
*data
)
2639 gimple_stmt_iterator bsi
;
2640 basic_block
*body
= get_loop_body (data
->current_loop
);
2644 for (i
= 0; i
< data
->current_loop
->num_nodes
; i
++)
2649 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2650 if (e
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)
2651 && !flow_bb_inside_loop_p (data
->current_loop
, e
->dest
))
2652 find_interesting_uses_outside (data
, e
);
2654 for (bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2655 find_interesting_uses_stmt (data
, gsi_stmt (bsi
));
2656 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2657 if (!is_gimple_debug (gsi_stmt (bsi
)))
2658 find_interesting_uses_stmt (data
, gsi_stmt (bsi
));
2661 split_address_groups (data
);
2663 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2667 fprintf (dump_file
, "\n<Invariant Vars>:\n");
2668 EXECUTE_IF_SET_IN_BITMAP (data
->relevant
, 0, i
, bi
)
2670 struct version_info
*info
= ver_info (data
, i
);
2673 fprintf (dump_file
, "Inv %d:\t", info
->inv_id
);
2674 print_generic_expr (dump_file
, info
->name
, TDF_SLIM
);
2675 fprintf (dump_file
, "%s\n",
2676 info
->has_nonlin_use
? "" : "\t(eliminable)");
2680 fprintf (dump_file
, "\n<IV Groups>:\n");
2681 dump_groups (dump_file
, data
);
2682 fprintf (dump_file
, "\n");
2688 /* Strips constant offsets from EXPR and stores them to OFFSET. If INSIDE_ADDR
2689 is true, assume we are inside an address. If TOP_COMPREF is true, assume
2690 we are at the top-level of the processed address. */
2693 strip_offset_1 (tree expr
, bool inside_addr
, bool top_compref
,
2694 HOST_WIDE_INT
*offset
)
2696 tree op0
= NULL_TREE
, op1
= NULL_TREE
, tmp
, step
;
2697 enum tree_code code
;
2698 tree type
, orig_type
= TREE_TYPE (expr
);
2699 HOST_WIDE_INT off0
, off1
, st
;
2700 tree orig_expr
= expr
;
2704 type
= TREE_TYPE (expr
);
2705 code
= TREE_CODE (expr
);
2711 if (!cst_and_fits_in_hwi (expr
)
2712 || integer_zerop (expr
))
2715 *offset
= int_cst_value (expr
);
2716 return build_int_cst (orig_type
, 0);
2718 case POINTER_PLUS_EXPR
:
2721 op0
= TREE_OPERAND (expr
, 0);
2722 op1
= TREE_OPERAND (expr
, 1);
2724 op0
= strip_offset_1 (op0
, false, false, &off0
);
2725 op1
= strip_offset_1 (op1
, false, false, &off1
);
2727 *offset
= (code
== MINUS_EXPR
? off0
- off1
: off0
+ off1
);
2728 if (op0
== TREE_OPERAND (expr
, 0)
2729 && op1
== TREE_OPERAND (expr
, 1))
2732 if (integer_zerop (op1
))
2734 else if (integer_zerop (op0
))
2736 if (code
== MINUS_EXPR
)
2737 expr
= fold_build1 (NEGATE_EXPR
, type
, op1
);
2742 expr
= fold_build2 (code
, type
, op0
, op1
);
2744 return fold_convert (orig_type
, expr
);
2747 op1
= TREE_OPERAND (expr
, 1);
2748 if (!cst_and_fits_in_hwi (op1
))
2751 op0
= TREE_OPERAND (expr
, 0);
2752 op0
= strip_offset_1 (op0
, false, false, &off0
);
2753 if (op0
== TREE_OPERAND (expr
, 0))
2756 *offset
= off0
* int_cst_value (op1
);
2757 if (integer_zerop (op0
))
2760 expr
= fold_build2 (MULT_EXPR
, type
, op0
, op1
);
2762 return fold_convert (orig_type
, expr
);
2765 case ARRAY_RANGE_REF
:
2769 step
= array_ref_element_size (expr
);
2770 if (!cst_and_fits_in_hwi (step
))
2773 st
= int_cst_value (step
);
2774 op1
= TREE_OPERAND (expr
, 1);
2775 op1
= strip_offset_1 (op1
, false, false, &off1
);
2776 *offset
= off1
* st
;
2779 && integer_zerop (op1
))
2781 /* Strip the component reference completely. */
2782 op0
= TREE_OPERAND (expr
, 0);
2783 op0
= strip_offset_1 (op0
, inside_addr
, top_compref
, &off0
);
2796 tmp
= component_ref_field_offset (expr
);
2797 field
= TREE_OPERAND (expr
, 1);
2799 && cst_and_fits_in_hwi (tmp
)
2800 && cst_and_fits_in_hwi (DECL_FIELD_BIT_OFFSET (field
)))
2802 HOST_WIDE_INT boffset
, abs_off
;
2804 /* Strip the component reference completely. */
2805 op0
= TREE_OPERAND (expr
, 0);
2806 op0
= strip_offset_1 (op0
, inside_addr
, top_compref
, &off0
);
2807 boffset
= int_cst_value (DECL_FIELD_BIT_OFFSET (field
));
2808 abs_off
= abs_hwi (boffset
) / BITS_PER_UNIT
;
2812 *offset
= off0
+ int_cst_value (tmp
) + abs_off
;
2819 op0
= TREE_OPERAND (expr
, 0);
2820 op0
= strip_offset_1 (op0
, true, true, &off0
);
2823 if (op0
== TREE_OPERAND (expr
, 0))
2826 expr
= build_fold_addr_expr (op0
);
2827 return fold_convert (orig_type
, expr
);
2830 /* ??? Offset operand? */
2831 inside_addr
= false;
2838 /* Default handling of expressions for that we want to recurse into
2839 the first operand. */
2840 op0
= TREE_OPERAND (expr
, 0);
2841 op0
= strip_offset_1 (op0
, inside_addr
, false, &off0
);
2844 if (op0
== TREE_OPERAND (expr
, 0)
2845 && (!op1
|| op1
== TREE_OPERAND (expr
, 1)))
2848 expr
= copy_node (expr
);
2849 TREE_OPERAND (expr
, 0) = op0
;
2851 TREE_OPERAND (expr
, 1) = op1
;
2853 /* Inside address, we might strip the top level component references,
2854 thus changing type of the expression. Handling of ADDR_EXPR
2856 expr
= fold_convert (orig_type
, expr
);
2861 /* Strips constant offsets from EXPR and stores them to OFFSET. */
2864 strip_offset (tree expr
, unsigned HOST_WIDE_INT
*offset
)
2867 tree core
= strip_offset_1 (expr
, false, false, &off
);
2872 /* Returns variant of TYPE that can be used as base for different uses.
2873 We return unsigned type with the same precision, which avoids problems
2877 generic_type_for (tree type
)
2879 if (POINTER_TYPE_P (type
))
2880 return unsigned_type_for (type
);
2882 if (TYPE_UNSIGNED (type
))
2885 return unsigned_type_for (type
);
2888 /* Records invariants in *EXPR_P. Callback for walk_tree. DATA contains
2889 the bitmap to that we should store it. */
2891 static struct ivopts_data
*fd_ivopts_data
;
2893 find_depends (tree
*expr_p
, int *ws ATTRIBUTE_UNUSED
, void *data
)
2895 bitmap
*depends_on
= (bitmap
*) data
;
2896 struct version_info
*info
;
2898 if (TREE_CODE (*expr_p
) != SSA_NAME
)
2900 info
= name_info (fd_ivopts_data
, *expr_p
);
2902 if (!info
->inv_id
|| info
->has_nonlin_use
)
2906 *depends_on
= BITMAP_ALLOC (NULL
);
2907 bitmap_set_bit (*depends_on
, info
->inv_id
);
2912 /* Adds a candidate BASE + STEP * i. Important field is set to IMPORTANT and
2913 position to POS. If USE is not NULL, the candidate is set as related to
2914 it. If both BASE and STEP are NULL, we add a pseudocandidate for the
2915 replacement of the final value of the iv by a direct computation. */
2917 static struct iv_cand
*
2918 add_candidate_1 (struct ivopts_data
*data
,
2919 tree base
, tree step
, bool important
, enum iv_position pos
,
2920 struct iv_use
*use
, gimple
*incremented_at
,
2921 struct iv
*orig_iv
= NULL
)
2924 struct iv_cand
*cand
= NULL
;
2925 tree type
, orig_type
;
2927 gcc_assert (base
&& step
);
2929 /* -fkeep-gc-roots-live means that we have to keep a real pointer
2930 live, but the ivopts code may replace a real pointer with one
2931 pointing before or after the memory block that is then adjusted
2932 into the memory block during the loop. FIXME: It would likely be
2933 better to actually force the pointer live and still use ivopts;
2934 for example, it would be enough to write the pointer into memory
2935 and keep it there until after the loop. */
2936 if (flag_keep_gc_roots_live
&& POINTER_TYPE_P (TREE_TYPE (base
)))
2939 /* For non-original variables, make sure their values are computed in a type
2940 that does not invoke undefined behavior on overflows (since in general,
2941 we cannot prove that these induction variables are non-wrapping). */
2942 if (pos
!= IP_ORIGINAL
)
2944 orig_type
= TREE_TYPE (base
);
2945 type
= generic_type_for (orig_type
);
2946 if (type
!= orig_type
)
2948 base
= fold_convert (type
, base
);
2949 step
= fold_convert (type
, step
);
2953 for (i
= 0; i
< data
->vcands
.length (); i
++)
2955 cand
= data
->vcands
[i
];
2957 if (cand
->pos
!= pos
)
2960 if (cand
->incremented_at
!= incremented_at
2961 || ((pos
== IP_AFTER_USE
|| pos
== IP_BEFORE_USE
)
2962 && cand
->ainc_use
!= use
))
2965 if (operand_equal_p (base
, cand
->iv
->base
, 0)
2966 && operand_equal_p (step
, cand
->iv
->step
, 0)
2967 && (TYPE_PRECISION (TREE_TYPE (base
))
2968 == TYPE_PRECISION (TREE_TYPE (cand
->iv
->base
))))
2972 if (i
== data
->vcands
.length ())
2974 cand
= XCNEW (struct iv_cand
);
2976 cand
->iv
= alloc_iv (data
, base
, step
);
2978 if (pos
!= IP_ORIGINAL
)
2980 cand
->var_before
= create_tmp_var_raw (TREE_TYPE (base
), "ivtmp");
2981 cand
->var_after
= cand
->var_before
;
2983 cand
->important
= important
;
2984 cand
->incremented_at
= incremented_at
;
2985 data
->vcands
.safe_push (cand
);
2987 if (TREE_CODE (step
) != INTEGER_CST
)
2989 fd_ivopts_data
= data
;
2990 walk_tree (&step
, find_depends
, &cand
->depends_on
, NULL
);
2993 if (pos
== IP_AFTER_USE
|| pos
== IP_BEFORE_USE
)
2994 cand
->ainc_use
= use
;
2996 cand
->ainc_use
= NULL
;
2998 cand
->orig_iv
= orig_iv
;
2999 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3000 dump_cand (dump_file
, cand
);
3003 cand
->important
|= important
;
3005 /* Relate candidate to the group for which it is added. */
3007 bitmap_set_bit (data
->vgroups
[use
->group_id
]->related_cands
, i
);
3012 /* Returns true if incrementing the induction variable at the end of the LOOP
3015 The purpose is to avoid splitting latch edge with a biv increment, thus
3016 creating a jump, possibly confusing other optimization passes and leaving
3017 less freedom to scheduler. So we allow IP_END_POS only if IP_NORMAL_POS
3018 is not available (so we do not have a better alternative), or if the latch
3019 edge is already nonempty. */
3022 allow_ip_end_pos_p (struct loop
*loop
)
3024 if (!ip_normal_pos (loop
))
3027 if (!empty_block_p (ip_end_pos (loop
)))
3033 /* If possible, adds autoincrement candidates BASE + STEP * i based on use USE.
3034 Important field is set to IMPORTANT. */
3037 add_autoinc_candidates (struct ivopts_data
*data
, tree base
, tree step
,
3038 bool important
, struct iv_use
*use
)
3040 basic_block use_bb
= gimple_bb (use
->stmt
);
3041 machine_mode mem_mode
;
3042 unsigned HOST_WIDE_INT cstepi
;
3044 /* If we insert the increment in any position other than the standard
3045 ones, we must ensure that it is incremented once per iteration.
3046 It must not be in an inner nested loop, or one side of an if
3048 if (use_bb
->loop_father
!= data
->current_loop
3049 || !dominated_by_p (CDI_DOMINATORS
, data
->current_loop
->latch
, use_bb
)
3050 || stmt_could_throw_p (use
->stmt
)
3051 || !cst_and_fits_in_hwi (step
))
3054 cstepi
= int_cst_value (step
);
3056 mem_mode
= TYPE_MODE (TREE_TYPE (*use
->op_p
));
3057 if (((USE_LOAD_PRE_INCREMENT (mem_mode
)
3058 || USE_STORE_PRE_INCREMENT (mem_mode
))
3059 && GET_MODE_SIZE (mem_mode
) == cstepi
)
3060 || ((USE_LOAD_PRE_DECREMENT (mem_mode
)
3061 || USE_STORE_PRE_DECREMENT (mem_mode
))
3062 && GET_MODE_SIZE (mem_mode
) == -cstepi
))
3064 enum tree_code code
= MINUS_EXPR
;
3066 tree new_step
= step
;
3068 if (POINTER_TYPE_P (TREE_TYPE (base
)))
3070 new_step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
3071 code
= POINTER_PLUS_EXPR
;
3074 new_step
= fold_convert (TREE_TYPE (base
), new_step
);
3075 new_base
= fold_build2 (code
, TREE_TYPE (base
), base
, new_step
);
3076 add_candidate_1 (data
, new_base
, step
, important
, IP_BEFORE_USE
, use
,
3079 if (((USE_LOAD_POST_INCREMENT (mem_mode
)
3080 || USE_STORE_POST_INCREMENT (mem_mode
))
3081 && GET_MODE_SIZE (mem_mode
) == cstepi
)
3082 || ((USE_LOAD_POST_DECREMENT (mem_mode
)
3083 || USE_STORE_POST_DECREMENT (mem_mode
))
3084 && GET_MODE_SIZE (mem_mode
) == -cstepi
))
3086 add_candidate_1 (data
, base
, step
, important
, IP_AFTER_USE
, use
,
3091 /* Adds a candidate BASE + STEP * i. Important field is set to IMPORTANT and
3092 position to POS. If USE is not NULL, the candidate is set as related to
3093 it. The candidate computation is scheduled before exit condition and at
3097 add_candidate (struct ivopts_data
*data
,
3098 tree base
, tree step
, bool important
, struct iv_use
*use
,
3099 struct iv
*orig_iv
= NULL
)
3101 if (ip_normal_pos (data
->current_loop
))
3102 add_candidate_1 (data
, base
, step
, important
,
3103 IP_NORMAL
, use
, NULL
, orig_iv
);
3104 if (ip_end_pos (data
->current_loop
)
3105 && allow_ip_end_pos_p (data
->current_loop
))
3106 add_candidate_1 (data
, base
, step
, important
, IP_END
, use
, NULL
, orig_iv
);
3109 /* Adds standard iv candidates. */
3112 add_standard_iv_candidates (struct ivopts_data
*data
)
3114 add_candidate (data
, integer_zero_node
, integer_one_node
, true, NULL
);
3116 /* The same for a double-integer type if it is still fast enough. */
3118 (long_integer_type_node
) > TYPE_PRECISION (integer_type_node
)
3119 && TYPE_PRECISION (long_integer_type_node
) <= BITS_PER_WORD
)
3120 add_candidate (data
, build_int_cst (long_integer_type_node
, 0),
3121 build_int_cst (long_integer_type_node
, 1), true, NULL
);
3123 /* The same for a double-integer type if it is still fast enough. */
3125 (long_long_integer_type_node
) > TYPE_PRECISION (long_integer_type_node
)
3126 && TYPE_PRECISION (long_long_integer_type_node
) <= BITS_PER_WORD
)
3127 add_candidate (data
, build_int_cst (long_long_integer_type_node
, 0),
3128 build_int_cst (long_long_integer_type_node
, 1), true, NULL
);
3132 /* Adds candidates bases on the old induction variable IV. */
3135 add_iv_candidate_for_biv (struct ivopts_data
*data
, struct iv
*iv
)
3139 struct iv_cand
*cand
;
3141 /* Check if this biv is used in address type use. */
3142 if (iv
->no_overflow
&& iv
->have_address_use
3143 && INTEGRAL_TYPE_P (TREE_TYPE (iv
->base
))
3144 && TYPE_PRECISION (TREE_TYPE (iv
->base
)) < TYPE_PRECISION (sizetype
))
3146 tree base
= fold_convert (sizetype
, iv
->base
);
3147 tree step
= fold_convert (sizetype
, iv
->step
);
3149 /* Add iv cand of same precision as index part in TARGET_MEM_REF. */
3150 add_candidate (data
, base
, step
, true, NULL
, iv
);
3151 /* Add iv cand of the original type only if it has nonlinear use. */
3153 add_candidate (data
, iv
->base
, iv
->step
, true, NULL
);
3156 add_candidate (data
, iv
->base
, iv
->step
, true, NULL
);
3158 /* The same, but with initial value zero. */
3159 if (POINTER_TYPE_P (TREE_TYPE (iv
->base
)))
3160 add_candidate (data
, size_int (0), iv
->step
, true, NULL
);
3162 add_candidate (data
, build_int_cst (TREE_TYPE (iv
->base
), 0),
3163 iv
->step
, true, NULL
);
3165 phi
= SSA_NAME_DEF_STMT (iv
->ssa_name
);
3166 if (gimple_code (phi
) == GIMPLE_PHI
)
3168 /* Additionally record the possibility of leaving the original iv
3170 def
= PHI_ARG_DEF_FROM_EDGE (phi
, loop_latch_edge (data
->current_loop
));
3171 /* Don't add candidate if it's from another PHI node because
3172 it's an affine iv appearing in the form of PEELED_CHREC. */
3173 phi
= SSA_NAME_DEF_STMT (def
);
3174 if (gimple_code (phi
) != GIMPLE_PHI
)
3176 cand
= add_candidate_1 (data
,
3177 iv
->base
, iv
->step
, true, IP_ORIGINAL
, NULL
,
3178 SSA_NAME_DEF_STMT (def
));
3181 cand
->var_before
= iv
->ssa_name
;
3182 cand
->var_after
= def
;
3186 gcc_assert (gimple_bb (phi
) == data
->current_loop
->header
);
3190 /* Adds candidates based on the old induction variables. */
3193 add_iv_candidate_for_bivs (struct ivopts_data
*data
)
3199 EXECUTE_IF_SET_IN_BITMAP (data
->relevant
, 0, i
, bi
)
3201 iv
= ver_info (data
, i
)->iv
;
3202 if (iv
&& iv
->biv_p
&& !integer_zerop (iv
->step
))
3203 add_iv_candidate_for_biv (data
, iv
);
3207 /* Record common candidate {BASE, STEP} derived from USE in hashtable. */
3210 record_common_cand (struct ivopts_data
*data
, tree base
,
3211 tree step
, struct iv_use
*use
)
3213 struct iv_common_cand ent
;
3214 struct iv_common_cand
**slot
;
3218 ent
.hash
= iterative_hash_expr (base
, 0);
3219 ent
.hash
= iterative_hash_expr (step
, ent
.hash
);
3221 slot
= data
->iv_common_cand_tab
->find_slot (&ent
, INSERT
);
3224 *slot
= new iv_common_cand ();
3225 (*slot
)->base
= base
;
3226 (*slot
)->step
= step
;
3227 (*slot
)->uses
.create (8);
3228 (*slot
)->hash
= ent
.hash
;
3229 data
->iv_common_cands
.safe_push ((*slot
));
3232 gcc_assert (use
!= NULL
);
3233 (*slot
)->uses
.safe_push (use
);
3237 /* Comparison function used to sort common candidates. */
3240 common_cand_cmp (const void *p1
, const void *p2
)
3243 const struct iv_common_cand
*const *const ccand1
3244 = (const struct iv_common_cand
*const *)p1
;
3245 const struct iv_common_cand
*const *const ccand2
3246 = (const struct iv_common_cand
*const *)p2
;
3248 n1
= (*ccand1
)->uses
.length ();
3249 n2
= (*ccand2
)->uses
.length ();
3253 /* Adds IV candidates based on common candidated recorded. */
3256 add_iv_candidate_derived_from_uses (struct ivopts_data
*data
)
3259 struct iv_cand
*cand_1
, *cand_2
;
3261 data
->iv_common_cands
.qsort (common_cand_cmp
);
3262 for (i
= 0; i
< data
->iv_common_cands
.length (); i
++)
3264 struct iv_common_cand
*ptr
= data
->iv_common_cands
[i
];
3266 /* Only add IV candidate if it's derived from multiple uses. */
3267 if (ptr
->uses
.length () <= 1)
3272 if (ip_normal_pos (data
->current_loop
))
3273 cand_1
= add_candidate_1 (data
, ptr
->base
, ptr
->step
,
3274 false, IP_NORMAL
, NULL
, NULL
);
3276 if (ip_end_pos (data
->current_loop
)
3277 && allow_ip_end_pos_p (data
->current_loop
))
3278 cand_2
= add_candidate_1 (data
, ptr
->base
, ptr
->step
,
3279 false, IP_END
, NULL
, NULL
);
3281 /* Bind deriving uses and the new candidates. */
3282 for (j
= 0; j
< ptr
->uses
.length (); j
++)
3284 struct iv_group
*group
= data
->vgroups
[ptr
->uses
[j
]->group_id
];
3286 bitmap_set_bit (group
->related_cands
, cand_1
->id
);
3288 bitmap_set_bit (group
->related_cands
, cand_2
->id
);
3292 /* Release data since it is useless from this point. */
3293 data
->iv_common_cand_tab
->empty ();
3294 data
->iv_common_cands
.truncate (0);
3297 /* Adds candidates based on the value of USE's iv. */
3300 add_iv_candidate_for_use (struct ivopts_data
*data
, struct iv_use
*use
)
3302 unsigned HOST_WIDE_INT offset
;
3305 struct iv
*iv
= use
->iv
;
3307 add_candidate (data
, iv
->base
, iv
->step
, false, use
);
3309 /* Record common candidate for use in case it can be shared by others. */
3310 record_common_cand (data
, iv
->base
, iv
->step
, use
);
3312 /* Record common candidate with initial value zero. */
3313 basetype
= TREE_TYPE (iv
->base
);
3314 if (POINTER_TYPE_P (basetype
))
3315 basetype
= sizetype
;
3316 record_common_cand (data
, build_int_cst (basetype
, 0), iv
->step
, use
);
3318 /* Record common candidate with constant offset stripped in base.
