1 /* Straight-line strength reduction.
2 Copyright (C) 2012-2016 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 "gimple-pretty-print.h"
48 #include "fold-const.h"
49 #include "gimple-iterator.h"
50 #include "gimplify-me.h"
51 #include "stor-layout.h"
56 #include "tree-ssa-address.h"
57 #include "tree-affine.h"
60 /* Information about a strength reduction candidate. Each statement
61 in the candidate table represents an expression of one of the
62 following forms (the special case of CAND_REF will be described
65 (CAND_MULT) S1: X = (B + i) * S
66 (CAND_ADD) S1: X = B + (i * S)
68 Here X and B are SSA names, i is an integer constant, and S is
69 either an SSA name or a constant. We call B the "base," i the
70 "index", and S the "stride."
72 Any statement S0 that dominates S1 and is of the form:
74 (CAND_MULT) S0: Y = (B + i') * S
75 (CAND_ADD) S0: Y = B + (i' * S)
77 is called a "basis" for S1. In both cases, S1 may be replaced by
79 S1': X = Y + (i - i') * S,
81 where (i - i') * S is folded to the extent possible.
83 All gimple statements are visited in dominator order, and each
84 statement that may contribute to one of the forms of S1 above is
85 given at least one entry in the candidate table. Such statements
86 include addition, pointer addition, subtraction, multiplication,
87 negation, copies, and nontrivial type casts. If a statement may
88 represent more than one expression of the forms of S1 above,
89 multiple "interpretations" are stored in the table and chained
92 * An add of two SSA names may treat either operand as the base.
93 * A multiply of two SSA names, likewise.
94 * A copy or cast may be thought of as either a CAND_MULT with
95 i = 0 and S = 1, or as a CAND_ADD with i = 0 or S = 0.
97 Candidate records are allocated from an obstack. They are addressed
98 both from a hash table keyed on S1, and from a vector of candidate
99 pointers arranged in predominator order.
103 Currently we don't recognize:
108 as a strength reduction opportunity, even though this S1 would
109 also be replaceable by the S1' above. This can be added if it
110 comes up in practice.
112 Strength reduction in addressing
113 --------------------------------
114 There is another kind of candidate known as CAND_REF. A CAND_REF
115 describes a statement containing a memory reference having
116 complex addressing that might benefit from strength reduction.
117 Specifically, we are interested in references for which
118 get_inner_reference returns a base address, offset, and bitpos as
121 base: MEM_REF (T1, C1)
122 offset: MULT_EXPR (PLUS_EXPR (T2, C2), C3)
123 bitpos: C4 * BITS_PER_UNIT
125 Here T1 and T2 are arbitrary trees, and C1, C2, C3, C4 are
126 arbitrary integer constants. Note that C2 may be zero, in which
127 case the offset will be MULT_EXPR (T2, C3).
129 When this pattern is recognized, the original memory reference
130 can be replaced with:
132 MEM_REF (POINTER_PLUS_EXPR (T1, MULT_EXPR (T2, C3)),
135 which distributes the multiply to allow constant folding. When
136 two or more addressing expressions can be represented by MEM_REFs
137 of this form, differing only in the constants C1, C2, and C4,
138 making this substitution produces more efficient addressing during
139 the RTL phases. When there are not at least two expressions with
140 the same values of T1, T2, and C3, there is nothing to be gained
143 Strength reduction of CAND_REFs uses the same infrastructure as
144 that used by CAND_MULTs and CAND_ADDs. We record T1 in the base (B)
145 field, MULT_EXPR (T2, C3) in the stride (S) field, and
146 C1 + (C2 * C3) + C4 in the index (i) field. A basis for a CAND_REF
147 is thus another CAND_REF with the same B and S values. When at
148 least two CAND_REFs are chained together using the basis relation,
149 each of them is replaced as above, resulting in improved code
150 generation for addressing.
152 Conditional candidates
153 ======================
155 Conditional candidates are best illustrated with an example.
156 Consider the code sequence:
159 (2) a_0 = x_0 * 5; MULT (B: x_0; i: 0; S: 5)
161 (3) x_1 = x_0 + 1; ADD (B: x_0, i: 1; S: 1)
162 (4) x_2 = PHI <x_0, x_1>; PHI (B: x_0, i: 0, S: 1)
163 (5) x_3 = x_2 + 1; ADD (B: x_2, i: 1, S: 1)
164 (6) a_1 = x_3 * 5; MULT (B: x_2, i: 1; S: 5)
166 Here strength reduction is complicated by the uncertain value of x_2.
167 A legitimate transformation is:
176 (4) [x_2 = PHI <x_0, x_1>;]
177 (4a) t_2 = PHI <a_0, t_1>;
181 where the bracketed instructions may go dead.
183 To recognize this opportunity, we have to observe that statement (6)
184 has a "hidden basis" (2). The hidden basis is unlike a normal basis
185 in that the statement and the hidden basis have different base SSA
186 names (x_2 and x_0, respectively). The relationship is established
187 when a statement's base name (x_2) is defined by a phi statement (4),
188 each argument of which (x_0, x_1) has an identical "derived base name."
189 If the argument is defined by a candidate (as x_1 is by (3)) that is a
190 CAND_ADD having a stride of 1, the derived base name of the argument is
191 the base name of the candidate (x_0). Otherwise, the argument itself
192 is its derived base name (as is the case with argument x_0).
194 The hidden basis for statement (6) is the nearest dominating candidate
195 whose base name is the derived base name (x_0) of the feeding phi (4),
196 and whose stride is identical to that of the statement. We can then
197 create the new "phi basis" (4a) and feeding adds along incoming arcs (3a),
198 allowing the final replacement of (6) by the strength-reduced (6r).
200 To facilitate this, a new kind of candidate (CAND_PHI) is introduced.
201 A CAND_PHI is not a candidate for replacement, but is maintained in the
202 candidate table to ease discovery of hidden bases. Any phi statement
203 whose arguments share a common derived base name is entered into the
204 table with the derived base name, an (arbitrary) index of zero, and a
205 stride of 1. A statement with a hidden basis can then be detected by
206 simply looking up its feeding phi definition in the candidate table,
207 extracting the derived base name, and searching for a basis in the
208 usual manner after substituting the derived base name.
210 Note that the transformation is only valid when the original phi and
211 the statements that define the phi's arguments are all at the same
212 position in the loop hierarchy. */
215 /* Index into the candidate vector, offset by 1. VECs are zero-based,
216 while cand_idx's are one-based, with zero indicating null. */
217 typedef unsigned cand_idx
;
219 /* The kind of candidate. */
230 /* The candidate statement S1. */
233 /* The base expression B: often an SSA name, but not always. */
239 /* The index constant i. */
242 /* The type of the candidate. This is normally the type of base_expr,
243 but casts may have occurred when combining feeding instructions.
244 A candidate can only be a basis for candidates of the same final type.
245 (For CAND_REFs, this is the type to be used for operand 1 of the
246 replacement MEM_REF.) */
249 /* The kind of candidate (CAND_MULT, etc.). */
252 /* Index of this candidate in the candidate vector. */
255 /* Index of the next candidate record for the same statement.
256 A statement may be useful in more than one way (e.g., due to
257 commutativity). So we can have multiple "interpretations"
259 cand_idx next_interp
;
261 /* Index of the basis statement S0, if any, in the candidate vector. */
264 /* First candidate for which this candidate is a basis, if one exists. */
267 /* Next candidate having the same basis as this one. */
270 /* If this is a conditional candidate, the CAND_PHI candidate
271 that defines the base SSA name B. */
274 /* Savings that can be expected from eliminating dead code if this
275 candidate is replaced. */
279 typedef struct slsr_cand_d slsr_cand
, *slsr_cand_t
;
280 typedef const struct slsr_cand_d
*const_slsr_cand_t
;
282 /* Pointers to candidates are chained together as part of a mapping
283 from base expressions to the candidates that use them. */
287 /* Base expression for the chain of candidates: often, but not
288 always, an SSA name. */
291 /* Pointer to a candidate. */
295 struct cand_chain_d
*next
;
299 typedef struct cand_chain_d cand_chain
, *cand_chain_t
;
300 typedef const struct cand_chain_d
*const_cand_chain_t
;
302 /* Information about a unique "increment" associated with candidates
303 having an SSA name for a stride. An increment is the difference
304 between the index of the candidate and the index of its basis,
305 i.e., (i - i') as discussed in the module commentary.
