1 /* Straight-line strength reduction.
2 Copyright (C) 2012-2015 Free Software Foundation, Inc.
3 Contributed by Bill Schmidt, IBM <wschmidt@linux.ibm.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 /* There are many algorithms for performing strength reduction on
22 loops. This is not one of them. IVOPTS handles strength reduction
23 of induction variables just fine. This pass is intended to pick
24 up the crumbs it leaves behind, by considering opportunities for
25 strength reduction along dominator paths.
27 Strength reduction addresses explicit multiplies, and certain
28 multiplies implicit in addressing expressions. It would also be
29 possible to apply strength reduction to divisions and modulos,
30 but such opportunities are relatively uncommon.
32 Strength reduction is also currently restricted to integer operations.
33 If desired, it could be extended to floating-point operations under
34 control of something like -funsafe-math-optimizations. */
38 #include "coretypes.h"
44 #include "tree-pass.h"
47 #include "insn-config.h"
49 #include "gimple-pretty-print.h"
51 #include "fold-const.h"
52 #include "internal-fn.h"
53 #include "gimple-iterator.h"
54 #include "gimplify-me.h"
55 #include "stor-layout.h"
67 #include "tree-ssa-address.h"
68 #include "tree-affine.h"
71 /* Information about a strength reduction candidate. Each statement
72 in the candidate table represents an expression of one of the
73 following forms (the special case of CAND_REF will be described
76 (CAND_MULT) S1: X = (B + i) * S
77 (CAND_ADD) S1: X = B + (i * S)
79 Here X and B are SSA names, i is an integer constant, and S is
80 either an SSA name or a constant. We call B the "base," i the
81 "index", and S the "stride."
83 Any statement S0 that dominates S1 and is of the form:
85 (CAND_MULT) S0: Y = (B + i') * S
86 (CAND_ADD) S0: Y = B + (i' * S)
88 is called a "basis" for S1. In both cases, S1 may be replaced by
90 S1': X = Y + (i - i') * S,
92 where (i - i') * S is folded to the extent possible.
94 All gimple statements are visited in dominator order, and each
95 statement that may contribute to one of the forms of S1 above is
96 given at least one entry in the candidate table. Such statements
97 include addition, pointer addition, subtraction, multiplication,
98 negation, copies, and nontrivial type casts. If a statement may
99 represent more than one expression of the forms of S1 above,
100 multiple "interpretations" are stored in the table and chained
103 * An add of two SSA names may treat either operand as the base.
104 * A multiply of two SSA names, likewise.
105 * A copy or cast may be thought of as either a CAND_MULT with
106 i = 0 and S = 1, or as a CAND_ADD with i = 0 or S = 0.
108 Candidate records are allocated from an obstack. They are addressed
109 both from a hash table keyed on S1, and from a vector of candidate
110 pointers arranged in predominator order.
114 Currently we don't recognize:
119 as a strength reduction opportunity, even though this S1 would
120 also be replaceable by the S1' above. This can be added if it
121 comes up in practice.
123 Strength reduction in addressing
124 --------------------------------
125 There is another kind of candidate known as CAND_REF. A CAND_REF
126 describes a statement containing a memory reference having
127 complex addressing that might benefit from strength reduction.
128 Specifically, we are interested in references for which
129 get_inner_reference returns a base address, offset, and bitpos as
132 base: MEM_REF (T1, C1)
133 offset: MULT_EXPR (PLUS_EXPR (T2, C2), C3)
134 bitpos: C4 * BITS_PER_UNIT
136 Here T1 and T2 are arbitrary trees, and C1, C2, C3, C4 are
137 arbitrary integer constants. Note that C2 may be zero, in which
138 case the offset will be MULT_EXPR (T2, C3).
140 When this pattern is recognized, the original memory reference
141 can be replaced with:
143 MEM_REF (POINTER_PLUS_EXPR (T1, MULT_EXPR (T2, C3)),
146 which distributes the multiply to allow constant folding. When
147 two or more addressing expressions can be represented by MEM_REFs
148 of this form, differing only in the constants C1, C2, and C4,
149 making this substitution produces more efficient addressing during
150 the RTL phases. When there are not at least two expressions with
151 the same values of T1, T2, and C3, there is nothing to be gained
154 Strength reduction of CAND_REFs uses the same infrastructure as
155 that used by CAND_MULTs and CAND_ADDs. We record T1 in the base (B)
156 field, MULT_EXPR (T2, C3) in the stride (S) field, and
157 C1 + (C2 * C3) + C4 in the index (i) field. A basis for a CAND_REF
158 is thus another CAND_REF with the same B and S values. When at
159 least two CAND_REFs are chained together using the basis relation,
160 each of them is replaced as above, resulting in improved code
161 generation for addressing.
163 Conditional candidates
164 ======================
166 Conditional candidates are best illustrated with an example.
167 Consider the code sequence:
170 (2) a_0 = x_0 * 5; MULT (B: x_0; i: 0; S: 5)
172 (3) x_1 = x_0 + 1; ADD (B: x_0, i: 1; S: 1)
173 (4) x_2 = PHI <x_0, x_1>; PHI (B: x_0, i: 0, S: 1)
174 (5) x_3 = x_2 + 1; ADD (B: x_2, i: 1, S: 1)
175 (6) a_1 = x_3 * 5; MULT (B: x_2, i: 1; S: 5)
177 Here strength reduction is complicated by the uncertain value of x_2.
178 A legitimate transformation is:
187 (4) [x_2 = PHI <x_0, x_1>;]
188 (4a) t_2 = PHI <a_0, t_1>;
192 where the bracketed instructions may go dead.
194 To recognize this opportunity, we have to observe that statement (6)
195 has a "hidden basis" (2). The hidden basis is unlike a normal basis
196 in that the statement and the hidden basis have different base SSA
197 names (x_2 and x_0, respectively). The relationship is established
198 when a statement's base name (x_2) is defined by a phi statement (4),
199 each argument of which (x_0, x_1) has an identical "derived base name."
200 If the argument is defined by a candidate (as x_1 is by (3)) that is a
201 CAND_ADD having a stride of 1, the derived base name of the argument is
202 the base name of the candidate (x_0). Otherwise, the argument itself
203 is its derived base name (as is the case with argument x_0).
205 The hidden basis for statement (6) is the nearest dominating candidate
206 whose base name is the derived base name (x_0) of the feeding phi (4),
207 and whose stride is identical to that of the statement. We can then
208 create the new "phi basis" (4a) and feeding adds along incoming arcs (3a),
209 allowing the final replacement of (6) by the strength-reduced (6r).
211 To facilitate this, a new kind of candidate (CAND_PHI) is introduced.
212 A CAND_PHI is not a candidate for replacement, but is maintained in the
213 candidate table to ease discovery of hidden bases. Any phi statement
214 whose arguments share a common derived base name is entered into the
215 table with the derived base name, an (arbitrary) index of zero, and a
216 stride of 1. A statement with a hidden basis can then be detected by
217 simply looking up its feeding phi definition in the candidate table,
218 extracting the derived base name, and searching for a basis in the
219 usual manner after substituting the derived base name.
221 Note that the transformation is only valid when the original phi and
222 the statements that define the phi's arguments are all at the same
223 position in the loop hierarchy. */
226 /* Index into the candidate vector, offset by 1. VECs are zero-based,
227 while cand_idx's are one-based, with zero indicating null. */
228 typedef unsigned cand_idx
;
230 /* The kind of candidate. */
241 /* The candidate statement S1. */
244 /* The base expression B: often an SSA name, but not always. */
250 /* The index constant i. */
253 /* The type of the candidate. This is normally the type of base_expr,
254 but casts may have occurred when combining feeding instructions.
255 A candidate can only be a basis for candidates of the same final type.
256 (For CAND_REFs, this is the type to be used for operand 1 of the
257 replacement MEM_REF.) */
260 /* The kind of candidate (CAND_MULT, etc.). */
263 /* Index of this candidate in the candidate vector. */
266 /* Index of the next candidate record for the same statement.
267 A statement may be useful in more than one way (e.g., due to
268 commutativity). So we can have multiple "interpretations"
270 cand_idx next_interp
;
272 /* Index of the basis statement S0, if any, in the candidate vector. */
275 /* First candidate for which this candidate is a basis, if one exists. */
278 /* Next candidate having the same basis as this one. */
281 /* If this is a conditional candidate, the CAND_PHI candidate
282 that defines the base SSA name B. */
285 /* Savings that can be expected from eliminating dead code if this
286 candidate is replaced. */
290 typedef struct slsr_cand_d slsr_cand
, *slsr_cand_t
;
291 typedef const struct slsr_cand_d
*const_slsr_cand_t
;
293 /* Pointers to candidates are chained together as part of a mapping
294 from base expressions to the candidates that use them. */
298 /* Base expression for the chain of candidates: often, but not
299 always, an SSA name. */
302 /* Pointer to a candidate. */
306 struct cand_chain_d
*next
;
310 typedef struct cand_chain_d cand_chain
, *cand_chain_t
;
311 typedef const struct cand_chain_d
*const_cand_chain_t
;
313 /* Information about a unique "increment" associated with candidates
314 having an SSA name for a stride. An increment is the difference
315 between the index of the candidate and the index of its basis,
316 i.e., (i - i') as discussed in the module commentary.
318 When we are not going to generate address arithmetic we treat
319 increments that differ only in sign as the same, allowing sharing
320 of the cost of initializers. The absolute value of the increment
321 is stored in the incr_info. */
325 /* The increment that relates a candidate to its basis. */
328 /* How many times the increment occurs in the candidate tree. */
331 /* Cost of replacing candidates using this increment. Negative and
332 zero costs indicate replacement should be performed. */
335 /* If this increment is profitable but is not -1, 0, or 1, it requires
336 an initializer T_0 = stride * incr to be found or introduced in the
337 nearest common dominator of all candidates. This field holds T_0
338 for subsequent use. */
341 /* If the initializer was found to already exist, this is the block
342 where it was found. */
346 typedef struct incr_info_d incr_info
, *incr_info_t
;
348 /* Candidates are maintained in a vector. If candidate X dominates
349 candidate Y, then X appears before Y in the vector; but the
350 converse does not necessarily hold. */
351 static vec
<slsr_cand_t
> cand_vec
;
365 enum phi_adjust_status
371 enum count_phis_status
377 /* Pointer map embodying a mapping from statements to candidates. */
378 static hash_map
<gimple
*, slsr_cand_t
> *stmt_cand_map
;
380 /* Obstack for candidates. */
381 static struct obstack cand_obstack
;
383 /* Obstack for candidate chains. */
384 static struct obstack chain_obstack
;
386 /* An array INCR_VEC of incr_infos is used during analysis of related
387 candidates having an SSA name for a stride. INCR_VEC_LEN describes
388 its current length. MAX_INCR_VEC_LEN is used to avoid costly
389 pathological cases. */
390 static incr_info_t incr_vec
;
391 static unsigned incr_vec_len
;
392 const int MAX_INCR_VEC_LEN
= 16;
394 /* For a chain of candidates with unknown stride, indicates whether or not
395 we must generate pointer arithmetic when replacing statements. */
396 static bool address_arithmetic_p
;
398 /* Forward function declarations. */
399 static slsr_cand_t
base_cand_from_table (tree
);
400 static tree
introduce_cast_before_cand (slsr_cand_t
, tree
, tree
);
401 static bool legal_cast_p_1 (tree
, tree
);
403 /* Produce a pointer to the IDX'th candidate in the candidate vector. */
406 lookup_cand (cand_idx idx
)
408 return cand_vec
[idx
- 1];
411 /* Helper for hashing a candidate chain header. */
413 struct cand_chain_hasher
: nofree_ptr_hash
<cand_chain
>
415 static inline hashval_t
hash (const cand_chain
*);
416 static inline bool equal (const cand_chain
*, const cand_chain
*);
420 cand_chain_hasher::hash (const cand_chain
*p
)
422 tree base_expr
= p
->base_expr
;
423 return iterative_hash_expr (base_expr
, 0);
427 cand_chain_hasher::equal (const cand_chain
*chain1
, const cand_chain
*chain2
)
429 return operand_equal_p (chain1
->base_expr
, chain2
->base_expr
, 0);
432 /* Hash table embodying a mapping from base exprs to chains of candidates. */
433 static hash_table
<cand_chain_hasher
> *base_cand_map
;
435 /* Pointer map used by tree_to_aff_combination_expand. */
436 static hash_map
<tree
, name_expansion
*> *name_expansions
;
437 /* Pointer map embodying a mapping from bases to alternative bases. */
438 static hash_map
<tree
, tree
> *alt_base_map
;
440 /* Given BASE, use the tree affine combiniation facilities to
441 find the underlying tree expression for BASE, with any
442 immediate offset excluded.
444 N.B. we should eliminate this backtracking with better forward
445 analysis in a future release. */
448 get_alternative_base (tree base
)
450 tree
*result
= alt_base_map
->get (base
);
457 tree_to_aff_combination_expand (base
, TREE_TYPE (base
),
458 &aff
, &name_expansions
);
460 expr
= aff_combination_to_tree (&aff
);
462 gcc_assert (!alt_base_map
->put (base
, base
== expr
? NULL
: expr
));
464 return expr
== base
? NULL
: expr
;
470 /* Look in the candidate table for a CAND_PHI that defines BASE and
471 return it if found; otherwise return NULL. */
474 find_phi_def (tree base
)
478 if (TREE_CODE (base
) != SSA_NAME
)
481 c
= base_cand_from_table (base
);
483 if (!c
|| c
->kind
!= CAND_PHI
)
489 /* Helper routine for find_basis_for_candidate. May be called twice:
490 once for the candidate's base expr, and optionally again either for
491 the candidate's phi definition or for a CAND_REF's alternative base
495 find_basis_for_base_expr (slsr_cand_t c
, tree base_expr
)
497 cand_chain mapping_key
;
499 slsr_cand_t basis
= NULL
;
501 // Limit potential of N^2 behavior for long candidate chains.