3319 Like the use itself, we also add candidate directly for it. */
3320 base
= strip_offset (iv
->base
, &offset
);
3321 if (offset
|| base
!= iv
->base
)
3323 record_common_cand (data
, base
, iv
->step
, use
);
3324 add_candidate (data
, base
, iv
->step
, false, use
);
3327 /* Record common candidate with base_object removed in base. */
3328 if (iv
->base_object
!= NULL
)
3332 tree step
, base_object
= iv
->base_object
;
3338 STRIP_NOPS (base_object
);
3339 tree_to_aff_combination (base
, TREE_TYPE (base
), &aff_base
);
3340 for (i
= 0; i
< aff_base
.n
; i
++)
3342 if (aff_base
.elts
[i
].coef
!= 1)
3345 if (operand_equal_p (aff_base
.elts
[i
].val
, base_object
, 0))
3350 aff_combination_remove_elt (&aff_base
, i
);
3351 base
= aff_combination_to_tree (&aff_base
);
3352 basetype
= TREE_TYPE (base
);
3353 if (POINTER_TYPE_P (basetype
))
3354 basetype
= sizetype
;
3356 step
= fold_convert (basetype
, step
);
3357 record_common_cand (data
, base
, step
, use
);
3358 /* Also record common candidate with offset stripped. */
3359 base
= strip_offset (base
, &offset
);
3361 record_common_cand (data
, base
, step
, use
);
3365 /* At last, add auto-incremental candidates. Make such variables
3366 important since other iv uses with same base object may be based
3368 if (use
!= NULL
&& use
->type
== USE_ADDRESS
)
3369 add_autoinc_candidates (data
, iv
->base
, iv
->step
, true, use
);
3372 /* Adds candidates based on the uses. */
3375 add_iv_candidate_for_groups (struct ivopts_data
*data
)
3379 /* Only add candidate for the first use in group. */
3380 for (i
= 0; i
< data
->vgroups
.length (); i
++)
3382 struct iv_group
*group
= data
->vgroups
[i
];
3384 gcc_assert (group
->vuses
[0] != NULL
);
3385 add_iv_candidate_for_use (data
, group
->vuses
[0]);
3387 add_iv_candidate_derived_from_uses (data
);
3390 /* Record important candidates and add them to related_cands bitmaps. */
3393 record_important_candidates (struct ivopts_data
*data
)
3396 struct iv_group
*group
;
3398 for (i
= 0; i
< data
->vcands
.length (); i
++)
3400 struct iv_cand
*cand
= data
->vcands
[i
];
3402 if (cand
->important
)
3403 bitmap_set_bit (data
->important_candidates
, i
);
3406 data
->consider_all_candidates
= (data
->vcands
.length ()
3407 <= CONSIDER_ALL_CANDIDATES_BOUND
);
3409 /* Add important candidates to groups' related_cands bitmaps. */
3410 for (i
= 0; i
< data
->vgroups
.length (); i
++)
3412 group
= data
->vgroups
[i
];
3413 bitmap_ior_into (group
->related_cands
, data
->important_candidates
);
3417 /* Allocates the data structure mapping the (use, candidate) pairs to costs.
3418 If consider_all_candidates is true, we use a two-dimensional array, otherwise
3419 we allocate a simple list to every use. */
3422 alloc_use_cost_map (struct ivopts_data
*data
)
3424 unsigned i
, size
, s
;
3426 for (i
= 0; i
< data
->vgroups
.length (); i
++)
3428 struct iv_group
*group
= data
->vgroups
[i
];
3430 if (data
->consider_all_candidates
)
3431 size
= data
->vcands
.length ();
3434 s
= bitmap_count_bits (group
->related_cands
);
3436 /* Round up to the power of two, so that moduling by it is fast. */
3437 size
= s
? (1 << ceil_log2 (s
)) : 1;
3440 group
->n_map_members
= size
;
3441 group
->cost_map
= XCNEWVEC (struct cost_pair
, size
);
3445 /* Sets cost of (GROUP, CAND) pair to COST and record that it depends
3446 on invariants DEPENDS_ON and that the value used in expressing it
3447 is VALUE, and in case of iv elimination the comparison operator is COMP. */
3450 set_group_iv_cost (struct ivopts_data
*data
,
3451 struct iv_group
*group
, struct iv_cand
*cand
,
3452 comp_cost cost
, bitmap depends_on
, tree value
,
3453 enum tree_code comp
, iv_inv_expr_ent
*inv_expr
)
3457 if (cost
.infinite_cost_p ())
3459 BITMAP_FREE (depends_on
);
3463 if (data
->consider_all_candidates
)
3465 group
->cost_map
[cand
->id
].cand
= cand
;
3466 group
->cost_map
[cand
->id
].cost
= cost
;
3467 group
->cost_map
[cand
->id
].depends_on
= depends_on
;
3468 group
->cost_map
[cand
->id
].value
= value
;
3469 group
->cost_map
[cand
->id
].comp
= comp
;
3470 group
->cost_map
[cand
->id
].inv_expr
= inv_expr
;
3474 /* n_map_members is a power of two, so this computes modulo. */
3475 s
= cand
->id
& (group
->n_map_members
- 1);
3476 for (i
= s
; i
< group
->n_map_members
; i
++)
3477 if (!group
->cost_map
[i
].cand
)
3479 for (i
= 0; i
< s
; i
++)
3480 if (!group
->cost_map
[i
].cand
)
3486 group
->cost_map
[i
].cand
= cand
;
3487 group
->cost_map
[i
].cost
= cost
;
3488 group
->cost_map
[i
].depends_on
= depends_on
;
3489 group
->cost_map
[i
].value
= value
;
3490 group
->cost_map
[i
].comp
= comp
;
3491 group
->cost_map
[i
].inv_expr
= inv_expr
;
3494 /* Gets cost of (GROUP, CAND) pair. */
3496 static struct cost_pair
*
3497 get_group_iv_cost (struct ivopts_data
*data
, struct iv_group
*group
,
3498 struct iv_cand
*cand
)
3501 struct cost_pair
*ret
;
3506 if (data
->consider_all_candidates
)
3508 ret
= group
->cost_map
+ cand
->id
;
3515 /* n_map_members is a power of two, so this computes modulo. */
3516 s
= cand
->id
& (group
->n_map_members
- 1);
3517 for (i
= s
; i
< group
->n_map_members
; i
++)
3518 if (group
->cost_map
[i
].cand
== cand
)
3519 return group
->cost_map
+ i
;
3520 else if (group
->cost_map
[i
].cand
== NULL
)
3522 for (i
= 0; i
< s
; i
++)
3523 if (group
->cost_map
[i
].cand
== cand
)
3524 return group
->cost_map
+ i
;
3525 else if (group
->cost_map
[i
].cand
== NULL
)
3531 /* Produce DECL_RTL for object obj so it looks like it is stored in memory. */
3533 produce_memory_decl_rtl (tree obj
, int *regno
)
3535 addr_space_t as
= TYPE_ADDR_SPACE (TREE_TYPE (obj
));
3536 machine_mode address_mode
= targetm
.addr_space
.address_mode (as
);
3540 if (TREE_STATIC (obj
) || DECL_EXTERNAL (obj
))
3542 const char *name
= IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (obj
));
3543 x
= gen_rtx_SYMBOL_REF (address_mode
, name
);
3544 SET_SYMBOL_REF_DECL (x
, obj
);
3545 x
= gen_rtx_MEM (DECL_MODE (obj
), x
);
3546 set_mem_addr_space (x
, as
);
3547 targetm
.encode_section_info (obj
, x
, true);
3551 x
= gen_raw_REG (address_mode
, (*regno
)++);
3552 x
= gen_rtx_MEM (DECL_MODE (obj
), x
);
3553 set_mem_addr_space (x
, as
);
3559 /* Prepares decl_rtl for variables referred in *EXPR_P. Callback for
3560 walk_tree. DATA contains the actual fake register number. */
3563 prepare_decl_rtl (tree
*expr_p
, int *ws
, void *data
)
3565 tree obj
= NULL_TREE
;
3567 int *regno
= (int *) data
;
3569 switch (TREE_CODE (*expr_p
))
3572 for (expr_p
= &TREE_OPERAND (*expr_p
, 0);
3573 handled_component_p (*expr_p
);
3574 expr_p
= &TREE_OPERAND (*expr_p
, 0))
3577 if (DECL_P (obj
) && HAS_RTL_P (obj
) && !DECL_RTL_SET_P (obj
))
3578 x
= produce_memory_decl_rtl (obj
, regno
);
3583 obj
= SSA_NAME_VAR (*expr_p
);
3584 /* Defer handling of anonymous SSA_NAMEs to the expander. */
3587 if (!DECL_RTL_SET_P (obj
))
3588 x
= gen_raw_REG (DECL_MODE (obj
), (*regno
)++);
3597 if (DECL_RTL_SET_P (obj
))
3600 if (DECL_MODE (obj
) == BLKmode
)
3601 x
= produce_memory_decl_rtl (obj
, regno
);
3603 x
= gen_raw_REG (DECL_MODE (obj
), (*regno
)++);
3613 decl_rtl_to_reset
.safe_push (obj
);
3614 SET_DECL_RTL (obj
, x
);
3620 /* Determines cost of the computation of EXPR. */
3623 computation_cost (tree expr
, bool speed
)
3627 tree type
= TREE_TYPE (expr
);
3629 /* Avoid using hard regs in ways which may be unsupported. */
3630 int regno
= LAST_VIRTUAL_REGISTER
+ 1;
3631 struct cgraph_node
*node
= cgraph_node::get (current_function_decl
);
3632 enum node_frequency real_frequency
= node
->frequency
;
3634 node
->frequency
= NODE_FREQUENCY_NORMAL
;
3635 crtl
->maybe_hot_insn_p
= speed
;
3636 walk_tree (&expr
, prepare_decl_rtl
, ®no
, NULL
);
3638 rslt
= expand_expr (expr
, NULL_RTX
, TYPE_MODE (type
), EXPAND_NORMAL
);
3641 default_rtl_profile ();
3642 node
->frequency
= real_frequency
;
3644 cost
= seq_cost (seq
, speed
);
3646 cost
+= address_cost (XEXP (rslt
, 0), TYPE_MODE (type
),
3647 TYPE_ADDR_SPACE (type
), speed
);
3648 else if (!REG_P (rslt
))
3649 cost
+= set_src_cost (rslt
, TYPE_MODE (type
), speed
);
3654 /* Returns variable containing the value of candidate CAND at statement AT. */
3657 var_at_stmt (struct loop
*loop
, struct iv_cand
*cand
, gimple
*stmt
)
3659 if (stmt_after_increment (loop
, cand
, stmt
))
3660 return cand
->var_after
;
3662 return cand
->var_before
;
3665 /* If A is (TYPE) BA and B is (TYPE) BB, and the types of BA and BB have the
3666 same precision that is at least as wide as the precision of TYPE, stores
3667 BA to A and BB to B, and returns the type of BA. Otherwise, returns the
3671 determine_common_wider_type (tree
*a
, tree
*b
)
3673 tree wider_type
= NULL
;
3675 tree atype
= TREE_TYPE (*a
);
3677 if (CONVERT_EXPR_P (*a
))
3679 suba
= TREE_OPERAND (*a
, 0);
3680 wider_type
= TREE_TYPE (suba
);
3681 if (TYPE_PRECISION (wider_type
) < TYPE_PRECISION (atype
))
3687 if (CONVERT_EXPR_P (*b
))
3689 subb
= TREE_OPERAND (*b
, 0);
3690 if (TYPE_PRECISION (wider_type
) != TYPE_PRECISION (TREE_TYPE (subb
)))
3701 /* Determines the expression by that USE is expressed from induction variable
3702 CAND at statement AT in LOOP. The expression is stored in a decomposed
3703 form into AFF. Returns false if USE cannot be expressed using CAND. */
3706 get_computation_aff (struct loop
*loop
,
3707 struct iv_use
*use
, struct iv_cand
*cand
, gimple
*at
,
3708 struct aff_tree
*aff
)
3710 tree ubase
= use
->iv
->base
;
3711 tree ustep
= use
->iv
->step
;
3712 tree cbase
= cand
->iv
->base
;
3713 tree cstep
= cand
->iv
->step
, cstep_common
;
3714 tree utype
= TREE_TYPE (ubase
), ctype
= TREE_TYPE (cbase
);
3715 tree common_type
, var
;
3717 aff_tree cbase_aff
, var_aff
;
3720 if (TYPE_PRECISION (utype
) > TYPE_PRECISION (ctype
))
3722 /* We do not have a precision to express the values of use. */
3726 var
= var_at_stmt (loop
, cand
, at
);
3727 uutype
= unsigned_type_for (utype
);
3729 /* If the conversion is not noop, perform it. */
3730 if (TYPE_PRECISION (utype
) < TYPE_PRECISION (ctype
))
3732 if (cand
->orig_iv
!= NULL
&& CONVERT_EXPR_P (cbase
)
3733 && (CONVERT_EXPR_P (cstep
) || TREE_CODE (cstep
) == INTEGER_CST
))
3735 tree inner_base
, inner_step
, inner_type
;
3736 inner_base
= TREE_OPERAND (cbase
, 0);
3737 if (CONVERT_EXPR_P (cstep
))
3738 inner_step
= TREE_OPERAND (cstep
, 0);
3742 inner_type
= TREE_TYPE (inner_base
);
3743 /* If candidate is added from a biv whose type is smaller than
3744 ctype, we know both candidate and the biv won't overflow.
3745 In this case, it's safe to skip the convertion in candidate.
3746 As an example, (unsigned short)((unsigned long)A) equals to
3747 (unsigned short)A, if A has a type no larger than short. */
3748 if (TYPE_PRECISION (inner_type
) <= TYPE_PRECISION (uutype
))
3754 cstep
= fold_convert (uutype
, cstep
);
3755 cbase
= fold_convert (uutype
, cbase
);
3756 var
= fold_convert (uutype
, var
);
3759 /* Ratio is 1 when computing the value of biv cand by itself.
3760 We can't rely on constant_multiple_of in this case because the
3761 use is created after the original biv is selected. The call
3762 could fail because of inconsistent fold behavior. See PR68021
3763 for more information. */
3764 if (cand
->pos
== IP_ORIGINAL
&& cand
->incremented_at
== use
->stmt
)
3766 gcc_assert (is_gimple_assign (use
->stmt
));
3767 gcc_assert (use
->iv
->ssa_name
== cand
->var_after
);
3768 gcc_assert (gimple_assign_lhs (use
->stmt
) == cand
->var_after
);
3771 else if (!constant_multiple_of (ustep
, cstep
, &rat
))
3774 /* In case both UBASE and CBASE are shortened to UUTYPE from some common
3775 type, we achieve better folding by computing their difference in this
3776 wider type, and cast the result to UUTYPE. We do not need to worry about
3777 overflows, as all the arithmetics will in the end be performed in UUTYPE
3779 common_type
= determine_common_wider_type (&ubase
, &cbase
);
3781 /* use = ubase - ratio * cbase + ratio * var. */
3782 tree_to_aff_combination (ubase
, common_type
, aff
);
3783 tree_to_aff_combination (cbase
, common_type
, &cbase_aff
);
3784 tree_to_aff_combination (var
, uutype
, &var_aff
);
3786 /* We need to shift the value if we are after the increment. */
3787 if (stmt_after_increment (loop
, cand
, at
))
3791 if (common_type
!= uutype
)
3792 cstep_common
= fold_convert (common_type
, cstep
);
3794 cstep_common
= cstep
;
3796 tree_to_aff_combination (cstep_common
, common_type
, &cstep_aff
);
3797 aff_combination_add (&cbase_aff
, &cstep_aff
);
3800 aff_combination_scale (&cbase_aff
, -rat
);
3801 aff_combination_add (aff
, &cbase_aff
);
3802 if (common_type
!= uutype
)
3803 aff_combination_convert (aff
, uutype
);
3805 aff_combination_scale (&var_aff
, rat
);
3806 aff_combination_add (aff
, &var_aff
);
3811 /* Return the type of USE. */
3814 get_use_type (struct iv_use
*use
)
3816 tree base_type
= TREE_TYPE (use
->iv
->base
);
3819 if (use
->type
== USE_ADDRESS
)
3821 /* The base_type may be a void pointer. Create a pointer type based on
3822 the mem_ref instead. */
3823 type
= build_pointer_type (TREE_TYPE (*use
->op_p
));
3824 gcc_assert (TYPE_ADDR_SPACE (TREE_TYPE (type
))
3825 == TYPE_ADDR_SPACE (TREE_TYPE (base_type
)));
3833 /* Determines the expression by that USE is expressed from induction variable
3834 CAND at statement AT in LOOP. The computation is unshared. */
3837 get_computation_at (struct loop
*loop
,
3838 struct iv_use
*use
, struct iv_cand
*cand
, gimple
*at
)
3841 tree type
= get_use_type (use
);
3843 if (!get_computation_aff (loop
, use
, cand
, at
, &aff
))
3845 unshare_aff_combination (&aff
);
3846 return fold_convert (type
, aff_combination_to_tree (&aff
));
3849 /* Determines the expression by that USE is expressed from induction variable
3850 CAND in LOOP. The computation is unshared. */
3853 get_computation (struct loop
*loop
, struct iv_use
*use
, struct iv_cand
*cand
)
3855 return get_computation_at (loop
, use
, cand
, use
->stmt
);
3858 /* Adjust the cost COST for being in loop setup rather than loop body.
3859 If we're optimizing for space, the loop setup overhead is constant;
3860 if we're optimizing for speed, amortize it over the per-iteration cost. */
3862 adjust_setup_cost (struct ivopts_data
*data
, unsigned cost
)
3866 else if (optimize_loop_for_speed_p (data
->current_loop
))
3867 return cost
/ avg_loop_niter (data
->current_loop
);
3872 /* Returns true if multiplying by RATIO is allowed in an address. Test the
3873 validity for a memory reference accessing memory of mode MODE in
3874 address space AS. */
3878 multiplier_allowed_in_address_p (HOST_WIDE_INT ratio
, machine_mode mode
,
3881 #define MAX_RATIO 128
3882 unsigned int data_index
= (int) as
* MAX_MACHINE_MODE
+ (int) mode
;
3883 static vec
<sbitmap
> valid_mult_list
;
3886 if (data_index
>= valid_mult_list
.length ())
3887 valid_mult_list
.safe_grow_cleared (data_index
+ 1);
3889 valid_mult
= valid_mult_list
[data_index
];
3892 machine_mode address_mode
= targetm
.addr_space
.address_mode (as
);
3893 rtx reg1
= gen_raw_REG (address_mode
, LAST_VIRTUAL_REGISTER
+ 1);
3894 rtx reg2
= gen_raw_REG (address_mode
, LAST_VIRTUAL_REGISTER
+ 2);
3898 valid_mult
= sbitmap_alloc (2 * MAX_RATIO
+ 1);
3899 bitmap_clear (valid_mult
);
3900 scaled
= gen_rtx_fmt_ee (MULT
, address_mode
, reg1
, NULL_RTX
);
3901 addr
= gen_rtx_fmt_ee (PLUS
, address_mode
, scaled
, reg2
);
3902 for (i
= -MAX_RATIO
; i
<= MAX_RATIO
; i
++)
3904 XEXP (scaled
, 1) = gen_int_mode (i
, address_mode
);
3905 if (memory_address_addr_space_p (mode
, addr
, as
)
3906 || memory_address_addr_space_p (mode
, scaled
, as
))
3907 bitmap_set_bit (valid_mult
, i
+ MAX_RATIO
);
3910 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3912 fprintf (dump_file
, " allowed multipliers:");
3913 for (i
= -MAX_RATIO
; i
<= MAX_RATIO
; i
++)
3914 if (bitmap_bit_p (valid_mult
, i
+ MAX_RATIO
))
3915 fprintf (dump_file
, " %d", (int) i
);
3916 fprintf (dump_file
, "\n");
3917 fprintf (dump_file
, "\n");
3920 valid_mult_list
[data_index
] = valid_mult
;
3923 if (ratio
> MAX_RATIO
|| ratio
< -MAX_RATIO
)
3926 return bitmap_bit_p (valid_mult
, ratio
+ MAX_RATIO
);
3929 /* Returns cost of address in shape symbol + var + OFFSET + RATIO * index.