307 When we are not going to generate address arithmetic we treat
308 increments that differ only in sign as the same, allowing sharing
309 of the cost of initializers. The absolute value of the increment
310 is stored in the incr_info. */
314 /* The increment that relates a candidate to its basis. */
317 /* How many times the increment occurs in the candidate tree. */
320 /* Cost of replacing candidates using this increment. Negative and
321 zero costs indicate replacement should be performed. */
324 /* If this increment is profitable but is not -1, 0, or 1, it requires
325 an initializer T_0 = stride * incr to be found or introduced in the
326 nearest common dominator of all candidates. This field holds T_0
327 for subsequent use. */
330 /* If the initializer was found to already exist, this is the block
331 where it was found. */
335 typedef struct incr_info_d incr_info
, *incr_info_t
;
337 /* Candidates are maintained in a vector. If candidate X dominates
338 candidate Y, then X appears before Y in the vector; but the
339 converse does not necessarily hold. */
340 static vec
<slsr_cand_t
> cand_vec
;
354 enum phi_adjust_status
360 enum count_phis_status
366 /* Pointer map embodying a mapping from statements to candidates. */
367 static hash_map
<gimple
*, slsr_cand_t
> *stmt_cand_map
;
369 /* Obstack for candidates. */
370 static struct obstack cand_obstack
;
372 /* Obstack for candidate chains. */
373 static struct obstack chain_obstack
;
375 /* An array INCR_VEC of incr_infos is used during analysis of related
376 candidates having an SSA name for a stride. INCR_VEC_LEN describes
377 its current length. MAX_INCR_VEC_LEN is used to avoid costly
378 pathological cases. */
379 static incr_info_t incr_vec
;
380 static unsigned incr_vec_len
;
381 const int MAX_INCR_VEC_LEN
= 16;
383 /* For a chain of candidates with unknown stride, indicates whether or not
384 we must generate pointer arithmetic when replacing statements. */
385 static bool address_arithmetic_p
;
387 /* Forward function declarations. */
388 static slsr_cand_t
base_cand_from_table (tree
);
389 static tree
introduce_cast_before_cand (slsr_cand_t
, tree
, tree
);
390 static bool legal_cast_p_1 (tree
, tree
);
392 /* Produce a pointer to the IDX'th candidate in the candidate vector. */
395 lookup_cand (cand_idx idx
)
397 return cand_vec
[idx
- 1];
400 /* Helper for hashing a candidate chain header. */
402 struct cand_chain_hasher
: nofree_ptr_hash
<cand_chain
>
404 static inline hashval_t
hash (const cand_chain
*);
405 static inline bool equal (const cand_chain
*, const cand_chain
*);
409 cand_chain_hasher::hash (const cand_chain
*p
)
411 tree base_expr
= p
->base_expr
;
412 return iterative_hash_expr (base_expr
, 0);
416 cand_chain_hasher::equal (const cand_chain
*chain1
, const cand_chain
*chain2
)
418 return operand_equal_p (chain1
->base_expr
, chain2
->base_expr
, 0);
421 /* Hash table embodying a mapping from base exprs to chains of candidates. */
422 static hash_table
<cand_chain_hasher
> *base_cand_map
;
424 /* Pointer map used by tree_to_aff_combination_expand. */
425 static hash_map
<tree
, name_expansion
*> *name_expansions
;
426 /* Pointer map embodying a mapping from bases to alternative bases. */
427 static hash_map
<tree
, tree
> *alt_base_map
;
429 /* Given BASE, use the tree affine combiniation facilities to
430 find the underlying tree expression for BASE, with any
431 immediate offset excluded.
433 N.B. we should eliminate this backtracking with better forward
434 analysis in a future release. */
437 get_alternative_base (tree base
)
439 tree
*result
= alt_base_map
->get (base
);
446 tree_to_aff_combination_expand (base
, TREE_TYPE (base
),
447 &aff
, &name_expansions
);
449 expr
= aff_combination_to_tree (&aff
);
451 gcc_assert (!alt_base_map
->put (base
, base
== expr
? NULL
: expr
));
453 return expr
== base
? NULL
: expr
;
459 /* Look in the candidate table for a CAND_PHI that defines BASE and
460 return it if found; otherwise return NULL. */
463 find_phi_def (tree base
)
467 if (TREE_CODE (base
) != SSA_NAME
)
470 c
= base_cand_from_table (base
);
472 if (!c
|| c
->kind
!= CAND_PHI
)
478 /* Helper routine for find_basis_for_candidate. May be called twice:
479 once for the candidate's base expr, and optionally again either for
480 the candidate's phi definition or for a CAND_REF's alternative base
484 find_basis_for_base_expr (slsr_cand_t c
, tree base_expr
)
486 cand_chain mapping_key
;
488 slsr_cand_t basis
= NULL
;
490 // Limit potential of N^2 behavior for long candidate chains.
492 int max_iters
= PARAM_VALUE (PARAM_MAX_SLSR_CANDIDATE_SCAN
);
494 mapping_key
.base_expr
= base_expr
;
495 chain
= base_cand_map
->find (&mapping_key
);
497 for (; chain
&& iters
< max_iters
; chain
= chain
->next
, ++iters
)
499 slsr_cand_t one_basis
= chain
->cand
;
501 if (one_basis
->kind
!= c
->kind
502 || one_basis
->cand_stmt
== c
->cand_stmt
503 || !operand_equal_p (one_basis
->stride
, c
->stride
, 0)
504 || !types_compatible_p (one_basis
->cand_type
, c
->cand_type
)
505 || !dominated_by_p (CDI_DOMINATORS
,
506 gimple_bb (c
->cand_stmt
),
507 gimple_bb (one_basis
->cand_stmt
)))
510 if (!basis
|| basis
->cand_num
< one_basis
->cand_num
)
517 /* Use the base expr from candidate C to look for possible candidates
518 that can serve as a basis for C. Each potential basis must also
519 appear in a block that dominates the candidate statement and have
520 the same stride and type. If more than one possible basis exists,
521 the one with highest index in the vector is chosen; this will be
522 the most immediately dominating basis. */
525 find_basis_for_candidate (slsr_cand_t c
)
527 slsr_cand_t basis
= find_basis_for_base_expr (c
, c
->base_expr
);
529 /* If a candidate doesn't have a basis using its base expression,
530 it may have a basis hidden by one or more intervening phis. */
531 if (!basis
&& c
->def_phi
)
533 basic_block basis_bb
, phi_bb
;
534 slsr_cand_t phi_cand
= lookup_cand (c
->def_phi
);
535 basis
= find_basis_for_base_expr (c
, phi_cand
->base_expr
);
539 /* A hidden basis must dominate the phi-definition of the
540 candidate's base name. */
541 phi_bb
= gimple_bb (phi_cand
->cand_stmt
);
542 basis_bb
= gimple_bb (basis
->cand_stmt
);
544 if (phi_bb
== basis_bb
545 || !dominated_by_p (CDI_DOMINATORS
, phi_bb
, basis_bb
))
551 /* If we found a hidden basis, estimate additional dead-code
552 savings if the phi and its feeding statements can be removed. */
553 if (basis
&& has_single_use (gimple_phi_result (phi_cand
->cand_stmt
)))
554 c
->dead_savings
+= phi_cand
->dead_savings
;
558 if (flag_expensive_optimizations
&& !basis
&& c
->kind
== CAND_REF
)
560 tree alt_base_expr
= get_alternative_base (c
->base_expr
);
562 basis
= find_basis_for_base_expr (c
, alt_base_expr
);
567 c
->sibling
= basis
->dependent
;
568 basis
->dependent
= c
->cand_num
;
569 return basis
->cand_num
;
575 /* Record a mapping from BASE to C, indicating that C may potentially serve
576 as a basis using that base expression. BASE may be the same as
577 C->BASE_EXPR; alternatively BASE can be a different tree that share the
578 underlining expression of C->BASE_EXPR. */
581 record_potential_basis (slsr_cand_t c
, tree base
)
588 node
= (cand_chain_t
) obstack_alloc (&chain_obstack
, sizeof (cand_chain
));
589 node
->base_expr
= base
;
592 slot
= base_cand_map
->find_slot (node
, INSERT
);
596 cand_chain_t head
= (cand_chain_t
) (*slot
);
597 node
->next
= head
->next
;
604 /* Allocate storage for a new candidate and initialize its fields.
605 Attempt to find a basis for the candidate.
607 For CAND_REF, an alternative base may also be recorded and used
608 to find a basis. This helps cases where the expression hidden
609 behind BASE (which is usually an SSA_NAME) has immediate offset,
613 a2[i + 20][j] = 2; */
616 alloc_cand_and_find_basis (enum cand_kind kind
, gimple
*gs
, tree base
,
617 const widest_int
&index
, tree stride
, tree ctype
,
620 slsr_cand_t c
= (slsr_cand_t
) obstack_alloc (&cand_obstack
,
626 c
->cand_type
= ctype
;
628 c
->cand_num
= cand_vec
.length () + 1;
632 c
->def_phi
= kind
== CAND_MULT
? find_phi_def (base
) : 0;
633 c
->dead_savings
= savings
;
635 cand_vec
.safe_push (c
);
637 if (kind
== CAND_PHI
)
640 c
->basis
= find_basis_for_candidate (c
);
642 record_potential_basis (c
, base
);
643 if (flag_expensive_optimizations
&& kind
== CAND_REF
)
645 tree alt_base
= get_alternative_base (base
);
647 record_potential_basis (c
, alt_base
);
653 /* Determine the target cost of statement GS when compiling according
657 stmt_cost (gimple
*gs
, bool speed
)
659 tree lhs
, rhs1
, rhs2
;
660 machine_mode lhs_mode
;
662 gcc_assert (is_gimple_assign (gs
));
663 lhs
= gimple_assign_lhs (gs
);
664 rhs1
= gimple_assign_rhs1 (gs
);
665 lhs_mode
= TYPE_MODE (TREE_TYPE (lhs
));
667 switch (gimple_assign_rhs_code (gs
))
670 rhs2
= gimple_assign_rhs2 (gs
);
672 if (tree_fits_shwi_p (rhs2
))
673 return mult_by_coeff_cost (tree_to_shwi (rhs2
), lhs_mode
, speed
);
675 gcc_assert (TREE_CODE (rhs1
) != INTEGER_CST
);
676 return mul_cost (speed
, lhs_mode
);
679 case POINTER_PLUS_EXPR
:
681 return add_cost (speed
, lhs_mode
);
684 return neg_cost (speed
, lhs_mode
);
687 return convert_cost (lhs_mode
, TYPE_MODE (TREE_TYPE (rhs1
)), speed
);
689 /* Note that we don't assign costs to copies that in most cases
702 /* Look up the defining statement for BASE_IN and return a pointer
703 to its candidate in the candidate table, if any; otherwise NULL.