503 int max_iters
= PARAM_VALUE (PARAM_MAX_SLSR_CANDIDATE_SCAN
);
505 mapping_key
.base_expr
= base_expr
;
506 chain
= base_cand_map
->find (&mapping_key
);
508 for (; chain
&& iters
< max_iters
; chain
= chain
->next
, ++iters
)
510 slsr_cand_t one_basis
= chain
->cand
;
512 if (one_basis
->kind
!= c
->kind
513 || one_basis
->cand_stmt
== c
->cand_stmt
514 || !operand_equal_p (one_basis
->stride
, c
->stride
, 0)
515 || !types_compatible_p (one_basis
->cand_type
, c
->cand_type
)
516 || !dominated_by_p (CDI_DOMINATORS
,
517 gimple_bb (c
->cand_stmt
),
518 gimple_bb (one_basis
->cand_stmt
)))
521 if (!basis
|| basis
->cand_num
< one_basis
->cand_num
)
528 /* Use the base expr from candidate C to look for possible candidates
529 that can serve as a basis for C. Each potential basis must also
530 appear in a block that dominates the candidate statement and have
531 the same stride and type. If more than one possible basis exists,
532 the one with highest index in the vector is chosen; this will be
533 the most immediately dominating basis. */
536 find_basis_for_candidate (slsr_cand_t c
)
538 slsr_cand_t basis
= find_basis_for_base_expr (c
, c
->base_expr
);
540 /* If a candidate doesn't have a basis using its base expression,
541 it may have a basis hidden by one or more intervening phis. */
542 if (!basis
&& c
->def_phi
)
544 basic_block basis_bb
, phi_bb
;
545 slsr_cand_t phi_cand
= lookup_cand (c
->def_phi
);
546 basis
= find_basis_for_base_expr (c
, phi_cand
->base_expr
);
550 /* A hidden basis must dominate the phi-definition of the
551 candidate's base name. */
552 phi_bb
= gimple_bb (phi_cand
->cand_stmt
);
553 basis_bb
= gimple_bb (basis
->cand_stmt
);
555 if (phi_bb
== basis_bb
556 || !dominated_by_p (CDI_DOMINATORS
, phi_bb
, basis_bb
))
562 /* If we found a hidden basis, estimate additional dead-code
563 savings if the phi and its feeding statements can be removed. */
564 if (basis
&& has_single_use (gimple_phi_result (phi_cand
->cand_stmt
)))
565 c
->dead_savings
+= phi_cand
->dead_savings
;
569 if (flag_expensive_optimizations
&& !basis
&& c
->kind
== CAND_REF
)
571 tree alt_base_expr
= get_alternative_base (c
->base_expr
);
573 basis
= find_basis_for_base_expr (c
, alt_base_expr
);
578 c
->sibling
= basis
->dependent
;
579 basis
->dependent
= c
->cand_num
;
580 return basis
->cand_num
;
586 /* Record a mapping from BASE to C, indicating that C may potentially serve
587 as a basis using that base expression. BASE may be the same as
588 C->BASE_EXPR; alternatively BASE can be a different tree that share the
589 underlining expression of C->BASE_EXPR. */
592 record_potential_basis (slsr_cand_t c
, tree base
)
599 node
= (cand_chain_t
) obstack_alloc (&chain_obstack
, sizeof (cand_chain
));
600 node
->base_expr
= base
;
603 slot
= base_cand_map
->find_slot (node
, INSERT
);
607 cand_chain_t head
= (cand_chain_t
) (*slot
);
608 node
->next
= head
->next
;
615 /* Allocate storage for a new candidate and initialize its fields.
616 Attempt to find a basis for the candidate.
618 For CAND_REF, an alternative base may also be recorded and used
619 to find a basis. This helps cases where the expression hidden
620 behind BASE (which is usually an SSA_NAME) has immediate offset,
624 a2[i + 20][j] = 2; */
627 alloc_cand_and_find_basis (enum cand_kind kind
, gimple
*gs
, tree base
,
628 const widest_int
&index
, tree stride
, tree ctype
,
631 slsr_cand_t c
= (slsr_cand_t
) obstack_alloc (&cand_obstack
,
637 c
->cand_type
= ctype
;
639 c
->cand_num
= cand_vec
.length () + 1;
643 c
->def_phi
= kind
== CAND_MULT
? find_phi_def (base
) : 0;
644 c
->dead_savings
= savings
;
646 cand_vec
.safe_push (c
);
648 if (kind
== CAND_PHI
)
651 c
->basis
= find_basis_for_candidate (c
);
653 record_potential_basis (c
, base
);
654 if (flag_expensive_optimizations
&& kind
== CAND_REF
)
656 tree alt_base
= get_alternative_base (base
);
658 record_potential_basis (c
, alt_base
);
664 /* Determine the target cost of statement GS when compiling according
668 stmt_cost (gimple
*gs
, bool speed
)
670 tree lhs
, rhs1
, rhs2
;
671 machine_mode lhs_mode
;
673 gcc_assert (is_gimple_assign (gs
));
674 lhs
= gimple_assign_lhs (gs
);
675 rhs1
= gimple_assign_rhs1 (gs
);
676 lhs_mode
= TYPE_MODE (TREE_TYPE (lhs
));
678 switch (gimple_assign_rhs_code (gs
))
681 rhs2
= gimple_assign_rhs2 (gs
);
683 if (tree_fits_shwi_p (rhs2
))
684 return mult_by_coeff_cost (tree_to_shwi (rhs2
), lhs_mode
, speed
);
686 gcc_assert (TREE_CODE (rhs1
) != INTEGER_CST
);
687 return mul_cost (speed
, lhs_mode
);
690 case POINTER_PLUS_EXPR
:
692 return add_cost (speed
, lhs_mode
);
695 return neg_cost (speed
, lhs_mode
);
698 return convert_cost (lhs_mode
, TYPE_MODE (TREE_TYPE (rhs1
)), speed
);
700 /* Note that we don't assign costs to copies that in most cases
710 /* Look up the defining statement for BASE_IN and return a pointer
711 to its candidate in the candidate table, if any; otherwise NULL.
712 Only CAND_ADD and CAND_MULT candidates are returned. */
715 base_cand_from_table (tree base_in
)
719 gimple
*def
= SSA_NAME_DEF_STMT (base_in
);
721 return (slsr_cand_t
) NULL
;
723 result
= stmt_cand_map
->get (def
);
725 if (result
&& (*result
)->kind
!= CAND_REF
)
728 return (slsr_cand_t
) NULL
;
731 /* Add an entry to the statement-to-candidate mapping. */
734 add_cand_for_stmt (gimple
*gs
, slsr_cand_t c
)
736 gcc_assert (!stmt_cand_map
->put (gs
, c
));
739 /* Given PHI which contains a phi statement, determine whether it
740 satisfies all the requirements of a phi candidate. If so, create
741 a candidate. Note that a CAND_PHI never has a basis itself, but
742 is used to help find a basis for subsequent candidates. */
745 slsr_process_phi (gphi
*phi
, bool speed
)
748 tree arg0_base
= NULL_TREE
, base_type
;
750 struct loop
*cand_loop
= gimple_bb (phi
)->loop_father
;
751 unsigned savings
= 0;
753 /* A CAND_PHI requires each of its arguments to have the same
754 derived base name. (See the module header commentary for a
755 definition of derived base names.) Furthermore, all feeding
756 definitions must be in the same position in the loop hierarchy
759 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
761 slsr_cand_t arg_cand
;
762 tree arg
= gimple_phi_arg_def (phi
, i
);
763 tree derived_base_name
= NULL_TREE
;
764 gimple
*arg_stmt
= NULL
;
765 basic_block arg_bb
= NULL
;
767 if (TREE_CODE (arg
) != SSA_NAME
)
770 arg_cand
= base_cand_from_table (arg
);
774 while (arg_cand
->kind
!= CAND_ADD
&& arg_cand
->kind
!= CAND_PHI
)
776 if (!arg_cand
->next_interp
)
779 arg_cand
= lookup_cand (arg_cand
->next_interp
);
782 if (!integer_onep (arg_cand
->stride
))
785 derived_base_name
= arg_cand
->base_expr
;
786 arg_stmt
= arg_cand
->cand_stmt
;
787 arg_bb
= gimple_bb (arg_stmt
);
789 /* Gather potential dead code savings if the phi statement
790 can be removed later on. */
791 if (has_single_use (arg
))
793 if (gimple_code (arg_stmt
) == GIMPLE_PHI
)
794 savings
+= arg_cand
->dead_savings
;
796 savings
+= stmt_cost (arg_stmt
, speed
);
801 derived_base_name
= arg
;
803 if (SSA_NAME_IS_DEFAULT_DEF (arg
))
804 arg_bb
= single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun
));
806 gimple_bb (SSA_NAME_DEF_STMT (arg
));
809 if (!arg_bb
|| arg_bb
->loop_father
!= cand_loop
)
813 arg0_base
= derived_base_name
;
814 else if (!operand_equal_p (derived_base_name
, arg0_base
, 0))
818 /* Create the candidate. "alloc_cand_and_find_basis" is named
819 misleadingly for this case, as no basis will be sought for a
821 base_type
= TREE_TYPE (arg0_base
);
823 c
= alloc_cand_and_find_basis (CAND_PHI
, phi
, arg0_base
,
824 0, integer_one_node
, base_type
, savings
);
826 /* Add the candidate to the statement-candidate mapping. */
827 add_cand_for_stmt (phi
, c
);
830 /* Given PBASE which is a pointer to tree, look up the defining
831 statement for it and check whether the candidate is in the
834 X = B + (1 * S), S is integer constant
835 X = B + (i * S), S is integer one
837 If so, set PBASE to the candidate's base_expr and return double
839 Otherwise, just return double int zero. */
842 backtrace_base_for_ref (tree
*pbase
)
844 tree base_in
= *pbase
;
845 slsr_cand_t base_cand
;
847 STRIP_NOPS (base_in
);
849 /* Strip off widening conversion(s) to handle cases where
850 e.g. 'B' is widened from an 'int' in order to calculate
852 if (CONVERT_EXPR_P (base_in
)
853 && legal_cast_p_1 (base_in
, TREE_OPERAND (base_in
, 0)))
854 base_in
= get_unwidened (base_in
, NULL_TREE
);
856 if (TREE_CODE (base_in
) != SSA_NAME
)
859 base_cand
= base_cand_from_table (base_in
);
861 while (base_cand
&& base_cand
->kind
!= CAND_PHI
)
863 if (base_cand
->kind
== CAND_ADD
864 && base_cand
->index
== 1
865 && TREE_CODE (base_cand
->stride
) == INTEGER_CST
)
867 /* X = B + (1 * S), S is integer constant. */
868 *pbase
= base_cand
->base_expr
;
869 return wi::to_widest (base_cand
->stride
);
871 else if (base_cand
->kind
== CAND_ADD
872 && TREE_CODE (base_cand
->stride
) == INTEGER_CST
873 && integer_onep (base_cand
->stride
))
875 /* X = B + (i * S), S is integer one. */
876 *pbase
= base_cand
->base_expr
;
877 return base_cand
->index
;
880 if (base_cand
->next_interp
)
881 base_cand
= lookup_cand (base_cand
->next_interp
);
889 /* Look for the following pattern:
891 *PBASE: MEM_REF (T1, C1)
893 *POFFSET: MULT_EXPR (T2, C3) [C2 is zero]