3930 If SYMBOL_PRESENT is false, symbol is omitted. If VAR_PRESENT is false,
3931 variable is omitted. Compute the cost for a memory reference that accesses
3932 a memory location of mode MEM_MODE in address space AS.
3934 MAY_AUTOINC is set to true if the autoincrement (increasing index by
3935 size of MEM_MODE / RATIO) is available. To make this determination, we
3936 look at the size of the increment to be made, which is given in CSTEP.
3937 CSTEP may be zero if the step is unknown.
3938 STMT_AFTER_INC is true iff the statement we're looking at is after the
3939 increment of the original biv.
3941 TODO -- there must be some better way. This all is quite crude. */
3945 AINC_PRE_INC
, /* Pre increment. */
3946 AINC_PRE_DEC
, /* Pre decrement. */
3947 AINC_POST_INC
, /* Post increment. */
3948 AINC_POST_DEC
, /* Post decrement. */
3949 AINC_NONE
/* Also the number of auto increment types. */
3952 struct address_cost_data
3954 HOST_WIDE_INT min_offset
, max_offset
;
3955 unsigned costs
[2][2][2][2];
3956 unsigned ainc_costs
[AINC_NONE
];
3961 get_address_cost (bool symbol_present
, bool var_present
,
3962 unsigned HOST_WIDE_INT offset
, HOST_WIDE_INT ratio
,
3963 HOST_WIDE_INT cstep
, machine_mode mem_mode
,
3964 addr_space_t as
, bool speed
,
3965 bool stmt_after_inc
, bool *may_autoinc
)
3967 machine_mode address_mode
= targetm
.addr_space
.address_mode (as
);
3968 static vec
<address_cost_data
*> address_cost_data_list
;
3969 unsigned int data_index
= (int) as
* MAX_MACHINE_MODE
+ (int) mem_mode
;
3970 address_cost_data
*data
;
3971 static bool has_preinc
[MAX_MACHINE_MODE
], has_postinc
[MAX_MACHINE_MODE
];
3972 static bool has_predec
[MAX_MACHINE_MODE
], has_postdec
[MAX_MACHINE_MODE
];
3973 unsigned cost
, acost
, complexity
;
3974 enum ainc_type autoinc_type
;
3975 bool offset_p
, ratio_p
, autoinc
;
3976 HOST_WIDE_INT s_offset
, autoinc_offset
, msize
;
3977 unsigned HOST_WIDE_INT mask
;
3980 if (data_index
>= address_cost_data_list
.length ())
3981 address_cost_data_list
.safe_grow_cleared (data_index
+ 1);
3983 data
= address_cost_data_list
[data_index
];
3987 HOST_WIDE_INT rat
, off
= 0;
3988 int old_cse_not_expected
, width
;
3989 unsigned sym_p
, var_p
, off_p
, rat_p
, add_c
;
3994 data
= (address_cost_data
*) xcalloc (1, sizeof (*data
));
3996 reg1
= gen_raw_REG (address_mode
, LAST_VIRTUAL_REGISTER
+ 1);
3998 width
= GET_MODE_BITSIZE (address_mode
) - 1;
3999 if (width
> (HOST_BITS_PER_WIDE_INT
- 1))
4000 width
= HOST_BITS_PER_WIDE_INT
- 1;
4001 addr
= gen_rtx_fmt_ee (PLUS
, address_mode
, reg1
, NULL_RTX
);
4003 for (i
= width
; i
>= 0; i
--)
4005 off
= -(HOST_WIDE_INT_1U
<< i
);
4006 XEXP (addr
, 1) = gen_int_mode (off
, address_mode
);
4007 if (memory_address_addr_space_p (mem_mode
, addr
, as
))
4010 data
->min_offset
= (i
== -1? 0 : off
);
4012 for (i
= width
; i
>= 0; i
--)
4014 off
= (HOST_WIDE_INT_1U
<< i
) - 1;
4015 XEXP (addr
, 1) = gen_int_mode (off
, address_mode
);
4016 if (memory_address_addr_space_p (mem_mode
, addr
, as
))
4018 /* For some strict-alignment targets, the offset must be naturally
4019 aligned. Try an aligned offset if mem_mode is not QImode. */
4020 off
= mem_mode
!= QImode
4021 ? (HOST_WIDE_INT_1U
<< i
)
4022 - GET_MODE_SIZE (mem_mode
)
4026 XEXP (addr
, 1) = gen_int_mode (off
, address_mode
);
4027 if (memory_address_addr_space_p (mem_mode
, addr
, as
))
4033 data
->max_offset
= off
;
4035 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4037 fprintf (dump_file
, "get_address_cost:\n");
4038 fprintf (dump_file
, " min offset %s " HOST_WIDE_INT_PRINT_DEC
"\n",
4039 GET_MODE_NAME (mem_mode
),
4041 fprintf (dump_file
, " max offset %s " HOST_WIDE_INT_PRINT_DEC
"\n",
4042 GET_MODE_NAME (mem_mode
),
4047 for (i
= 2; i
<= MAX_RATIO
; i
++)
4048 if (multiplier_allowed_in_address_p (i
, mem_mode
, as
))
4054 /* Compute the cost of various addressing modes. */
4056 reg0
= gen_raw_REG (address_mode
, LAST_VIRTUAL_REGISTER
+ 1);
4057 reg1
= gen_raw_REG (address_mode
, LAST_VIRTUAL_REGISTER
+ 2);
4059 if (USE_LOAD_PRE_DECREMENT (mem_mode
)
4060 || USE_STORE_PRE_DECREMENT (mem_mode
))
4062 addr
= gen_rtx_PRE_DEC (address_mode
, reg0
);
4063 has_predec
[mem_mode
]
4064 = memory_address_addr_space_p (mem_mode
, addr
, as
);
4066 if (has_predec
[mem_mode
])
4067 data
->ainc_costs
[AINC_PRE_DEC
]
4068 = address_cost (addr
, mem_mode
, as
, speed
);
4070 if (USE_LOAD_POST_DECREMENT (mem_mode
)
4071 || USE_STORE_POST_DECREMENT (mem_mode
))
4073 addr
= gen_rtx_POST_DEC (address_mode
, reg0
);
4074 has_postdec
[mem_mode
]
4075 = memory_address_addr_space_p (mem_mode
, addr
, as
);
4077 if (has_postdec
[mem_mode
])
4078 data
->ainc_costs
[AINC_POST_DEC
]
4079 = address_cost (addr
, mem_mode
, as
, speed
);
4081 if (USE_LOAD_PRE_INCREMENT (mem_mode
)
4082 || USE_STORE_PRE_DECREMENT (mem_mode
))
4084 addr
= gen_rtx_PRE_INC (address_mode
, reg0
);
4085 has_preinc
[mem_mode
]
4086 = memory_address_addr_space_p (mem_mode
, addr
, as
);
4088 if (has_preinc
[mem_mode
])
4089 data
->ainc_costs
[AINC_PRE_INC
]
4090 = address_cost (addr
, mem_mode
, as
, speed
);
4092 if (USE_LOAD_POST_INCREMENT (mem_mode
)
4093 || USE_STORE_POST_INCREMENT (mem_mode
))
4095 addr
= gen_rtx_POST_INC (address_mode
, reg0
);
4096 has_postinc
[mem_mode
]
4097 = memory_address_addr_space_p (mem_mode
, addr
, as
);
4099 if (has_postinc
[mem_mode
])
4100 data
->ainc_costs
[AINC_POST_INC
]
4101 = address_cost (addr
, mem_mode
, as
, speed
);
4103 for (i
= 0; i
< 16; i
++)
4106 var_p
= (i
>> 1) & 1;
4107 off_p
= (i
>> 2) & 1;
4108 rat_p
= (i
>> 3) & 1;
4112 addr
= gen_rtx_fmt_ee (MULT
, address_mode
, addr
,
4113 gen_int_mode (rat
, address_mode
));
4116 addr
= gen_rtx_fmt_ee (PLUS
, address_mode
, addr
, reg1
);
4120 base
= gen_rtx_SYMBOL_REF (address_mode
, ggc_strdup (""));
4121 /* ??? We can run into trouble with some backends by presenting
4122 it with symbols which haven't been properly passed through
4123 targetm.encode_section_info. By setting the local bit, we
4124 enhance the probability of things working. */
4125 SYMBOL_REF_FLAGS (base
) = SYMBOL_FLAG_LOCAL
;
4128 base
= gen_rtx_fmt_e (CONST
, address_mode
,
4130 (PLUS
, address_mode
, base
,
4131 gen_int_mode (off
, address_mode
)));
4134 base
= gen_int_mode (off
, address_mode
);
4139 addr
= gen_rtx_fmt_ee (PLUS
, address_mode
, addr
, base
);
4142 /* To avoid splitting addressing modes, pretend that no cse will
4144 old_cse_not_expected
= cse_not_expected
;
4145 cse_not_expected
= true;
4146 addr
= memory_address_addr_space (mem_mode
, addr
, as
);
4147 cse_not_expected
= old_cse_not_expected
;
4151 acost
= seq_cost (seq
, speed
);
4152 acost
+= address_cost (addr
, mem_mode
, as
, speed
);
4156 data
->costs
[sym_p
][var_p
][off_p
][rat_p
] = acost
;
4159 /* On some targets, it is quite expensive to load symbol to a register,
4160 which makes addresses that contain symbols look much more expensive.
4161 However, the symbol will have to be loaded in any case before the
4162 loop (and quite likely we have it in register already), so it does not
4163 make much sense to penalize them too heavily. So make some final
4164 tweaks for the SYMBOL_PRESENT modes:
4166 If VAR_PRESENT is false, and the mode obtained by changing symbol to
4167 var is cheaper, use this mode with small penalty.
4168 If VAR_PRESENT is true, try whether the mode with
4169 SYMBOL_PRESENT = false is cheaper even with cost of addition, and
4170 if this is the case, use it. */
4171 add_c
= add_cost (speed
, address_mode
);
4172 for (i
= 0; i
< 8; i
++)
4175 off_p
= (i
>> 1) & 1;
4176 rat_p
= (i
>> 2) & 1;
4178 acost
= data
->costs
[0][1][off_p
][rat_p
] + 1;
4182 if (acost
< data
->costs
[1][var_p
][off_p
][rat_p
])
4183 data
->costs
[1][var_p
][off_p
][rat_p
] = acost
;
4186 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4188 fprintf (dump_file
, "<Address Costs>:\n");
4190 for (i
= 0; i
< 16; i
++)
4193 var_p
= (i
>> 1) & 1;
4194 off_p
= (i
>> 2) & 1;
4195 rat_p
= (i
>> 3) & 1;
4197 fprintf (dump_file
, " ");
4199 fprintf (dump_file
, "sym + ");
4201 fprintf (dump_file
, "var + ");
4203 fprintf (dump_file
, "cst + ");
4205 fprintf (dump_file
, "rat * ");
4207 acost
= data
->costs
[sym_p
][var_p
][off_p
][rat_p
];
4208 fprintf (dump_file
, "index costs %d\n", acost
);
4210 if (has_predec
[mem_mode
] || has_postdec
[mem_mode
]
4211 || has_preinc
[mem_mode
] || has_postinc
[mem_mode
])
4212 fprintf (dump_file
, " May include autoinc/dec\n");
4213 fprintf (dump_file
, "\n");
4216 address_cost_data_list
[data_index
] = data
;
4219 bits
= GET_MODE_BITSIZE (address_mode
);
4220 mask
= ~(HOST_WIDE_INT_M1U
<< (bits
- 1) << 1);
4222 if ((offset
>> (bits
- 1) & 1))
4227 autoinc_type
= AINC_NONE
;
4228 msize
= GET_MODE_SIZE (mem_mode
);
4229 autoinc_offset
= offset
;
4231 autoinc_offset
+= ratio
* cstep
;
4232 if (symbol_present
|| var_present
|| ratio
!= 1)
4236 if (has_postinc
[mem_mode
] && autoinc_offset
== 0
4238 autoinc_type
= AINC_POST_INC
;
4239 else if (has_postdec
[mem_mode
] && autoinc_offset
== 0
4241 autoinc_type
= AINC_POST_DEC
;
4242 else if (has_preinc
[mem_mode
] && autoinc_offset
== msize
4244 autoinc_type
= AINC_PRE_INC
;
4245 else if (has_predec
[mem_mode
] && autoinc_offset
== -msize
4247 autoinc_type
= AINC_PRE_DEC
;
4249 if (autoinc_type
!= AINC_NONE
)
4254 offset_p
= (s_offset
!= 0
4255 && data
->min_offset
<= s_offset
4256 && s_offset
<= data
->max_offset
);
4257 ratio_p
= (ratio
!= 1
4258 && multiplier_allowed_in_address_p (ratio
, mem_mode
, as
));
4260 if (ratio
!= 1 && !ratio_p
)
4261 cost
+= mult_by_coeff_cost (ratio
, address_mode
, speed
);
4263 if (s_offset
&& !offset_p
&& !symbol_present
)
4264 cost
+= add_cost (speed
, address_mode
);
4267 *may_autoinc
= autoinc
;
4269 acost
= data
->ainc_costs
[autoinc_type
];
4271 acost
= data
->costs
[symbol_present
][var_present
][offset_p
][ratio_p
];
4272 complexity
= (symbol_present
!= 0) + (var_present
!= 0) + offset_p
+ ratio_p
;
4273 return comp_cost (cost
+ acost
, complexity
);
4276 /* Calculate the SPEED or size cost of shiftadd EXPR in MODE. MULT is the
4277 EXPR operand holding the shift. COST0 and COST1 are the costs for
4278 calculating the operands of EXPR. Returns true if successful, and returns
4279 the cost in COST. */
4282 get_shiftadd_cost (tree expr
, machine_mode mode
, comp_cost cost0
,
4283 comp_cost cost1
, tree mult
, bool speed
, comp_cost
*cost
)
4286 tree op1
= TREE_OPERAND (expr
, 1);
4287 tree cst
= TREE_OPERAND (mult
, 1);
4288 tree multop
= TREE_OPERAND (mult
, 0);
4289 int m
= exact_log2 (int_cst_value (cst
));
4290 int maxm
= MIN (BITS_PER_WORD
, GET_MODE_BITSIZE (mode
));
4291 int as_cost
, sa_cost
;
4294 if (!(m
>= 0 && m
< maxm
))
4298 mult_in_op1
= operand_equal_p (op1
, mult
, 0);
4300 as_cost
= add_cost (speed
, mode
) + shift_cost (speed
, mode
, m
);
4302 /* If the target has a cheap shift-and-add or shift-and-sub instruction,
4303 use that in preference to a shift insn followed by an add insn. */
4304 sa_cost
= (TREE_CODE (expr
) != MINUS_EXPR
4305 ? shiftadd_cost (speed
, mode
, m
)
4307 ? shiftsub1_cost (speed
, mode
, m
)
4308 : shiftsub0_cost (speed
, mode
, m
)));
4310 res
= comp_cost (MIN (as_cost
, sa_cost
), 0);
4311 res
+= (mult_in_op1
? cost0
: cost1
);
4313 STRIP_NOPS (multop
);
4314 if (!is_gimple_val (multop
))
4315 res
+= force_expr_to_var_cost (multop
, speed
);
4321 /* Estimates cost of forcing expression EXPR into a variable. */
4324 force_expr_to_var_cost (tree expr
, bool speed
)
4326 static bool costs_initialized
= false;
4327 static unsigned integer_cost
[2];
4328 static unsigned symbol_cost
[2];
4329 static unsigned address_cost
[2];
4331 comp_cost cost0
, cost1
, cost
;
4334 if (!costs_initialized
)
4336 tree type
= build_pointer_type (integer_type_node
);
4341 var
= create_tmp_var_raw (integer_type_node
, "test_var");
4342 TREE_STATIC (var
) = 1;
4343 x
= produce_memory_decl_rtl (var
, NULL
);
4344 SET_DECL_RTL (var
, x
);
4346 addr
= build1 (ADDR_EXPR
, type
, var
);
4349 for (i
= 0; i
< 2; i
++)
4351 integer_cost
[i
] = computation_cost (build_int_cst (integer_type_node
,
4354 symbol_cost
[i
] = computation_cost (addr
, i
) + 1;
4357 = computation_cost (fold_build_pointer_plus_hwi (addr
, 2000), i
) + 1;
4358 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4360 fprintf (dump_file
, "force_expr_to_var_cost %s costs:\n", i
? "speed" : "size");
4361 fprintf (dump_file
, " integer %d\n", (int) integer_cost
[i
]);
4362 fprintf (dump_file
, " symbol %d\n", (int) symbol_cost
[i
]);
4363 fprintf (dump_file
, " address %d\n", (int) address_cost
[i
]);
4364 fprintf (dump_file
, " other %d\n", (int) target_spill_cost
[i
]);
4365 fprintf (dump_file
, "\n");
4369 costs_initialized
= true;
4374 if (SSA_VAR_P (expr
))
4377 if (is_gimple_min_invariant (expr
))
4379 if (TREE_CODE (expr
) == INTEGER_CST
)
4380 return comp_cost (integer_cost
[speed
], 0);
4382 if (TREE_CODE (expr
) == ADDR_EXPR
)
4384 tree obj
= TREE_OPERAND (expr
, 0);
4386 if (TREE_CODE (obj
) == VAR_DECL
4387 || TREE_CODE (obj
) == PARM_DECL
4388 || TREE_CODE (obj
) == RESULT_DECL
)
4389 return comp_cost (symbol_cost
[speed
], 0);
4392 return comp_cost (address_cost
[speed
], 0);
4395 switch (TREE_CODE (expr
))
4397 case POINTER_PLUS_EXPR
:
4401 op0
= TREE_OPERAND (expr
, 0);
4402 op1
= TREE_OPERAND (expr
, 1);
4409 op0
= TREE_OPERAND (expr
, 0);
4415 /* Just an arbitrary value, FIXME. */
4416 return comp_cost (target_spill_cost
[speed
], 0);
4419 if (op0
== NULL_TREE
4420 || TREE_CODE (op0
) == SSA_NAME
|| CONSTANT_CLASS_P (op0
))
4423 cost0
= force_expr_to_var_cost (op0
, speed
);
4425 if (op1
== NULL_TREE
4426 || TREE_CODE (op1
) == SSA_NAME
|| CONSTANT_CLASS_P (op1
))
4429 cost1
= force_expr_to_var_cost (op1
, speed
);
4431 mode
= TYPE_MODE (TREE_TYPE (expr
));
4432 switch (TREE_CODE (expr
))
4434 case POINTER_PLUS_EXPR
:
4438 cost
= comp_cost (add_cost (speed
, mode
), 0);
4439 if (TREE_CODE (expr
) != NEGATE_EXPR
)
4441 tree mult
= NULL_TREE
;
4443 if (TREE_CODE (op1
) == MULT_EXPR
)
4445 else if (TREE_CODE (op0
) == MULT_EXPR
)
4448 if (mult
!= NULL_TREE
4449 && cst_and_fits_in_hwi (TREE_OPERAND (mult
, 1))
4450 && get_shiftadd_cost (expr
, mode
, cost0
, cost1
, mult
,
4458 tree inner_mode
, outer_mode
;
4459 outer_mode
= TREE_TYPE (expr
);
4460 inner_mode
= TREE_TYPE (op0
);
4461 cost
= comp_cost (convert_cost (TYPE_MODE (outer_mode
),
4462 TYPE_MODE (inner_mode
), speed
), 0);
4467 if (cst_and_fits_in_hwi (op0
))
4468 cost
= comp_cost (mult_by_coeff_cost (int_cst_value (op0
),
4470 else if (cst_and_fits_in_hwi (op1
))
4471 cost
= comp_cost (mult_by_coeff_cost (int_cst_value (op1
),
4474 return comp_cost (target_spill_cost
[speed
], 0);
4484 /* Bound the cost by target_spill_cost. The parts of complicated
4485 computations often are either loop invariant or at least can
4486 be shared between several iv uses, so letting this grow without
4487 limits would not give reasonable results. */
4488 if (cost
.cost
> (int) target_spill_cost
[speed
])
4489 cost
.cost
= target_spill_cost
[speed
];
4494 /* Estimates cost of forcing EXPR into a variable. DEPENDS_ON is a set of the
4495 invariants the computation depends on. */
4498 force_var_cost (struct ivopts_data
*data
,
4499 tree expr
, bitmap
*depends_on
)
4503 fd_ivopts_data
= data
;
4504 walk_tree (&expr
, find_depends
, depends_on
, NULL
);
4507 return force_expr_to_var_cost (expr
, data
->speed
);
4510 /* Estimates cost of expressing address ADDR as var + symbol + offset. The
4511 value of offset is added to OFFSET, SYMBOL_PRESENT and VAR_PRESENT are set
4512 to false if the corresponding part is missing. DEPENDS_ON is a set of the
4513 invariants the computation depends on. */
4516 split_address_cost (struct ivopts_data
*data
,
4517 tree addr
, bool *symbol_present
, bool *var_present
,
4518 unsigned HOST_WIDE_INT
*offset
, bitmap
*depends_on
)
4521 HOST_WIDE_INT bitsize
;
4522 HOST_WIDE_INT bitpos
;
4525 int unsignedp
, reversep
, volatilep
;
4527 core
= get_inner_reference (addr
, &bitsize
, &bitpos
, &toffset
, &mode
,
4528 &unsignedp
, &reversep
, &volatilep
);
4531 || bitpos
% BITS_PER_UNIT
!= 0
4533 || TREE_CODE (core
) != VAR_DECL
)
4535 *symbol_present
= false;
4536 *var_present
= true;
4537 fd_ivopts_data
= data
;
4539 walk_tree (&addr
, find_depends
, depends_on
, NULL
);
4541 return comp_cost (target_spill_cost
[data
->speed
], 0);
4544 *offset
+= bitpos
/ BITS_PER_UNIT
;
4545 if (TREE_STATIC (core
)
4546 || DECL_EXTERNAL (core
))
4548 *symbol_present
= true;
4549 *var_present
= false;
4553 *symbol_present
= false;
4554 *var_present
= true;
4558 /* Estimates cost of expressing difference of addresses E1 - E2 as
4559 var + symbol + offset. The value of offset is added to OFFSET,
4560 SYMBOL_PRESENT and VAR_PRESENT are set to false if the corresponding
4561 part is missing. DEPENDS_ON is a set of the invariants the computation
4565 ptr_difference_cost (struct ivopts_data
*data
,
4566 tree e1
, tree e2
, bool *symbol_present
, bool *var_present
,
4567 unsigned HOST_WIDE_INT
*offset
, bitmap
*depends_on
)
4569 HOST_WIDE_INT diff
= 0;
4570 aff_tree aff_e1
, aff_e2
;
4573 gcc_assert (TREE_CODE (e1
) == ADDR_EXPR
);
4575 if (ptr_difference_const (e1
, e2
, &diff
))
4578 *symbol_present
= false;
4579 *var_present
= false;
4583 if (integer_zerop (e2
))
4584 return split_address_cost (data
, TREE_OPERAND (e1
, 0),
4585 symbol_present
, var_present
, offset
, depends_on
);
4587 *symbol_present
= false;
4588 *var_present
= true;
4590 type
= signed_type_for (TREE_TYPE (e1
));
4591 tree_to_aff_combination (e1
, type
, &aff_e1
);
4592 tree_to_aff_combination (e2
, type
, &aff_e2
);
4593 aff_combination_scale (&aff_e2
, -1);
4594 aff_combination_add (&aff_e1
, &aff_e2
);
4596 return force_var_cost (data
, aff_combination_to_tree (&aff_e1
), depends_on
);
4599 /* Estimates cost of expressing difference E1 - E2 as
4600 var + symbol + offset. The value of offset is added to OFFSET,
4601 SYMBOL_PRESENT and VAR_PRESENT are set to false if the corresponding
4602 part is missing. DEPENDS_ON is a set of the invariants the computation
4606 difference_cost (struct ivopts_data
*data
,
4607 tree e1
, tree e2
, bool *symbol_present
, bool *var_present
,
4608 unsigned HOST_WIDE_INT
*offset
, bitmap
*depends_on
)
4610 machine_mode mode
= TYPE_MODE (TREE_TYPE (e1
));
4611 unsigned HOST_WIDE_INT off1
, off2
;
4612 aff_tree aff_e1
, aff_e2
;
4615 e1
= strip_offset (e1
, &off1
);
4616 e2
= strip_offset (e2
, &off2
);
4617 *offset
+= off1
- off2
;
4622 if (TREE_CODE (e1
) == ADDR_EXPR
)
4623 return ptr_difference_cost (data
, e1
, e2
, symbol_present
, var_present
,
4624 offset
, depends_on
);
4625 *symbol_present
= false;
4627 if (operand_equal_p (e1
, e2
, 0))
4629 *var_present
= false;
4633 *var_present
= true;
4635 if (integer_zerop (e2
))
4636 return force_var_cost (data
, e1
, depends_on
);
4638 if (integer_zerop (e1
))
4640 comp_cost cost
= force_var_cost (data
, e2
, depends_on
);
4641 cost
+= mult_by_coeff_cost (-1, mode
, data
->speed
);
4645 type
= signed_type_for (TREE_TYPE (e1
));
4646 tree_to_aff_combination (e1
, type
, &aff_e1
);
4647 tree_to_aff_combination (e2
, type
, &aff_e2
);
4648 aff_combination_scale (&aff_e2
, -1);
4649 aff_combination_add (&aff_e1
, &aff_e2
);
4651 return force_var_cost (data
, aff_combination_to_tree (&aff_e1
), depends_on
);
4654 /* Returns true if AFF1 and AFF2 are identical. */
4657 compare_aff_trees (aff_tree
*aff1
, aff_tree
*aff2
)
4661 if (aff1
->n
!= aff2
->n
)
4664 for (i
= 0; i
< aff1
->n
; i
++)
4666 if (aff1
->elts
[i
].coef
!= aff2
->elts
[i
].coef
)
4669 if (!operand_equal_p (aff1
->elts
[i
].val
, aff2
->elts
[i
].val
, 0))
4675 /* Stores EXPR in DATA->inv_expr_tab, return pointer to iv_inv_expr_ent. */
4677 static iv_inv_expr_ent
*
4678 record_inv_expr (struct ivopts_data
*data
, tree expr
)
4680 struct iv_inv_expr_ent ent
;
4681 struct iv_inv_expr_ent
**slot
;
4684 ent
.hash
= iterative_hash_expr (expr
, 0);
4685 slot
= data
->inv_expr_tab
->find_slot (&ent
, INSERT
);
4689 *slot
= XNEW (struct iv_inv_expr_ent
);
4690 (*slot
)->expr
= expr
;
4691 (*slot
)->hash
= ent
.hash
;
4692 (*slot
)->id
= data
->max_inv_expr_id
++;
4698 /* Returns the invariant expression if expression UBASE - RATIO * CBASE
4699 requires a new compiler generated temporary. Returns -1 otherwise.