704 Only CAND_ADD and CAND_MULT candidates are returned. */
707 base_cand_from_table (tree base_in
)
711 gimple
*def
= SSA_NAME_DEF_STMT (base_in
);
713 return (slsr_cand_t
) NULL
;
715 result
= stmt_cand_map
->get (def
);
717 if (result
&& (*result
)->kind
!= CAND_REF
)
720 return (slsr_cand_t
) NULL
;
723 /* Add an entry to the statement-to-candidate mapping. */
726 add_cand_for_stmt (gimple
*gs
, slsr_cand_t c
)
728 gcc_assert (!stmt_cand_map
->put (gs
, c
));
731 /* Given PHI which contains a phi statement, determine whether it
732 satisfies all the requirements of a phi candidate. If so, create
733 a candidate. Note that a CAND_PHI never has a basis itself, but
734 is used to help find a basis for subsequent candidates. */
737 slsr_process_phi (gphi
*phi
, bool speed
)
740 tree arg0_base
= NULL_TREE
, base_type
;
742 struct loop
*cand_loop
= gimple_bb (phi
)->loop_father
;
743 unsigned savings
= 0;
745 /* A CAND_PHI requires each of its arguments to have the same
746 derived base name. (See the module header commentary for a
747 definition of derived base names.) Furthermore, all feeding
748 definitions must be in the same position in the loop hierarchy
751 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
753 slsr_cand_t arg_cand
;
754 tree arg
= gimple_phi_arg_def (phi
, i
);
755 tree derived_base_name
= NULL_TREE
;
756 gimple
*arg_stmt
= NULL
;
757 basic_block arg_bb
= NULL
;
759 if (TREE_CODE (arg
) != SSA_NAME
)
762 arg_cand
= base_cand_from_table (arg
);
766 while (arg_cand
->kind
!= CAND_ADD
&& arg_cand
->kind
!= CAND_PHI
)
768 if (!arg_cand
->next_interp
)
771 arg_cand
= lookup_cand (arg_cand
->next_interp
);
774 if (!integer_onep (arg_cand
->stride
))
777 derived_base_name
= arg_cand
->base_expr
;
778 arg_stmt
= arg_cand
->cand_stmt
;
779 arg_bb
= gimple_bb (arg_stmt
);
781 /* Gather potential dead code savings if the phi statement
782 can be removed later on. */
783 if (has_single_use (arg
))
785 if (gimple_code (arg_stmt
) == GIMPLE_PHI
)
786 savings
+= arg_cand
->dead_savings
;
788 savings
+= stmt_cost (arg_stmt
, speed
);
791 else if (SSA_NAME_IS_DEFAULT_DEF (arg
))
793 derived_base_name
= arg
;
794 arg_bb
= single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun
));
797 if (!arg_bb
|| arg_bb
->loop_father
!= cand_loop
)
801 arg0_base
= derived_base_name
;
802 else if (!operand_equal_p (derived_base_name
, arg0_base
, 0))
806 /* Create the candidate. "alloc_cand_and_find_basis" is named
807 misleadingly for this case, as no basis will be sought for a
809 base_type
= TREE_TYPE (arg0_base
);
811 c
= alloc_cand_and_find_basis (CAND_PHI
, phi
, arg0_base
,
812 0, integer_one_node
, base_type
, savings
);
814 /* Add the candidate to the statement-candidate mapping. */
815 add_cand_for_stmt (phi
, c
);
818 /* Given PBASE which is a pointer to tree, look up the defining
819 statement for it and check whether the candidate is in the
822 X = B + (1 * S), S is integer constant
823 X = B + (i * S), S is integer one
825 If so, set PBASE to the candidate's base_expr and return double
827 Otherwise, just return double int zero. */
830 backtrace_base_for_ref (tree
*pbase
)
832 tree base_in
= *pbase
;
833 slsr_cand_t base_cand
;
835 STRIP_NOPS (base_in
);
837 /* Strip off widening conversion(s) to handle cases where
838 e.g. 'B' is widened from an 'int' in order to calculate
840 if (CONVERT_EXPR_P (base_in
)
841 && legal_cast_p_1 (base_in
, TREE_OPERAND (base_in
, 0)))
842 base_in
= get_unwidened (base_in
, NULL_TREE
);
844 if (TREE_CODE (base_in
) != SSA_NAME
)
847 base_cand
= base_cand_from_table (base_in
);
849 while (base_cand
&& base_cand
->kind
!= CAND_PHI
)
851 if (base_cand
->kind
== CAND_ADD
852 && base_cand
->index
== 1
853 && TREE_CODE (base_cand
->stride
) == INTEGER_CST
)
855 /* X = B + (1 * S), S is integer constant. */
856 *pbase
= base_cand
->base_expr
;
857 return wi::to_widest (base_cand
->stride
);
859 else if (base_cand
->kind
== CAND_ADD
860 && TREE_CODE (base_cand
->stride
) == INTEGER_CST
861 && integer_onep (base_cand
->stride
))
863 /* X = B + (i * S), S is integer one. */
864 *pbase
= base_cand
->base_expr
;
865 return base_cand
->index
;
868 if (base_cand
->next_interp
)
869 base_cand
= lookup_cand (base_cand
->next_interp
);
877 /* Look for the following pattern:
879 *PBASE: MEM_REF (T1, C1)
881 *POFFSET: MULT_EXPR (T2, C3) [C2 is zero]
883 MULT_EXPR (PLUS_EXPR (T2, C2), C3)
885 MULT_EXPR (MINUS_EXPR (T2, -C2), C3)
887 *PINDEX: C4 * BITS_PER_UNIT
889 If not present, leave the input values unchanged and return FALSE.
890 Otherwise, modify the input values as follows and return TRUE:
893 *POFFSET: MULT_EXPR (T2, C3)
894 *PINDEX: C1 + (C2 * C3) + C4
896 When T2 is recorded by a CAND_ADD in the form of (T2' + C5), it
897 will be further restructured to:
900 *POFFSET: MULT_EXPR (T2', C3)
901 *PINDEX: C1 + (C2 * C3) + C4 + (C5 * C3) */
904 restructure_reference (tree
*pbase
, tree
*poffset
, widest_int
*pindex
,
907 tree base
= *pbase
, offset
= *poffset
;
908 widest_int index
= *pindex
;
909 tree mult_op0
, t1
, t2
, type
;
910 widest_int c1
, c2
, c3
, c4
, c5
;
914 || TREE_CODE (base
) != MEM_REF
915 || TREE_CODE (offset
) != MULT_EXPR
916 || TREE_CODE (TREE_OPERAND (offset
, 1)) != INTEGER_CST
917 || wi::umod_floor (index
, BITS_PER_UNIT
) != 0)
920 t1
= TREE_OPERAND (base
, 0);
921 c1
= widest_int::from (mem_ref_offset (base
), SIGNED
);
922 type
= TREE_TYPE (TREE_OPERAND (base
, 1));
924 mult_op0
= TREE_OPERAND (offset
, 0);
925 c3
= wi::to_widest (TREE_OPERAND (offset
, 1));
927 if (TREE_CODE (mult_op0
) == PLUS_EXPR
)
929 if (TREE_CODE (TREE_OPERAND (mult_op0
, 1)) == INTEGER_CST
)
931 t2
= TREE_OPERAND (mult_op0
, 0);
932 c2
= wi::to_widest (TREE_OPERAND (mult_op0
, 1));
937 else if (TREE_CODE (mult_op0
) == MINUS_EXPR
)
939 if (TREE_CODE (TREE_OPERAND (mult_op0
, 1)) == INTEGER_CST
)
941 t2
= TREE_OPERAND (mult_op0
, 0);
942 c2
= -wi::to_widest (TREE_OPERAND (mult_op0
, 1));
953 c4
= index
>> LOG2_BITS_PER_UNIT
;
954 c5
= backtrace_base_for_ref (&t2
);
957 *poffset
= fold_build2 (MULT_EXPR
, sizetype
, fold_convert (sizetype
, t2
),
958 wide_int_to_tree (sizetype
, c3
));
959 *pindex
= c1
+ c2
* c3
+ c4
+ c5
* c3
;
965 /* Given GS which contains a data reference, create a CAND_REF entry in
966 the candidate table and attempt to find a basis. */
969 slsr_process_ref (gimple
*gs
)
971 tree ref_expr
, base
, offset
, type
;
972 HOST_WIDE_INT bitsize
, bitpos
;
974 int unsignedp
, reversep
, volatilep
;
977 if (gimple_vdef (gs
))
978 ref_expr
= gimple_assign_lhs (gs