895 MULT_EXPR (PLUS_EXPR (T2, C2), C3)
897 MULT_EXPR (MINUS_EXPR (T2, -C2), C3)
899 *PINDEX: C4 * BITS_PER_UNIT
901 If not present, leave the input values unchanged and return FALSE.
902 Otherwise, modify the input values as follows and return TRUE:
905 *POFFSET: MULT_EXPR (T2, C3)
906 *PINDEX: C1 + (C2 * C3) + C4
908 When T2 is recorded by a CAND_ADD in the form of (T2' + C5), it
909 will be further restructured to:
912 *POFFSET: MULT_EXPR (T2', C3)
913 *PINDEX: C1 + (C2 * C3) + C4 + (C5 * C3) */
916 restructure_reference (tree
*pbase
, tree
*poffset
, widest_int
*pindex
,
919 tree base
= *pbase
, offset
= *poffset
;
920 widest_int index
= *pindex
;
921 tree mult_op0
, t1
, t2
, type
;
922 widest_int c1
, c2
, c3
, c4
, c5
;
926 || TREE_CODE (base
) != MEM_REF
927 || TREE_CODE (offset
) != MULT_EXPR
928 || TREE_CODE (TREE_OPERAND (offset
, 1)) != INTEGER_CST
929 || wi::umod_floor (index
, BITS_PER_UNIT
) != 0)
932 t1
= TREE_OPERAND (base
, 0);
933 c1
= widest_int::from (mem_ref_offset (base
), SIGNED
);
934 type
= TREE_TYPE (TREE_OPERAND (base
, 1));
936 mult_op0
= TREE_OPERAND (offset
, 0);
937 c3
= wi::to_widest (TREE_OPERAND (offset
, 1));
939 if (TREE_CODE (mult_op0
) == PLUS_EXPR
)
941 if (TREE_CODE (TREE_OPERAND (mult_op0
, 1)) == INTEGER_CST
)
943 t2
= TREE_OPERAND (mult_op0
, 0);
944 c2
= wi::to_widest (TREE_OPERAND (mult_op0
, 1));
949 else if (TREE_CODE (mult_op0
) == MINUS_EXPR
)
951 if (TREE_CODE (TREE_OPERAND (mult_op0
, 1)) == INTEGER_CST
)
953 t2
= TREE_OPERAND (mult_op0
, 0);
954 c2
= -wi::to_widest (TREE_OPERAND (mult_op0
, 1));
965 c4
= wi::lrshift (index
, LOG2_BITS_PER_UNIT
);
966 c5
= backtrace_base_for_ref (&t2
);
969 *poffset
= fold_build2 (MULT_EXPR
, sizetype
, fold_convert (sizetype
, t2
),
970 wide_int_to_tree (sizetype
, c3
));
971 *pindex
= c1
+ c2
* c3
+ c4
+ c5
* c3
;
977 /* Given GS which contains a data reference, create a CAND_REF entry in
978 the candidate table and attempt to find a basis. */
981 slsr_process_ref (gimple
*gs
)
983 tree ref_expr
, base
, offset
, type
;
984 HOST_WIDE_INT bitsize
, bitpos
;
986 int unsignedp
, volatilep
;
989 if (gimple_vdef (gs
))
990 ref_expr
= gimple_assign_lhs (gs
);
992 ref_expr
= gimple_assign_rhs1 (gs
);
994 if (!handled_component_p (ref_expr
)
995 || TREE_CODE (ref_expr
) == BIT_FIELD_REF
996 || (TREE_CODE (ref_expr
) == COMPONENT_REF
997 && DECL_BIT_FIELD (TREE_OPERAND (ref_expr
, 1))))
1000 base
= get_inner_reference (ref_expr
, &bitsize
, &bitpos
, &offset
, &mode
,
1001 &unsignedp
, &volatilep
, false);
1002 widest_int index
= bitpos
;
1004 if (!restructure_reference (&base
, &offset
, &index
, &type
))
1007 c
= alloc_cand_and_find_basis (CAND_REF
, gs
, base
, index
, offset
,
1010 /* Add the candidate to the statement-candidate mapping. */
1011 add_cand_for_stmt (gs
, c
);
1014 /* Create a candidate entry for a statement GS, where GS multiplies
1015 two SSA names BASE_IN and STRIDE_IN. Propagate any known information
1016 about the two SSA names into the new candidate. Return the new
1020 create_mul_ssa_cand (gimple
*gs
, tree base_in
, tree stride_in
, bool speed
)
1022 tree base
= NULL_TREE
, stride
= NULL_TREE
, ctype
= NULL_TREE
;
1024 unsigned savings
= 0;
1026 slsr_cand_t base_cand
= base_cand_from_table (base_in
);
1028 /* Look at all interpretations of the base candidate, if necessary,
1029 to find information to propagate into this candidate. */
1030 while (base_cand
&& !base
&& base_cand
->kind
!= CAND_PHI
)
1033 if (base_cand
->kind
== CAND_MULT
&& integer_onep (base_cand
->stride
))
1039 base
= base_cand
->base_expr
;
1040 index
= base_cand
->index
;
1042 ctype
= base_cand
->cand_type
;
1043 if (has_single_use (base_in
))
1044 savings
= (base_cand
->dead_savings
1045 + stmt_cost (base_cand
->cand_stmt
, speed
));
1047 else if (base_cand
->kind
== CAND_ADD
1048 && TREE_CODE (base_cand
->stride
) == INTEGER_CST
)
1050 /* Y = B + (i' * S), S constant
1052 ============================
1053 X = B + ((i' * S) * Z) */
1054 base
= base_cand
->base_expr
;
1055 index
= base_cand
->index
* wi::to_widest (base_cand
->stride
);
1057 ctype
= base_cand
->cand_type
;
1058 if (has_single_use (base_in
))
1059 savings
= (base_cand
->dead_savings
1060 + stmt_cost (base_cand
->cand_stmt
, speed
));
1063 if (base_cand
->next_interp
)
1064 base_cand
= lookup_cand (base_cand
->next_interp
);
1071 /* No interpretations had anything useful to propagate, so
1072 produce X = (Y + 0) * Z. */
1076 ctype
= TREE_TYPE (base_in
);
1079 c
= alloc_cand_and_find_basis (CAND_MULT
, gs
, base
, index
, stride
,
1084 /* Create a candidate entry for a statement GS, where GS multiplies
1085 SSA name BASE_IN by constant STRIDE_IN. Propagate any known
1086 information about BASE_IN into the new candidate. Return the new
1090 create_mul_imm_cand (gimple
*gs
, tree base_in
, tree stride_in
, bool speed
)
1092 tree base
= NULL_TREE
, stride
= NULL_TREE
, ctype
= NULL_TREE
;
1093 widest_int index
, temp
;
1094 unsigned savings
= 0;
1096 slsr_cand_t base_cand
= base_cand_from_table (base_in
);
1098 /* Look at all interpretations of the base candidate, if necessary,
1099 to find information to propagate into this candidate. */
1100 while (base_cand
&& !base
&& base_cand
->kind
!= CAND_PHI
)
1102 if (base_cand
->kind
== CAND_MULT
1103 && TREE_CODE (base_cand
->stride
) == INTEGER_CST
)
1105 /* Y = (B + i') * S, S constant
1107 ============================
1108 X = (B + i') * (S * c) */
1109 temp
= wi::to_widest (base_cand
->stride
) * wi::to_widest (stride_in
);
1110 if (wi::fits_to_tree_p (temp
, TREE_TYPE (stride_in
)))
1112 base
= base_cand
->base_expr
;
1113 index
= base_cand
->index
;
1114 stride
= wide_int_to_tree (TREE_TYPE (stride_in
), temp
);
1115 ctype
= base_cand
->cand_type
;
1116 if (has_single_use (base_in
))
1117 savings
= (base_cand
->dead_savings
1118 + stmt_cost (base_cand
->cand_stmt
, speed
));
1121 else if (base_cand
->kind
== CAND_ADD
&& integer_onep (base_cand
->stride
))
1125 ===========================
1127 base
= base_cand
->base_expr
;
1128 index
= base_cand
->index
;
1130 ctype
= base_cand
->cand_type
;
1131 if (has_single_use (base_in
))
1132 savings
= (base_cand
->dead_savings
1133 + stmt_cost (base_cand
->cand_stmt
, speed
));
1135 else if (base_cand
->kind
== CAND_ADD
1136 && base_cand
->index
== 1
1137 && TREE_CODE (base_cand
->stride
) == INTEGER_CST
)
1139 /* Y = B + (1 * S), S constant
1141 ===========================
1143 base
= base_cand
->base_expr
;
1144 index
= wi::to_widest (base_cand
->stride
);
1146 ctype
= base_cand
->cand_type
;
1147 if (has_single_use (base_in
))
1148 savings
= (base_cand
->dead_savings
1149 + stmt_cost (base_cand
->cand_stmt
, speed
));
1152 if (base_cand
->next_interp
)
1153 base_cand
= lookup_cand (base_cand
->next_interp
);
1160 /* No interpretations had anything useful to propagate, so
1161 produce X = (Y + 0) * c. */
1165 ctype
= TREE_TYPE (base_in
);
1168 c
= alloc_cand_and_find_basis (CAND_MULT
, gs
, base
, index
, stride
,
1173 /* Given GS which is a multiply of scalar integers, make an appropriate
1174 entry in the candidate table. If this is a multiply of two SSA names,
1175 create two CAND_MULT interpretations and attempt to find a basis for
1176 each of them. Otherwise, create a single CAND_MULT and attempt to
1180 slsr_process_mul (gimple
*gs
, tree rhs1
, tree rhs2
, bool speed
)
1184 /* If this is a multiply of an SSA name with itself, it is highly
1185 unlikely that we will get a strength reduction opportunity, so
1186 don't record it as a candidate. This simplifies the logic for
1187 finding a basis, so if this is removed that must be considered. */
1191 if (TREE_CODE (rhs2
) == SSA_NAME
)
1193 /* Record an interpretation of this statement in the candidate table
1194 assuming RHS1 is the base expression and RHS2 is the stride. */
1195 c
= create_mul_ssa_cand (gs
, rhs1
, rhs2
, speed
);
1197 /* Add the first interpretation to the statement-candidate mapping. */
1198 add_cand_for_stmt (gs
, c
);
1200 /* Record another interpretation of this statement assuming RHS1
1201 is the stride and RHS2 is the base expression. */
1202 c2
= create_mul_ssa_cand (gs
, rhs2
, rhs1
, speed
);
1203 c
->next_interp
= c2
->cand_num
;
1207 /* Record an interpretation for the multiply-immediate. */
1208 c
= create_mul_imm_cand (gs
, rhs1
, rhs2
, speed
);
1210 /* Add the interpretation to the statement-candidate mapping. */
1211 add_cand_for_stmt (gs
, c
);
1215 /* Create a candidate entry for a statement GS, where GS adds two
1216 SSA names BASE_IN and ADDEND_IN if SUBTRACT_P is false, and
1217 subtracts ADDEND_IN from BASE_IN otherwise. Propagate any known
1218 information about the two SSA names into the new candidate.
1219 Return the new candidate. */
1222 create_add_ssa_cand (gimple
*gs
, tree base_in
, tree addend_in
,
1223 bool subtract_p
, bool speed
)
1225 tree base
= NULL_TREE
, stride
= NULL_TREE
, ctype
= NULL
;
1227 unsigned savings
= 0;
1229 slsr_cand_t base_cand
= base_cand_from_table (base_in
);
1230 slsr_cand_t addend_cand
= base_cand_from_table (addend_in
);
1232 /* The most useful transformation is a multiply-immediate feeding
1233 an add or subtract. Look for that first. */
1234 while (addend_cand
&& !base
&& addend_cand
->kind
!= CAND_PHI
)
1236 if (addend_cand
->kind
== CAND_MULT
1237 && addend_cand
->index
== 0
1238 && TREE_CODE (addend_cand
->stride
) == INTEGER_CST
)
1240 /* Z = (B + 0) * S, S constant
1242 ===========================
1243 X = Y + ((+/-1 * S) * B) */
1245 index
= wi::to_widest (addend_cand
->stride
);
1248 stride
= addend_cand
->base_expr
;
1249 ctype
= TREE_TYPE (base_in
);
1250 if (has_single_use (addend_in
))
1251 savings
= (addend_cand
->dead_savings
1252 + stmt_cost (addend_cand
->cand_stmt
, speed
));
1255 if (addend_cand
->next_interp
)
1256 addend_cand
= lookup_cand (addend_cand
->next_interp
);
1261 while (base_cand
&& !base
&& base_cand
->kind
!= CAND_PHI
)
1263 if (base_cand
->kind
== CAND_ADD
1264 && (base_cand
->index
== 0
1265 || operand_equal_p (base_cand
->stride
,
1266 integer_zero_node
, 0)))
1268 /* Y = B + (i' * S), i' * S = 0
1270 ============================
1271 X = B + (+/-1 * Z) */
1272 base
= base_cand
->base_expr
;
1273 index
= subtract_p
? -1 : 1;
1275 ctype
= base_cand
->cand_type
;
1276 if (has_single_use (base_in
))
1277 savings
= (base_cand
->dead_savings
1278 + stmt_cost (base_cand
->cand_stmt
, speed
));
1280 else if (subtract_p
)
1282 slsr_cand_t subtrahend_cand
= base_cand_from_table (addend_in
);
1284 while (subtrahend_cand
&& !base
&& subtrahend_cand
->kind
!= CAND_PHI
)
1286 if (subtrahend_cand
->kind
== CAND_MULT
1287 && subtrahend_cand
->index
== 0
1288 && TREE_CODE (subtrahend_cand
->stride
) == INTEGER_CST
)
1290 /* Z = (B + 0) * S, S constant
1292 ===========================
1293 Value: X = Y + ((-1 * S) * B) */
1295 index
= wi::to_widest (subtrahend_cand
->stride
);
1297 stride
= subtrahend_cand
->base_expr
;
1298 ctype
= TREE_TYPE (base_in
);
1299 if (has_single_use (addend_in
))
1300 savings
= (subtrahend_cand
->dead_savings
1301 + stmt_cost (subtrahend_cand
->cand_stmt
, speed
));
1304 if (subtrahend_cand
->next_interp
)
1305 subtrahend_cand
= lookup_cand (subtrahend_cand
->next_interp
);
1307 subtrahend_cand
= NULL
;
1311 if (base_cand
->next_interp
)
1312 base_cand
= lookup_cand (base_cand
->next_interp
);
1319 /* No interpretations had anything useful to propagate, so
1320 produce X = Y + (1 * Z). */
1322 index
= subtract_p
? -1 : 1;
1324 ctype
= TREE_TYPE (base_in
);
1327 c
= alloc_cand_and_find_basis (CAND_ADD
, gs
, base
, index
, stride
,
1332 /* Create a candidate entry for a statement GS, where GS adds SSA
1333 name BASE_IN to constant INDEX_IN. Propagate any known information
1334 about BASE_IN into the new candidate. Return the new candidate. */
1337 create_add_imm_cand (gimple
*gs
, tree base_in
, const widest_int
&index_in
,
1340 enum cand_kind kind
= CAND_ADD
;
1341 tree base
= NULL_TREE
, stride
= NULL_TREE
, ctype
= NULL_TREE
;
1342 widest_int index
, multiple
;
1343 unsigned savings
= 0;
1345 slsr_cand_t base_cand
= base_cand_from_table (base_in
);
1347 while (base_cand
&& !base
&& base_cand
->kind
!= CAND_PHI
)
1349 signop sign
= TYPE_SIGN (TREE_TYPE (base_cand
->stride
));
1351 if (TREE_CODE (base_cand
->stride
) == INTEGER_CST
1352 && wi::multiple_of_p (index_in
, wi::to_widest (base_cand
->stride
),
1355 /* Y = (B + i') * S, S constant, c = kS for some integer k
1357 ============================
1358 X = (B + (i'+ k)) * S
1360 Y = B + (i' * S), S constant, c = kS for some integer k
1362 ============================
1363 X = (B + (i'+ k)) * S */
1364 kind
= base_cand
->kind
;
1365 base
= base_cand
->base_expr
;
1366 index
= base_cand
->index
+ multiple
;
1367 stride
= base_cand
->stride
;
1368 ctype
= base_cand
->cand_type
;
1369 if (has_single_use (base_in
))
1370 savings
= (base_cand
->dead_savings
1371 + stmt_cost (base_cand
->cand_stmt
, speed
));
1374 if (base_cand
->next_interp
)
1375 base_cand
= lookup_cand (base_cand
->next_interp
);
1382 /* No interpretations had anything useful to propagate, so
1383 produce X = Y + (c * 1). */
1387 stride
= integer_one_node
;
1388 ctype
= TREE_TYPE (base_in
);
1391 c
= alloc_cand_and_find_basis (kind
, gs
, base
, index
, stride
,
1396 /* Given GS which is an add or subtract of scalar integers or pointers,
1397 make at least one appropriate entry in the candidate table. */
1400 slsr_process_add (gimple
*gs
, tree rhs1
, tree rhs2
, bool speed
)
1402 bool subtract_p
= gimple_assign_rhs_code (gs
) == MINUS_EXPR
;
1403 slsr_cand_t c
= NULL
, c2
;
1405 if (TREE_CODE (rhs2
) == SSA_NAME
)
1407 /* First record an interpretation assuming RHS1 is the base expression
1408 and RHS2 is the stride. But it doesn't make sense for the
1409 stride to be a pointer, so don't record a candidate in that case. */
1410 if (!POINTER_TYPE_P (TREE_TYPE (rhs2
)))
1412 c
= create_add_ssa_cand (gs
, rhs1
, rhs2
, subtract_p
, speed
);
1414 /* Add the first interpretation to the statement-candidate
1416 add_cand_for_stmt (gs
, c
);
1419 /* If the two RHS operands are identical, or this is a subtract,
1421 if (operand_equal_p (rhs1
, rhs2
, 0) || subtract_p
)
1424 /* Otherwise, record another interpretation assuming RHS2 is the
1425 base expression and RHS1 is the stride, again provided that the
1426 stride is not a pointer. */
1427 if (!POINTER_TYPE_P (TREE_TYPE (rhs1
)))
1429 c2
= create_add_ssa_cand (gs
, rhs2
, rhs1
, false, speed
);
1431 c
->next_interp
= c2
->cand_num
;
1433 add_cand_for_stmt (gs
, c2
);
1438 /* Record an interpretation for the add-immediate. */
1439 widest_int index
= wi::to_widest (rhs2
);
1443 c
= create_add_imm_cand (gs
, rhs1
, index
, speed
);
1445 /* Add the interpretation to the statement-candidate mapping. */
1446 add_cand_for_stmt (gs
, c
);
1450 /* Given GS which is a negate of a scalar integer, make an appropriate
1451 entry in the candidate table. A negate is equivalent to a multiply
1455 slsr_process_neg (gimple
*gs
, tree rhs1
, bool speed
)
1457 /* Record a CAND_MULT interpretation for the multiply by -1. */
1458 slsr_cand_t c
= create_mul_imm_cand (gs
, rhs1
, integer_minus_one_node
, speed
);
1460 /* Add the interpretation to the statement-candidate mapping. */
1461 add_cand_for_stmt (gs
, c
);
1464 /* Help function for legal_cast_p, operating on two trees. Checks
1465 whether it's allowable to cast from RHS to LHS. See legal_cast_p
1466 for more details. */
1469 legal_cast_p_1 (tree lhs
, tree rhs
)
1471 tree lhs_type
, rhs_type
;
1472 unsigned lhs_size
, rhs_size
;
1473 bool lhs_wraps
, rhs_wraps
;
1475 lhs_type
= TREE_TYPE (lhs
);
1476 rhs_type
= TREE_TYPE (rhs
);
1477 lhs_size
= TYPE_PRECISION (lhs_type
);
1478 rhs_size
= TYPE_PRECISION (rhs_type
);
1479 lhs_wraps
= ANY_INTEGRAL_TYPE_P (lhs_type
) && TYPE_OVERFLOW_WRAPS (lhs_type
);
1480 rhs_wraps
= ANY_INTEGRAL_TYPE_P (rhs_type
) && TYPE_OVERFLOW_WRAPS (rhs_type
);
1482 if (lhs_size
< rhs_size
1483 || (rhs_wraps
&& !lhs_wraps
)
1484 || (rhs_wraps
&& lhs_wraps
&& rhs_size
!= lhs_size
))
1490 /* Return TRUE if GS is a statement that defines an SSA name from
1491 a conversion and is legal for us to combine with an add and multiply
1492 in the candidate table. For example, suppose we have:
1498 Without the type-cast, we would create a CAND_MULT for D with base B,
1499 index i, and stride S. We want to record this candidate only if it
1500 is equivalent to apply the type cast following the multiply:
1506 We will record the type with the candidate for D. This allows us
1507 to use a similar previous candidate as a basis. If we have earlier seen
1513 we can replace D with
1515 D = D' + (i - i') * S;
1517 But if moving the type-cast would change semantics, we mustn't do this.
1519 This is legitimate for casts from a non-wrapping integral type to
1520 any integral type of the same or larger size. It is not legitimate
1521 to convert a wrapping type to a non-wrapping type, or to a wrapping
1522 type of a different size. I.e., with a wrapping type, we must
1523 assume that the addition B + i could wrap, in which case performing
1524 the multiply before or after one of the "illegal" type casts will
1525 have different semantics. */
1528 legal_cast_p (gimple
*gs
, tree rhs
)
1530 if (!is_gimple_assign (gs
)
1531 || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (gs
)))
1534 return legal_cast_p_1 (gimple_assign_lhs (gs
), rhs
);
1537 /* Given GS which is a cast to a scalar integer type, determine whether
1538 the cast is legal for strength reduction. If so, make at least one
1539 appropriate entry in the candidate table. */
1542 slsr_process_cast (gimple
*gs
, tree rhs1
, bool speed
)
1545 slsr_cand_t base_cand
, c
, c2
;
1546 unsigned savings
= 0;
1548 if (!legal_cast_p (gs
, rhs1
))
1551 lhs
= gimple_assign_lhs (gs
);
1552 base_cand
= base_cand_from_table (rhs1
);
1553 ctype
= TREE_TYPE (lhs
);
1555 if (base_cand
&& base_cand
->kind
!= CAND_PHI
)
1559 /* Propagate all data from the base candidate except the type,
1560 which comes from the cast, and the base candidate's cast,
1561 which is no longer applicable. */
1562 if (has_single_use (rhs1
))
1563 savings
= (base_cand
->dead_savings
1564 + stmt_cost (base_cand
->cand_stmt
, speed
));
1566 c
= alloc_cand_and_find_basis (base_cand
->kind
, gs
,
1567 base_cand
->base_expr
,
1568 base_cand
->index
, base_cand
->stride
,
1570 if (base_cand
->next_interp
)
1571 base_cand
= lookup_cand (base_cand
->next_interp
);
1578 /* If nothing is known about the RHS, create fresh CAND_ADD and
1579 CAND_MULT interpretations:
1584 The first of these is somewhat arbitrary, but the choice of
1585 1 for the stride simplifies the logic for propagating casts
1587 c
= alloc_cand_and_find_basis (CAND_ADD
, gs
, rhs1
,
1588 0, integer_one_node
, ctype
, 0);
1589 c2
= alloc_cand_and_find_basis (CAND_MULT
, gs
, rhs1
,
1590 0, integer_one_node
, ctype
, 0);
1591 c
->next_interp
= c2
->cand_num
;
1594 /* Add the first (or only) interpretation to the statement-candidate
1596 add_cand_for_stmt (gs
, c
);
1599 /* Given GS which is a copy of a scalar integer type, make at least one
1600 appropriate entry in the candidate table.
1602 This interface is included for completeness, but is unnecessary
1603 if this pass immediately follows a pass that performs copy
1604 propagation, such as DOM. */
1607 slsr_process_copy (gimple
*gs
, tree rhs1
, bool speed
)
1609 slsr_cand_t base_cand
, c
, c2
;
1610 unsigned savings
= 0;
1612 base_cand
= base_cand_from_table (rhs1
);
1614 if (base_cand
&& base_cand
->kind
!= CAND_PHI
)
1618 /* Propagate all data from the base candidate. */
1619 if (has_single_use (rhs1
))
1620 savings
= (base_cand
->dead_savings
1621 + stmt_cost (base_cand
->cand_stmt
, speed
));
1623 c
= alloc_cand_and_find_basis (base_cand
->kind
, gs
,
1624 base_cand
->base_expr
,
1625 base_cand
->index
, base_cand
->stride
,
1626 base_cand
->cand_type
, savings
);
1627 if (base_cand
->next_interp
)
1628 base_cand
= lookup_cand (base_cand
->next_interp
);
1635 /* If nothing is known about the RHS, create fresh CAND_ADD and
1636 CAND_MULT interpretations:
1641 The first of these is somewhat arbitrary, but the choice of
1642 1 for the stride simplifies the logic for propagating casts
1644 c
= alloc_cand_and_find_basis (CAND_ADD
, gs
, rhs1
,
1645 0, integer_one_node
, TREE_TYPE (rhs1
), 0);
1646 c2
= alloc_cand_and_find_basis (CAND_MULT
, gs
, rhs1
,
1647 0, integer_one_node
, TREE_TYPE (rhs1
), 0);
1648 c
->next_interp
= c2
->cand_num
;
1651 /* Add the first (or only) interpretation to the statement-candidate
1653 add_cand_for_stmt (gs
, c
);
1656 class find_candidates_dom_walker
: public dom_walker
1659 find_candidates_dom_walker (cdi_direction direction
)
1660 : dom_walker (direction
) {}
1661 virtual void before_dom_children (basic_block
);
1664 /* Find strength-reduction candidates in block BB. */
1667 find_candidates_dom_walker::before_dom_children (basic_block bb
)
1669 bool speed
= optimize_bb_for_speed_p (bb
);
1671 for (gphi_iterator gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
);
1673 slsr_process_phi (gsi
.phi (), speed
);
1675 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);
1678 gimple
*gs
= gsi_stmt (gsi
);
1680 if (gimple_vuse (gs
) && gimple_assign_single_p (gs
))
1681 slsr_process_ref (gs
);
1683 else if (is_gimple_assign (gs
)
1684 && SCALAR_INT_MODE_P
1685 (TYPE_MODE (TREE_TYPE (gimple_assign_lhs (gs
)))))
1687 tree rhs1
= NULL_TREE
, rhs2
= NULL_TREE
;
1689 switch (gimple_assign_rhs_code (gs
))
1693 rhs1
= gimple_assign_rhs1 (gs
);
1694 rhs2
= gimple_assign_rhs2 (gs
);
1695 /* Should never happen, but currently some buggy situations
1696 in earlier phases put constants in rhs1. */
1697 if (TREE_CODE (rhs1
) != SSA_NAME
)
1701 /* Possible future opportunity: rhs1 of a ptr+ can be
1703 case POINTER_PLUS_EXPR
:
1705 rhs2
= gimple_assign_rhs2 (gs
);
1711 rhs1
= gimple_assign_rhs1 (gs
);
1712 if (TREE_CODE (rhs1
) != SSA_NAME
)
1720 switch (gimple_assign_rhs_code (gs
))
1723 slsr_process_mul (gs
, rhs1
, rhs2
, speed
);
1727 case POINTER_PLUS_EXPR
:
1729 slsr_process_add (gs
, rhs1
, rhs2
, speed
);
1733 slsr_process_neg (gs
, rhs1
, speed
);
1737 slsr_process_cast (gs
, rhs1
, speed
);
1741 slsr_process_copy (gs
, rhs1
, speed
);
1751 /* Dump a candidate for debug. */
1754 dump_candidate (slsr_cand_t c
)
1756 fprintf (dump_file
, "%3d [%d] ", c
->cand_num
,
1757 gimple_bb (c
->cand_stmt
)->index
);
1758 print_gimple_stmt (dump_file
, c
->cand_stmt
, 0, 0);
1762 fputs (" MULT : (", dump_file
);
1763 print_generic_expr (dump_file
, c
->base_expr
, 0);
1764 fputs (" + ", dump_file
);
1765 print_decs (c
->index
, dump_file
);
1766 fputs (") * ", dump_file
);
1767 print_generic_expr (dump_file
, c
->stride
, 0);
1768 fputs (" : ", dump_file
);
1771 fputs (" ADD : ", dump_file
);
1772 print_generic_expr (dump_file
, c
->base_expr
, 0);
1773 fputs (" + (", dump_file
);
1774 print_decs (c
->index
, dump_file
);
1775 fputs (" * ", dump_file
);
1776 print_generic_expr (dump_file
, c
->stride
, 0);
1777 fputs (") : ", dump_file
);
1780 fputs (" REF : ", dump_file
);
1781 print_generic_expr (dump_file
, c
->base_expr
, 0);
1782 fputs (" + (", dump_file
);
1783 print_generic_expr (dump_file
, c
->stride
, 0);
1784 fputs (") + ", dump_file
);
1785 print_decs (c
->index
, dump_file
);
1786 fputs (" : ", dump_file
);
1789 fputs (" PHI : ", dump_file
);
1790 print_generic_expr (dump_file
, c
->base_expr
, 0);
1791 fputs (" + (unknown * ", dump_file
);
1792 print_generic_expr (dump_file
, c
->stride
, 0);
1793 fputs (") : ", dump_file
);
1798 print_generic_expr (dump_file
, c
->cand_type
, 0);
1799 fprintf (dump_file
, "\n basis: %d dependent: %d sibling: %d\n",
1800 c
->basis
, c
->dependent
, c
->sibling
);
1801 fprintf (dump_file
, " next-interp: %d dead-savings: %d\n",
1802 c
->next_interp
, c
->dead_savings
);
1804 fprintf (dump_file
, " phi: %d\n", c
->def_phi
);
1805 fputs ("\n", dump_file
);
1808 /* Dump the candidate vector for debug. */
1811 dump_cand_vec (void)
1816 fprintf (dump_file
, "\nStrength reduction candidate vector:\n\n");
1818 FOR_EACH_VEC_ELT (cand_vec
, i
, c
)
1822 /* Callback used to dump the candidate chains hash table. */
1825 ssa_base_cand_dump_callback (cand_chain
**slot
, void *ignored ATTRIBUTE_UNUSED
)
1827 const_cand_chain_t chain
= *slot
;
1830 print_generic_expr (dump_file
, chain
->base_expr
, 0);
1831 fprintf (dump_file
, " -> %d", chain
->cand
->cand_num
);
1833 for (p
= chain
->next
; p
; p
= p
->next
)
1834 fprintf (dump_file
, " -> %d", p
->cand
->cand_num
);
1836 fputs ("\n", dump_file
);
1840 /* Dump the candidate chains. */
1843 dump_cand_chains (void)
1845 fprintf (dump_file
, "\nStrength reduction candidate chains:\n\n");
1846 base_cand_map
->traverse_noresize
<void *, ssa_base_cand_dump_callback
>
1848 fputs ("\n", dump_file
);
1851 /* Dump the increment vector for debug. */
1854 dump_incr_vec (void)
1856 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1860 fprintf (dump_file
, "\nIncrement vector:\n\n");
1862 for (i
= 0; i
< incr_vec_len
; i
++)
1864 fprintf (dump_file
, "%3d increment: ", i
);
1865 print_decs (incr_vec
[i
].incr
, dump_file
);
1866 fprintf (dump_file
, "\n count: %d", incr_vec
[i
].count
);
1867 fprintf (dump_file
, "\n cost: %d", incr_vec
[i
].cost
);
1868 fputs ("\n initializer: ", dump_file
);
1869 print_generic_expr (dump_file
, incr_vec
[i
].initializer
, 0);
1870 fputs ("\n\n", dump_file
);
1875 /* Replace *EXPR in candidate C with an equivalent strength-reduced
1879 replace_ref (tree
*expr
, slsr_cand_t c
)
1881 tree add_expr
, mem_ref
, acc_type
= TREE_TYPE (*expr
);
1882 unsigned HOST_WIDE_INT misalign
;
1885 /* Ensure the memory reference carries the minimum alignment
1886 requirement for the data type. See PR58041. */
1887 get_object_alignment_1 (*expr
, &align
, &misalign
);
1889 align
= (misalign
& -misalign
);
1890 if (align
< TYPE_ALIGN (acc_type
))
1891 acc_type
= build_aligned_type (acc_type
, align
);
1893 add_expr
= fold_build2 (POINTER_PLUS_EXPR
, TREE_TYPE (c
->base_expr
),
1894 c
->base_expr
, c
->stride
);
1895 mem_ref
= fold_build2 (MEM_REF
, acc_type
, add_expr
,
1896 wide_int_to_tree (c
->cand_type
, c
->index
));
1898 /* Gimplify the base addressing expression for the new MEM_REF tree. */
1899 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
1900 TREE_OPERAND (mem_ref
, 0)
1901 = force_gimple_operand_gsi (&gsi
, TREE_OPERAND (mem_ref
, 0),
1902 /*simple_p=*/true, NULL
,
1903 /*before=*/true, GSI_SAME_STMT
);
1904 copy_ref_info (mem_ref
, *expr
);
1906 update_stmt (c
->cand_stmt
);
1909 /* Replace CAND_REF candidate C, each sibling of candidate C, and each
1910 dependent of candidate C with an equivalent strength-reduced data
1914 replace_refs (slsr_cand_t c
)
1916 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1918 fputs ("Replacing reference: ", dump_file
);
1919 print_gimple_stmt (dump_file
, c
->cand_stmt
, 0, 0);
1922 if (gimple_vdef (c
->cand_stmt
))
1924 tree
*lhs
= gimple_assign_lhs_ptr (c
->cand_stmt
);
1925 replace_ref (lhs
, c
);
1929 tree
*rhs
= gimple_assign_rhs1_ptr (c
->cand_stmt
);
1930 replace_ref (rhs
, c
);
1933 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1935 fputs ("With: ", dump_file
);
1936 print_gimple_stmt (dump_file
, c
->cand_stmt
, 0, 0);
1937 fputs ("\n", dump_file
);
1941 replace_refs (lookup_cand (c
->sibling
));
1944 replace_refs (lookup_cand (c
->dependent
));
1947 /* Return TRUE if candidate C is dependent upon a PHI. */
1950 phi_dependent_cand_p (slsr_cand_t c
)
1952 /* A candidate is not necessarily dependent upon a PHI just because
1953 it has a phi definition for its base name. It may have a basis
1954 that relies upon the same phi definition, in which case the PHI
1955 is irrelevant to this candidate. */
1958 && lookup_cand (c
->basis
)->def_phi
!= c
->def_phi
);
1961 /* Calculate the increment required for candidate C relative to
1965 cand_increment (slsr_cand_t c
)
1969 /* If the candidate doesn't have a basis, just return its own
1970 index. This is useful in record_increments to help us find
1971 an existing initializer. Also, if the candidate's basis is
1972 hidden by a phi, then its own index will be the increment
1973 from the newly introduced phi basis. */
1974 if (!c
->basis
|| phi_dependent_cand_p (c
))
1977 basis
= lookup_cand (c
->basis
);
1978 gcc_assert (operand_equal_p (c
->base_expr
, basis
->base_expr
, 0));
1979 return c
->index
- basis
->index
;
1982 /* Calculate the increment required for candidate C relative to
1983 its basis. If we aren't going to generate pointer arithmetic
1984 for this candidate, return the absolute value of that increment