4700 ADDRESS_P is a flag indicating if the expression is for address
4703 static iv_inv_expr_ent
*
4704 get_loop_invariant_expr (struct ivopts_data
*data
, tree ubase
,
4705 tree cbase
, HOST_WIDE_INT ratio
,
4708 aff_tree ubase_aff
, cbase_aff
;
4716 if ((TREE_CODE (ubase
) == INTEGER_CST
)
4717 && (TREE_CODE (cbase
) == INTEGER_CST
))
4720 /* Strips the constant part. */
4721 if (TREE_CODE (ubase
) == PLUS_EXPR
4722 || TREE_CODE (ubase
) == MINUS_EXPR
4723 || TREE_CODE (ubase
) == POINTER_PLUS_EXPR
)
4725 if (TREE_CODE (TREE_OPERAND (ubase
, 1)) == INTEGER_CST
)
4726 ubase
= TREE_OPERAND (ubase
, 0);
4729 /* Strips the constant part. */
4730 if (TREE_CODE (cbase
) == PLUS_EXPR
4731 || TREE_CODE (cbase
) == MINUS_EXPR
4732 || TREE_CODE (cbase
) == POINTER_PLUS_EXPR
)
4734 if (TREE_CODE (TREE_OPERAND (cbase
, 1)) == INTEGER_CST
)
4735 cbase
= TREE_OPERAND (cbase
, 0);
4740 if (((TREE_CODE (ubase
) == SSA_NAME
)
4741 || (TREE_CODE (ubase
) == ADDR_EXPR
4742 && is_gimple_min_invariant (ubase
)))
4743 && (TREE_CODE (cbase
) == INTEGER_CST
))
4746 if (((TREE_CODE (cbase
) == SSA_NAME
)
4747 || (TREE_CODE (cbase
) == ADDR_EXPR
4748 && is_gimple_min_invariant (cbase
)))
4749 && (TREE_CODE (ubase
) == INTEGER_CST
))
4755 if (operand_equal_p (ubase
, cbase
, 0))
4758 if (TREE_CODE (ubase
) == ADDR_EXPR
4759 && TREE_CODE (cbase
) == ADDR_EXPR
)
4763 usym
= TREE_OPERAND (ubase
, 0);
4764 csym
= TREE_OPERAND (cbase
, 0);
4765 if (TREE_CODE (usym
) == ARRAY_REF
)
4767 tree ind
= TREE_OPERAND (usym
, 1);
4768 if (TREE_CODE (ind
) == INTEGER_CST
4769 && tree_fits_shwi_p (ind
)
4770 && tree_to_shwi (ind
) == 0)
4771 usym
= TREE_OPERAND (usym
, 0);
4773 if (TREE_CODE (csym
) == ARRAY_REF
)
4775 tree ind
= TREE_OPERAND (csym
, 1);
4776 if (TREE_CODE (ind
) == INTEGER_CST
4777 && tree_fits_shwi_p (ind
)
4778 && tree_to_shwi (ind
) == 0)
4779 csym
= TREE_OPERAND (csym
, 0);
4781 if (operand_equal_p (usym
, csym
, 0))
4784 /* Now do more complex comparison */
4785 tree_to_aff_combination (ubase
, TREE_TYPE (ubase
), &ubase_aff
);
4786 tree_to_aff_combination (cbase
, TREE_TYPE (cbase
), &cbase_aff
);
4787 if (compare_aff_trees (&ubase_aff
, &cbase_aff
))
4791 tree_to_aff_combination (ub
, TREE_TYPE (ub
), &ubase_aff
);
4792 tree_to_aff_combination (cb
, TREE_TYPE (cb
), &cbase_aff
);
4794 aff_combination_scale (&cbase_aff
, -1 * ratio
);
4795 aff_combination_add (&ubase_aff
, &cbase_aff
);
4796 expr
= aff_combination_to_tree (&ubase_aff
);
4797 return record_inv_expr (data
, expr
);
4800 /* Scale (multiply) the computed COST (except scratch part that should be
4801 hoisted out a loop) by header->frequency / AT->frequency,
4802 which makes expected cost more accurate. */
4805 get_scaled_computation_cost_at (ivopts_data
*data
, gimple
*at
, iv_cand
*cand
,
4808 int loop_freq
= data
->current_loop
->header
->frequency
;
4809 int bb_freq
= at
->bb
->frequency
;
4812 gcc_assert (cost
.scratch
<= cost
.cost
);
4814 = cost
.scratch
+ (cost
.cost
- cost
.scratch
) * bb_freq
/ loop_freq
;
4816 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4817 fprintf (dump_file
, "Scaling iv_use based on cand %d "
4818 "by %2.2f: %d (scratch: %d) -> %d (%d/%d)\n",
4819 cand
->id
, 1.0f
* bb_freq
/ loop_freq
, cost
.cost
,
4820 cost
.scratch
, scaled_cost
, bb_freq
, loop_freq
);
4822 cost
.cost
= scaled_cost
;
4828 /* Determines the cost of the computation by that USE is expressed
4829 from induction variable CAND. If ADDRESS_P is true, we just need
4830 to create an address from it, otherwise we want to get it into
4831 register. A set of invariants we depend on is stored in
4832 DEPENDS_ON. AT is the statement at that the value is computed.
4833 If CAN_AUTOINC is nonnull, use it to record whether autoinc
4834 addressing is likely. */
4837 get_computation_cost_at (struct ivopts_data
*data
,
4838 struct iv_use
*use
, struct iv_cand
*cand
,
4839 bool address_p
, bitmap
*depends_on
, gimple
*at
,
4841 iv_inv_expr_ent
**inv_expr
)
4843 tree ubase
= use
->iv
->base
, ustep
= use
->iv
->step
;
4845 tree utype
= TREE_TYPE (ubase
), ctype
;
4846 unsigned HOST_WIDE_INT cstepi
, offset
= 0;
4847 HOST_WIDE_INT ratio
, aratio
;
4848 bool var_present
, symbol_present
, stmt_is_after_inc
;
4851 bool speed
= optimize_bb_for_speed_p (gimple_bb (at
));
4852 machine_mode mem_mode
= (address_p
4853 ? TYPE_MODE (TREE_TYPE (*use
->op_p
))
4859 /* Only consider real candidates. */
4861 return infinite_cost
;
4863 cbase
= cand
->iv
->base
;
4864 cstep
= cand
->iv
->step
;
4865 ctype
= TREE_TYPE (cbase
);
4867 if (TYPE_PRECISION (utype
) > TYPE_PRECISION (ctype
))
4869 /* We do not have a precision to express the values of use. */
4870 return infinite_cost
;
4874 || (use
->iv
->base_object
4875 && cand
->iv
->base_object
4876 && POINTER_TYPE_P (TREE_TYPE (use
->iv
->base_object
))
4877 && POINTER_TYPE_P (TREE_TYPE (cand
->iv
->base_object
))))
4879 /* Do not try to express address of an object with computation based
4880 on address of a different object. This may cause problems in rtl
4881 level alias analysis (that does not expect this to be happening,
4882 as this is illegal in C), and would be unlikely to be useful
4884 if (use
->iv
->base_object
4885 && cand
->iv
->base_object
4886 && !operand_equal_p (use
->iv
->base_object
, cand
->iv
->base_object
, 0))
4887 return infinite_cost
;
4890 if (TYPE_PRECISION (utype
) < TYPE_PRECISION (ctype
))
4892 /* TODO -- add direct handling of this case. */
4896 /* CSTEPI is removed from the offset in case statement is after the
4897 increment. If the step is not constant, we use zero instead.
4898 This is a bit imprecise (there is the extra addition), but
4899 redundancy elimination is likely to transform the code so that
4900 it uses value of the variable before increment anyway,
4901 so it is not that much unrealistic. */
4902 if (cst_and_fits_in_hwi (cstep
))
4903 cstepi
= int_cst_value (cstep
);
4907 if (!constant_multiple_of (ustep
, cstep
, &rat
))
4908 return infinite_cost
;
4910 if (wi::fits_shwi_p (rat
))
4911 ratio
= rat
.to_shwi ();
4913 return infinite_cost
;
4916 ctype
= TREE_TYPE (cbase
);
4918 stmt_is_after_inc
= stmt_after_increment (data
->current_loop
, cand
, at
);
4920 /* use = ubase + ratio * (var - cbase). If either cbase is a constant
4921 or ratio == 1, it is better to handle this like
4923 ubase - ratio * cbase + ratio * var
4925 (also holds in the case ratio == -1, TODO. */
4927 if (cst_and_fits_in_hwi (cbase
))
4929 offset
= - ratio
* (unsigned HOST_WIDE_INT
) int_cst_value (cbase
);
4930 cost
= difference_cost (data
,
4931 ubase
, build_int_cst (utype
, 0),
4932 &symbol_present
, &var_present
, &offset
,
4934 cost
/= avg_loop_niter (data
->current_loop
);
4936 else if (ratio
== 1)
4938 tree real_cbase
= cbase
;
4940 /* Check to see if any adjustment is needed. */
4941 if (cstepi
== 0 && stmt_is_after_inc
)
4943 aff_tree real_cbase_aff
;
4946 tree_to_aff_combination (cbase
, TREE_TYPE (real_cbase
),
4948 tree_to_aff_combination (cstep
, TREE_TYPE (cstep
), &cstep_aff
);
4950 aff_combination_add (&real_cbase_aff
, &cstep_aff
);
4951 real_cbase
= aff_combination_to_tree (&real_cbase_aff
);
4954 cost
= difference_cost (data
,
4956 &symbol_present
, &var_present
, &offset
,
4958 cost
/= avg_loop_niter (data
->current_loop
);
4961 && !POINTER_TYPE_P (ctype
)
4962 && multiplier_allowed_in_address_p
4964 TYPE_ADDR_SPACE (TREE_TYPE (utype
))))
4966 tree real_cbase
= cbase
;
4968 if (cstepi
== 0 && stmt_is_after_inc
)
4970 if (POINTER_TYPE_P (ctype
))
4971 real_cbase
= fold_build2 (POINTER_PLUS_EXPR
, ctype
, cbase
, cstep
);
4973 real_cbase
= fold_build2 (PLUS_EXPR
, ctype
, cbase
, cstep
);
4975 real_cbase
= fold_build2 (MULT_EXPR
, ctype
, real_cbase
,
4976 build_int_cst (ctype
, ratio
));
4977 cost
= difference_cost (data
,
4979 &symbol_present
, &var_present
, &offset
,
4981 cost
/= avg_loop_niter (data
->current_loop
);
4985 cost
= force_var_cost (data
, cbase
, depends_on
);
4986 cost
+= difference_cost (data
, ubase
, build_int_cst (utype
, 0),
4987 &symbol_present
, &var_present
, &offset
,
4989 cost
/= avg_loop_niter (data
->current_loop
);
4990 cost
+= add_cost (data
->speed
, TYPE_MODE (ctype
));
4993 /* Record setup cost in scratch field. */
4994 cost
.scratch
= cost
.cost
;
4996 if (inv_expr
&& depends_on
&& *depends_on
)
4998 *inv_expr
= get_loop_invariant_expr (data
, ubase
, cbase
, ratio
,
5000 /* Clear depends on. */
5001 if (*inv_expr
!= NULL
)
5002 bitmap_clear (*depends_on
);
5005 /* If we are after the increment, the value of the candidate is higher by
5007 if (stmt_is_after_inc
)
5008 offset
-= ratio
* cstepi
;
5010 /* Now the computation is in shape symbol + var1 + const + ratio * var2.
5011 (symbol/var1/const parts may be omitted). If we are looking for an
5012 address, find the cost of addressing this. */
5015 cost
+= get_address_cost (symbol_present
, var_present
,
5016 offset
, ratio
, cstepi
,
5018 TYPE_ADDR_SPACE (TREE_TYPE (utype
)),
5019 speed
, stmt_is_after_inc
, can_autoinc
);
5020 return get_scaled_computation_cost_at (data
, at
, cand
, cost
);
5023 /* Otherwise estimate the costs for computing the expression. */
5024 if (!symbol_present
&& !var_present
&& !offset
)
5027 cost
+= mult_by_coeff_cost (ratio
, TYPE_MODE (ctype
), speed
);
5028 return get_scaled_computation_cost_at (data
, at
, cand
, cost
);
5031 /* Symbol + offset should be compile-time computable so consider that they
5032 are added once to the variable, if present. */
5033 if (var_present
&& (symbol_present
|| offset
))
5034 cost
+= adjust_setup_cost (data
,
5035 add_cost (speed
, TYPE_MODE (ctype
)));
5037 /* Having offset does not affect runtime cost in case it is added to
5038 symbol, but it increases complexity. */
5042 cost
+= add_cost (speed
, TYPE_MODE (ctype
));
5044 aratio
= ratio
> 0 ? ratio
: -ratio
;
5046 cost
+= mult_by_coeff_cost (aratio
, TYPE_MODE (ctype
), speed
);
5048 return get_scaled_computation_cost_at (data
, at
, cand
, cost
);
5052 *can_autoinc
= false;
5054 /* Just get the expression, expand it and measure the cost. */
5055 tree comp
= get_computation_at (data
->current_loop
, use
, cand
, at
);
5058 return infinite_cost
;
5061 comp
= build_simple_mem_ref (comp
);
5063 cost
= comp_cost (computation_cost (comp
, speed
), 0);
5065 return get_scaled_computation_cost_at (data
, at
, cand
, cost
);
5068 /* Determines the cost of the computation by that USE is expressed
5069 from induction variable CAND. If ADDRESS_P is true, we just need
5070 to create an address from it, otherwise we want to get it into
5071 register. A set of invariants we depend on is stored in
5072 DEPENDS_ON. If CAN_AUTOINC is nonnull, use it to record whether
5073 autoinc addressing is likely. */
5076 get_computation_cost (struct ivopts_data
*data
,
5077 struct iv_use
*use
, struct iv_cand
*cand
,
5078 bool address_p
, bitmap
*depends_on
,
5079 bool *can_autoinc
, iv_inv_expr_ent
**inv_expr
)
5081 return get_computation_cost_at (data
,
5082 use
, cand
, address_p
, depends_on
, use
->stmt
,
5083 can_autoinc
, inv_expr
);
5086 /* Determines cost of computing the use in GROUP with CAND in a generic
5090 determine_group_iv_cost_generic (struct ivopts_data
*data
,
5091 struct iv_group
*group
, struct iv_cand
*cand
)
5094 iv_inv_expr_ent
*inv_expr
= NULL
;
5095 bitmap depends_on
= NULL
;
5096 struct iv_use
*use
= group
->vuses
[0];
5098 /* The simple case first -- if we need to express value of the preserved
5099 original biv, the cost is 0. This also prevents us from counting the
5100 cost of increment twice -- once at this use and once in the cost of
5102 if (cand
->pos
== IP_ORIGINAL
&& cand
->incremented_at
== use
->stmt
)
5105 cost
= get_computation_cost (data
, use
, cand
, false,
5106 &depends_on
, NULL
, &inv_expr
);
5108 set_group_iv_cost (data
, group
, cand
, cost
, depends_on
,
5109 NULL_TREE
, ERROR_MARK
, inv_expr
);
5110 return !cost
.infinite_cost_p ();
5113 /* Determines cost of computing uses in GROUP with CAND in addresses. */
5116 determine_group_iv_cost_address (struct ivopts_data
*data
,
5117 struct iv_group
*group
, struct iv_cand
*cand
)
5122 iv_inv_expr_ent
*inv_expr
= NULL
;
5123 struct iv_use
*use
= group
->vuses
[0];
5124 comp_cost sum_cost
= no_cost
, cost
;
5126 cost
= get_computation_cost (data
, use
, cand
, true,
5127 &depends_on
, &can_autoinc
, &inv_expr
);
5130 if (!sum_cost
.infinite_cost_p () && cand
->ainc_use
== use
)
5133 sum_cost
-= cand
->cost_step
;
5134 /* If we generated the candidate solely for exploiting autoincrement
5135 opportunities, and it turns out it can't be used, set the cost to
5136 infinity to make sure we ignore it. */
5137 else if (cand
->pos
== IP_AFTER_USE
|| cand
->pos
== IP_BEFORE_USE
)
5138 sum_cost
= infinite_cost
;
5141 /* Uses in a group can share setup code, so only add setup cost once. */
5142 cost
-= cost
.scratch
;
5143 /* Compute and add costs for rest uses of this group. */
5144 for (i
= 1; i
< group
->vuses
.length () && !sum_cost
.infinite_cost_p (); i
++)
5146 struct iv_use
*next
= group
->vuses
[i
];
5148 /* TODO: We could skip computing cost for sub iv_use when it has the
5149 same cost as the first iv_use, but the cost really depends on the
5150 offset and where the iv_use is. */
5151 cost
= get_computation_cost (data
, next
, cand
, true,
5152 NULL
, &can_autoinc
, NULL
);
5155 set_group_iv_cost (data
, group
, cand
, sum_cost
, depends_on
,
5156 NULL_TREE
, ERROR_MARK
, inv_expr
);
5158 return !sum_cost
.infinite_cost_p ();
5161 /* Computes value of candidate CAND at position AT in iteration NITER, and
5162 stores it to VAL. */
5165 cand_value_at (struct loop
*loop
, struct iv_cand
*cand
, gimple
*at
, tree niter
,
5168 aff_tree step
, delta
, nit
;
5169 struct iv
*iv
= cand
->iv
;
5170 tree type
= TREE_TYPE (iv
->base
);
5172 if (POINTER_TYPE_P (type
))
5173 steptype
= sizetype
;
5175 steptype
= unsigned_type_for (type
);
5177 tree_to_aff_combination (iv
->step
, TREE_TYPE (iv
->step
), &step
);
5178 aff_combination_convert (&step
, steptype
);
5179 tree_to_aff_combination (niter
, TREE_TYPE (niter
), &nit
);
5180 aff_combination_convert (&nit
, steptype
);
5181 aff_combination_mult (&nit
, &step
, &delta
);
5182 if (stmt_after_increment (loop
, cand
, at
))
5183 aff_combination_add (&delta
, &step
);
5185 tree_to_aff_combination (iv
->base
, type
, val
);
5186 if (!POINTER_TYPE_P (type
))
5187 aff_combination_convert (val
, steptype
);
5188 aff_combination_add (val
, &delta
);
5191 /* Returns period of induction variable iv. */
5194 iv_period (struct iv
*iv
)
5196 tree step
= iv
->step
, period
, type
;
5199 gcc_assert (step
&& TREE_CODE (step
) == INTEGER_CST
);
5201 type
= unsigned_type_for (TREE_TYPE (step
));
5202 /* Period of the iv is lcm (step, type_range)/step -1,
5203 i.e., N*type_range/step - 1. Since type range is power
5204 of two, N == (step >> num_of_ending_zeros_binary (step),
5205 so the final result is
5207 (type_range >> num_of_ending_zeros_binary (step)) - 1
5210 pow2div
= num_ending_zeros (step
);
5212 period
= build_low_bits_mask (type
,
5213 (TYPE_PRECISION (type
)
5214 - tree_to_uhwi (pow2div
)));
5219 /* Returns the comparison operator used when eliminating the iv USE. */
5221 static enum tree_code
5222 iv_elimination_compare (struct ivopts_data
*data
, struct iv_use
*use
)
5224 struct loop
*loop
= data
->current_loop
;
5228 ex_bb
= gimple_bb (use
->stmt
);
5229 exit
= EDGE_SUCC (ex_bb
, 0);
5230 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
5231 exit
= EDGE_SUCC (ex_bb
, 1);
5233 return (exit
->flags
& EDGE_TRUE_VALUE
? EQ_EXPR
: NE_EXPR
);
5236 /* Returns true if we can prove that BASE - OFFSET does not overflow. For now,
5237 we only detect the situation that BASE = SOMETHING + OFFSET, where the
5238 calculation is performed in non-wrapping type.