);
980 ref_expr
= gimple_assign_rhs1 (gs
);
982 if (!handled_component_p (ref_expr
)
983 || TREE_CODE (ref_expr
) == BIT_FIELD_REF
984 || (TREE_CODE (ref_expr
) == COMPONENT_REF
985 && DECL_BIT_FIELD (TREE_OPERAND (ref_expr
, 1))))
988 base
= get_inner_reference (ref_expr
, &bitsize
, &bitpos
, &offset
, &mode
,
989 &unsignedp
, &reversep
, &volatilep
);
992 widest_int index
= bitpos
;
994 if (!restructure_reference (&base
, &offset
, &index
, &type
))
997 c
= alloc_cand_and_find_basis (CAND_REF
, gs
, base
, index
, offset
,
1000 /* Add the candidate to the statement-candidate mapping. */
1001 add_cand_for_stmt (gs
, c
);
1004 /* Create a candidate entry for a statement GS, where GS multiplies
1005 two SSA names BASE_IN and STRIDE_IN. Propagate any known information
1006 about the two SSA names into the new candidate. Return the new
1010 create_mul_ssa_cand (gimple
*gs
, tree base_in
, tree stride_in
, bool speed
)
1012 tree base
= NULL_TREE
, stride
= NULL_TREE
, ctype
= NULL_TREE
;
1014 unsigned savings
= 0;
1016 slsr_cand_t base_cand
= base_cand_from_table (base_in
);
1018 /* Look at all interpretations of the base candidate, if necessary,
1019 to find information to propagate into this candidate. */
1020 while (base_cand
&& !base
&& base_cand
->kind
!= CAND_PHI
)
1023 if (base_cand
->kind
== CAND_MULT
&& integer_onep (base_cand
->stride
))
1029 base
= base_cand
->base_expr
;
1030 index
= base_cand
->index
;
1032 ctype
= base_cand
->cand_type
;
1033 if (has_single_use (base_in
))
1034 savings
= (base_cand
->dead_savings
1035 + stmt_cost (base_cand
->cand_stmt
, speed
));
1037 else if (base_cand
->kind
== CAND_ADD
1038 && TREE_CODE (base_cand
->stride
) == INTEGER_CST
)
1040 /* Y = B + (i' * S), S constant
1042 ============================
1043 X = B + ((i' * S) * Z) */
1044 base
= base_cand
->base_expr
;
1045 index
= base_cand
->index
* wi::to_widest (base_cand
->stride
);
1047 ctype
= base_cand
->cand_type
;
1048 if (has_single_use (base_in
))
1049 savings
= (base_cand
->dead_savings
1050 + stmt_cost (base_cand
->cand_stmt
, speed
));
1053 if (base_cand
->next_interp
)
1054 base_cand
= lookup_cand (base_cand
->next_interp
);
1061 /* No interpretations had anything useful to propagate, so
1062 produce X = (Y + 0) * Z. */
1066 ctype
= TREE_TYPE (base_in
);
1069 c
= alloc_cand_and_find_basis (CAND_MULT
, gs
, base
, index
, stride
,
1074 /* Create a candidate entry for a statement GS, where GS multiplies
1075 SSA name BASE_IN by constant STRIDE_IN. Propagate any known
1076 information about BASE_IN into the new candidate. Return the new
1080 create_mul_imm_cand (gimple
*gs
, tree base_in
, tree stride_in
, bool speed
)
1082 tree base
= NULL_TREE
, stride
= NULL_TREE
, ctype
= NULL_TREE
;
1083 widest_int index
, temp
;
1084 unsigned savings
= 0;
1086 slsr_cand_t base_cand
= base_cand_from_table (base_in
);
1088 /* Look at all interpretations of the base candidate, if necessary,
1089 to find information to propagate into this candidate. */
1090 while (base_cand
&& !base
&& base_cand
->kind
!= CAND_PHI
)
1092 if (base_cand
->kind
== CAND_MULT
1093 && TREE_CODE (base_cand
->stride
) == INTEGER_CST
)
1095 /* Y = (B + i') * S, S constant
1097 ============================
1098 X = (B + i') * (S * c) */
1099 temp
= wi::to_widest (base_cand
->stride
) * wi::to_widest (stride_in
);
1100 if (wi::fits_to_tree_p (temp
, TREE_TYPE (stride_in
)))
1102 base
= base_cand
->base_expr
;
1103 index
= base_cand
->index
;
1104 stride
= wide_int_to_tree (TREE_TYPE (stride_in
), temp
);
1105 ctype
= base_cand
->cand_type
;
1106 if (has_single_use (base_in
))
1107 savings
= (base_cand
->dead_savings
1108 + stmt_cost (base_cand
->cand_stmt
, speed
));
1111 else if (base_cand
->kind
== CAND_ADD
&& integer_onep (base_cand
->stride
))
1115 ===========================
1117 base
= base_cand
->base_expr
;
1118 index
= base_cand
->index
;
1120 ctype
= base_cand
->cand_type
;
1121 if (has_single_use (base_in
))
1122 savings
= (base_cand
->dead_savings
1123 + stmt_cost (base_cand
->cand_stmt
, speed
));
1125 else if (base_cand
->kind
== CAND_ADD
1126 && base_cand
->index
== 1
1127 && TREE_CODE (base_cand
->stride
) == INTEGER_CST
)
1129 /* Y = B + (1 * S), S constant
1131 ===========================
1133 base
= base_cand
->base_expr
;
1134 index
= wi::to_widest (base_cand
->stride
);
1136 ctype
= base_cand
->cand_type
;
1137 if (has_single_use (base_in
))
1138 savings
= (base_cand
->dead_savings
1139 + stmt_cost (base_cand
->cand_stmt
, speed
));
1142 if (base_cand
->next_interp
)
1143 base_cand
= lookup_cand (base_cand
->next_interp
);
1150 /* No interpretations had anything useful to propagate, so
1151 produce X = (Y + 0) * c. */
1155 ctype
= TREE_TYPE (base_in
);
1158 c
= alloc_cand_and_find_basis (CAND_MULT
, gs
, base
, index
, stride
,
1163 /* Given GS which is a multiply of scalar integers, make an appropriate
1164 entry in the candidate table. If this is a multiply of two SSA names,
1165 create two CAND_MULT interpretations and attempt to find a basis for
1166 each of them. Otherwise, create a single CAND_MULT and attempt to
1170 slsr_process_mul (gimple
*gs
, tree rhs1
, tree rhs2
, bool speed
)
1174 /* If this is a multiply of an SSA name with itself, it is highly
1175 unlikely that we will get a strength reduction opportunity, so
1176 don't record it as a candidate. This simplifies the logic for
1177 finding a basis, so if this is removed that must be considered. */
1181 if (TREE_CODE (rhs2
) == SSA_NAME
)
1183 /* Record an interpretation of this statement in the candidate table
1184 assuming RHS1 is the base expression and RHS2 is the stride. */
1185 c
= create_mul_ssa_cand (gs
, rhs1
, rhs2
, speed
);
1187 /* Add the first interpretation to the statement-candidate mapping. */
1188 add_cand_for_stmt (gs
, c
);
1190 /* Record another interpretation of this statement assuming RHS1
1191 is the stride and RHS2 is the base expression. */
1192 c2
= create_mul_ssa_cand (gs
, rhs2
, rhs1
, speed
);
1193 c
->next_interp
= c2
->cand_num
;
1197 /* Record an interpretation for the multiply-immediate. */
1198 c
= create_mul_imm_cand (gs
, rhs1
, rhs2
, speed
);
1200 /* Add the interpretation to the statement-candidate mapping. */
1201 add_cand_for_stmt (gs
, c
);
1205 /* Create a candidate entry for a statement GS, where GS adds two
1206 SSA names BASE_IN and ADDEND_IN if SUBTRACT_P is false, and
1207 subtracts ADDEND_IN from BASE_IN otherwise. Propagate any known
1208 information about the two SSA names into the new candidate.