1987 static inline widest_int
1988 cand_abs_increment (slsr_cand_t c
)
1990 widest_int increment
= cand_increment (c
);
1992 if (!address_arithmetic_p
&& wi::neg_p (increment
))
1993 increment
= -increment
;
1998 /* Return TRUE iff candidate C has already been replaced under
1999 another interpretation. */
2002 cand_already_replaced (slsr_cand_t c
)
2004 return (gimple_bb (c
->cand_stmt
) == 0);
2007 /* Common logic used by replace_unconditional_candidate and
2008 replace_conditional_candidate. */
2011 replace_mult_candidate (slsr_cand_t c
, tree basis_name
, widest_int bump
)
2013 tree target_type
= TREE_TYPE (gimple_assign_lhs (c
->cand_stmt
));
2014 enum tree_code cand_code
= gimple_assign_rhs_code (c
->cand_stmt
);
2016 /* It is highly unlikely, but possible, that the resulting
2017 bump doesn't fit in a HWI. Abandon the replacement
2018 in this case. This does not affect siblings or dependents
2019 of C. Restriction to signed HWI is conservative for unsigned
2020 types but allows for safe negation without twisted logic. */
2021 if (wi::fits_shwi_p (bump
)
2022 && bump
.to_shwi () != HOST_WIDE_INT_MIN
2023 /* It is not useful to replace casts, copies, or adds of
2024 an SSA name and a constant. */
2025 && cand_code
!= MODIFY_EXPR
2026 && !CONVERT_EXPR_CODE_P (cand_code
)
2027 && cand_code
!= PLUS_EXPR
2028 && cand_code
!= POINTER_PLUS_EXPR
2029 && cand_code
!= MINUS_EXPR
)
2031 enum tree_code code
= PLUS_EXPR
;
2033 gimple
*stmt_to_print
= NULL
;
2035 /* If the basis name and the candidate's LHS have incompatible
2036 types, introduce a cast. */
2037 if (!useless_type_conversion_p (target_type
, TREE_TYPE (basis_name
)))
2038 basis_name
= introduce_cast_before_cand (c
, target_type
, basis_name
);
2039 if (wi::neg_p (bump
))
2045 bump_tree
= wide_int_to_tree (target_type
, bump
);
2047 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2049 fputs ("Replacing: ", dump_file
);
2050 print_gimple_stmt (dump_file
, c
->cand_stmt
, 0, 0);
2055 tree lhs
= gimple_assign_lhs (c
->cand_stmt
);
2056 gassign
*copy_stmt
= gimple_build_assign (lhs
, basis_name
);
2057 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
2058 gimple_set_location (copy_stmt
, gimple_location (c
->cand_stmt
));
2059 gsi_replace (&gsi
, copy_stmt
, false);
2060 c
->cand_stmt
= copy_stmt
;
2061 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2062 stmt_to_print
= copy_stmt
;
2067 if (cand_code
!= NEGATE_EXPR
) {
2068 rhs1
= gimple_assign_rhs1 (c
->cand_stmt
);
2069 rhs2
= gimple_assign_rhs2 (c
->cand_stmt
);
2071 if (cand_code
!= NEGATE_EXPR
2072 && ((operand_equal_p (rhs1
, basis_name
, 0)
2073 && operand_equal_p (rhs2
, bump_tree
, 0))
2074 || (operand_equal_p (rhs1
, bump_tree
, 0)
2075 && operand_equal_p (rhs2
, basis_name
, 0))))
2077 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2079 fputs ("(duplicate, not actually replacing)", dump_file
);
2080 stmt_to_print
= c
->cand_stmt
;
2085 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
2086 gimple_assign_set_rhs_with_ops (&gsi
, code
,
2087 basis_name
, bump_tree
);
2088 update_stmt (gsi_stmt (gsi
));
2089 c
->cand_stmt
= gsi_stmt (gsi
);
2090 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2091 stmt_to_print
= gsi_stmt (gsi
);
2095 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2097 fputs ("With: ", dump_file
);
2098 print_gimple_stmt (dump_file
, stmt_to_print
, 0, 0);
2099 fputs ("\n", dump_file
);
2104 /* Replace candidate C with an add or subtract. Note that we only
2105 operate on CAND_MULTs with known strides, so we will never generate
2106 a POINTER_PLUS_EXPR. Each candidate X = (B + i) * S is replaced by
2107 X = Y + ((i - i') * S), as described in the module commentary. The
2108 folded value ((i - i') * S) is referred to here as the "bump." */
2111 replace_unconditional_candidate (slsr_cand_t c
)
2115 if (cand_already_replaced (c
))
2118 basis
= lookup_cand (c
->basis
);
2119 widest_int bump
= cand_increment (c
) * wi::to_widest (c
->stride
);
2121 replace_mult_candidate (c
, gimple_assign_lhs (basis
->cand_stmt
), bump
);
2124 /* Return the index in the increment vector of the given INCREMENT,
2125 or -1 if not found. The latter can occur if more than
2126 MAX_INCR_VEC_LEN increments have been found. */
2129 incr_vec_index (const widest_int
&increment
)
2133 for (i
= 0; i
< incr_vec_len
&& increment
!= incr_vec
[i
].incr
; i
++)
2136 if (i
< incr_vec_len
)
2142 /* Create a new statement along edge E to add BASIS_NAME to the product
2143 of INCREMENT and the stride of candidate C. Create and return a new
2144 SSA name from *VAR to be used as the LHS of the new statement.
2145 KNOWN_STRIDE is true iff C's stride is a constant. */
2148 create_add_on_incoming_edge (slsr_cand_t c
, tree basis_name
,
2149 widest_int increment
, edge e
, location_t loc
,
2152 basic_block insert_bb
;
2153 gimple_stmt_iterator gsi
;
2154 tree lhs
, basis_type
;
2157 /* If the add candidate along this incoming edge has the same
2158 index as C's hidden basis, the hidden basis represents this
2163 basis_type
= TREE_TYPE (basis_name
);
2164 lhs
= make_temp_ssa_name (basis_type
, NULL
, "slsr");
2169 enum tree_code code
= PLUS_EXPR
;
2170 widest_int bump
= increment
* wi::to_widest (c
->stride
);
2171 if (wi::neg_p (bump
))
2177 bump_tree
= wide_int_to_tree (basis_type
, bump
);
2178 new_stmt
= gimple_build_assign (lhs
, code
, basis_name
, bump_tree
);
2183 bool negate_incr
= (!address_arithmetic_p
&& wi::neg_p (increment
));
2184 i
= incr_vec_index (negate_incr
? -increment
: increment
);
2185 gcc_assert (i
>= 0);
2187 if (incr_vec
[i
].initializer
)
2189 enum tree_code code
= negate_incr
? MINUS_EXPR
: PLUS_EXPR
;
2190 new_stmt
= gimple_build_assign (lhs
, code
, basis_name
,
2191 incr_vec
[i
].initializer
);
2193 else if (increment
== 1)
2194 new_stmt
= gimple_build_assign (lhs
, PLUS_EXPR
, basis_name
, c
->stride
);
2195 else if (increment
== -1)
2196 new_stmt
= gimple_build_assign (lhs
, MINUS_EXPR
, basis_name
,
2202 insert_bb
= single_succ_p (e
->src
) ? e
->src
: split_edge (e
);
2203 gsi
= gsi_last_bb (insert_bb
);
2205 if (!gsi_end_p (gsi
) && is_ctrl_stmt (gsi_stmt (gsi
)))
2206 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
2208 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
2210 gimple_set_location (new_stmt
, loc
);
2212 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2214 fprintf (dump_file
, "Inserting in block %d: ", insert_bb
->index
);
2215 print_gimple_stmt (dump_file
, new_stmt
, 0, 0);
2221 /* Given a candidate C with BASIS_NAME being the LHS of C's basis which
2222 is hidden by the phi node FROM_PHI, create a new phi node in the same
2223 block as FROM_PHI. The new phi is suitable for use as a basis by C,
2224 with its phi arguments representing conditional adjustments to the
2225 hidden basis along conditional incoming paths. Those adjustments are
2226 made by creating add statements (and sometimes recursively creating
2227 phis) along those incoming paths. LOC is the location to attach to
2228 the introduced statements. KNOWN_STRIDE is true iff C's stride is a
2232 create_phi_basis (slsr_cand_t c
, gimple
*from_phi
, tree basis_name
,
2233 location_t loc
, bool known_stride
)
2239 slsr_cand_t basis
= lookup_cand (c
->basis
);
2240 int nargs
= gimple_phi_num_args (from_phi
);
2241 basic_block phi_bb
= gimple_bb (from_phi
);
2242 slsr_cand_t phi_cand
= base_cand_from_table (gimple_phi_result (from_phi
));
2243 phi_args
.create (nargs
);
2245 /* Process each argument of the existing phi that represents
2246 conditionally-executed add candidates. */
2247 for (i
= 0; i
< nargs
; i
++)
2249 edge e
= (*phi_bb
->preds
)[i
];
2250 tree arg
= gimple_phi_arg_def (from_phi
, i
);
2253 /* If the phi argument is the base name of the CAND_PHI, then
2254 this incoming arc should use the hidden basis. */
2255 if (operand_equal_p (arg
, phi_cand
->base_expr
, 0))
2256 if (basis
->index
== 0)
2257 feeding_def
= gimple_assign_lhs (basis
->cand_stmt
);
2260 widest_int incr
= -basis
->index
;
2261 feeding_def
= create_add_on_incoming_edge (c
, basis_name
, incr
,
2262 e
, loc
, known_stride
);
2266 gimple
*arg_def
= SSA_NAME_DEF_STMT (arg
);
2268 /* If there is another phi along this incoming edge, we must
2269 process it in the same fashion to ensure that all basis
2270 adjustments are made along its incoming edges. */
2271 if (gimple_code (arg_def
) == GIMPLE_PHI
)
2272 feeding_def
= create_phi_basis (c
, arg_def
, basis_name
,
2276 slsr_cand_t arg_cand
= base_cand_from_table (arg
);
2277 widest_int diff
= arg_cand
->index
- basis
->index
;
2278 feeding_def
= create_add_on_incoming_edge (c
, basis_name
, diff
,
2279 e
, loc
, known_stride
);
2283 /* Because of recursion, we need to save the arguments in a vector
2284 so we can create the PHI statement all at once. Otherwise the
2285 storage for the half-created PHI can be reclaimed. */
2286 phi_args
.safe_push (feeding_def
);
2289 /* Create the new phi basis. */
2290 name
= make_temp_ssa_name (TREE_TYPE (basis_name
), NULL
, "slsr");
2291 phi
= create_phi_node (name
, phi_bb
);
2292 SSA_NAME_DEF_STMT (name
) = phi
;
2294 FOR_EACH_VEC_ELT (phi_args
, i
, phi_arg
)
2296 edge e
= (*phi_bb
->preds
)[i
];
2297 add_phi_arg (phi
, phi_arg
, e
, loc
);
2302 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2304 fputs ("Introducing new phi basis: ", dump_file
);
2305 print_gimple_stmt (dump_file
, phi
, 0, 0);
2311 /* Given a candidate C whose basis is hidden by at least one intervening
2312 phi, introduce a matching number of new phis to represent its basis
2313 adjusted by conditional increments along possible incoming paths. Then
2314 replace C as though it were an unconditional candidate, using the new
2318 replace_conditional_candidate (slsr_cand_t c
)
2320 tree basis_name
, name
;
2324 /* Look up the LHS SSA name from C's basis. This will be the
2325 RHS1 of the adds we will introduce to create new phi arguments. */
2326 basis
= lookup_cand (c
->basis
);
2327 basis_name
= gimple_assign_lhs (basis
->cand_stmt
);
2329 /* Create a new phi statement which will represent C's true basis
2330 after the transformation is complete. */
2331 loc
= gimple_location (c
->cand_stmt
);
2332 name
= create_phi_basis (c
, lookup_cand (c
->def_phi
)->cand_stmt
,
2333 basis_name
, loc
, KNOWN_STRIDE
);
2334 /* Replace C with an add of the new basis phi and a constant. */
2335 widest_int bump
= c
->index
* wi::to_widest (c
->stride
);
2337 replace_mult_candidate (c
, name
, bump
);
2340 /* Compute the expected costs of inserting basis adjustments for
2341 candidate C with phi-definition PHI. The cost of inserting
2342 one adjustment is given by ONE_ADD_COST. If PHI has arguments
2343 which are themselves phi results, recursively calculate costs
2344 for those phis as well. */
2347 phi_add_costs (gimple
*phi
, slsr_cand_t c
, int one_add_cost
)
2351 slsr_cand_t phi_cand
= base_cand_from_table (gimple_phi_result (phi
));
2353 /* If we work our way back to a phi that isn't dominated by the hidden
2354 basis, this isn't a candidate for replacement. Indicate this by
2355 returning an unreasonably high cost. It's not easy to detect
2356 these situations when determining the basis, so we defer the
2357 decision until now. */
2358 basic_block phi_bb
= gimple_bb (phi
);
2359 slsr_cand_t basis
= lookup_cand (c
->basis
);
2360 basic_block basis_bb
= gimple_bb (basis
->cand_stmt
);
2362 if (phi_bb
== basis_bb
|| !dominated_by_p (CDI_DOMINATORS
, phi_bb
, basis_bb
))
2363 return COST_INFINITE
;
2365 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2367 tree arg
= gimple_phi_arg_def (phi
, i
);
2369 if (arg
!= phi_cand
->base_expr
)
2371 gimple
*arg_def
= SSA_NAME_DEF_STMT (arg
);
2373 if (gimple_code (arg_def
) == GIMPLE_PHI
)
2374 cost
+= phi_add_costs (arg_def
, c
, one_add_cost
);
2377 slsr_cand_t arg_cand
= base_cand_from_table (arg
);
2379 if (arg_cand
->index
!= c
->index
)
2380 cost
+= one_add_cost
;
2388 /* For candidate C, each sibling of candidate C, and each dependent of
2389 candidate C, determine whether the candidate is dependent upon a
2390 phi that hides its basis. If not, replace the candidate unconditionally.