5240 TODO: More generally, we could test for the situation that
5241 BASE = SOMETHING + OFFSET' and OFFSET is between OFFSET' and zero.
5242 This would require knowing the sign of OFFSET. */
5245 difference_cannot_overflow_p (struct ivopts_data
*data
, tree base
, tree offset
)
5247 enum tree_code code
;
5249 aff_tree aff_e1
, aff_e2
, aff_offset
;
5251 if (!nowrap_type_p (TREE_TYPE (base
)))
5254 base
= expand_simple_operations (base
);
5256 if (TREE_CODE (base
) == SSA_NAME
)
5258 gimple
*stmt
= SSA_NAME_DEF_STMT (base
);
5260 if (gimple_code (stmt
) != GIMPLE_ASSIGN
)
5263 code
= gimple_assign_rhs_code (stmt
);
5264 if (get_gimple_rhs_class (code
) != GIMPLE_BINARY_RHS
)
5267 e1
= gimple_assign_rhs1 (stmt
);
5268 e2
= gimple_assign_rhs2 (stmt
);
5272 code
= TREE_CODE (base
);
5273 if (get_gimple_rhs_class (code
) != GIMPLE_BINARY_RHS
)
5275 e1
= TREE_OPERAND (base
, 0);
5276 e2
= TREE_OPERAND (base
, 1);
5279 /* Use affine expansion as deeper inspection to prove the equality. */
5280 tree_to_aff_combination_expand (e2
, TREE_TYPE (e2
),
5281 &aff_e2
, &data
->name_expansion_cache
);
5282 tree_to_aff_combination_expand (offset
, TREE_TYPE (offset
),
5283 &aff_offset
, &data
->name_expansion_cache
);
5284 aff_combination_scale (&aff_offset
, -1);
5288 aff_combination_add (&aff_e2
, &aff_offset
);
5289 if (aff_combination_zero_p (&aff_e2
))
5292 tree_to_aff_combination_expand (e1
, TREE_TYPE (e1
),
5293 &aff_e1
, &data
->name_expansion_cache
);
5294 aff_combination_add (&aff_e1
, &aff_offset
);
5295 return aff_combination_zero_p (&aff_e1
);
5297 case POINTER_PLUS_EXPR
:
5298 aff_combination_add (&aff_e2
, &aff_offset
);
5299 return aff_combination_zero_p (&aff_e2
);
5306 /* Tries to replace loop exit by one formulated in terms of a LT_EXPR
5307 comparison with CAND. NITER describes the number of iterations of
5308 the loops. If successful, the comparison in COMP_P is altered accordingly.
5310 We aim to handle the following situation:
5326 Here, the number of iterations of the loop is (a + 1 > b) ? 0 : b - a - 1.
5327 We aim to optimize this to
5335 while (p < p_0 - a + b);
5337 This preserves the correctness, since the pointer arithmetics does not
5338 overflow. More precisely:
5340 1) if a + 1 <= b, then p_0 - a + b is the final value of p, hence there is no
5341 overflow in computing it or the values of p.
5342 2) if a + 1 > b, then we need to verify that the expression p_0 - a does not
5343 overflow. To prove this, we use the fact that p_0 = base + a. */
5346 iv_elimination_compare_lt (struct ivopts_data
*data
,
5347 struct iv_cand
*cand
, enum tree_code
*comp_p
,
5348 struct tree_niter_desc
*niter
)
5350 tree cand_type
, a
, b
, mbz
, nit_type
= TREE_TYPE (niter
->niter
), offset
;
5351 struct aff_tree nit
, tmpa
, tmpb
;
5352 enum tree_code comp
;
5355 /* We need to know that the candidate induction variable does not overflow.
5356 While more complex analysis may be used to prove this, for now just
5357 check that the variable appears in the original program and that it
5358 is computed in a type that guarantees no overflows. */
5359 cand_type
= TREE_TYPE (cand
->iv
->base
);
5360 if (cand
->pos
!= IP_ORIGINAL
|| !nowrap_type_p (cand_type
))
5363 /* Make sure that the loop iterates till the loop bound is hit, as otherwise
5364 the calculation of the BOUND could overflow, making the comparison
5366 if (!data
->loop_single_exit_p
)
5369 /* We need to be able to decide whether candidate is increasing or decreasing
5370 in order to choose the right comparison operator. */
5371 if (!cst_and_fits_in_hwi (cand
->iv
->step
))
5373 step
= int_cst_value (cand
->iv
->step
);
5375 /* Check that the number of iterations matches the expected pattern:
5376 a + 1 > b ? 0 : b - a - 1. */
5377 mbz
= niter
->may_be_zero
;
5378 if (TREE_CODE (mbz
) == GT_EXPR
)
5380 /* Handle a + 1 > b. */
5381 tree op0
= TREE_OPERAND (mbz
, 0);
5382 if (TREE_CODE (op0
) == PLUS_EXPR
&& integer_onep (TREE_OPERAND (op0
, 1)))
5384 a
= TREE_OPERAND (op0
, 0);
5385 b
= TREE_OPERAND (mbz
, 1);
5390 else if (TREE_CODE (mbz
) == LT_EXPR
)
5392 tree op1
= TREE_OPERAND (mbz
, 1);
5394 /* Handle b < a + 1. */
5395 if (TREE_CODE (op1
) == PLUS_EXPR
&& integer_onep (TREE_OPERAND (op1
, 1)))
5397 a
= TREE_OPERAND (op1
, 0);
5398 b
= TREE_OPERAND (mbz
, 0);
5406 /* Expected number of iterations is B - A - 1. Check that it matches
5407 the actual number, i.e., that B - A - NITER = 1. */
5408 tree_to_aff_combination (niter
->niter
, nit_type
, &nit
);
5409 tree_to_aff_combination (fold_convert (nit_type
, a
), nit_type
, &tmpa
);
5410 tree_to_aff_combination (fold_convert (nit_type
, b
), nit_type
, &tmpb
);
5411 aff_combination_scale (&nit
, -1);
5412 aff_combination_scale (&tmpa
, -1);
5413 aff_combination_add (&tmpb
, &tmpa
);
5414 aff_combination_add (&tmpb
, &nit
);
5415 if (tmpb
.n
!= 0 || tmpb
.offset
!= 1)
5418 /* Finally, check that CAND->IV->BASE - CAND->IV->STEP * A does not
5420 offset
= fold_build2 (MULT_EXPR
, TREE_TYPE (cand
->iv
->step
),
5422 fold_convert (TREE_TYPE (cand
->iv
->step
), a
));
5423 if (!difference_cannot_overflow_p (data
, cand
->iv
->base
, offset
))
5426 /* Determine the new comparison operator. */
5427 comp
= step
< 0 ? GT_EXPR
: LT_EXPR
;
5428 if (*comp_p
== NE_EXPR
)
5430 else if (*comp_p
== EQ_EXPR
)
5431 *comp_p
= invert_tree_comparison (comp
, false);
5438 /* Check whether it is possible to express the condition in USE by comparison
5439 of candidate CAND. If so, store the value compared with to BOUND, and the
5440 comparison operator to COMP. */
5443 may_eliminate_iv (struct ivopts_data
*data
,
5444 struct iv_use
*use
, struct iv_cand
*cand
, tree
*bound
,
5445 enum tree_code
*comp
)
5450 struct loop
*loop
= data
->current_loop
;
5452 struct tree_niter_desc
*desc
= NULL
;
5454 if (TREE_CODE (cand
->iv
->step
) != INTEGER_CST
)
5457 /* For now works only for exits that dominate the loop latch.
5458 TODO: extend to other conditions inside loop body. */
5459 ex_bb
= gimple_bb (use
->stmt
);
5460 if (use
->stmt
!= last_stmt (ex_bb
)
5461 || gimple_code (use
->stmt
) != GIMPLE_COND
5462 || !dominated_by_p (CDI_DOMINATORS
, loop
->latch
, ex_bb
))
5465 exit
= EDGE_SUCC (ex_bb
, 0);
5466 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
5467 exit
= EDGE_SUCC (ex_bb
, 1);
5468 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
5471 desc
= niter_for_exit (data
, exit
);
5475 /* Determine whether we can use the variable to test the exit condition.
5476 This is the case iff the period of the induction variable is greater
5477 than the number of iterations for which the exit condition is true. */
5478 period
= iv_period (cand
->iv
);
5480 /* If the number of iterations is constant, compare against it directly. */
5481 if (TREE_CODE (desc
->niter
) == INTEGER_CST
)
5483 /* See cand_value_at. */
5484 if (stmt_after_increment (loop
, cand
, use
->stmt
))
5486 if (!tree_int_cst_lt (desc
->niter
, period
))
5491 if (tree_int_cst_lt (period
, desc
->niter
))
5496 /* If not, and if this is the only possible exit of the loop, see whether
5497 we can get a conservative estimate on the number of iterations of the
5498 entire loop and compare against that instead. */
5501 widest_int period_value
, max_niter
;
5503 max_niter
= desc
->max
;
5504 if (stmt_after_increment (loop
, cand
, use
->stmt
))
5506 period_value
= wi::to_widest (period
);
5507 if (wi::gtu_p (max_niter
, period_value
))
5509 /* See if we can take advantage of inferred loop bound
5511 if (data
->loop_single_exit_p
)
5513 if (!max_loop_iterations (loop
, &max_niter
))
5515 /* The loop bound is already adjusted by adding 1. */
5516 if (wi::gtu_p (max_niter
, period_value
))
5524 cand_value_at (loop
, cand
, use
->stmt
, desc
->niter
, &bnd
);
5526 *bound
= fold_convert (TREE_TYPE (cand
->iv
->base
),
5527 aff_combination_to_tree (&bnd
));
5528 *comp
= iv_elimination_compare (data
, use
);
5530 /* It is unlikely that computing the number of iterations using division
5531 would be more profitable than keeping the original induction variable. */
5532 if (expression_expensive_p (*bound
))
5535 /* Sometimes, it is possible to handle the situation that the number of
5536 iterations may be zero unless additional assumtions by using <
5537 instead of != in the exit condition.
5539 TODO: we could also calculate the value MAY_BE_ZERO ? 0 : NITER and
5540 base the exit condition on it. However, that is often too
5542 if (!integer_zerop (desc
->may_be_zero
))
5543 return iv_elimination_compare_lt (data
, cand
, comp
, desc
);
5548 /* Calculates the cost of BOUND, if it is a PARM_DECL. A PARM_DECL must
5549 be copied, if it is used in the loop body and DATA->body_includes_call. */
5552 parm_decl_cost (struct ivopts_data
*data
, tree bound
)
5554 tree sbound
= bound
;
5555 STRIP_NOPS (sbound
);
5557 if (TREE_CODE (sbound
) == SSA_NAME
5558 && SSA_NAME_IS_DEFAULT_DEF (sbound
)
5559 && TREE_CODE (SSA_NAME_VAR (sbound
)) == PARM_DECL
5560 && data
->body_includes_call
)
5561 return COSTS_N_INSNS (1);
5566 /* Determines cost of computing the use in GROUP with CAND in a condition. */
5569 determine_group_iv_cost_cond (struct ivopts_data
*data
,
5570 struct iv_group
*group
, struct iv_cand
*cand
)
5572 tree bound
= NULL_TREE
;
5574 bitmap depends_on_elim
= NULL
, depends_on_express
= NULL
, depends_on
;
5575 comp_cost elim_cost
, express_cost
, cost
, bound_cost
;
5577 iv_inv_expr_ent
*elim_inv_expr
= NULL
, *express_inv_expr
= NULL
, *inv_expr
;
5578 tree
*control_var
, *bound_cst
;
5579 enum tree_code comp
= ERROR_MARK
;
5580 struct iv_use
*use
= group
->vuses
[0];
5582 gcc_assert (cand
->iv
);
5584 /* Try iv elimination. */
5585 if (may_eliminate_iv (data
, use
, cand
, &bound
, &comp
))
5587 elim_cost
= force_var_cost (data
, bound
, &depends_on_elim
);
5588 if (elim_cost
.cost
== 0)
5589 elim_cost
.cost
= parm_decl_cost (data
, bound
);
5590 else if (TREE_CODE (bound
) == INTEGER_CST
)
5592 /* If we replace a loop condition 'i < n' with 'p < base + n',
5593 depends_on_elim will have 'base' and 'n' set, which implies
5594 that both 'base' and 'n' will be live during the loop. More likely,
5595 'base + n' will be loop invariant, resulting in only one live value
5596 during the loop. So in that case we clear depends_on_elim and set
5597 elim_inv_expr_id instead. */
5598 if (depends_on_elim
&& bitmap_count_bits (depends_on_elim
) > 1)
5600 elim_inv_expr
= record_inv_expr (data
, bound
);
5601 bitmap_clear (depends_on_elim
);
5603 /* The bound is a loop invariant, so it will be only computed
5605 elim_cost
.cost
= adjust_setup_cost (data
, elim_cost
.cost
);
5608 elim_cost
= infinite_cost
;
5610 /* Try expressing the original giv. If it is compared with an invariant,
5611 note that we cannot get rid of it. */
5612 ok
= extract_cond_operands (data
, use
->stmt
, &control_var
, &bound_cst
,
5616 /* When the condition is a comparison of the candidate IV against
5617 zero, prefer this IV.