1209 Return the new candidate. */
1212 create_add_ssa_cand (gimple
*gs
, tree base_in
, tree addend_in
,
1213 bool subtract_p
, bool speed
)
1215 tree base
= NULL_TREE
, stride
= NULL_TREE
, ctype
= NULL
;
1217 unsigned savings
= 0;
1219 slsr_cand_t base_cand
= base_cand_from_table (base_in
);
1220 slsr_cand_t addend_cand
= base_cand_from_table (addend_in
);
1222 /* The most useful transformation is a multiply-immediate feeding
1223 an add or subtract. Look for that first. */
1224 while (addend_cand
&& !base
&& addend_cand
->kind
!= CAND_PHI
)
1226 if (addend_cand
->kind
== CAND_MULT
1227 && addend_cand
->index
== 0
1228 && TREE_CODE (addend_cand
->stride
) == INTEGER_CST
)
1230 /* Z = (B + 0) * S, S constant
1232 ===========================
1233 X = Y + ((+/-1 * S) * B) */
1235 index
= wi::to_widest (addend_cand
->stride
);
1238 stride
= addend_cand
->base_expr
;
1239 ctype
= TREE_TYPE (base_in
);
1240 if (has_single_use (addend_in
))
1241 savings
= (addend_cand
->dead_savings
1242 + stmt_cost (addend_cand
->cand_stmt
, speed
));
1245 if (addend_cand
->next_interp
)
1246 addend_cand
= lookup_cand (addend_cand
->next_interp
);
1251 while (base_cand
&& !base
&& base_cand
->kind
!= CAND_PHI
)
1253 if (base_cand
->kind
== CAND_ADD
1254 && (base_cand
->index
== 0
1255 || operand_equal_p (base_cand
->stride
,
1256 integer_zero_node
, 0)))
1258 /* Y = B + (i' * S), i' * S = 0
1260 ============================
1261 X = B + (+/-1 * Z) */
1262 base
= base_cand
->base_expr
;
1263 index
= subtract_p
? -1 : 1;
1265 ctype
= base_cand
->cand_type
;
1266 if (has_single_use (base_in
))
1267 savings
= (base_cand
->dead_savings
1268 + stmt_cost (base_cand
->cand_stmt
, speed
));
1270 else if (subtract_p
)
1272 slsr_cand_t subtrahend_cand
= base_cand_from_table (addend_in
);
1274 while (subtrahend_cand
&& !base
&& subtrahend_cand
->kind
!= CAND_PHI
)
1276 if (subtrahend_cand
->kind
== CAND_MULT
1277 && subtrahend_cand
->index
== 0
1278 && TREE_CODE (subtrahend_cand
->stride
) == INTEGER_CST
)
1280 /* Z = (B + 0) * S, S constant
1282 ===========================
1283 Value: X = Y + ((-1 * S) * B) */
1285 index
= wi::to_widest (subtrahend_cand
->stride
);
1287 stride
= subtrahend_cand
->base_expr
;
1288 ctype
= TREE_TYPE (base_in
);
1289 if (has_single_use (addend_in
))
1290 savings
= (subtrahend_cand
->dead_savings
1291 + stmt_cost (subtrahend_cand
->cand_stmt
, speed
));
1294 if (subtrahend_cand
->next_interp
)
1295 subtrahend_cand
= lookup_cand (subtrahend_cand
->next_interp
);
1297 subtrahend_cand
= NULL
;
1301 if (base_cand
->next_interp
)
1302 base_cand
= lookup_cand (base_cand
->next_interp
);
1309 /* No interpretations had anything useful to propagate, so
1310 produce X = Y + (1 * Z). */
1312 index
= subtract_p
? -1 : 1;
1314 ctype
= TREE_TYPE (base_in
);
1317 c
= alloc_cand_and_find_basis (CAND_ADD
, gs
, base
, index
, stride
,
1322 /* Create a candidate entry for a statement GS, where GS adds SSA
1323 name BASE_IN to constant INDEX_IN. Propagate any known information
1324 about BASE_IN into the new candidate. Return the new candidate. */
1327 create_add_imm_cand (gimple
*gs
, tree base_in
, const widest_int
&index_in
,
1330 enum cand_kind kind
= CAND_ADD
;
1331 tree base
= NULL_TREE
, stride
= NULL_TREE
, ctype
= NULL_TREE
;
1332 widest_int index
, multiple
;
1333 unsigned savings
= 0;
1335 slsr_cand_t base_cand
= base_cand_from_table (base_in
);
1337 while (base_cand
&& !base
&& base_cand
->kind
!= CAND_PHI
)
1339 signop sign
= TYPE_SIGN (TREE_TYPE (base_cand
->stride
));
1341 if (TREE_CODE (base_cand
->stride
) == INTEGER_CST
1342 && wi::multiple_of_p (index_in
, wi::to_widest (base_cand
->stride
),
1345 /* Y = (B + i') * S, S constant, c = kS for some integer k
1347 ============================
1348 X = (B + (i'+ k)) * S
1350 Y = B + (i' * S), S constant, c = kS for some integer k
1352 ============================
1353 X = (B + (i'+ k)) * S */
1354 kind
= base_cand
->kind
;
1355 base
= base_cand
->base_expr
;
1356 index
= base_cand
->index
+ multiple
;
1357 stride
= base_cand
->stride
;
1358 ctype
= base_cand
->cand_type
;
1359 if (has_single_use (base_in
))
1360 savings
= (base_cand
->dead_savings
1361 + stmt_cost (base_cand
->cand_stmt
, speed
));
1364 if (base_cand
->next_interp
)
1365 base_cand
= lookup_cand (base_cand
->next_interp
);
1372 /* No interpretations had anything useful to propagate, so
1373 produce X = Y + (c * 1). */
1377 stride
= integer_one_node
;
1378 ctype
= TREE_TYPE (base_in
);
1381 c
= alloc_cand_and_find_basis (kind
, gs
, base
, index
, stride
,
1386 /* Given GS which is an add or subtract of scalar integers or pointers,
1387 make at least one appropriate entry in the candidate table. */
1390 slsr_process_add (gimple
*gs
, tree rhs1
, tree rhs2
, bool speed
)
1392 bool subtract_p
= gimple_assign_rhs_code (gs
) == MINUS_EXPR
;
1393 slsr_cand_t c
= NULL
, c2
;
1395 if (TREE_CODE (rhs2
) == SSA_NAME
)
1397 /* First record an interpretation assuming RHS1 is the base expression
1398 and RHS2 is the stride. But it doesn't make sense for the
1399 stride to be a pointer, so don't record a candidate in that case. */
1400 if (!POINTER_TYPE_P (TREE_TYPE (rhs2
)))
1402 c
= create_add_ssa_cand (gs
, rhs1
, rhs2
, subtract_p
, speed
);
1404 /* Add the first interpretation to the statement-candidate
1406 add_cand_for_stmt (gs
, c
);
1409 /* If the two RHS operands are identical, or this is a subtract,
1411 if (operand_equal_p (rhs1
, rhs2
, 0) || subtract_p
)
1414 /* Otherwise, record another interpretation assuming RHS2 is the
1415 base expression and RHS1 is the stride, again provided that the
1416 stride is not a pointer. */
1417 if (!POINTER_TYPE_P (TREE_TYPE (rhs1
)))
1419 c2
= create_add_ssa_cand (gs
, rhs2
, rhs1
, false, speed
);
1421 c
->next_interp
= c2
->cand_num
;
1423 add_cand_for_stmt (gs
, c2
);
1428 /* Record an interpretation for the add-immediate. */
1429 widest_int index
= wi::to_widest (rhs2
);
1433 c
= create_add_imm_cand (gs
, rhs1
, index
, speed
);
1435 /* Add the interpretation to the statement-candidate mapping. */
1436 add_cand_for_stmt (gs
, c
);
1440 /* Given GS which is a negate of a scalar integer, make an appropriate
1441 entry in the candidate table. A negate is equivalent to a multiply
1445 slsr_process_neg (gimple
*gs
, tree rhs1
, bool speed
)
1447 /* Record a CAND_MULT interpretation for the multiply by -1. */
1448 slsr_cand_t c
= create_mul_imm_cand (gs
, rhs1
, integer_minus_one_node
, speed
);
1450 /* Add the interpretation to the statement-candidate mapping. */
1451 add_cand_for_stmt (gs
, c
);
1454 /* Help function for legal_cast_p, operating on two trees. Checks
1455 whether it's allowable to cast from RHS to LHS. See legal_cast_p
1456 for more details. */
1459 legal_cast_p_1 (tree lhs
, tree rhs
)
1461 tree lhs_type
, rhs_type
;
1462 unsigned lhs_size
, rhs_size
;
1463 bool lhs_wraps
, rhs_wraps
;
1465 lhs_type
= TREE_TYPE (lhs
);
1466 rhs_type
= TREE_TYPE (rhs
);
1467 lhs_size
= TYPE_PRECISION (lhs_type
);
1468 rhs_size
= TYPE_PRECISION (rhs_type
);
1469 lhs_wraps
= ANY_INTEGRAL_TYPE_P (lhs_type
) && TYPE_OVERFLOW_WRAPS (lhs_type
);
1470 rhs_wraps
= ANY_INTEGRAL_TYPE_P (rhs_type
) && TYPE_OVERFLOW_WRAPS (rhs_type
);
1472 if (lhs_size
< rhs_size
1473 || (rhs_wraps
&& !lhs_wraps
)
1474 || (rhs_wraps
&& lhs_wraps
&& rhs_size
!= lhs_size
))
1480 /* Return TRUE if GS is a statement that defines an SSA name from
1481 a conversion and is legal for us to combine with an add and multiply
1482 in the candidate table. For example, suppose we have:
1488 Without the type-cast, we would create a CAND_MULT for D with base B,
1489 index i, and stride S. We want to record this candidate only if it
1490 is equivalent to apply the type cast following the multiply:
1496 We will record the type with the candidate for D. This allows us
1497 to use a similar previous candidate as a basis. If we have earlier seen
1503 we can replace D with
1505 D = D' + (i - i') * S;
1507 But if moving the type-cast would change semantics, we mustn't do this.
1509 This is legitimate for casts from a non-wrapping integral type to
1510 any integral type of the same or larger size. It is not legitimate
1511 to convert a wrapping type to a non-wrapping type, or to a wrapping
1512 type of a different size. I.e., with a wrapping type, we must
1513 assume that the addition B + i could wrap, in which case performing
1514 the multiply before or after one of the "illegal" type casts will
1515 have different semantics. */
1518 legal_cast_p (gimple
*gs
, tree rhs
)
1520 if (!is_gimple_assign (gs
)
1521 || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (gs
)))
1524 return legal_cast_p_1 (gimple_assign_lhs (gs
), rhs
);
1527 /* Given GS which is a cast to a scalar integer type, determine whether
1528 the cast is legal for strength reduction. If so, make at least one
1529 appropriate entry in the candidate table. */
1532 slsr_process_cast (gimple
*gs
, tree rhs1
, bool speed
)
1535 slsr_cand_t base_cand
, c
= NULL
, c2
;
1536 unsigned savings
= 0;
1538 if (!legal_cast_p (gs
, rhs1
))
1541 lhs
= gimple_assign_lhs (gs
);
1542 base_cand
= base_cand_from_table (rhs1
);
1543 ctype
= TREE_TYPE (lhs
);
1545 if (base_cand
&& base_cand
->kind
!= CAND_PHI
)
1549 /* Propagate all data from the base candidate except the type,
1550 which comes from the cast, and the base candidate's cast,
1551 which is no longer applicable. */
1552 if (has_single_use (rhs1
))
1553 savings
= (base_cand
->dead_savings
1554 + stmt_cost (base_cand
->cand_stmt
, speed
));
1556 c
= alloc_cand_and_find_basis (base_cand
->kind
, gs
,
1557 base_cand
->base_expr
,
1558 base_cand
->index
, base_cand
->stride
,
1560 if (base_cand
->next_interp
)
1561 base_cand
= lookup_cand (base_cand
->next_interp
);
1568 /* If nothing is known about the RHS, create fresh CAND_ADD and
1569 CAND_MULT interpretations:
1574 The first of these is somewhat arbitrary, but the choice of
1575 1 for the stride simplifies the logic for propagating casts
1577 c
= alloc_cand_and_find_basis (CAND_ADD
, gs
, rhs1
,
1578 0, integer_one_node
, ctype
, 0);
1579 c2
= alloc_cand_and_find_basis (CAND_MULT
, gs
, rhs1
,
1580 0, integer_one_node
, ctype
, 0);
1581 c
->next_interp
= c2
->cand_num
;
1584 /* Add the first (or only) interpretation to the statement-candidate
1586 add_cand_for_stmt (gs
, c
);
1589 /* Given GS which is a copy of a scalar integer type, make at least one
1590 appropriate entry in the candidate table.