2391 Otherwise, determine whether the cost of introducing compensation code
2392 for the candidate is offset by the gains from strength reduction. If
2393 so, replace the candidate and introduce the compensation code. */
2396 replace_uncond_cands_and_profitable_phis (slsr_cand_t c
)
2398 if (phi_dependent_cand_p (c
))
2400 if (c
->kind
== CAND_MULT
)
2402 /* A candidate dependent upon a phi will replace a multiply by
2403 a constant with an add, and will insert at most one add for
2404 each phi argument. Add these costs with the potential dead-code
2405 savings to determine profitability. */
2406 bool speed
= optimize_bb_for_speed_p (gimple_bb (c
->cand_stmt
));
2407 int mult_savings
= stmt_cost (c
->cand_stmt
, speed
);
2408 gimple
*phi
= lookup_cand (c
->def_phi
)->cand_stmt
;
2409 tree phi_result
= gimple_phi_result (phi
);
2410 int one_add_cost
= add_cost (speed
,
2411 TYPE_MODE (TREE_TYPE (phi_result
)));
2412 int add_costs
= one_add_cost
+ phi_add_costs (phi
, c
, one_add_cost
);
2413 int cost
= add_costs
- mult_savings
- c
->dead_savings
;
2415 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2417 fprintf (dump_file
, " Conditional candidate %d:\n", c
->cand_num
);
2418 fprintf (dump_file
, " add_costs = %d\n", add_costs
);
2419 fprintf (dump_file
, " mult_savings = %d\n", mult_savings
);
2420 fprintf (dump_file
, " dead_savings = %d\n", c
->dead_savings
);
2421 fprintf (dump_file
, " cost = %d\n", cost
);
2422 if (cost
<= COST_NEUTRAL
)
2423 fputs (" Replacing...\n", dump_file
);
2425 fputs (" Not replaced.\n", dump_file
);
2428 if (cost
<= COST_NEUTRAL
)
2429 replace_conditional_candidate (c
);
2433 replace_unconditional_candidate (c
);
2436 replace_uncond_cands_and_profitable_phis (lookup_cand (c
->sibling
));
2439 replace_uncond_cands_and_profitable_phis (lookup_cand (c
->dependent
));
2442 /* Count the number of candidates in the tree rooted at C that have
2443 not already been replaced under other interpretations. */
2446 count_candidates (slsr_cand_t c
)
2448 unsigned count
= cand_already_replaced (c
) ? 0 : 1;
2451 count
+= count_candidates (lookup_cand (c
->sibling
));
2454 count
+= count_candidates (lookup_cand (c
->dependent
));
2459 /* Increase the count of INCREMENT by one in the increment vector.
2460 INCREMENT is associated with candidate C. If INCREMENT is to be
2461 conditionally executed as part of a conditional candidate replacement,
2462 IS_PHI_ADJUST is true, otherwise false. If an initializer
2463 T_0 = stride * I is provided by a candidate that dominates all
2464 candidates with the same increment, also record T_0 for subsequent use. */
2467 record_increment (slsr_cand_t c
, widest_int increment
, bool is_phi_adjust
)
2472 /* Treat increments that differ only in sign as identical so as to
2473 share initializers, unless we are generating pointer arithmetic. */
2474 if (!address_arithmetic_p
&& wi::neg_p (increment
))
2475 increment
= -increment
;
2477 for (i
= 0; i
< incr_vec_len
; i
++)
2479 if (incr_vec
[i
].incr
== increment
)
2481 incr_vec
[i
].count
++;
2484 /* If we previously recorded an initializer that doesn't
2485 dominate this candidate, it's not going to be useful to
2487 if (incr_vec
[i
].initializer
2488 && !dominated_by_p (CDI_DOMINATORS
,
2489 gimple_bb (c
->cand_stmt
),
2490 incr_vec
[i
].init_bb
))
2492 incr_vec
[i
].initializer
= NULL_TREE
;
2493 incr_vec
[i
].init_bb
= NULL
;
2500 if (!found
&& incr_vec_len
< MAX_INCR_VEC_LEN
- 1)
2502 /* The first time we see an increment, create the entry for it.
2503 If this is the root candidate which doesn't have a basis, set
2504 the count to zero. We're only processing it so it can possibly
2505 provide an initializer for other candidates. */
2506 incr_vec
[incr_vec_len
].incr
= increment
;
2507 incr_vec
[incr_vec_len
].count
= c
->basis
|| is_phi_adjust
? 1 : 0;
2508 incr_vec
[incr_vec_len
].cost
= COST_INFINITE
;
2510 /* Optimistically record the first occurrence of this increment
2511 as providing an initializer (if it does); we will revise this
2512 opinion later if it doesn't dominate all other occurrences.
2513 Exception: increments of -1, 0, 1 never need initializers;
2514 and phi adjustments don't ever provide initializers. */
2515 if (c
->kind
== CAND_ADD
2517 && c
->index
== increment
2518 && (wi::gts_p (increment
, 1)
2519 || wi::lts_p (increment
, -1))
2520 && (gimple_assign_rhs_code (c
->cand_stmt
) == PLUS_EXPR
2521 || gimple_assign_rhs_code (c
->cand_stmt
) == POINTER_PLUS_EXPR
))
2523 tree t0
= NULL_TREE
;
2524 tree rhs1
= gimple_assign_rhs1 (c
->cand_stmt
);
2525 tree rhs2
= gimple_assign_rhs2 (c
->cand_stmt
);
2526 if (operand_equal_p (rhs1
, c
->base_expr
, 0))
2528 else if (operand_equal_p (rhs2
, c
->base_expr
, 0))
2531 && SSA_NAME_DEF_STMT (t0
)
2532 && gimple_bb (SSA_NAME_DEF_STMT (t0
)))
2534 incr_vec
[incr_vec_len
].initializer
= t0
;
2535 incr_vec
[incr_vec_len
++].init_bb
2536 = gimple_bb (SSA_NAME_DEF_STMT (t0
));
2540 incr_vec
[incr_vec_len
].initializer
= NULL_TREE
;
2541 incr_vec
[incr_vec_len
++].init_bb
= NULL
;
2546 incr_vec
[incr_vec_len
].initializer
= NULL_TREE
;
2547 incr_vec
[incr_vec_len
++].init_bb
= NULL
;
2552 /* Given phi statement PHI that hides a candidate from its BASIS, find
2553 the increments along each incoming arc (recursively handling additional
2554 phis that may be present) and record them. These increments are the
2555 difference in index between the index-adjusting statements and the
2556 index of the basis. */
2559 record_phi_increments (slsr_cand_t basis
, gimple
*phi
)
2562 slsr_cand_t phi_cand
= base_cand_from_table (gimple_phi_result (phi
));
2564 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2566 tree arg
= gimple_phi_arg_def (phi
, i
);
2568 if (!operand_equal_p (arg
, phi_cand
->base_expr
, 0))
2570 gimple
*arg_def
= SSA_NAME_DEF_STMT (arg
);
2572 if (gimple_code (arg_def
) == GIMPLE_PHI
)
2573 record_phi_increments (basis
, arg_def
);
2576 slsr_cand_t arg_cand
= base_cand_from_table (arg
);
2577 widest_int diff
= arg_cand
->index
- basis
->index
;
2578 record_increment (arg_cand
, diff
, PHI_ADJUST
);
2584 /* Determine how many times each unique increment occurs in the set
2585 of candidates rooted at C's parent, recording the data in the
2586 increment vector. For each unique increment I, if an initializer
2587 T_0 = stride * I is provided by a candidate that dominates all
2588 candidates with the same increment, also record T_0 for subsequent
2592 record_increments (slsr_cand_t c
)
2594 if (!cand_already_replaced (c
))
2596 if (!phi_dependent_cand_p (c
))
2597 record_increment (c
, cand_increment (c
), NOT_PHI_ADJUST
);
2600 /* A candidate with a basis hidden by a phi will have one
2601 increment for its relationship to the index represented by
2602 the phi, and potentially additional increments along each
2603 incoming edge. For the root of the dependency tree (which
2604 has no basis), process just the initial index in case it has
2605 an initializer that can be used by subsequent candidates. */
2606 record_increment (c
, c
->index
, NOT_PHI_ADJUST
);
2609 record_phi_increments (lookup_cand (c
->basis
),
2610 lookup_cand (c
->def_phi
)->cand_stmt
);
2615 record_increments (lookup_cand (c
->sibling
));
2618 record_increments (lookup_cand (c
->dependent
));
2621 /* Add up and return the costs of introducing add statements that
2622 require the increment INCR on behalf of candidate C and phi
2623 statement PHI. Accumulate into *SAVINGS the potential savings
2624 from removing existing statements that feed PHI and have no other
2628 phi_incr_cost (slsr_cand_t c
, const widest_int
&incr
, gimple
*phi
,
2633 slsr_cand_t basis
= lookup_cand (c
->basis
);
2634 slsr_cand_t phi_cand
= base_cand_from_table (gimple_phi_result (phi
));
2636 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2638 tree arg
= gimple_phi_arg_def (phi
, i
);
2640 if (!operand_equal_p (arg
, phi_cand
->base_expr
, 0))
2642 gimple
*arg_def
= SSA_NAME_DEF_STMT (arg
);
2644 if (gimple_code (arg_def
) == GIMPLE_PHI
)
2646 int feeding_savings
= 0;
2647 cost
+= phi_incr_cost (c
, incr
, arg_def
, &feeding_savings
);
2648 if (has_single_use (gimple_phi_result (arg_def
)))
2649 *savings
+= feeding_savings
;
2653 slsr_cand_t arg_cand
= base_cand_from_table (arg
);
2654 widest_int diff
= arg_cand
->index
- basis
->index
;
2658 tree basis_lhs
= gimple_assign_lhs (basis
->cand_stmt
);
2659 tree lhs
= gimple_assign_lhs (arg_cand
->cand_stmt
);
2660 cost
+= add_cost (true, TYPE_MODE (TREE_TYPE (basis_lhs
)));
2661 if (has_single_use (lhs
))
2662 *savings
+= stmt_cost (arg_cand
->cand_stmt
, true);
2671 /* Return the first candidate in the tree rooted at C that has not
2672 already been replaced, favoring siblings over dependents. */
2675 unreplaced_cand_in_tree (slsr_cand_t c
)
2677 if (!cand_already_replaced (c
))
2682 slsr_cand_t sib
= unreplaced_cand_in_tree (lookup_cand (c
->sibling
));
2689 slsr_cand_t dep
= unreplaced_cand_in_tree (lookup_cand (c
->dependent
));
2697 /* Return TRUE if the candidates in the tree rooted at C should be
2698 optimized for speed, else FALSE. We estimate this based on the block
2699 containing the most dominant candidate in the tree that has not yet
2703 optimize_cands_for_speed_p (slsr_cand_t c
)
2705 slsr_cand_t c2
= unreplaced_cand_in_tree (c
);
2707 return optimize_bb_for_speed_p (gimple_bb (c2
->cand_stmt
));
2710 /* Add COST_IN to the lowest cost of any dependent path starting at
2711 candidate C or any of its siblings, counting only candidates along
2712 such paths with increment INCR. Assume that replacing a candidate
2713 reduces cost by REPL_SAVINGS. Also account for savings from any
2714 statements that would go dead. If COUNT_PHIS is true, include
2715 costs of introducing feeding statements for conditional candidates. */
2718 lowest_cost_path (int cost_in
, int repl_savings
, slsr_cand_t c
,
2719 const widest_int
&incr
, bool count_phis
)
2721 int local_cost
, sib_cost
, savings
= 0;
2722 widest_int cand_incr
= cand_abs_increment (c
);
2724 if (cand_already_replaced (c
))
2725 local_cost
= cost_in
;
2726 else if (incr
== cand_incr
)
2727 local_cost
= cost_in
- repl_savings
- c
->dead_savings
;
2729 local_cost
= cost_in
- c
->dead_savings
;
2732 && phi_dependent_cand_p (c
)
2733 && !cand_already_replaced (c
))
2735 gimple
*phi
= lookup_cand (c
->def_phi
)->cand_stmt
;
2736 local_cost
+= phi_incr_cost (c
, incr
, phi
, &savings
);
2738 if (has_single_use (gimple_phi_result (phi
)))
2739 local_cost
-= savings
;
2743 local_cost
= lowest_cost_path (local_cost
, repl_savings
,
2744 lookup_cand (c
->dependent
), incr
,
2749 sib_cost
= lowest_cost_path (cost_in
, repl_savings
,
2750 lookup_cand (c
->sibling
), incr
,
2752 local_cost
= MIN (local_cost
, sib_cost
);
2758 /* Compute the total savings that would accrue from all replacements
2759 in the candidate tree rooted at C, counting only candidates with
2760 increment INCR. Assume that replacing a candidate reduces cost
2761 by REPL_SAVINGS. Also account for savings from statements that
2765 total_savings (int repl_savings
, slsr_cand_t c
, const widest_int
&incr
,
2769 widest_int cand_incr
= cand_abs_increment (c
);
2771 if (incr
== cand_incr
&& !cand_already_replaced (c
))
2772 savings
+= repl_savings
+ c
->dead_savings
;
2775 && phi_dependent_cand_p (c
)
2776 && !cand_already_replaced (c
))
2778 int phi_savings
= 0;
2779 gimple
*phi
= lookup_cand (c
->def_phi
)->cand_stmt
;
2780 savings
-= phi_incr_cost (c
, incr
, phi
, &phi_savings
);
2782 if (has_single_use (gimple_phi_result (phi
)))
2783 savings
+= phi_savings
;
2787 savings
+= total_savings (repl_savings
, lookup_cand (c
->dependent
), incr
,
2791 savings
+= total_savings (repl_savings
, lookup_cand (c
->sibling
), incr
,
2797 /* Use target-specific costs to determine and record which increments
2798 in the current candidate tree are profitable to replace, assuming
2799 MODE and SPEED. FIRST_DEP is the first dependent of the root of
2802 One slight limitation here is that we don't account for the possible
2803 introduction of casts in some cases. See replace_one_candidate for
2804 the cases where these are introduced. This should probably be cleaned
2808 analyze_increments (slsr_cand_t first_dep
, machine_mode mode
, bool speed
)
2812 for (i
= 0; i
< incr_vec_len
; i
++)
2814 HOST_WIDE_INT incr
= incr_vec
[i
].incr
.to_shwi ();
2816 /* If somehow this increment is bigger than a HWI, we won't
2817 be optimizing candidates that use it. And if the increment
2818 has a count of zero, nothing will be done with it. */
2819 if (!wi::fits_shwi_p (incr_vec
[i
].incr
) || !incr_vec
[i
].count
)
2820 incr_vec
[i
].cost
= COST_INFINITE
;
2822 /* Increments of 0, 1, and -1 are always profitable to replace,
2823 because they always replace a multiply or add with an add or
2824 copy, and may cause one or more existing instructions to go
2825 dead. Exception: -1 can't be assumed to be profitable for
2826 pointer addition. */
2830 && (gimple_assign_rhs_code (first_dep
->cand_stmt
)
2831 != POINTER_PLUS_EXPR
)))
2832 incr_vec
[i
].cost
= COST_NEUTRAL
;
2834 /* FORNOW: If we need to add an initializer, give up if a cast from
2835 the candidate's type to its stride's type can lose precision.