5619 TODO: The constant that we're subtracting from the cost should
5620 be target-dependent. This information should be added to the
5621 target costs for each backend. */
5622 if (!elim_cost
.infinite_cost_p () /* Do not try to decrease infinite! */
5623 && integer_zerop (*bound_cst
)
5624 && (operand_equal_p (*control_var
, cand
->var_after
, 0)
5625 || operand_equal_p (*control_var
, cand
->var_before
, 0)))
5628 express_cost
= get_computation_cost (data
, use
, cand
, false,
5629 &depends_on_express
, NULL
,
5631 fd_ivopts_data
= data
;
5632 walk_tree (&cmp_iv
->base
, find_depends
, &depends_on_express
, NULL
);
5634 /* Count the cost of the original bound as well. */
5635 bound_cost
= force_var_cost (data
, *bound_cst
, NULL
);
5636 if (bound_cost
.cost
== 0)
5637 bound_cost
.cost
= parm_decl_cost (data
, *bound_cst
);
5638 else if (TREE_CODE (*bound_cst
) == INTEGER_CST
)
5639 bound_cost
.cost
= 0;
5640 express_cost
+= bound_cost
;
5642 /* Choose the better approach, preferring the eliminated IV. */
5643 if (elim_cost
<= express_cost
)
5646 depends_on
= depends_on_elim
;
5647 depends_on_elim
= NULL
;
5648 inv_expr
= elim_inv_expr
;
5652 cost
= express_cost
;
5653 depends_on
= depends_on_express
;
5654 depends_on_express
= NULL
;
5657 inv_expr
= express_inv_expr
;
5660 set_group_iv_cost (data
, group
, cand
, cost
,
5661 depends_on
, bound
, comp
, inv_expr
);
5663 if (depends_on_elim
)
5664 BITMAP_FREE (depends_on_elim
);
5665 if (depends_on_express
)
5666 BITMAP_FREE (depends_on_express
);
5668 return !cost
.infinite_cost_p ();
5671 /* Determines cost of computing uses in GROUP with CAND. Returns false
5672 if USE cannot be represented with CAND. */
5675 determine_group_iv_cost (struct ivopts_data
*data
,
5676 struct iv_group
*group
, struct iv_cand
*cand
)
5678 switch (group
->type
)
5680 case USE_NONLINEAR_EXPR
:
5681 return determine_group_iv_cost_generic (data
, group
, cand
);
5684 return determine_group_iv_cost_address (data
, group
, cand
);
5687 return determine_group_iv_cost_cond (data
, group
, cand
);
5694 /* Return true if get_computation_cost indicates that autoincrement is
5695 a possibility for the pair of USE and CAND, false otherwise. */
5698 autoinc_possible_for_pair (struct ivopts_data
*data
, struct iv_use
*use
,
5699 struct iv_cand
*cand
)
5705 if (use
->type
!= USE_ADDRESS
)
5708 cost
= get_computation_cost (data
, use
, cand
, true, &depends_on
,
5709 &can_autoinc
, NULL
);
5711 BITMAP_FREE (depends_on
);
5713 return !cost
.infinite_cost_p () && can_autoinc
;
5716 /* Examine IP_ORIGINAL candidates to see if they are incremented next to a
5717 use that allows autoincrement, and set their AINC_USE if possible. */
5720 set_autoinc_for_original_candidates (struct ivopts_data
*data
)
5724 for (i
= 0; i
< data
->vcands
.length (); i
++)
5726 struct iv_cand
*cand
= data
->vcands
[i
];
5727 struct iv_use
*closest_before
= NULL
;
5728 struct iv_use
*closest_after
= NULL
;
5729 if (cand
->pos
!= IP_ORIGINAL
)
5732 for (j
= 0; j
< data
->vgroups
.length (); j
++)
5734 struct iv_group
*group
= data
->vgroups
[j
];
5735 struct iv_use
*use
= group
->vuses
[0];
5736 unsigned uid
= gimple_uid (use
->stmt
);
5738 if (gimple_bb (use
->stmt
) != gimple_bb (cand
->incremented_at
))
5741 if (uid
< gimple_uid (cand
->incremented_at
)
5742 && (closest_before
== NULL
5743 || uid
> gimple_uid (closest_before
->stmt
)))
5744 closest_before
= use
;
5746 if (uid
> gimple_uid (cand
->incremented_at
)
5747 && (closest_after
== NULL
5748 || uid
< gimple_uid (closest_after
->stmt
)))
5749 closest_after
= use
;
5752 if (closest_before
!= NULL
5753 && autoinc_possible_for_pair (data
, closest_before
, cand
))
5754 cand
->ainc_use
= closest_before
;
5755 else if (closest_after
!= NULL
5756 && autoinc_possible_for_pair (data
, closest_after
, cand
))
5757 cand
->ainc_use
= closest_after
;
5761 /* Finds the candidates for the induction variables. */
5764 find_iv_candidates (struct ivopts_data
*data
)
5766 /* Add commonly used ivs. */
5767 add_standard_iv_candidates (data
);
5769 /* Add old induction variables. */
5770 add_iv_candidate_for_bivs (data
);
5772 /* Add induction variables derived from uses. */
5773 add_iv_candidate_for_groups (data
);
5775 set_autoinc_for_original_candidates (data
);
5777 /* Record the important candidates. */
5778 record_important_candidates (data
);
5780 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
5784 fprintf (dump_file
, "\n<Important Candidates>:\t");
5785 for (i
= 0; i
< data
->vcands
.length (); i
++)
5786 if (data
->vcands
[i
]->important
)
5787 fprintf (dump_file
, " %d,", data
->vcands
[i
]->id
);
5788 fprintf (dump_file
, "\n");
5790 fprintf (dump_file
, "\n<Group, Cand> Related:\n");
5791 for (i
= 0; i
< data
->vgroups
.length (); i
++)
5793 struct iv_group
*group
= data
->vgroups
[i
];
5795 if (group
->related_cands
)
5797 fprintf (dump_file
, " Group %d:\t", group
->id
);
5798 dump_bitmap (dump_file
, group
->related_cands
);
5801 fprintf (dump_file
, "\n");
5805 /* Determines costs of computing use of iv with an iv candidate. */
5808 determine_group_iv_costs (struct ivopts_data
*data
)
5811 struct iv_cand
*cand
;
5812 struct iv_group
*group
;
5813 bitmap to_clear
= BITMAP_ALLOC (NULL
);
5815 alloc_use_cost_map (data
);
5817 for (i
= 0; i
< data
->vgroups
.length (); i
++)
5819 group
= data
->vgroups
[i
];
5821 if (data
->consider_all_candidates
)
5823 for (j
= 0; j
< data
->vcands
.length (); j
++)
5825 cand
= data
->vcands
[j
];
5826 determine_group_iv_cost (data
, group
, cand
);
5833 EXECUTE_IF_SET_IN_BITMAP (group
->related_cands
, 0, j
, bi
)
5835 cand
= data
->vcands
[j
];
5836 if (!determine_group_iv_cost (data
, group
, cand
))
5837 bitmap_set_bit (to_clear
, j
);
5840 /* Remove the candidates for that the cost is infinite from
5841 the list of related candidates. */
5842 bitmap_and_compl_into (group
->related_cands
, to_clear
);
5843 bitmap_clear (to_clear
);
5847 BITMAP_FREE (to_clear
);
5849 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
5851 fprintf (dump_file
, "\n<Invariant Expressions>:\n");
5852 auto_vec
<iv_inv_expr_ent
*> list (data
->inv_expr_tab
->elements ());
5854 for (hash_table
<iv_inv_expr_hasher
>::iterator it
5855 = data
->inv_expr_tab
->begin (); it
!= data
->inv_expr_tab
->end ();
5857 list
.safe_push (*it
);
5859 list
.qsort (sort_iv_inv_expr_ent
);
5861 for (i
= 0; i
< list
.length (); ++i
)
5863 fprintf (dump_file
, "inv_expr %d: \t", i
);
5864 print_generic_expr (dump_file
, list
[i
]->expr
, TDF_SLIM
);
5865 fprintf (dump_file
, "\n");
5868 fprintf (dump_file
, "\n<Group-candidate Costs>:\n");
5870 for (i
= 0; i
< data
->vgroups
.length (); i
++)
5872 group
= data
->vgroups
[i
];
5874 fprintf (dump_file
, "Group %d:\n", i
);
5875 fprintf (dump_file
, " cand\tcost\tcompl.\tinv.ex.\tdepends on\n");
5876 for (j
= 0; j
< group
->n_map_members
; j
++)
5878 if (!group
->cost_map
[j
].cand
5879 || group
->cost_map
[j
].cost
.infinite_cost_p ())
5882 fprintf (dump_file
, " %d\t%d\t%d\t",
5883 group
->cost_map
[j
].cand
->id
,
5884 group
->cost_map
[j
].cost
.cost
,
5885 group
->cost_map
[j
].cost
.complexity
);
5886 if (group
->cost_map
[j
].inv_expr
!= NULL
)
5887 fprintf (dump_file
, "%d\t",
5888 group
->cost_map
[j
].inv_expr
->id
);
5890 fprintf (dump_file
, "\t");
5891 if (group
->cost_map
[j
].depends_on
)
5892 bitmap_print (dump_file
,
5893 group
->cost_map
[j
].depends_on
, "","");
5894 fprintf (dump_file
, "\n");
5897 fprintf (dump_file
, "\n");
5899 fprintf (dump_file
, "\n");
5903 /* Determines cost of the candidate CAND. */
5906 determine_iv_cost (struct ivopts_data
*data
, struct iv_cand
*cand
)
5908 comp_cost cost_base
;
5909 unsigned cost
, cost_step
;
5918 /* There are two costs associated with the candidate -- its increment
5919 and its initialization. The second is almost negligible for any loop
5920 that rolls enough, so we take it just very little into account. */
5922 base
= cand
->iv
->base
;
5923 cost_base
= force_var_cost (data
, base
, NULL
);
5924 /* It will be exceptional that the iv register happens to be initialized with
5925 the proper value at no cost. In general, there will at least be a regcopy
5927 if (cost_base
.cost
== 0)
5928 cost_base
.cost
= COSTS_N_INSNS (1);
5929 cost_step
= add_cost (data
->speed
, TYPE_MODE (TREE_TYPE (base
)));
5931 cost
= cost_step
+ adjust_setup_cost (data
, cost_base
.cost
);
5933 /* Prefer the original ivs unless we may gain something by replacing it.
5934 The reason is to make debugging simpler; so this is not relevant for
5935 artificial ivs created by other optimization passes. */
5936 if (cand
->pos
!= IP_ORIGINAL
5937 || !SSA_NAME_VAR (cand
->var_before
)
5938 || DECL_ARTIFICIAL (SSA_NAME_VAR (cand
->var_before
)))
5941 /* Prefer not to insert statements into latch unless there are some
5942 already (so that we do not create unnecessary jumps). */
5943 if (cand
->pos
== IP_END
5944 && empty_block_p (ip_end_pos (data
->current_loop
)))
5948 cand
->cost_step
= cost_step
;
5951 /* Determines costs of computation of the candidates. */
5954 determine_iv_costs (struct ivopts_data
*data
)
5958 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
5960 fprintf (dump_file
, "<Candidate Costs>:\n");
5961 fprintf (dump_file
, " cand\tcost\n");
5964 for (i
= 0; i
< data
->vcands
.length (); i
++)
5966 struct iv_cand
*cand
= data
->vcands
[i
];
5968 determine_iv_cost (data
, cand
);
5970 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
5971 fprintf (dump_file
, " %d\t%d\n", i
, cand
->cost
);
5974 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
5975 fprintf (dump_file
, "\n");
5978 /* Calculates cost for having SIZE induction variables. */
5981 ivopts_global_cost_for_size (struct ivopts_data
*data
, unsigned size
)
5983 /* We add size to the cost, so that we prefer eliminating ivs
5985 return size
+ estimate_reg_pressure_cost (size
, data
->regs_used
, data
->speed
,
5986 data
->body_includes_call
);
5989 /* For each size of the induction variable set determine the penalty. */
5992 determine_set_costs (struct ivopts_data
*data
)
5998 struct loop
*loop
= data
->current_loop
;
6001 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
6003 fprintf (dump_file
, "<Global Costs>:\n");
6004 fprintf (dump_file
, " target_avail_regs %d\n", target_avail_regs
);
6005 fprintf (dump_file
, " target_clobbered_regs %d\n", target_clobbered_regs
);
6006 fprintf (dump_file
, " target_reg_cost %d\n", target_reg_cost
[data
->speed
]);
6007 fprintf (dump_file
, " target_spill_cost %d\n", target_spill_cost
[data
->speed
]);
6011 for (psi
= gsi_start_phis (loop
->header
); !gsi_end_p (psi
); gsi_next (&psi
))
6014 op
= PHI_RESULT (phi
);
6016 if (virtual_operand_p (op
))
6019 if (get_iv (data
, op
))
6025 EXECUTE_IF_SET_IN_BITMAP (data
->relevant
, 0, j
, bi
)
6027 struct version_info
*info
= ver_info (data
, j
);
6029 if (info
->inv_id
&& info
->has_nonlin_use
)
6033 data
->regs_used
= n
;
6034 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
6035 fprintf (dump_file
, " regs_used %d\n", n
);
6037 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
6039 fprintf (dump_file
, " cost for size:\n");
6040 fprintf (dump_file
, " ivs\tcost\n");
6041 for (j
= 0; j
<= 2 * target_avail_regs
; j
++)
6042 fprintf (dump_file
, " %d\t%d\n", j
,
6043 ivopts_global_cost_for_size (data
, j
));
6044 fprintf (dump_file
, "\n");
6048 /* Returns true if A is a cheaper cost pair than B. */
6051 cheaper_cost_pair (struct cost_pair
*a
, struct cost_pair
*b
)
6059 if (a
->cost
< b
->cost
)
6062 if (b
->cost
< a
->cost
)
6065 /* In case the costs are the same, prefer the cheaper candidate. */
6066 if (a
->cand
->cost
< b
->cand
->cost
)
6073 /* Returns candidate by that USE is expressed in IVS. */
6075 static struct cost_pair
*
6076 iv_ca_cand_for_group (struct iv_ca
*ivs
, struct iv_group
*group
)
6078 return ivs
->cand_for_group
[group
->id
];
6081 /* Computes the cost field of IVS structure. */
6084 iv_ca_recount_cost (struct ivopts_data
*data
, struct iv_ca
*ivs
)
6086 comp_cost cost
= ivs
->cand_use_cost
;
6088 cost
+= ivs
->cand_cost
;
6090 cost
+= ivopts_global_cost_for_size (data
,
6092 + ivs
->used_inv_exprs
->elements ());
6097 /* Remove invariants in set INVS to set IVS. */
6100 iv_ca_set_remove_invariants (struct iv_ca
*ivs
, bitmap invs
)
6108 EXECUTE_IF_SET_IN_BITMAP (invs
, 0, iid
, bi
)
6110 ivs
->n_invariant_uses
[iid
]--;
6111 if (ivs
->n_invariant_uses
[iid
] == 0)
6116 /* Set USE not to be expressed by any candidate in IVS. */
6119 iv_ca_set_no_cp (struct ivopts_data
*data
, struct iv_ca
*ivs
,
6120 struct iv_group
*group
)
6122 unsigned gid
= group
->id
, cid
;
6123 struct cost_pair
*cp
;
6125 cp
= ivs
->cand_for_group
[gid
];
6131 ivs
->cand_for_group
[gid
] = NULL
;
6132 ivs
->n_cand_uses
[cid
]--;
6134 if (ivs
->n_cand_uses
[cid
] == 0)
6136 bitmap_clear_bit (ivs
->cands
, cid
);
6137 /* Do not count the pseudocandidates. */
6141 ivs
->cand_cost
-= cp
->cand
->cost
;
6143 iv_ca_set_remove_invariants (ivs
, cp
->cand
->depends_on
);
6146 ivs
->cand_use_cost
-= cp
->cost
;
6148 iv_ca_set_remove_invariants (ivs
, cp
->depends_on
);
6150 if (cp
->inv_expr
!= NULL
)
6152 unsigned *slot
= ivs
->used_inv_exprs
->get (cp
->inv_expr
);
6155 ivs
->used_inv_exprs
->remove (cp
->inv_expr
);
6157 iv_ca_recount_cost (data
, ivs
);
6160 /* Add invariants in set INVS to set IVS. */
6163 iv_ca_set_add_invariants (struct iv_ca
*ivs
, bitmap invs
)
6171 EXECUTE_IF_SET_IN_BITMAP (invs
, 0, iid
, bi
)
6173 ivs
->n_invariant_uses
[iid
]++;
6174 if (ivs
->n_invariant_uses
[iid
] == 1)
6179 /* Set cost pair for GROUP in set IVS to CP. */
6182 iv_ca_set_cp (struct ivopts_data
*data
, struct iv_ca
*ivs
,
6183 struct iv_group
*group
, struct cost_pair
*cp
)
6185 unsigned gid
= group
->id
, cid
;
6187 if (ivs
->cand_for_group
[gid
] == cp
)
6190 if (ivs
->cand_for_group
[gid
])
6191 iv_ca_set_no_cp (data
, ivs
, group
);
6198 ivs
->cand_for_group
[gid
] = cp
;
6199 ivs
->n_cand_uses
[cid
]++;
6200 if (ivs
->n_cand_uses
[cid
] == 1)
6202 bitmap_set_bit (ivs
->cands
, cid
);
6203 /* Do not count the pseudocandidates. */
6207 ivs
->cand_cost
+= cp
->cand
->cost
;
6209 iv_ca_set_add_invariants (ivs
, cp
->cand
->depends_on
);
6212 ivs
->cand_use_cost
+= cp
->cost
;
6213 iv_ca_set_add_invariants (ivs
, cp
->depends_on
);
6215 if (cp
->inv_expr
!= NULL
)
6217 unsigned *slot
= &ivs
->used_inv_exprs
->get_or_insert (cp
->inv_expr
);
6220 iv_ca_recount_cost (data
, ivs
);
6224 /* Extend set IVS by expressing USE by some of the candidates in it
6225 if possible. Consider all important candidates if candidates in
6226 set IVS don't give any result. */
6229 iv_ca_add_group (struct ivopts_data
*data
, struct iv_ca
*ivs
,
6230 struct iv_group
*group
)
6232 struct cost_pair
*best_cp
= NULL
, *cp
;
6235 struct iv_cand
*cand
;
6237 gcc_assert (ivs
->upto
>= group
->id
);
6241 EXECUTE_IF_SET_IN_BITMAP (ivs
->cands
, 0, i
, bi
)
6243 cand
= data
->vcands
[i
];
6244 cp
= get_group_iv_cost (data
, group
, cand
);
6245 if (cheaper_cost_pair (cp
, best_cp
))
6249 if (best_cp
== NULL
)
6251 EXECUTE_IF_SET_IN_BITMAP (data
->important_candidates
, 0, i
, bi
)
6253 cand
= data
->vcands
[i
];
6254 cp
= get_group_iv_cost (data
, group
, cand
);
6255 if (cheaper_cost_pair (cp
, best_cp
))
6260 iv_ca_set_cp (data
, ivs
, group
, best_cp
);
6263 /* Get cost for assignment IVS. */
6266 iv_ca_cost (struct iv_ca
*ivs
)
6268 /* This was a conditional expression but it triggered a bug in
6270 if (ivs
->bad_groups
)
6271 return infinite_cost
;
6276 /* Returns true if all dependences of CP are among invariants in IVS. */
6279 iv_ca_has_deps (struct iv_ca
*ivs
, struct cost_pair
*cp
)
6284 if (!cp
->depends_on
)
6287 EXECUTE_IF_SET_IN_BITMAP (cp
->depends_on
, 0, i
, bi
)
6289 if (ivs
->n_invariant_uses
[i
] == 0)
6296 /* Creates change of expressing GROUP by NEW_CP instead of OLD_CP and chains
6299 static struct iv_ca_delta
*
6300 iv_ca_delta_add (struct iv_group
*group
, struct cost_pair
*old_cp
,
6301 struct cost_pair
*new_cp
, struct iv_ca_delta
*next
)
6303 struct iv_ca_delta
*change
= XNEW (struct iv_ca_delta
);
6305 change
->group
= group
;
6306 change
->old_cp
= old_cp
;
6307 change
->new_cp
= new_cp
;
6308 change
->next
= next
;
6313 /* Joins two lists of changes L1 and L2. Destructive -- old lists
6316 static struct iv_ca_delta
*
6317 iv_ca_delta_join (struct iv_ca_delta
*l1
, struct iv_ca_delta
*l2
)
6319 struct iv_ca_delta
*last
;
6327 for (last
= l1
; last
->next
; last
= last
->next
)
6334 /* Reverse the list of changes DELTA, forming the inverse to it. */
6336 static struct iv_ca_delta
*
6337 iv_ca_delta_reverse (struct iv_ca_delta
*delta
)
6339 struct iv_ca_delta
*act
, *next
, *prev
= NULL
;
6341 for (act
= delta
; act
; act
= next
)
6347 std::swap (act
->old_cp
, act
->new_cp
);
6353 /* Commit changes in DELTA to IVS. If FORWARD is false, the changes are
6354 reverted instead. */
6357 iv_ca_delta_commit (struct ivopts_data
*data
, struct iv_ca
*ivs
,
6358 struct iv_ca_delta
*delta
, bool forward
)
6360 struct cost_pair
*from
, *to
;
6361 struct iv_ca_delta
*act
;
6364 delta
= iv_ca_delta_reverse (delta
);
6366 for (act
= delta
; act
; act
= act
->next
)
6370 gcc_assert (iv_ca_cand_for_group (ivs
, act
->group
) == from
);
6371 iv_ca_set_cp (data
, ivs
, act
->group
, to
);
6375 iv_ca_delta_reverse (delta
);
6378 /* Returns true if CAND is used in IVS. */
6381 iv_ca_cand_used_p (struct iv_ca
*ivs
, struct iv_cand
*cand
)
6383 return ivs
->n_cand_uses
[cand
->id
] > 0;
6386 /* Returns number of induction variable candidates in the set IVS. */
6389 iv_ca_n_cands (struct iv_ca
*ivs
)
6391 return ivs
->n_cands
;
6394 /* Free the list of changes DELTA. */
6397 iv_ca_delta_free (struct iv_ca_delta
**delta
)
6399 struct iv_ca_delta
*act
, *next
;
6401 for (act
= *delta
; act
; act
= next
)
6410 /* Allocates new iv candidates assignment. */
6412 static struct iv_ca
*
6413 iv_ca_new (struct ivopts_data
*data
)
6415 struct iv_ca
*nw
= XNEW (struct iv_ca
);
6419 nw
->cand_for_group
= XCNEWVEC (struct cost_pair
*,
6420 data
->vgroups
.length ());
6421 nw
->n_cand_uses
= XCNEWVEC (unsigned, data
->vcands
.length ());
6422 nw
->cands
= BITMAP_ALLOC (NULL
);
6425 nw
->cand_use_cost
= no_cost
;
6427 nw
->n_invariant_uses
= XCNEWVEC (unsigned, data
->max_inv_id
+ 1);
6428 nw
->used_inv_exprs
= new hash_map
<iv_inv_expr_ent
*, unsigned> (13);
6434 /* Free memory occupied by the set IVS. */
6437 iv_ca_free (struct iv_ca
**ivs
)
6439 free ((*ivs
)->cand_for_group
);
6440 free ((*ivs
)->n_cand_uses
);
6441 BITMAP_FREE ((*ivs
)->cands
);
6442 free ((*ivs
)->n_invariant_uses
);
6443 delete ((*ivs