1592 This interface is included for completeness, but is unnecessary
1593 if this pass immediately follows a pass that performs copy
1594 propagation, such as DOM. */
1597 slsr_process_copy (gimple
*gs
, tree rhs1
, bool speed
)
1599 slsr_cand_t base_cand
, c
= NULL
, c2
;
1600 unsigned savings
= 0;
1602 base_cand
= base_cand_from_table (rhs1
);
1604 if (base_cand
&& base_cand
->kind
!= CAND_PHI
)
1608 /* Propagate all data from the base candidate. */
1609 if (has_single_use (rhs1
))
1610 savings
= (base_cand
->dead_savings
1611 + stmt_cost (base_cand
->cand_stmt
, speed
));
1613 c
= alloc_cand_and_find_basis (base_cand
->kind
, gs
,
1614 base_cand
->base_expr
,
1615 base_cand
->index
, base_cand
->stride
,
1616 base_cand
->cand_type
, savings
);
1617 if (base_cand
->next_interp
)
1618 base_cand
= lookup_cand (base_cand
->next_interp
);
1625 /* If nothing is known about the RHS, create fresh CAND_ADD and
1626 CAND_MULT interpretations:
1631 The first of these is somewhat arbitrary, but the choice of
1632 1 for the stride simplifies the logic for propagating casts
1634 c
= alloc_cand_and_find_basis (CAND_ADD
, gs
, rhs1
,
1635 0, integer_one_node
, TREE_TYPE (rhs1
), 0);
1636 c2
= alloc_cand_and_find_basis (CAND_MULT
, gs
, rhs1
,
1637 0, integer_one_node
, TREE_TYPE (rhs1
), 0);
1638 c
->next_interp
= c2
->cand_num
;
1641 /* Add the first (or only) interpretation to the statement-candidate
1643 add_cand_for_stmt (gs
, c
);
1646 class find_candidates_dom_walker
: public dom_walker
1649 find_candidates_dom_walker (cdi_direction direction
)
1650 : dom_walker (direction
) {}
1651 virtual edge
before_dom_children (basic_block
);
1654 /* Find strength-reduction candidates in block BB. */
1657 find_candidates_dom_walker::before_dom_children (basic_block bb
)
1659 bool speed
= optimize_bb_for_speed_p (bb
);
1661 for (gphi_iterator gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
);
1663 slsr_process_phi (gsi
.phi (), speed
);
1665 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);
1668 gimple
*gs
= gsi_stmt (gsi
);
1670 if (gimple_vuse (gs
) && gimple_assign_single_p (gs
))
1671 slsr_process_ref (gs
);
1673 else if (is_gimple_assign (gs
)
1674 && SCALAR_INT_MODE_P
1675 (TYPE_MODE (TREE_TYPE (gimple_assign_lhs (gs
)))))
1677 tree rhs1
= NULL_TREE
, rhs2
= NULL_TREE
;
1679 switch (gimple_assign_rhs_code (gs
))
1683 rhs1
= gimple_assign_rhs1 (gs
);
1684 rhs2
= gimple_assign_rhs2 (gs
);
1685 /* Should never happen, but currently some buggy situations
1686 in earlier phases put constants in rhs1. */
1687 if (TREE_CODE (rhs1
) != SSA_NAME
)
1691 /* Possible future opportunity: rhs1 of a ptr+ can be
1693 case POINTER_PLUS_EXPR
:
1695 rhs2
= gimple_assign_rhs2 (gs
);
1701 rhs1
= gimple_assign_rhs1 (gs
);
1702 if (TREE_CODE (rhs1
) != SSA_NAME
)
1710 switch (gimple_assign_rhs_code (gs
))
1713 slsr_process_mul (gs
, rhs1
, rhs2
, speed
);
1717 case POINTER_PLUS_EXPR
:
1719 slsr_process_add (gs
, rhs1
, rhs2
, speed
);
1723 slsr_process_neg (gs
, rhs1
, speed
);
1727 slsr_process_cast (gs
, rhs1
, speed
);
1731 slsr_process_copy (gs
, rhs1
, speed
);
1742 /* Dump a candidate for debug. */
1745 dump_candidate (slsr_cand_t c
)
1747 fprintf (dump_file
, "%3d [%d] ", c
->cand_num
,
1748 gimple_bb (c
->cand_stmt
)->index
);
1749 print_gimple_stmt (dump_file
, c
->cand_stmt
, 0, 0);
1753 fputs (" MULT : (", dump_file
);
1754 print_generic_expr (dump_file
, c
->base_expr
, 0);
1755 fputs (" + ", dump_file
);
1756 print_decs (c
->index
, dump_file
);
1757 fputs (") * ", dump_file
);
1758 print_generic_expr (dump_file
, c
->stride
, 0);
1759 fputs (" : ", dump_file
);
1762 fputs (" ADD : ", 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 (" REF : ", dump_file
);
1772 print_generic_expr (dump_file
, c
->base_expr
, 0);
1773 fputs (" + (", dump_file
);
1774 print_generic_expr (dump_file
, c
->stride
, 0);
1775 fputs (") + ", dump_file
);
1776 print_decs (c
->index
, dump_file
);
1777 fputs (" : ", dump_file
);
1780 fputs (" PHI : ", dump_file
);
1781 print_generic_expr (dump_file
, c
->base_expr
, 0);
1782 fputs (" + (unknown * ", dump_file
);
1783 print_generic_expr (dump_file
, c
->stride
, 0);
1784 fputs (") : ", dump_file
);
1789 print_generic_expr (dump_file
, c
->cand_type
, 0);
1790 fprintf (dump_file
, "\n basis: %d dependent: %d sibling: %d\n",
1791 c
->basis
, c
->dependent
, c
->sibling
);
1792 fprintf (dump_file
, " next-interp: %d dead-savings: %d\n",
1793 c
->next_interp
, c
->dead_savings
);
1795 fprintf (dump_file
, " phi: %d\n", c
->def_phi
);
1796 fputs ("\n", dump_file
);
1799 /* Dump the candidate vector for debug. */
1802 dump_cand_vec (void)
1807 fprintf (dump_file
, "\nStrength reduction candidate vector:\n\n");
1809 FOR_EACH_VEC_ELT (cand_vec
, i
, c
)
1813 /* Callback used to dump the candidate chains hash table. */
1816 ssa_base_cand_dump_callback (cand_chain
**slot
, void *ignored ATTRIBUTE_UNUSED
)
1818 const_cand_chain_t chain
= *slot
;
1821 print_generic_expr (dump_file
, chain
->base_expr
, 0);
1822 fprintf (dump_file
, " -> %d", chain
->cand
->cand_num
);
1824 for (p
= chain
->next
; p
; p
= p
->next
)
1825 fprintf (dump_file
, " -> %d", p
->cand
->cand_num
);
1827 fputs ("\n", dump_file
);
1831 /* Dump the candidate chains. */
1834 dump_cand_chains (void)
1836 fprintf (dump_file
, "\nStrength reduction candidate chains:\n\n");
1837 base_cand_map
->traverse_noresize
<void *, ssa_base_cand_dump_callback
>
1839 fputs ("\n", dump_file
);
1842 /* Dump the increment vector for debug. */
1845 dump_incr_vec (void)
1847 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1851 fprintf (dump_file
, "\nIncrement vector:\n\n");
1853 for (i
= 0; i
< incr_vec_len
; i
++)
1855 fprintf (dump_file
, "%3d increment: ", i
);
1856 print_decs (incr_vec
[i
].incr
, dump_file
);
1857 fprintf (dump_file
, "\n count: %d", incr_vec
[i
].count
);
1858 fprintf (dump_file
, "\n cost: %d", incr_vec
[i
].cost
);
1859 fputs ("\n initializer: ", dump_file
);
1860 print_generic_expr (dump_file
, incr_vec
[i
].initializer
, 0);
1861 fputs ("\n\n", dump_file
);
1866 /* Replace *EXPR in candidate C with an equivalent strength-reduced
1870 replace_ref (tree
*expr
, slsr_cand_t c
)
1872 tree add_expr
, mem_ref
, acc_type
= TREE_TYPE (*expr
);
1873 unsigned HOST_WIDE_INT misalign
;
1876 /* Ensure the memory reference carries the minimum alignment
1877 requirement for the data type. See PR58041. */
1878 get_object_alignment_1 (*expr
, &align
, &misalign
);
1880 align
= least_bit_hwi (misalign
);
1881 if (align
< TYPE_ALIGN (acc_type
))
1882 acc_type
= build_aligned_type (acc_type
, align
);
1884 add_expr
= fold_build2 (POINTER_PLUS_EXPR
, TREE_TYPE (c
->base_expr
),
1885 c
->base_expr
, c
->stride
);
1886 mem_ref
= fold_build2 (MEM_REF
, acc_type
, add_expr
,
1887 wide_int_to_tree (c
->cand_type
, c
->index
));