2836 This could eventually be handled better by expressly retaining the
2837 result of a cast to a wider type in the stride. Example:
2842 _4 = x + _3; ADD: x + (10 * _1) : int
2844 _6 = x + _3; ADD: x + (15 * _1) : int
2846 Right now replacing _6 would cause insertion of an initializer
2847 of the form "short int T = _1 * 5;" followed by a cast to
2848 int, which could overflow incorrectly. Had we recorded _2 or
2849 (int)_1 as the stride, this wouldn't happen. However, doing
2850 this breaks other opportunities, so this will require some
2852 else if (!incr_vec
[i
].initializer
2853 && TREE_CODE (first_dep
->stride
) != INTEGER_CST
2854 && !legal_cast_p_1 (first_dep
->stride
,
2855 gimple_assign_lhs (first_dep
->cand_stmt
)))
2857 incr_vec
[i
].cost
= COST_INFINITE
;
2859 /* If we need to add an initializer, make sure we don't introduce
2860 a multiply by a pointer type, which can happen in certain cast
2861 scenarios. FIXME: When cleaning up these cast issues, we can
2862 afford to introduce the multiply provided we cast out to an
2863 unsigned int of appropriate size. */
2864 else if (!incr_vec
[i
].initializer
2865 && TREE_CODE (first_dep
->stride
) != INTEGER_CST
2866 && POINTER_TYPE_P (TREE_TYPE (first_dep
->stride
)))
2868 incr_vec
[i
].cost
= COST_INFINITE
;
2870 /* For any other increment, if this is a multiply candidate, we
2871 must introduce a temporary T and initialize it with
2872 T_0 = stride * increment. When optimizing for speed, walk the
2873 candidate tree to calculate the best cost reduction along any
2874 path; if it offsets the fixed cost of inserting the initializer,
2875 replacing the increment is profitable. When optimizing for
2876 size, instead calculate the total cost reduction from replacing
2877 all candidates with this increment. */
2878 else if (first_dep
->kind
== CAND_MULT
)
2880 int cost
= mult_by_coeff_cost (incr
, mode
, speed
);
2881 int repl_savings
= mul_cost (speed
, mode
) - add_cost (speed
, mode
);
2883 cost
= lowest_cost_path (cost
, repl_savings
, first_dep
,
2884 incr_vec
[i
].incr
, COUNT_PHIS
);
2886 cost
-= total_savings (repl_savings
, first_dep
, incr_vec
[i
].incr
,
2889 incr_vec
[i
].cost
= cost
;
2892 /* If this is an add candidate, the initializer may already
2893 exist, so only calculate the cost of the initializer if it
2894 doesn't. We are replacing one add with another here, so the
2895 known replacement savings is zero. We will account for removal
2896 of dead instructions in lowest_cost_path or total_savings. */
2900 if (!incr_vec
[i
].initializer
)
2901 cost
= mult_by_coeff_cost (incr
, mode
, speed
);
2904 cost
= lowest_cost_path (cost
, 0, first_dep
, incr_vec
[i
].incr
,
2907 cost
-= total_savings (0, first_dep
, incr_vec
[i
].incr
,
2910 incr_vec
[i
].cost
= cost
;
2915 /* Return the nearest common dominator of BB1 and BB2. If the blocks
2916 are identical, return the earlier of C1 and C2 in *WHERE. Otherwise,
2917 if the NCD matches BB1, return C1 in *WHERE; if the NCD matches BB2,
2918 return C2 in *WHERE; and if the NCD matches neither, return NULL in
2919 *WHERE. Note: It is possible for one of C1 and C2 to be NULL. */
2922 ncd_for_two_cands (basic_block bb1
, basic_block bb2
,
2923 slsr_cand_t c1
, slsr_cand_t c2
, slsr_cand_t
*where
)
2939 ncd
= nearest_common_dominator (CDI_DOMINATORS
, bb1
, bb2
);
2941 /* If both candidates are in the same block, the earlier
2943 if (bb1
== ncd
&& bb2
== ncd
)
2945 if (!c1
|| (c2
&& c2
->cand_num
< c1
->cand_num
))
2951 /* Otherwise, if one of them produced a candidate in the
2952 dominator, that one wins. */
2953 else if (bb1
== ncd
)
2956 else if (bb2
== ncd
)
2959 /* If neither matches the dominator, neither wins. */
2966 /* Consider all candidates that feed PHI. Find the nearest common
2967 dominator of those candidates requiring the given increment INCR.
2968 Further find and return the nearest common dominator of this result
2969 with block NCD. If the returned block contains one or more of the
2970 candidates, return the earliest candidate in the block in *WHERE. */
2973 ncd_with_phi (slsr_cand_t c
, const widest_int
&incr
, gphi
*phi
,
2974 basic_block ncd
, slsr_cand_t
*where
)
2977 slsr_cand_t basis
= lookup_cand (c
->basis
);
2978 slsr_cand_t phi_cand
= base_cand_from_table (gimple_phi_result (phi
));
2980 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2982 tree arg
= gimple_phi_arg_def (phi
, i
);
2984 if (!operand_equal_p (arg
, phi_cand
->base_expr
, 0))
2986 gimple
*arg_def
= SSA_NAME_DEF_STMT (arg
);
2988 if (gimple_code (arg_def
) == GIMPLE_PHI
)
2989 ncd
= ncd_with_phi (c
, incr
, as_a
<gphi
*> (arg_def
), ncd
,
2993 slsr_cand_t arg_cand
= base_cand_from_table (arg
);
2994 widest_int diff
= arg_cand
->index
- basis
->index
;
2995 basic_block pred
= gimple_phi_arg_edge (phi
, i
)->src
;
2997 if ((incr
== diff
) || (!address_arithmetic_p
&& incr
== -diff
))
2998 ncd
= ncd_for_two_cands (ncd
, pred
, *where
, NULL
, where
);
3006 /* Consider the candidate C together with any candidates that feed
3007 C's phi dependence (if any). Find and return the nearest common
3008 dominator of those candidates requiring the given increment INCR.
3009 If the returned block contains one or more of the candidates,
3010 return the earliest candidate in the block in *WHERE. */
3013 ncd_of_cand_and_phis (slsr_cand_t c
, const widest_int
&incr
, slsr_cand_t
*where
)
3015 basic_block ncd
= NULL
;
3017 if (cand_abs_increment (c
) == incr
)
3019 ncd
= gimple_bb (c
->cand_stmt
);
3023 if (phi_dependent_cand_p (c
))
3024 ncd
= ncd_with_phi (c
, incr
,
3025 as_a
<gphi
*> (lookup_cand (c
->def_phi
)->cand_stmt
),
3031 /* Consider all candidates in the tree rooted at C for which INCR
3032 represents the required increment of C relative to its basis.
3033 Find and return the basic block that most nearly dominates all
3034 such candidates. If the returned block contains one or more of
3035 the candidates, return the earliest candidate in the block in
3039 nearest_common_dominator_for_cands (slsr_cand_t c
, const widest_int
&incr
,
3042 basic_block sib_ncd
= NULL
, dep_ncd
= NULL
, this_ncd
= NULL
, ncd
;
3043 slsr_cand_t sib_where
= NULL
, dep_where
= NULL
, this_where
= NULL
, new_where
;
3045 /* First find the NCD of all siblings and dependents. */
3047 sib_ncd
= nearest_common_dominator_for_cands (lookup_cand (c
->sibling
),
3050 dep_ncd
= nearest_common_dominator_for_cands (lookup_cand (c
->dependent
),
3052 if (!sib_ncd
&& !dep_ncd
)
3057 else if (sib_ncd
&& !dep_ncd
)
3059 new_where
= sib_where
;
3062 else if (dep_ncd
&& !sib_ncd
)
3064 new_where
= dep_where
;
3068 ncd
= ncd_for_two_cands (sib_ncd
, dep_ncd
, sib_where
,
3069 dep_where
, &new_where
);
3071 /* If the candidate's increment doesn't match the one we're interested
3072 in (and nor do any increments for feeding defs of a phi-dependence),
3073 then the result depends only on siblings and dependents. */
3074 this_ncd
= ncd_of_cand_and_phis (c
, incr
, &this_where
);
3076 if (!this_ncd
|| cand_already_replaced (c
))
3082 /* Otherwise, compare this candidate with the result from all siblings
3084 ncd
= ncd_for_two_cands (ncd
, this_ncd
, new_where
, this_where
, where
);
3089 /* Return TRUE if the increment indexed by INDEX is profitable to replace. */
3092 profitable_increment_p (unsigned index
)
3094 return (incr_vec
[index
].cost
<= COST_NEUTRAL
);
3097 /* For each profitable increment in the increment vector not equal to
3098 0 or 1 (or -1, for non-pointer arithmetic), find the nearest common
3099 dominator of all statements in the candidate chain rooted at C
3100 that require that increment, and insert an initializer
3101 T_0 = stride * increment at that location. Record T_0 with the
3102 increment record. */
3105 insert_initializers (slsr_cand_t c
)
3109 for (i
= 0; i
< incr_vec_len
; i
++)
3112 slsr_cand_t where
= NULL
;
3114 tree stride_type
, new_name
, incr_tree
;
3115 widest_int incr
= incr_vec
[i
].incr
;
3117 if (!profitable_increment_p (i
)
3120 && gimple_assign_rhs_code (c
->cand_stmt
) != POINTER_PLUS_EXPR
)
3124 /* We may have already identified an existing initializer that
3126 if (incr_vec
[i
].initializer
)
3128 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3130 fputs ("Using existing initializer: ", dump_file
);
3131 print_gimple_stmt (dump_file
,
3132 SSA_NAME_DEF_STMT (incr_vec
[i
].initializer
),
3138 /* Find the block that most closely dominates all candidates
3139 with this increment. If there is at least one candidate in
3140 that block, the earliest one will be returned in WHERE. */
3141 bb
= nearest_common_dominator_for_cands (c
, incr
, &where
);
3143 /* Create a new SSA name to hold the initializer's value. */
3144 stride_type
= TREE_TYPE (c
->stride
);
3145 new_name
= make_temp_ssa_name (stride_type
, NULL
, "slsr");
3146 incr_vec
[i
].initializer
= new_name
;
3148 /* Create the initializer and insert it in the latest possible
3149 dominating position. */
3150 incr_tree
= wide_int_to_tree (stride_type
, incr
);
3151 init_stmt
= gimple_build_assign (new_name
, MULT_EXPR
,
3152 c
->stride
, incr_tree
);
3155 gimple_stmt_iterator gsi
= gsi_for_stmt (where
->cand_stmt
);
3156 gsi_insert_before (&gsi
, init_stmt
, GSI_SAME_STMT
);
3157 gimple_set_location (init_stmt
, gimple_location (where
->cand_stmt
));
3161 gimple_stmt_iterator gsi
= gsi_last_bb (bb
);
3162 gimple
*basis_stmt
= lookup_cand (c
->basis
)->cand_stmt
;
3164 if (!gsi_end_p (gsi
) && is_ctrl_stmt (gsi_stmt (gsi
)))
3165 gsi_insert_before (&gsi
, init_stmt
, GSI_SAME_STMT
);
3167 gsi_insert_after (&gsi
, init_stmt
, GSI_SAME_STMT
);
3169 gimple_set_location (init_stmt
, gimple_location (basis_stmt
));
3172 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3174 fputs ("Inserting initializer: ", dump_file
);
3175 print_gimple_stmt (dump_file
, init_stmt
, 0, 0);
3180 /* Return TRUE iff all required increments for candidates feeding PHI
3181 are profitable to replace on behalf of candidate C. */
3184 all_phi_incrs_profitable (slsr_cand_t c
, gimple
*phi
)
3187 slsr_cand_t basis
= lookup_cand (c
->basis
);
3188 slsr_cand_t phi_cand
= base_cand_from_table (gimple_phi_result (phi
));
3190 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
3192 tree arg
= gimple_phi_arg_def (phi
, i
);
3194 if (!operand_equal_p (arg
, phi_cand
->base_expr
, 0))
3196 gimple
*arg_def
= SSA_NAME_DEF_STMT (arg
);
3198 if (gimple_code (arg_def
) == GIMPLE_PHI
)
3200 if (!all_phi_incrs_profitable (c
, arg_def
))
3206 slsr_cand_t arg_cand
= base_cand_from_table (arg
);
3207 widest_int increment
= arg_cand
->index
- basis
->index
;
3209 if (!address_arithmetic_p
&& wi::neg_p (increment
))
3210 increment
= -increment
;
3212 j
= incr_vec_index (increment
);
3214 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3216 fprintf (dump_file
, " Conditional candidate %d, phi: ",
3218 print_gimple_stmt (dump_file
, phi
, 0, 0);
3219 fputs (" increment: ", dump_file
);
3220 print_decs (increment
, dump_file
);
3223 "\n Not replaced; incr_vec overflow.\n");
3225 fprintf (dump_file
, "\n cost: %d\n", incr_vec
[j
].cost
);
3226 if (profitable_increment_p (j
))
3227 fputs (" Replacing...\n", dump_file
);
3229 fputs (" Not replaced.\n", dump_file
);
3233 if (j
< 0 || !profitable_increment_p (j
))
3242 /* Create a NOP_EXPR that copies FROM_EXPR into a new SSA name of
3243 type TO_TYPE, and insert it in front of the statement represented
3244 by candidate C. Use *NEW_VAR to create the new SSA name. Return
3245 the new SSA name. */
3248 introduce_cast_before_cand (slsr_cand_t c
, tree to_type
, tree from_expr
)
3252 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
3254 cast_lhs
= make_temp_ssa_name (to_type
, NULL
, "slsr");
3255 cast_stmt
= gimple_build_assign (cast_lhs
, NOP_EXPR
, from_expr
);
3256 gimple_set_location (cast_stmt
, gimple_location (c
->cand_stmt
));
3257 gsi_insert_before (&gsi
, cast_stmt
, GSI_SAME_STMT
);
3259 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3261 fputs (" Inserting: ", dump_file
);
3262 print_gimple_stmt (dump_file
, cast_stmt
, 0, 0);
3268 /* Replace the RHS of the statement represented by candidate C with
3269 NEW_CODE, NEW_RHS1, and NEW_RHS2, provided that to do so doesn't
3270 leave C unchanged or just interchange its operands. The original
3271 operation and operands are in OLD_CODE, OLD_RHS1, and OLD_RHS2.