)->used_inv_exprs
);
6448 /* Dumps IVS to FILE. */
6451 iv_ca_dump (struct ivopts_data
*data
, FILE *file
, struct iv_ca
*ivs
)
6454 comp_cost cost
= iv_ca_cost (ivs
);
6456 fprintf (file
, " cost: %d (complexity %d)\n", cost
.cost
,
6458 fprintf (file
, " cand_cost: %d\n cand_group_cost: %d (complexity %d)\n",
6459 ivs
->cand_cost
, ivs
->cand_use_cost
.cost
,
6460 ivs
->cand_use_cost
.complexity
);
6461 bitmap_print (file
, ivs
->cands
, " candidates: ","\n");
6463 for (i
= 0; i
< ivs
->upto
; i
++)
6465 struct iv_group
*group
= data
->vgroups
[i
];
6466 struct cost_pair
*cp
= iv_ca_cand_for_group (ivs
, group
);
6468 fprintf (file
, " group:%d --> iv_cand:%d, cost=(%d,%d)\n",
6469 group
->id
, cp
->cand
->id
, cp
->cost
.cost
,
6470 cp
->cost
.complexity
);
6472 fprintf (file
, " group:%d --> ??\n", group
->id
);
6475 const char *pref
= "";
6476 fprintf (file
, " invariant variables: ");
6477 for (i
= 1; i
<= data
->max_inv_id
; i
++)
6478 if (ivs
->n_invariant_uses
[i
])
6480 fprintf (file
, "%s%d", pref
, i
);
6485 fprintf (file
, "\n invariant expressions: ");
6486 for (hash_map
<iv_inv_expr_ent
*, unsigned>::iterator it
6487 = ivs
->used_inv_exprs
->begin (); it
!= ivs
->used_inv_exprs
->end (); ++it
)
6489 fprintf (file
, "%s%d", pref
, (*it
).first
->id
);
6493 fprintf (file
, "\n\n");
6496 /* Try changing candidate in IVS to CAND for each use. Return cost of the
6497 new set, and store differences in DELTA. Number of induction variables
6498 in the new set is stored to N_IVS. MIN_NCAND is a flag. When it is true
6499 the function will try to find a solution with mimimal iv candidates. */
6502 iv_ca_extend (struct ivopts_data
*data
, struct iv_ca
*ivs
,
6503 struct iv_cand
*cand
, struct iv_ca_delta
**delta
,
6504 unsigned *n_ivs
, bool min_ncand
)
6508 struct iv_group
*group
;
6509 struct cost_pair
*old_cp
, *new_cp
;
6512 for (i
= 0; i
< ivs
->upto
; i
++)
6514 group
= data
->vgroups
[i
];
6515 old_cp
= iv_ca_cand_for_group (ivs
, group
);
6518 && old_cp
->cand
== cand
)
6521 new_cp
= get_group_iv_cost (data
, group
, cand
);
6525 if (!min_ncand
&& !iv_ca_has_deps (ivs
, new_cp
))
6528 if (!min_ncand
&& !cheaper_cost_pair (new_cp
, old_cp
))
6531 *delta
= iv_ca_delta_add (group
, old_cp
, new_cp
, *delta
);
6534 iv_ca_delta_commit (data
, ivs
, *delta
, true);
6535 cost
= iv_ca_cost (ivs
);
6537 *n_ivs
= iv_ca_n_cands (ivs
);
6538 iv_ca_delta_commit (data
, ivs
, *delta
, false);
6543 /* Try narrowing set IVS by removing CAND. Return the cost of
6544 the new set and store the differences in DELTA. START is
6545 the candidate with which we start narrowing. */
6548 iv_ca_narrow (struct ivopts_data
*data
, struct iv_ca
*ivs
,
6549 struct iv_cand
*cand
, struct iv_cand
*start
,
6550 struct iv_ca_delta
**delta
)
6553 struct iv_group
*group
;
6554 struct cost_pair
*old_cp
, *new_cp
, *cp
;
6556 struct iv_cand
*cnd
;
6557 comp_cost cost
, best_cost
, acost
;
6560 for (i
= 0; i
< data
->vgroups
.length (); i
++)
6562 group
= data
->vgroups
[i
];
6564 old_cp
= iv_ca_cand_for_group (ivs
, group
);
6565 if (old_cp
->cand
!= cand
)
6568 best_cost
= iv_ca_cost (ivs
);
6569 /* Start narrowing with START. */
6570 new_cp
= get_group_iv_cost (data
, group
, start
);
6572 if (data
->consider_all_candidates
)
6574 EXECUTE_IF_SET_IN_BITMAP (ivs
->cands
, 0, ci
, bi
)
6576 if (ci
== cand
->id
|| (start
&& ci
== start
->id
))
6579 cnd
= data
->vcands
[ci
];
6581 cp
= get_group_iv_cost (data
, group
, cnd
);
6585 iv_ca_set_cp (data
, ivs
, group
, cp
);
6586 acost
= iv_ca_cost (ivs
);
6588 if (acost
< best_cost
)
6597 EXECUTE_IF_AND_IN_BITMAP (group
->related_cands
, ivs
->cands
, 0, ci
, bi
)
6599 if (ci
== cand
->id
|| (start
&& ci
== start
->id
))
6602 cnd
= data
->vcands
[ci
];
6604 cp
= get_group_iv_cost (data
, group
, cnd
);
6608 iv_ca_set_cp (data
, ivs
, group
, cp
);
6609 acost
= iv_ca_cost (ivs
);
6611 if (acost
< best_cost
)
6618 /* Restore to old cp for use. */
6619 iv_ca_set_cp (data
, ivs
, group
, old_cp
);
6623 iv_ca_delta_free (delta
);
6624 return infinite_cost
;
6627 *delta
= iv_ca_delta_add (group
, old_cp
, new_cp
, *delta
);
6630 iv_ca_delta_commit (data
, ivs
, *delta
, true);
6631 cost
= iv_ca_cost (ivs
);
6632 iv_ca_delta_commit (data
, ivs
, *delta
, false);
6637 /* Try optimizing the set of candidates IVS by removing candidates different
6638 from to EXCEPT_CAND from it. Return cost of the new set, and store
6639 differences in DELTA. */
6642 iv_ca_prune (struct ivopts_data
*data
, struct iv_ca
*ivs
,
6643 struct iv_cand
*except_cand
, struct iv_ca_delta
**delta
)
6646 struct iv_ca_delta
*act_delta
, *best_delta
;
6648 comp_cost best_cost
, acost
;
6649 struct iv_cand
*cand
;
6652 best_cost
= iv_ca_cost (ivs
);
6654 EXECUTE_IF_SET_IN_BITMAP (ivs
->cands
, 0, i
, bi
)
6656 cand
= data
->vcands
[i
];
6658 if (cand
== except_cand
)
6661 acost
= iv_ca_narrow (data
, ivs
, cand
, except_cand
, &act_delta
);
6663 if (acost
< best_cost
)
6666 iv_ca_delta_free (&best_delta
);
6667 best_delta
= act_delta
;
6670 iv_ca_delta_free (&act_delta
);
6679 /* Recurse to possibly remove other unnecessary ivs. */
6680 iv_ca_delta_commit (data
, ivs
, best_delta
, true);
6681 best_cost
= iv_ca_prune (data
, ivs
, except_cand
, delta
);
6682 iv_ca_delta_commit (data
, ivs
, best_delta
, false);
6683 *delta
= iv_ca_delta_join (best_delta
, *delta
);
6687 /* Check if CAND_IDX is a candidate other than OLD_CAND and has
6688 cheaper local cost for GROUP than BEST_CP. Return pointer to
6689 the corresponding cost_pair, otherwise just return BEST_CP. */
6691 static struct cost_pair
*
6692 cheaper_cost_with_cand (struct ivopts_data
*data
, struct iv_group
*group
,
6693 unsigned int cand_idx
, struct iv_cand
*old_cand
,
6694 struct cost_pair
*best_cp
)
6696 struct iv_cand
*cand
;
6697 struct cost_pair
*cp
;
6699 gcc_assert (old_cand
!= NULL
&& best_cp
!= NULL
);
6700 if (cand_idx
== old_cand
->id
)
6703 cand
= data
->vcands
[cand_idx
];
6704 cp
= get_group_iv_cost (data
, group
, cand
);
6705 if (cp
!= NULL
&& cheaper_cost_pair (cp
, best_cp
))
6711 /* Try breaking local optimal fixed-point for IVS by replacing candidates
6712 which are used by more than one iv uses. For each of those candidates,
6713 this function tries to represent iv uses under that candidate using
6714 other ones with lower local cost, then tries to prune the new set.
6715 If the new set has lower cost, It returns the new cost after recording
6716 candidate replacement in list DELTA. */
6719 iv_ca_replace (struct ivopts_data
*data
, struct iv_ca
*ivs
,
6720 struct iv_ca_delta
**delta
)
6722 bitmap_iterator bi
, bj
;
6723 unsigned int i
, j
, k
;
6724 struct iv_cand
*cand
;
6725 comp_cost orig_cost
, acost
;
6726 struct iv_ca_delta
*act_delta
, *tmp_delta
;
6727 struct cost_pair
*old_cp
, *best_cp
= NULL
;
6730 orig_cost
= iv_ca_cost (ivs
);
6732 EXECUTE_IF_SET_IN_BITMAP (ivs
->cands
, 0, i
, bi
)
6734 if (ivs
->n_cand_uses
[i
] == 1
6735 || ivs
->n_cand_uses
[i
] > ALWAYS_PRUNE_CAND_SET_BOUND
)
6738 cand
= data
->vcands
[i
];
6741 /* Represent uses under current candidate using other ones with
6742 lower local cost. */
6743 for (j
= 0; j
< ivs
->upto
; j
++)
6745 struct iv_group
*group
= data
->vgroups
[j
];
6746 old_cp
= iv_ca_cand_for_group (ivs
, group
);
6748 if (old_cp
->cand
!= cand
)
6752 if (data
->consider_all_candidates
)
6753 for (k
= 0; k
< data
->vcands
.length (); k
++)
6754 best_cp
= cheaper_cost_with_cand (data
, group
, k
,
6755 old_cp
->cand
, best_cp
);
6757 EXECUTE_IF_SET_IN_BITMAP (group
->related_cands
, 0, k
, bj
)
6758 best_cp
= cheaper_cost_with_cand (data
, group
, k
,
6759 old_cp
->cand
, best_cp
);
6761 if (best_cp
== old_cp
)
6764 act_delta
= iv_ca_delta_add (group
, old_cp
, best_cp
, act_delta
);
6766 /* No need for further prune. */
6770 /* Prune the new candidate set. */
6771 iv_ca_delta_commit (data
, ivs
, act_delta
, true);
6772 acost
= iv_ca_prune (data
, ivs
, NULL
, &tmp_delta
);
6773 iv_ca_delta_commit (data
, ivs
, act_delta
, false);
6774 act_delta
= iv_ca_delta_join (act_delta
, tmp_delta
);
6776 if (acost
< orig_cost
)
6782 iv_ca_delta_free (&act_delta
);
6788 /* Tries to extend the sets IVS in the best possible way in order to
6789 express the GROUP. If ORIGINALP is true, prefer candidates from
6790 the original set of IVs, otherwise favor important candidates not
6791 based on any memory object. */
6794 try_add_cand_for (struct ivopts_data
*data
, struct iv_ca
*ivs
,
6795 struct iv_group
*group
, bool originalp
)
6797 comp_cost best_cost
, act_cost
;
6800 struct iv_cand
*cand
;
6801 struct iv_ca_delta
*best_delta
= NULL
, *act_delta
;
6802 struct cost_pair
*cp
;
6804 iv_ca_add_group (data
, ivs
, group
);
6805 best_cost
= iv_ca_cost (ivs
);
6806 cp
= iv_ca_cand_for_group (ivs
, group
);
6809 best_delta
= iv_ca_delta_add (group
, NULL
, cp
, NULL
);
6810 iv_ca_set_no_cp (data
, ivs
, group
);
6813 /* If ORIGINALP is true, try to find the original IV for the use. Otherwise
6814 first try important candidates not based on any memory object. Only if
6815 this fails, try the specific ones. Rationale -- in loops with many
6816 variables the best choice often is to use just one generic biv. If we
6817 added here many ivs specific to the uses, the optimization algorithm later
6818 would be likely to get stuck in a local minimum, thus causing us to create
6819 too many ivs. The approach from few ivs to more seems more likely to be
6820 successful -- starting from few ivs, replacing an expensive use by a
6821 specific iv should always be a win. */
6822 EXECUTE_IF_SET_IN_BITMAP (group
->related_cands
, 0, i
, bi
)
6824 cand
= data
->vcands
[i
];
6826 if (originalp
&& cand
->pos
!=IP_ORIGINAL
)
6829 if (!originalp
&& cand
->iv
->base_object
!= NULL_TREE
)
6832 if (iv_ca_cand_used_p (ivs
, cand
))
6835 cp
= get_group_iv_cost (data
, group
, cand
);
6839 iv_ca_set_cp (data
, ivs
, group
, cp
);
6840 act_cost
= iv_ca_extend (data
, ivs
, cand
, &act_delta
, NULL
,
6842 iv_ca_set_no_cp (data
, ivs
, group
);
6843 act_delta
= iv_ca_delta_add (group
, NULL
, cp
, act_delta
);
6845 if (act_cost
< best_cost
)
6847 best_cost
= act_cost
;
6849 iv_ca_delta_free (&best_delta
);
6850 best_delta
= act_delta
;
6853 iv_ca_delta_free (&act_delta
);
6856 if (best_cost
.infinite_cost_p ())
6858 for (i
= 0; i
< group
->n_map_members
; i
++)
6860 cp
= group
->cost_map
+ i
;
6865 /* Already tried this. */
6866 if (cand
->important
)
6868 if (originalp
&& cand
->pos
== IP_ORIGINAL
)
6870 if (!originalp
&& cand
->iv
->base_object
== NULL_TREE
)
6874 if (iv_ca_cand_used_p (ivs
, cand
))
6878 iv_ca_set_cp (data
, ivs
, group
, cp
);
6879 act_cost
= iv_ca_extend (data
, ivs
, cand
, &act_delta
, NULL
, true);
6880 iv_ca_set_no_cp (data
, ivs
, group
);
6881 act_delta
= iv_ca_delta_add (group
,
6882 iv_ca_cand_for_group (ivs
, group
),
6885 if (act_cost
< best_cost
)
6887 best_cost
= act_cost
;
6890 iv_ca_delta_free (&best_delta
);
6891 best_delta
= act_delta
;
6894 iv_ca_delta_free (&act_delta
);
6898 iv_ca_delta_commit (data
, ivs
, best_delta
, true);
6899 iv_ca_delta_free (&best_delta
);
6901 return !best_cost
.infinite_cost_p ();
6904 /* Finds an initial assignment of candidates to uses. */
6906 static struct iv_ca
*
6907 get_initial_solution (struct ivopts_data
*data
, bool originalp
)
6910 struct iv_ca
*ivs
= iv_ca_new (data
);
6912 for (i
= 0; i
< data
->vgroups
.length (); i
++)
6913 if (!try_add_cand_for (data
, ivs
, data
->vgroups
[i
], originalp
))
6922 /* Tries to improve set of induction variables IVS. TRY_REPLACE_P
6923 points to a bool variable, this function tries to break local
6924 optimal fixed-point by replacing candidates in IVS if it's true. */
6927 try_improve_iv_set (struct ivopts_data
*data
,
6928 struct iv_ca
*ivs
, bool *try_replace_p
)
6931 comp_cost acost
, best_cost
= iv_ca_cost (ivs
);
6932 struct iv_ca_delta
*best_delta
= NULL
, *act_delta
, *tmp_delta
;
6933 struct iv_cand
*cand
;
6935 /* Try extending the set of induction variables by one. */
6936 for (i
= 0; i
< data
->vcands
.length (); i
++)
6938 cand
= data
->vcands
[i
];
6940 if (iv_ca_cand_used_p (ivs
, cand
))
6943 acost
= iv_ca_extend (data
, ivs
, cand
, &act_delta
, &n_ivs
, false);
6947 /* If we successfully added the candidate and the set is small enough,
6948 try optimizing it by removing other candidates. */
6949 if (n_ivs
<= ALWAYS_PRUNE_CAND_SET_BOUND
)
6951 iv_ca_delta_commit (data
, ivs
, act_delta
, true);
6952 acost
= iv_ca_prune (data
, ivs
, cand
, &tmp_delta
);
6953 iv_ca_delta_commit (data
, ivs
, act_delta
, false);
6954 act_delta
= iv_ca_delta_join (act_delta
, tmp_delta
);
6957 if (acost
< best_cost
)
6960 iv_ca_delta_free (&best_delta
);
6961 best_delta
= act_delta
;
6964 iv_ca_delta_free (&act_delta
);
6969 /* Try removing the candidates from the set instead. */
6970 best_cost
= iv_ca_prune (data
, ivs
, NULL
, &best_delta
);
6972 if (!best_delta
&& *try_replace_p
)
6974 *try_replace_p
= false;
6975 /* So far candidate selecting algorithm tends to choose fewer IVs
6976 so that it can handle cases in which loops have many variables
6977 but the best choice is often to use only one general biv. One
6978 weakness is it can't handle opposite cases, in which different
6979 candidates should be chosen with respect to each use. To solve
6980 the problem, we replace candidates in a manner described by the
6981 comments of iv_ca_replace, thus give general algorithm a chance
6982 to break local optimal fixed-point in these cases. */
6983 best_cost
= iv_ca_replace (data
, ivs
, &best_delta
);
6990 iv_ca_delta_commit (data
, ivs
, best_delta
, true);
6991 gcc_assert (best_cost
== iv_ca_cost (ivs
));
6992 iv_ca_delta_free (&best_delta
);
6996 /* Attempts to find the optimal set of induction variables. We do simple
6997 greedy heuristic -- we try to replace at most one candidate in the selected
6998 solution and remove the unused ivs while this improves the cost. */
7000 static struct iv_ca
*
7001 find_optimal_iv_set_1 (struct ivopts_data
*data
, bool originalp
)
7004 bool try_replace_p
= true;
7006 /* Get the initial solution. */
7007 set
= get_initial_solution (data
, originalp
);
7010 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
7011 fprintf (dump_file
, "Unable to substitute for ivs, failed.\n");
7015 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
7017 fprintf (dump_file
, "Initial set of candidates:\n");
7018 iv_ca_dump (data
, dump_file
, set
);
7021 while (try_improve_iv_set (data
, set
, &try_replace_p
))
7023 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
7025 fprintf (dump_file
, "Improved to:\n");
7026 iv_ca_dump (data
, dump_file
, set
);
7033 static struct iv_ca
*
7034 find_optimal_iv_set (struct ivopts_data
*data
)
7037 comp_cost cost
, origcost
;
7038 struct iv_ca
*set
, *origset
;
7040 /* Determine the cost based on a strategy that starts with original IVs,
7041 and try again using a strategy that prefers candidates not based
7043 origset
= find_optimal_iv_set_1 (data
, true);
7044 set
= find_optimal_iv_set_1 (data
, false);
7046 if (!origset
&& !set
)
7049 origcost
= origset
? iv_ca_cost (origset
) : infinite_cost
;
7050 cost
= set
? iv_ca_cost (set
) : infinite_cost
;
7052 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
7054 fprintf (dump_file
, "Original cost %d (complexity %d)\n\n",
7055 origcost
.cost
, origcost
.complexity
);
7056 fprintf (dump_file
, "Final cost %d (complexity %d)\n\n",
7057 cost
.cost
, cost
.complexity
);
7060 /* Choose the one with the best cost. */
7061 if (origcost
<= cost
)
7068 iv_ca_free (&origset
);
7070 for (i
= 0; i
< data
->vgroups
.length (); i
++)
7072 struct iv_group
*group
= data
->vgroups
[i
];
7073 group
->selected
= iv_ca_cand_for_group (set
, group
)->cand
;
7079 /* Creates a new induction variable corresponding to CAND. */
7082 create_new_iv (struct ivopts_data
*data
, struct iv_cand
*cand
)
7084 gimple_stmt_iterator incr_pos
;
7087 struct iv_group
*group
;
7096 incr_pos
= gsi_last_bb (ip_normal_pos (data
->current_loop
));
7100 incr_pos
= gsi_last_bb (ip_end_pos (data
->current_loop
));
7108 incr_pos
= gsi_for_stmt (cand
->incremented_at
);
7112 /* Mark that the iv is preserved. */
7113 name_info (data
, cand
->var_before
)->preserve_biv
= true;
7114 name_info (data
, cand
->var_after
)->preserve_biv
= true;
7116 /* Rewrite the increment so that it uses var_before directly. */
7117 use
= find_interesting_uses_op (data
, cand
->var_after
);
7118 group
= data
->vgroups
[use
->group_id
];
7119 group
->selected
= cand
;
7123 gimple_add_tmp_var (cand
->var_before
);
7125 base
= unshare_expr (cand
->iv
->base
);
7127 create_iv (base
, unshare_expr (cand
->iv
->step
),
7128 cand
->var_before
, data
->current_loop
,
7129 &incr_pos
, after
, &cand
->var_before
, &cand
->var_after
);
7132 /* Creates new induction variables described in SET. */
7135 create_new_ivs (struct ivopts_data
*data
, struct iv_ca
*set
)
7138 struct iv_cand
*cand
;
7141 EXECUTE_IF_SET_IN_BITMAP (set
->cands
, 0, i
, bi
)
7143 cand
= data
->vcands
[i
];
7144 create_new_iv (data
, cand
);
7147 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
7149 fprintf (dump_file
, "Selected IV set for loop %d",
7150 data
->current_loop
->num
);
7151 if (data
->loop_loc
!= UNKNOWN_LOCATION
)
7152 fprintf (dump_file
, " at %s:%d", LOCATION_FILE (data
->loop_loc
),
7153 LOCATION_LINE (data
->loop_loc
));
7154 fprintf (dump_file
, ", " HOST_WIDE_INT_PRINT_DEC
" avg niters",
7155 avg_loop_niter (data
->current_loop
));
7156 fprintf (dump_file
, ", " HOST_WIDE_INT_PRINT_UNSIGNED
" expressions",
7157 (unsigned HOST_WIDE_INT
) set
->used_inv_exprs
->elements ());
7158 fprintf (dump_file
, ", %lu IVs:\n", bitmap_count_bits (set
->cands
));
7159 EXECUTE_IF_SET_IN_BITMAP (set
->cands
, 0, i
, bi
)
7161 cand
= data
->vcands
[i
];
7162 dump_cand (dump_file
, cand
);
7164 fprintf (dump_file
, "\n");
7168 /* Rewrites USE (definition of iv used in a nonlinear expression)
7169 using candidate CAND. */
7172 rewrite_use_nonlinear_expr (struct ivopts_data
*data
,
7173 struct iv_use
*use
, struct iv_cand
*cand
)
7178 gimple_stmt_iterator bsi
;
7180 /* An important special case -- if we are asked to express value of
7181 the original iv by itself, just exit; there is no need to
7182 introduce a new computation (that might also need casting the
7183 variable to unsigned and back). */
7184 if (cand
->pos
== IP_ORIGINAL
7185 && cand
->incremented_at
== use
->stmt
)
7187 enum tree_code stmt_code
;
7189 gcc_assert (is_gimple_assign (use
->stmt
));
7190 gcc_assert (gimple_assign_lhs (use
->stmt
) == cand
->var_after
);
7192 /* Check whether we may leave the computation unchanged.