1889 /* Gimplify the base addressing expression for the new MEM_REF tree. */
1890 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
1891 TREE_OPERAND (mem_ref
, 0)
1892 = force_gimple_operand_gsi (&gsi
, TREE_OPERAND (mem_ref
, 0),
1893 /*simple_p=*/true, NULL
,
1894 /*before=*/true, GSI_SAME_STMT
);
1895 copy_ref_info (mem_ref
, *expr
);
1897 update_stmt (c
->cand_stmt
);
1900 /* Replace CAND_REF candidate C, each sibling of candidate C, and each
1901 dependent of candidate C with an equivalent strength-reduced data
1905 replace_refs (slsr_cand_t c
)
1907 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1909 fputs ("Replacing reference: ", dump_file
);
1910 print_gimple_stmt (dump_file
, c
->cand_stmt
, 0, 0);
1913 if (gimple_vdef (c
->cand_stmt
))
1915 tree
*lhs
= gimple_assign_lhs_ptr (c
->cand_stmt
);
1916 replace_ref (lhs
, c
);
1920 tree
*rhs
= gimple_assign_rhs1_ptr (c
->cand_stmt
);
1921 replace_ref (rhs
, c
);
1924 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1926 fputs ("With: ", dump_file
);
1927 print_gimple_stmt (dump_file
, c
->cand_stmt
, 0, 0);
1928 fputs ("\n", dump_file
);
1932 replace_refs (lookup_cand (c
->sibling
));
1935 replace_refs (lookup_cand (c
->dependent
));
1938 /* Return TRUE if candidate C is dependent upon a PHI. */
1941 phi_dependent_cand_p (slsr_cand_t c
)
1943 /* A candidate is not necessarily dependent upon a PHI just because
1944 it has a phi definition for its base name. It may have a basis
1945 that relies upon the same phi definition, in which case the PHI
1946 is irrelevant to this candidate. */
1949 && lookup_cand (c
->basis
)->def_phi
!= c
->def_phi
);
1952 /* Calculate the increment required for candidate C relative to
1956 cand_increment (slsr_cand_t c
)
1960 /* If the candidate doesn't have a basis, just return its own
1961 index. This is useful in record_increments to help us find
1962 an existing initializer. Also, if the candidate's basis is
1963 hidden by a phi, then its own index will be the increment
1964 from the newly introduced phi basis. */
1965 if (!c
->basis
|| phi_dependent_cand_p (c
))
1968 basis
= lookup_cand (c
->basis
);
1969 gcc_assert (operand_equal_p (c
->base_expr
, basis
->base_expr
, 0));
1970 return c
->index
- basis
->index
;
1973 /* Calculate the increment required for candidate C relative to
1974 its basis. If we aren't going to generate pointer arithmetic
1975 for this candidate, return the absolute value of that increment
1978 static inline widest_int
1979 cand_abs_increment (slsr_cand_t c
)
1981 widest_int increment
= cand_increment (c
);
1983 if (!address_arithmetic_p
&& wi::neg_p (increment
))
1984 increment
= -increment
;
1989 /* Return TRUE iff candidate C has already been replaced under
1990 another interpretation. */
1993 cand_already_replaced (slsr_cand_t c
)
1995 return (gimple_bb (c
->cand_stmt
) == 0);
1998 /* Common logic used by replace_unconditional_candidate and
1999 replace_conditional_candidate. */
2002 replace_mult_candidate (slsr_cand_t c
, tree basis_name
, widest_int bump
)
2004 tree target_type
= TREE_TYPE (gimple_assign_lhs (c
->cand_stmt
));
2005 enum tree_code cand_code
= gimple_assign_rhs_code (c
->cand_stmt
);
2007 /* It is highly unlikely, but possible, that the resulting
2008 bump doesn't fit in a HWI. Abandon the replacement
2009 in this case. This does not affect siblings or dependents
2010 of C. Restriction to signed HWI is conservative for unsigned
2011 types but allows for safe negation without twisted logic. */
2012 if (wi::fits_shwi_p (bump
)
2013 && bump
.to_shwi () != HOST_WIDE_INT_MIN
2014 /* It is not useful to replace casts, copies, or adds of
2015 an SSA name and a constant. */
2016 && cand_code
!= SSA_NAME
2017 && !CONVERT_EXPR_CODE_P (cand_code
)
2018 && cand_code
!= PLUS_EXPR
2019 && cand_code
!= POINTER_PLUS_EXPR
2020 && cand_code
!= MINUS_EXPR
)
2022 enum tree_code code
= PLUS_EXPR
;
2024 gimple
*stmt_to_print
= NULL
;
2026 /* If the basis name and the candidate's LHS have incompatible
2027 types, introduce a cast. */
2028 if (!useless_type_conversion_p (target_type
, TREE_TYPE (basis_name
)))
2029 basis_name
= introduce_cast_before_cand (c
, target_type
, basis_name
);
2030 if (wi::neg_p (bump
))
2036 bump_tree
= wide_int_to_tree (target_type
, bump
);
2038 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2040 fputs ("Replacing: ", dump_file
);
2041 print_gimple_stmt (dump_file
, c
->cand_stmt
, 0, 0);
2046 tree lhs
= gimple_assign_lhs (c
->cand_stmt
);
2047 gassign
*copy_stmt
= gimple_build_assign (lhs
, basis_name
);
2048 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
2049 gimple_set_location (copy_stmt
, gimple_location (c
->cand_stmt
));
2050 gsi_replace (&gsi
, copy_stmt
, false);
2051 c
->cand_stmt
= copy_stmt
;
2052 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2053 stmt_to_print
= copy_stmt
;
2058 if (cand_code
!= NEGATE_EXPR
) {
2059 rhs1
= gimple_assign_rhs1 (c
->cand_stmt
);
2060 rhs2
= gimple_assign_rhs2 (c
->cand_stmt
);
2062 if (cand_code
!= NEGATE_EXPR
2063 && ((operand_equal_p (rhs1
, basis_name
, 0)
2064 && operand_equal_p (rhs2
, bump_tree
, 0))
2065 || (operand_equal_p (rhs1
, bump_tree
, 0)
2066 && operand_equal_p (rhs2
, basis_name
, 0))))
2068 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2070 fputs ("(duplicate, not actually replacing)", dump_file
);
2071 stmt_to_print
= c
->cand_stmt
;
2076 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
2077 gimple_assign_set_rhs_with_ops (&gsi
, code
,
2078 basis_name
, bump_tree
);
2079 update_stmt (gsi_stmt (gsi
));
2080 c
->cand_stmt
= gsi_stmt (gsi
);
2081 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2082 stmt_to_print
= gsi_stmt (gsi
);
2086 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2088 fputs ("With: ", dump_file
);
2089 print_gimple_stmt (dump_file
, stmt_to_print
, 0, 0);
2090 fputs ("\n", dump_file
);
2095 /* Replace candidate C with an add or subtract. Note that we only
2096 operate on CAND_MULTs with known strides, so we will never generate
2097 a POINTER_PLUS_EXPR. Each candidate X = (B + i) * S is replaced by
2098 X = Y + ((i - i') * S), as described in the module commentary. The
2099 folded value ((i - i') * S) is referred to here as the "bump." */
2102 replace_unconditional_candidate (slsr_cand_t c
)
2106 if (cand_already_replaced (c
))
2109 basis
= lookup_cand (c
->basis
);
2110 widest_int bump
= cand_increment (c
) * wi::to_widest (c
->stride
);
2112 replace_mult_candidate (c
, gimple_assign_lhs (basis
->cand_stmt
), bump
);
2115 /* Return the index in the increment vector of the given INCREMENT,
2116 or -1 if not found. The latter can occur if more than
2117 MAX_INCR_VEC_LEN increments have been found. */
2120 incr_vec_index (const widest_int
&increment
)
2124 for (i
= 0; i
< incr_vec_len
&& increment
!= incr_vec
[i
].incr
; i
++)
2127 if (i
< incr_vec_len
)