3272 If the replacement was made and we are doing a details dump,
3273 return the revised statement, else NULL. */
3276 replace_rhs_if_not_dup (enum tree_code new_code
, tree new_rhs1
, tree new_rhs2
,
3277 enum tree_code old_code
, tree old_rhs1
, tree old_rhs2
,
3280 if (new_code
!= old_code
3281 || ((!operand_equal_p (new_rhs1
, old_rhs1
, 0)
3282 || !operand_equal_p (new_rhs2
, old_rhs2
, 0))
3283 && (!operand_equal_p (new_rhs1
, old_rhs2
, 0)
3284 || !operand_equal_p (new_rhs2
, old_rhs1
, 0))))
3286 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
3287 gimple_assign_set_rhs_with_ops (&gsi
, new_code
, new_rhs1
, new_rhs2
);
3288 update_stmt (gsi_stmt (gsi
));
3289 c
->cand_stmt
= gsi_stmt (gsi
);
3291 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3292 return gsi_stmt (gsi
);
3295 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3296 fputs (" (duplicate, not actually replacing)\n", dump_file
);
3301 /* Strength-reduce the statement represented by candidate C by replacing
3302 it with an equivalent addition or subtraction. I is the index into
3303 the increment vector identifying C's increment. NEW_VAR is used to
3304 create a new SSA name if a cast needs to be introduced. BASIS_NAME
3305 is the rhs1 to use in creating the add/subtract. */
3308 replace_one_candidate (slsr_cand_t c
, unsigned i
, tree basis_name
)
3310 gimple
*stmt_to_print
= NULL
;
3311 tree orig_rhs1
, orig_rhs2
;
3313 enum tree_code orig_code
, repl_code
;
3314 widest_int cand_incr
;
3316 orig_code
= gimple_assign_rhs_code (c
->cand_stmt
);
3317 orig_rhs1
= gimple_assign_rhs1 (c
->cand_stmt
);
3318 orig_rhs2
= gimple_assign_rhs2 (c
->cand_stmt
);
3319 cand_incr
= cand_increment (c
);
3321 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3323 fputs ("Replacing: ", dump_file
);
3324 print_gimple_stmt (dump_file
, c
->cand_stmt
, 0, 0);
3325 stmt_to_print
= c
->cand_stmt
;
3328 if (address_arithmetic_p
)
3329 repl_code
= POINTER_PLUS_EXPR
;
3331 repl_code
= PLUS_EXPR
;
3333 /* If the increment has an initializer T_0, replace the candidate
3334 statement with an add of the basis name and the initializer. */
3335 if (incr_vec
[i
].initializer
)
3337 tree init_type
= TREE_TYPE (incr_vec
[i
].initializer
);
3338 tree orig_type
= TREE_TYPE (orig_rhs2
);
3340 if (types_compatible_p (orig_type
, init_type
))
3341 rhs2
= incr_vec
[i
].initializer
;
3343 rhs2
= introduce_cast_before_cand (c
, orig_type
,
3344 incr_vec
[i
].initializer
);
3346 if (incr_vec
[i
].incr
!= cand_incr
)
3348 gcc_assert (repl_code
== PLUS_EXPR
);
3349 repl_code
= MINUS_EXPR
;
3352 stmt_to_print
= replace_rhs_if_not_dup (repl_code
, basis_name
, rhs2
,
3353 orig_code
, orig_rhs1
, orig_rhs2
,
3357 /* Otherwise, the increment is one of -1, 0, and 1. Replace
3358 with a subtract of the stride from the basis name, a copy
3359 from the basis name, or an add of the stride to the basis
3360 name, respectively. It may be necessary to introduce a
3361 cast (or reuse an existing cast). */
3362 else if (cand_incr
== 1)
3364 tree stride_type
= TREE_TYPE (c
->stride
);
3365 tree orig_type
= TREE_TYPE (orig_rhs2
);
3367 if (types_compatible_p (orig_type
, stride_type
))
3370 rhs2
= introduce_cast_before_cand (c
, orig_type
, c
->stride
);
3372 stmt_to_print
= replace_rhs_if_not_dup (repl_code
, basis_name
, rhs2
,
3373 orig_code
, orig_rhs1
, orig_rhs2
,
3377 else if (cand_incr
== -1)
3379 tree stride_type
= TREE_TYPE (c
->stride
);
3380 tree orig_type
= TREE_TYPE (orig_rhs2
);
3381 gcc_assert (repl_code
!= POINTER_PLUS_EXPR
);
3383 if (types_compatible_p (orig_type
, stride_type
))
3386 rhs2
= introduce_cast_before_cand (c
, orig_type
, c
->stride
);
3388 if (orig_code
!= MINUS_EXPR
3389 || !operand_equal_p (basis_name
, orig_rhs1
, 0)
3390 || !operand_equal_p (rhs2
, orig_rhs2
, 0))
3392 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
3393 gimple_assign_set_rhs_with_ops (&gsi
, MINUS_EXPR
, basis_name
, rhs2
);
3394 update_stmt (gsi_stmt (gsi
));
3395 c
->cand_stmt
= gsi_stmt (gsi
);
3397 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3398 stmt_to_print
= gsi_stmt (gsi
);
3400 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3401 fputs (" (duplicate, not actually replacing)\n", dump_file
);
3404 else if (cand_incr
== 0)
3406 tree lhs
= gimple_assign_lhs (c
->cand_stmt
);
3407 tree lhs_type
= TREE_TYPE (lhs
);
3408 tree basis_type
= TREE_TYPE (basis_name
);
3410 if (types_compatible_p (lhs_type
, basis_type
))
3412 gassign
*copy_stmt
= gimple_build_assign (lhs
, basis_name
);
3413 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
3414 gimple_set_location (copy_stmt
, gimple_location (c
->cand_stmt
));
3415 gsi_replace (&gsi
, copy_stmt
, false);
3416 c
->cand_stmt
= copy_stmt
;
3418 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3419 stmt_to_print
= copy_stmt
;
3423 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
3424 gassign
*cast_stmt
= gimple_build_assign (lhs
, NOP_EXPR
, basis_name
);
3425 gimple_set_location (cast_stmt
, gimple_location (c
->cand_stmt
));
3426 gsi_replace (&gsi
, cast_stmt
, false);
3427 c
->cand_stmt
= cast_stmt
;
3429 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3430 stmt_to_print
= cast_stmt
;
3436 if (dump_file
&& (dump_flags
& TDF_DETAILS
) && stmt_to_print
)
3438 fputs ("With: ", dump_file
);
3439 print_gimple_stmt (dump_file
, stmt_to_print
, 0, 0);
3440 fputs ("\n", dump_file
);
3444 /* For each candidate in the tree rooted at C, replace it with
3445 an increment if such has been shown to be profitable. */
3448 replace_profitable_candidates (slsr_cand_t c
)
3450 if (!cand_already_replaced (c
))
3452 widest_int increment
= cand_abs_increment (c
);
3453 enum tree_code orig_code
= gimple_assign_rhs_code (c
->cand_stmt
);
3456 i
= incr_vec_index (increment
);
3458 /* Only process profitable increments. Nothing useful can be done
3459 to a cast or copy. */
3461 && profitable_increment_p (i
)
3462 && orig_code
!= MODIFY_EXPR
3463 && !CONVERT_EXPR_CODE_P (orig_code
))
3465 if (phi_dependent_cand_p (c
))
3467 gimple
*phi
= lookup_cand (c
->def_phi
)->cand_stmt
;
3469 if (all_phi_incrs_profitable (c
, phi
))
3471 /* Look up the LHS SSA name from C's basis. This will be
3472 the RHS1 of the adds we will introduce to create new
3474 slsr_cand_t basis
= lookup_cand (c
->basis
);
3475 tree basis_name
= gimple_assign_lhs (basis
->cand_stmt
);
3477 /* Create a new phi statement that will represent C's true
3478 basis after the transformation is complete. */
3479 location_t loc
= gimple_location (c
->cand_stmt
);
3480 tree name
= create_phi_basis (c
, phi
, basis_name
,
3481 loc
, UNKNOWN_STRIDE
);
3483 /* Replace C with an add of the new basis phi and the
3485 replace_one_candidate (c
, i
, name
);
3490 slsr_cand_t basis
= lookup_cand (c
->basis
);
3491 tree basis_name
= gimple_assign_lhs (basis
->cand_stmt
);
3492 replace_one_candidate (c
, i
, basis_name
);
3498 replace_profitable_candidates (lookup_cand (c
->sibling
));
3501 replace_profitable_candidates (lookup_cand (c
->dependent
));
3504 /* Analyze costs of related candidates in the candidate vector,
3505 and make beneficial replacements. */
3508 analyze_candidates_and_replace (void)
3513 /* Each candidate that has a null basis and a non-null
3514 dependent is the root of a tree of related statements.
3515 Analyze each tree to determine a subset of those
3516 statements that can be replaced with maximum benefit. */
3517 FOR_EACH_VEC_ELT (cand_vec
, i
, c
)
3519 slsr_cand_t first_dep
;
3521 if (c
->basis
!= 0 || c
->dependent
== 0)
3524 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3525 fprintf (dump_file
, "\nProcessing dependency tree rooted at %d.\n",
3528 first_dep
= lookup_cand (c
->dependent
);
3530 /* If this is a chain of CAND_REFs, unconditionally replace
3531 each of them with a strength-reduced data reference. */
3532 if (c
->kind
== CAND_REF
)
3535 /* If the common stride of all related candidates is a known
3536 constant, each candidate without a phi-dependence can be
3537 profitably replaced. Each replaces a multiply by a single
3538 add, with the possibility that a feeding add also goes dead.
3539 A candidate with a phi-dependence is replaced only if the
3540 compensation code it requires is offset by the strength
3541 reduction savings. */
3542 else if (TREE_CODE (c
->stride
) == INTEGER_CST
)
3543 replace_uncond_cands_and_profitable_phis (first_dep
);
3545 /* When the stride is an SSA name, it may still be profitable
3546 to replace some or all of the dependent candidates, depending
3547 on whether the introduced increments can be reused, or are
3548 less expensive to calculate than the replaced statements. */
3554 /* Determine whether we'll be generating pointer arithmetic
3555 when replacing candidates. */
3556 address_arithmetic_p
= (c
->kind
== CAND_ADD
3557 && POINTER_TYPE_P (c
->cand_type
));
3559 /* If all candidates have already been replaced under other
3560 interpretations, nothing remains to be done. */
3561 if (!count_candidates (c
))
3564 /* Construct an array of increments for this candidate chain. */
3565 incr_vec
= XNEWVEC (incr_info
, MAX_INCR_VEC_LEN
);
3567 record_increments (c
);
3569 /* Determine which increments are profitable to replace. */
3570 mode
= TYPE_MODE (TREE_TYPE (gimple_assign_lhs (c
->cand_stmt
)));
3571 speed
= optimize_cands_for_speed_p (c
);
3572 analyze_increments (first_dep
, mode
, speed
);
3574 /* Insert initializers of the form T_0 = stride * increment
3575 for use in profitable replacements. */
3576 insert_initializers (first_dep
);
3579 /* Perform the replacements. */
3580 replace_profitable_candidates (first_dep
);
3588 const pass_data pass_data_strength_reduction
=
3590 GIMPLE_PASS
, /* type */
3592 OPTGROUP_NONE
, /* optinfo_flags */
3593 TV_GIMPLE_SLSR
, /* tv_id */
3594 ( PROP_cfg
| PROP_ssa
), /* properties_required */
3595 0, /* properties_provided */
3596 0, /* properties_destroyed */
3597 0, /* todo_flags_start */
3598 0, /* todo_flags_finish */
3601 class pass_strength_reduction
: public gimple_opt_pass
3604 pass_strength_reduction (gcc::context
*ctxt
)
3605 : gimple_opt_pass (pass_data_strength_reduction
, ctxt
)
3608 /* opt_pass methods: */
3609 virtual bool gate (function
*) { return flag_tree_slsr
; }
3610 virtual unsigned int execute (function
*);
3612 }; // class pass_strength_reduction
3615 pass_strength_reduction::execute (function
*fun
)
3617 /* Create the obstack where candidates will reside. */
3618 gcc_obstack_init (&cand_obstack
);
3620 /* Allocate the candidate vector. */
3621 cand_vec
.create (128);
3623 /* Allocate the mapping from statements to candidate indices. */
3624 stmt_cand_map
= new hash_map
<gimple
*, slsr_cand_t
>;
3626 /* Create the obstack where candidate chains will reside. */
3627 gcc_obstack_init (&chain_obstack
);
3629 /* Allocate the mapping from base expressions to candidate chains. */
3630 base_cand_map
= new hash_table
<cand_chain_hasher
> (500);
3632 /* Allocate the mapping from bases to alternative bases. */
3633 alt_base_map
= new hash_map
<tree
, tree
>;
3635 /* Initialize the loop optimizer. We need to detect flow across
3636 back edges, and this gives us dominator information as well. */
3637 loop_optimizer_init (AVOID_CFG_MODIFICATIONS
);
3639 /* Walk the CFG in predominator order looking for strength reduction
3641 find_candidates_dom_walker (CDI_DOMINATORS
)
3642 .walk (fun
->cfg
->x_entry_block_ptr
);
3644 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3647 dump_cand_chains ();
3650 delete alt_base_map
;
3651 free_affine_expand_cache (&name_expansions
);
3653 /* Analyze costs and make appropriate replacements. */
3654 analyze_candidates_and_replace ();
3656 loop_optimizer_finalize ();
3657 delete base_cand_map
;
3658 base_cand_map
= NULL
;
3659 obstack_free (&chain_obstack
, NULL
);
3660 delete stmt_cand_map
;
3661 cand_vec
.release ();
3662 obstack_free (&cand_obstack
, NULL
);
3670 make_pass_strength_reduction (gcc::context
*ctxt
)
3672 return new pass_strength_reduction (ctxt
);