7193 This is the case only if it does not rely on other
7194 computations in the loop -- otherwise, the computation
7195 we rely upon may be removed in remove_unused_ivs,
7196 thus leading to ICE. */
7197 stmt_code
= gimple_assign_rhs_code (use
->stmt
);
7198 if (stmt_code
== PLUS_EXPR
7199 || stmt_code
== MINUS_EXPR
7200 || stmt_code
== POINTER_PLUS_EXPR
)
7202 if (gimple_assign_rhs1 (use
->stmt
) == cand
->var_before
)
7203 op
= gimple_assign_rhs2 (use
->stmt
);
7204 else if (gimple_assign_rhs2 (use
->stmt
) == cand
->var_before
)
7205 op
= gimple_assign_rhs1 (use
->stmt
);
7212 if (op
&& expr_invariant_in_loop_p (data
->current_loop
, op
))
7216 comp
= get_computation (data
->current_loop
, use
, cand
);
7217 gcc_assert (comp
!= NULL_TREE
);
7219 switch (gimple_code (use
->stmt
))
7222 tgt
= PHI_RESULT (use
->stmt
);
7224 /* If we should keep the biv, do not replace it. */
7225 if (name_info (data
, tgt
)->preserve_biv
)
7228 bsi
= gsi_after_labels (gimple_bb (use
->stmt
));
7232 tgt
= gimple_assign_lhs (use
->stmt
);
7233 bsi
= gsi_for_stmt (use
->stmt
);
7240 if (!valid_gimple_rhs_p (comp
)
7241 || (gimple_code (use
->stmt
) != GIMPLE_PHI
7242 /* We can't allow re-allocating the stmt as it might be pointed
7244 && (get_gimple_rhs_num_ops (TREE_CODE (comp
))
7245 >= gimple_num_ops (gsi_stmt (bsi
)))))
7247 comp
= force_gimple_operand_gsi (&bsi
, comp
, true, NULL_TREE
,
7248 true, GSI_SAME_STMT
);
7249 if (POINTER_TYPE_P (TREE_TYPE (tgt
)))
7251 duplicate_ssa_name_ptr_info (comp
, SSA_NAME_PTR_INFO (tgt
));
7252 /* As this isn't a plain copy we have to reset alignment
7254 if (SSA_NAME_PTR_INFO (comp
))
7255 mark_ptr_info_alignment_unknown (SSA_NAME_PTR_INFO (comp
));
7259 if (gimple_code (use
->stmt
) == GIMPLE_PHI
)
7261 ass
= gimple_build_assign (tgt
, comp
);
7262 gsi_insert_before (&bsi
, ass
, GSI_SAME_STMT
);
7264 bsi
= gsi_for_stmt (use
->stmt
);
7265 remove_phi_node (&bsi
, false);
7269 gimple_assign_set_rhs_from_tree (&bsi
, comp
);
7270 use
->stmt
= gsi_stmt (bsi
);
7274 /* Performs a peephole optimization to reorder the iv update statement with
7275 a mem ref to enable instruction combining in later phases. The mem ref uses
7276 the iv value before the update, so the reordering transformation requires
7277 adjustment of the offset. CAND is the selected IV_CAND.
7281 t = MEM_REF (base, iv1, 8, 16); // base, index, stride, offset
7289 directly propagating t over to (1) will introduce overlapping live range
7290 thus increase register pressure. This peephole transform it into:
7294 t = MEM_REF (base, iv2, 8, 8);
7301 adjust_iv_update_pos (struct iv_cand
*cand
, struct iv_use
*use
)
7304 gimple
*iv_update
, *stmt
;
7306 gimple_stmt_iterator gsi
, gsi_iv
;
7308 if (cand
->pos
!= IP_NORMAL
)
7311 var_after
= cand
->var_after
;
7312 iv_update
= SSA_NAME_DEF_STMT (var_after
);
7314 bb
= gimple_bb (iv_update
);
7315 gsi
= gsi_last_nondebug_bb (bb
);
7316 stmt
= gsi_stmt (gsi
);
7318 /* Only handle conditional statement for now. */
7319 if (gimple_code (stmt
) != GIMPLE_COND
)
7322 gsi_prev_nondebug (&gsi
);
7323 stmt
= gsi_stmt (gsi
);
7324 if (stmt
!= iv_update
)
7327 gsi_prev_nondebug (&gsi
);
7328 if (gsi_end_p (gsi
))
7331 stmt
= gsi_stmt (gsi
);
7332 if (gimple_code (stmt
) != GIMPLE_ASSIGN
)
7335 if (stmt
!= use
->stmt
)
7338 if (TREE_CODE (gimple_assign_lhs (stmt
)) != SSA_NAME
)
7341 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
7343 fprintf (dump_file
, "Reordering \n");
7344 print_gimple_stmt (dump_file
, iv_update
, 0, 0);
7345 print_gimple_stmt (dump_file
, use
->stmt
, 0, 0);
7346 fprintf (dump_file
, "\n");
7349 gsi
= gsi_for_stmt (use
->stmt
);
7350 gsi_iv
= gsi_for_stmt (iv_update
);
7351 gsi_move_before (&gsi_iv
, &gsi
);
7353 cand
->pos
= IP_BEFORE_USE
;
7354 cand
->incremented_at
= use
->stmt
;
7357 /* Rewrites USE (address that is an iv) using candidate CAND. */
7360 rewrite_use_address (struct ivopts_data
*data
,
7361 struct iv_use
*use
, struct iv_cand
*cand
)
7364 gimple_stmt_iterator bsi
= gsi_for_stmt (use
->stmt
);
7365 tree base_hint
= NULL_TREE
;
7369 adjust_iv_update_pos (cand
, use
);
7370 ok
= get_computation_aff (data
->current_loop
, use
, cand
, use
->stmt
, &aff
);
7372 unshare_aff_combination (&aff
);
7374 /* To avoid undefined overflow problems, all IV candidates use unsigned
7375 integer types. The drawback is that this makes it impossible for
7376 create_mem_ref to distinguish an IV that is based on a memory object
7377 from one that represents simply an offset.
7379 To work around this problem, we pass a hint to create_mem_ref that
7380 indicates which variable (if any) in aff is an IV based on a memory
7381 object. Note that we only consider the candidate. If this is not
7382 based on an object, the base of the reference is in some subexpression
7383 of the use -- but these will use pointer types, so they are recognized
7384 by the create_mem_ref heuristics anyway. */
7385 if (cand
->iv
->base_object
)
7386 base_hint
= var_at_stmt (data
->current_loop
, cand
, use
->stmt
);
7388 iv
= var_at_stmt (data
->current_loop
, cand
, use
->stmt
);
7389 ref
= create_mem_ref (&bsi
, TREE_TYPE (*use
->op_p
), &aff
,
7390 reference_alias_ptr_type (*use
->op_p
),
7391 iv
, base_hint
, data
->speed
);
7392 copy_ref_info (ref
, *use
->op_p
);
7396 /* Rewrites USE (the condition such that one of the arguments is an iv) using
7400 rewrite_use_compare (struct ivopts_data
*data
,
7401 struct iv_use
*use
, struct iv_cand
*cand
)
7403 tree comp
, *var_p
, op
, bound
;
7404 gimple_stmt_iterator bsi
= gsi_for_stmt (use
->stmt
);
7405 enum tree_code compare
;
7406 struct iv_group
*group
= data
->vgroups
[use
->group_id
];
7407 struct cost_pair
*cp
= get_group_iv_cost (data
, group
, cand
);
7413 tree var
= var_at_stmt (data
->current_loop
, cand
, use
->stmt
);
7414 tree var_type
= TREE_TYPE (var
);
7417 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
7419 fprintf (dump_file
, "Replacing exit test: ");
7420 print_gimple_stmt (dump_file
, use
->stmt
, 0, TDF_SLIM
);
7423 bound
= unshare_expr (fold_convert (var_type
, bound
));
7424 op
= force_gimple_operand (bound
, &stmts
, true, NULL_TREE
);
7426 gsi_insert_seq_on_edge_immediate (
7427 loop_preheader_edge (data
->current_loop
),
7430 gcond
*cond_stmt
= as_a
<gcond
*> (use
->stmt
);
7431 gimple_cond_set_lhs (cond_stmt
, var
);
7432 gimple_cond_set_code (cond_stmt
, compare
);
7433 gimple_cond_set_rhs (cond_stmt
, op
);
7437 /* The induction variable elimination failed; just express the original
7439 comp
= get_computation (data
->current_loop
, use
, cand
);
7440 gcc_assert (comp
!= NULL_TREE
);
7442 ok
= extract_cond_operands (data
, use
->stmt
, &var_p
, NULL
, NULL
, NULL
);
7445 *var_p
= force_gimple_operand_gsi (&bsi
, comp
, true, SSA_NAME_VAR (*var_p
),
7446 true, GSI_SAME_STMT
);
7449 /* Rewrite the groups using the selected induction variables. */
7452 rewrite_groups (struct ivopts_data
*data
)
7456 for (i
= 0; i
< data
->vgroups
.length (); i
++)
7458 struct iv_group
*group
= data
->vgroups
[i
];
7459 struct iv_cand
*cand
= group
->selected
;
7463 if (group
->type
== USE_NONLINEAR_EXPR
)
7465 for (j
= 0; j
< group
->vuses
.length (); j
++)
7467 rewrite_use_nonlinear_expr (data
, group
->vuses
[j
], cand
);
7468 update_stmt (group
->vuses
[j
]->stmt
);
7471 else if (group
->type
== USE_ADDRESS
)
7473 for (j
= 0; j
< group
->vuses
.length (); j
++)
7475 rewrite_use_address (data
, group
->vuses
[j
], cand
);
7476 update_stmt (group
->vuses
[j
]->stmt
);
7481 gcc_assert (group
->type
== USE_COMPARE
);
7483 for (j
= 0; j
< group
->vuses
.length (); j
++)
7485 rewrite_use_compare (data
, group
->vuses
[j
], cand
);
7486 update_stmt (group
->vuses
[j
]->stmt
);
7492 /* Removes the ivs that are not used after rewriting. */
7495 remove_unused_ivs (struct ivopts_data
*data
)
7499 bitmap toremove
= BITMAP_ALLOC (NULL
);
7501 /* Figure out an order in which to release SSA DEFs so that we don't
7502 release something that we'd have to propagate into a debug stmt
7504 EXECUTE_IF_SET_IN_BITMAP (data
->relevant
, 0, j
, bi
)
7506 struct version_info
*info
;
7508 info
= ver_info (data
, j
);
7510 && !integer_zerop (info
->iv
->step
)
7512 && !info
->iv
->nonlin_use
7513 && !info
->preserve_biv
)
7515 bitmap_set_bit (toremove
, SSA_NAME_VERSION (info
->iv
->ssa_name
));
7517 tree def
= info
->iv
->ssa_name
;
7519 if (MAY_HAVE_DEBUG_STMTS
&& SSA_NAME_DEF_STMT (def
))
7521 imm_use_iterator imm_iter
;
7522 use_operand_p use_p
;
7526 FOR_EACH_IMM_USE_STMT (stmt
, imm_iter
, def
)
7528 if (!gimple_debug_bind_p (stmt
))
7531 /* We just want to determine whether to do nothing
7532 (count == 0), to substitute the computed
7533 expression into a single use of the SSA DEF by
7534 itself (count == 1), or to use a debug temp
7535 because the SSA DEF is used multiple times or as
7536 part of a larger expression (count > 1). */
7538 if (gimple_debug_bind_get_value (stmt
) != def
)
7542 BREAK_FROM_IMM_USE_STMT (imm_iter
);
7548 struct iv_use dummy_use
;
7549 struct iv_cand
*best_cand
= NULL
, *cand
;
7550 unsigned i
, best_pref
= 0, cand_pref
;
7552 memset (&dummy_use
, 0, sizeof (dummy_use
));
7553 dummy_use
.iv
= info
->iv
;
7554 for (i
= 0; i
< data
->vgroups
.length () && i
< 64; i
++)
7556 cand
= data
->vgroups
[i
]->selected
;
7557 if (cand
== best_cand
)
7559 cand_pref
= operand_equal_p (cand
->iv
->step
,
7563 += TYPE_MODE (TREE_TYPE (cand
->iv
->base
))
7564 == TYPE_MODE (TREE_TYPE (info
->iv
->base
))
7567 += TREE_CODE (cand
->iv
->base
) == INTEGER_CST
7569 if (best_cand
== NULL
|| best_pref
< cand_pref
)
7572 best_pref
= cand_pref
;
7579 tree comp
= get_computation_at (data
->current_loop
,
7580 &dummy_use
, best_cand
,
7581 SSA_NAME_DEF_STMT (def
));
7587 tree vexpr
= make_node (DEBUG_EXPR_DECL
);
7588 DECL_ARTIFICIAL (vexpr
) = 1;
7589 TREE_TYPE (vexpr
) = TREE_TYPE (comp
);
7590 if (SSA_NAME_VAR (def
))
7591 DECL_MODE (vexpr
) = DECL_MODE (SSA_NAME_VAR (def
));
7593 DECL_MODE (vexpr
) = TYPE_MODE (TREE_TYPE (vexpr
));
7595 = gimple_build_debug_bind (vexpr
, comp
, NULL
);
7596 gimple_stmt_iterator gsi
;
7598 if (gimple_code (SSA_NAME_DEF_STMT (def
)) == GIMPLE_PHI
)
7599 gsi
= gsi_after_labels (gimple_bb
7600 (SSA_NAME_DEF_STMT (def
)));
7602 gsi
= gsi_for_stmt (SSA_NAME_DEF_STMT (def
));
7604 gsi_insert_before (&gsi
, def_temp
, GSI_SAME_STMT
);
7608 FOR_EACH_IMM_USE_STMT (stmt
, imm_iter
, def
)
7610 if (!gimple_debug_bind_p (stmt
))
7613 FOR_EACH_IMM_USE_ON_STMT (use_p
, imm_iter
)
7614 SET_USE (use_p
, comp
);
7622 release_defs_bitset (toremove
);
7624 BITMAP_FREE (toremove
);
7627 /* Frees memory occupied by struct tree_niter_desc in *VALUE. Callback
7628 for hash_map::traverse. */
7631 free_tree_niter_desc (edge
const &, tree_niter_desc
*const &value
, void *)
7637 /* Frees data allocated by the optimization of a single loop. */
7640 free_loop_data (struct ivopts_data
*data
)
7648 data
->niters
->traverse
<void *, free_tree_niter_desc
> (NULL
);
7649 delete data
->niters
;
7650 data
->niters
= NULL
;
7653 EXECUTE_IF_SET_IN_BITMAP (data
->relevant
, 0, i
, bi
)
7655 struct version_info
*info
;
7657 info
= ver_info (data
, i
);
7659 info
->has_nonlin_use
= false;
7660 info
->preserve_biv
= false;
7663 bitmap_clear (data
->relevant
);
7664 bitmap_clear (data
->important_candidates
);
7666 for (i
= 0; i
< data
->vgroups
.length (); i
++)
7668 struct iv_group
*group
= data
->vgroups
[i
];
7670 for (j
= 0; j
< group
->vuses
.length (); j
++)
7671 free (group
->vuses
[j
]);
7672 group
->vuses
.release ();
7674 BITMAP_FREE (group
->related_cands
);
7675 for (j
= 0; j
< group
->n_map_members
; j
++)
7676 if (group
->cost_map
[j
].depends_on
)
7677 BITMAP_FREE (group
->cost_map
[j
].depends_on
);
7679 free (group
->cost_map
);
7682 data
->vgroups
.truncate (0);
7684 for (i
= 0; i
< data
->vcands
.length (); i
++)
7686 struct iv_cand
*cand
= data
->vcands
[i
];
7688 if (cand
->depends_on
)
7689 BITMAP_FREE (cand
->depends_on
);
7692 data
->vcands
.truncate (0);
7694 if (data
->version_info_size
< num_ssa_names
)
7696 data
->version_info_size
= 2 * num_ssa_names
;
7697 free (data
->version_info
);
7698 data
->version_info
= XCNEWVEC (struct version_info
, data
->version_info_size
);
7701 data
->max_inv_id
= 0;
7703 FOR_EACH_VEC_ELT (decl_rtl_to_reset
, i
, obj
)
7704 SET_DECL_RTL (obj
, NULL_RTX
);
7706 decl_rtl_to_reset
.truncate (0);
7708 data
->inv_expr_tab
->empty ();
7709 data
->max_inv_expr_id
= 0;
7711 data
->iv_common_cand_tab
->empty ();
7712 data
->iv_common_cands
.truncate (0);
7715 /* Finalizes data structures used by the iv optimization pass. LOOPS is the
7719 tree_ssa_iv_optimize_finalize (struct ivopts_data
*data
)
7721 free_loop_data (data
);
7722 free (data
->version_info
);
7723 BITMAP_FREE (data
->relevant
);
7724 BITMAP_FREE (data
->important_candidates
);
7726 decl_rtl_to_reset
.release ();
7727 data
->vgroups
.release ();
7728 data
->vcands
.release ();
7729 delete data
->inv_expr_tab
;
7730 data
->inv_expr_tab
= NULL
;
7731 free_affine_expand_cache (&data
->name_expansion_cache
);
7732 delete data
->iv_common_cand_tab
;
7733 data
->iv_common_cand_tab
= NULL
;
7734 data
->iv_common_cands
.release ();
7735 obstack_free (&data
->iv_obstack
, NULL
);
7738 /* Returns true if the loop body BODY includes any function calls. */
7741 loop_body_includes_call (basic_block
*body
, unsigned num_nodes
)
7743 gimple_stmt_iterator gsi
;
7746 for (i
= 0; i
< num_nodes
; i
++)
7747 for (gsi
= gsi_start_bb (body
[i
]); !gsi_end_p (gsi
); gsi_next (&gsi
))
7749 gimple
*stmt
= gsi_stmt (gsi
);
7750 if (is_gimple_call (stmt
)
7751 && !gimple_call_internal_p (stmt
)
7752 && !is_inexpensive_builtin (gimple_call_fndecl (stmt
)))
7758 /* Optimizes the LOOP. Returns true if anything changed. */
7761 tree_ssa_iv_optimize_loop (struct ivopts_data
*data
, struct loop
*loop
)
7763 bool changed
= false;
7764 struct iv_ca
*iv_ca
;
7765 edge exit
= single_dom_exit (loop
);
7768 gcc_assert (!data
->niters
);
7769 data
->current_loop
= loop
;
7770 data
->loop_loc
= find_loop_location (loop
);
7771 data
->speed
= optimize_loop_for_speed_p (loop
);
7773 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
7775 fprintf (dump_file
, "Processing loop %d", loop
->num
);
7776 if (data
->loop_loc
!= UNKNOWN_LOCATION
)
7777 fprintf (dump_file
, " at %s:%d", LOCATION_FILE (data
->loop_loc
),
7778 LOCATION_LINE (data
->loop_loc
));
7779 fprintf (dump_file
, "\n");
7783 fprintf (dump_file
, " single exit %d -> %d, exit condition ",
7784 exit
->src
->index
, exit
->dest
->index
);
7785 print_gimple_stmt (dump_file
, last_stmt (exit
->src
), 0, TDF_SLIM
);
7786 fprintf (dump_file
, "\n");
7789 fprintf (dump_file
, "\n");
7792 body
= get_loop_body (loop
);
7793 data
->body_includes_call
= loop_body_includes_call (body
, loop
->num_nodes
);
7794 renumber_gimple_stmt_uids_in_blocks (body
, loop
->num_nodes
);
7797 data
->loop_single_exit_p
= exit
!= NULL
&& loop_only_exit_p (loop
, exit
);
7799 /* For each ssa name determines whether it behaves as an induction variable
7801 if (!find_induction_variables (data
))
7804 /* Finds interesting uses (item 1). */
7805 find_interesting_uses (data
);
7806 if (data
->vgroups
.length () > MAX_CONSIDERED_GROUPS
)
7809 /* Finds candidates for the induction variables (item 2). */
7810 find_iv_candidates (data
);
7812 /* Calculates the costs (item 3, part 1). */
7813 determine_iv_costs (data
);
7814 determine_group_iv_costs (data
);
7815 determine_set_costs (data
);
7817 /* Find the optimal set of induction variables (item 3, part 2). */
7818 iv_ca
= find_optimal_iv_set (data
);
7823 /* Create the new induction variables (item 4, part 1). */
7824 create_new_ivs (data
, iv_ca
);
7825 iv_ca_free (&iv_ca
);
7827 /* Rewrite the uses (item 4, part 2). */
7828 rewrite_groups (data
);
7830 /* Remove the ivs that are unused after rewriting. */
7831 remove_unused_ivs (data
);
7833 /* We have changed the structure of induction variables; it might happen
7834 that definitions in the scev database refer to some of them that were
7839 free_loop_data (data
);
7844 /* Main entry point. Optimizes induction variables in loops. */
7847 tree_ssa_iv_optimize (void)
7850 struct ivopts_data data
;
7852 tree_ssa_iv_optimize_init (&data
);
7854 /* Optimize the loops starting with the innermost ones. */
7855 FOR_EACH_LOOP (loop
, LI_FROM_INNERMOST
)
7857 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
7858 flow_loop_dump (loop
, dump_file
, NULL
, 1);
7860 tree_ssa_iv_optimize_loop (&data
, loop
);
7863 tree_ssa_iv_optimize_finalize (&data
);