2133 /* Create a new statement along edge E to add BASIS_NAME to the product
2134 of INCREMENT and the stride of candidate C. Create and return a new
2135 SSA name from *VAR to be used as the LHS of the new statement.
2136 KNOWN_STRIDE is true iff C's stride is a constant. */
2139 create_add_on_incoming_edge (slsr_cand_t c
, tree basis_name
,
2140 widest_int increment
, edge e
, location_t loc
,
2143 basic_block insert_bb
;
2144 gimple_stmt_iterator gsi
;
2145 tree lhs
, basis_type
;
2148 /* If the add candidate along this incoming edge has the same
2149 index as C's hidden basis, the hidden basis represents this
2154 basis_type
= TREE_TYPE (basis_name
);
2155 lhs
= make_temp_ssa_name (basis_type
, NULL
, "slsr");
2160 enum tree_code code
= PLUS_EXPR
;
2161 widest_int bump
= increment
* wi::to_widest (c
->stride
);
2162 if (wi::neg_p (bump
))
2168 bump_tree
= wide_int_to_tree (basis_type
, bump
);
2169 new_stmt
= gimple_build_assign (lhs
, code
, basis_name
, bump_tree
);
2174 bool negate_incr
= (!address_arithmetic_p
&& wi::neg_p (increment
));
2175 i
= incr_vec_index (negate_incr
? -increment
: increment
);
2176 gcc_assert (i
>= 0);
2178 if (incr_vec
[i
].initializer
)
2180 enum tree_code code
= negate_incr
? MINUS_EXPR
: PLUS_EXPR
;
2181 new_stmt
= gimple_build_assign (lhs
, code
, basis_name
,
2182 incr_vec
[i
].initializer
);
2184 else if (increment
== 1)
2185 new_stmt
= gimple_build_assign (lhs
, PLUS_EXPR
, basis_name
, c
->stride
);
2186 else if (increment
== -1)
2187 new_stmt
= gimple_build_assign (lhs
, MINUS_EXPR
, basis_name
,
2193 insert_bb
= single_succ_p (e
->src
) ? e
->src
: split_edge (e
);
2194 gsi
= gsi_last_bb (insert_bb
);
2196 if (!gsi_end_p (gsi
) && is_ctrl_stmt (gsi_stmt (gsi
)))
2197 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
2199 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
2201 gimple_set_location (new_stmt
, loc
);
2203 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2205 fprintf (dump_file
, "Inserting in block %d: ", insert_bb
->index
);
2206 print_gimple_stmt (dump_file
, new_stmt
, 0, 0);
2212 /* Given a candidate C with BASIS_NAME being the LHS of C's basis which
2213 is hidden by the phi node FROM_PHI, create a new phi node in the same
2214 block as FROM_PHI. The new phi is suitable for use as a basis by C,
2215 with its phi arguments representing conditional adjustments to the
2216 hidden basis along conditional incoming paths. Those adjustments are
2217 made by creating add statements (and sometimes recursively creating
2218 phis) along those incoming paths. LOC is the location to attach to
2219 the introduced statements. KNOWN_STRIDE is true iff C's stride is a
2223 create_phi_basis (slsr_cand_t c
, gimple
*from_phi
, tree basis_name
,
2224 location_t loc
, bool known_stride
)
2229 slsr_cand_t basis
= lookup_cand (c
->basis
);
2230 int nargs
= gimple_phi_num_args (from_phi
);
2231 basic_block phi_bb
= gimple_bb (from_phi
);
2232 slsr_cand_t phi_cand
= *stmt_cand_map
->get (from_phi
);
2233 auto_vec
<tree
> phi_args (nargs
);
2235 /* Process each argument of the existing phi that represents
2236 conditionally-executed add candidates. */
2237 for (i
= 0; i
< nargs
; i
++)
2239 edge e
= (*phi_bb
->preds
)[i
];
2240 tree arg
= gimple_phi_arg_def (from_phi
, i
);
2243 /* If the phi argument is the base name of the CAND_PHI, then
2244 this incoming arc should use the hidden basis. */
2245 if (operand_equal_p (arg
, phi_cand
->base_expr
, 0))
2246 if (basis
->index
== 0)
2247 feeding_def
= gimple_assign_lhs (basis
->cand_stmt
);
2250 widest_int incr
= -basis
->index
;
2251 feeding_def
= create_add_on_incoming_edge (c
, basis_name
, incr
,
2252 e
, loc
, known_stride
);
2256 gimple
*arg_def
= SSA_NAME_DEF_STMT (arg
);
2258 /* If there is another phi along this incoming edge, we must
2259 process it in the same fashion to ensure that all basis
2260 adjustments are made along its incoming edges. */
2261 if (gimple_code (arg_def
) == GIMPLE_PHI
)
2262 feeding_def
= create_phi_basis (c
, arg_def
, basis_name
,
2266 slsr_cand_t arg_cand
= base_cand_from_table (arg
);
2267 widest_int diff
= arg_cand
->index
- basis
->index
;
2268 feeding_def
= create_add_on_incoming_edge (c
, basis_name
, diff
,
2269 e
, loc
, known_stride
);
2273 /* Because of recursion, we need to save the arguments in a vector
2274 so we can create the PHI statement all at once. Otherwise the
2275 storage for the half-created PHI can be reclaimed. */
2276 phi_args
.safe_push (feeding_def
);
2279 /* Create the new phi basis. */
2280 name
= make_temp_ssa_name (TREE_TYPE (basis_name
), NULL
, "slsr");
2281 phi
= create_phi_node (name
, phi_bb
);
2282 SSA_NAME_DEF_STMT (name
) = phi
;
2284 FOR_EACH_VEC_ELT (phi_args
, i
, phi_arg
)
2286 edge e
= (*phi_bb
->preds
)[i
];
2287 add_phi_arg (phi
, phi_arg
, e
, loc
);
2292 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2294 fputs ("Introducing new phi basis: ", dump_file
);
2295 print_gimple_stmt (dump_file
, phi
, 0, 0);
2301 /* Given a candidate C whose basis is hidden by at least one intervening
2302 phi, introduce a matching number of new phis to represent its basis
2303 adjusted by conditional increments along possible incoming paths. Then
2304 replace C as though it were an unconditional candidate, using the new
2308 replace_conditional_candidate (slsr_cand_t c
)
2310 tree basis_name
, name
;
2314 /* Look up the LHS SSA name from C's basis. This will be the
2315 RHS1 of the adds we will introduce to create new phi arguments. */
2316 basis
= lookup_cand (c
->basis
);
2317 basis_name
= gimple_assign_lhs (basis
->cand_stmt
);
2319 /* Create a new phi statement which will represent C's true basis
2320 after the transformation is complete. */
2321 loc
= gimple_location (c
->cand_stmt
);
2322 name
= create_phi_basis (c
, lookup_cand (c
->def_phi
)->cand_stmt
,
2323 basis_name
, loc
, KNOWN_STRIDE
);
2324 /* Replace C with an add of the new basis phi and a constant. */
2325 widest_int bump
= c
->index
* wi::to_widest (c
->stride
);
2327 replace_mult_candidate (c
, name
, bump
);
2330 /* Compute the expected costs of inserting basis adjustments for
2331 candidate C with phi-definition PHI. The cost of inserting
2332 one adjustment is given by ONE_ADD_COST. If PHI has arguments
2333 which are themselves phi results, recursively calculate costs
2334 for those phis as well. */
2337 phi_add_costs (gimple
*phi
, slsr_cand_t c
, int one_add_cost
)
2341 slsr_cand_t phi_cand
= *stmt_cand_map
->get (phi
);
2343 /* If we work our way back to a phi that isn't dominated by the hidden
2344 basis, this isn't a candidate for replacement. Indicate this by
2345 returning an unreasonably high cost. It's not easy to detect
2346 these situations when determining the basis, so we defer the
2347 decision until now. */
2348 basic_block phi_bb
= gimple_bb (phi
);
2349 slsr_cand_t basis
= lookup_cand (c
->basis
);
2350 basic_block basis_bb
= gimple_bb (basis
->cand_stmt
);
2352 if (phi_bb
== basis_bb
|| !dominated_by_p (CDI_DOMINATORS
, phi_bb
, basis_bb
))
2353 return COST_INFINITE
;
2355 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2357 tree arg
= gimple_phi_arg_def (phi
, i
);
2359 if (arg
!= phi_cand
->base_expr
)
2361 gimple
*arg_def
= SSA_NAME_DEF_STMT (arg
);
2363 if (gimple_code (arg_def
) == GIMPLE_PHI
)
2364 cost
+= phi_add_costs (arg_def
, c
, one_add_cost
);
2367 slsr_cand_t arg_cand
= base_cand_from_table (arg
);
2369 if (arg_cand
->index
!= c
->index
)
2370 cost
+= one_add_cost
;
2378 /* For candidate C, each sibling of candidate C, and each dependent of
2379 candidate C, determine whether the candidate is dependent upon a
2380 phi that hides its basis. If not, replace the candidate unconditionally.
2381 Otherwise, determine whether the cost of introducing compensation code
2382 for the candidate is offset by the gains from strength reduction. If
2383 so, replace the candidate and introduce the compensation code. */
2386 replace_uncond_cands_and_profitable_phis (slsr_cand_t c
)
2388 if (phi_dependent_cand_p (c
))
2390 if (c
->kind
== CAND_MULT
)
2392 /* A candidate dependent upon a phi will replace a multiply by
2393 a constant with an add, and will insert at most one add for
2394 each phi argument. Add these costs with the potential dead-code
2395 savings to determine profitability. */
2396 bool speed
= optimize_bb_for_speed_p (gimple_bb (c
->cand_stmt
));
2397 int mult_savings
= stmt_cost (c
->cand_stmt
, speed
);
2398 gimple
*phi
= lookup_cand (c
->def_phi
)->cand_stmt
;
2399 tree phi_result
= gimple_phi_result (phi
);
2400 int one_add_cost
= add_cost (speed
,
2401 TYPE_MODE (TREE_TYPE (phi_result
)));
2402 int add_costs
= one_add_cost
+ phi_add_costs (phi
, c
, one_add_cost
);
2403 int cost
= add_costs
- mult_savings
- c
->dead_savings
;
2405 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2407 fprintf (dump_file
, " Conditional candidate %d:\n", c
->cand_num
);
2408 fprintf (dump_file
, " add_costs = %d\n", add_costs
);
2409 fprintf (dump_file
, " mult_savings = %d\n", mult_savings
);
2410 fprintf (dump_file
, " dead_savings = %d\n", c
->dead_savings
);
2411 fprintf (dump_file
, " cost = %d\n", cost
);
2412 if (cost
<= COST_NEUTRAL
)
2413 fputs (" Replacing...\n", dump_file
);
2415 fputs (" Not replaced.\n", dump_file
);
2418 if (cost
<= COST_NEUTRAL
)
2419 replace_conditional_candidate (c
);
2423 replace_unconditional_candidate (c
);
2426 replace_uncond_cands_and_profitable_phis (lookup_cand (c
->sibling
));
2429 replace_uncond_cands_and_profitable_phis (lookup_cand (c
->dependent
));
2432 /* Count the number of candidates in the tree rooted at C that have
2433 not already been replaced under other interpretations. */
2436 count_candidates (slsr_cand_t c
)
2438 unsigned count
= cand_already_replaced (c
) ? 0 : 1;
2441 count
+= count_candidates (lookup_cand (c
->sibling
));
2444 count
+= count_candidates (lookup_cand (c
->dependent
));
2449 /* Increase the count of INCREMENT by one in the increment vector.
2450 INCREMENT is associated with candidate C. If INCREMENT is to be
2451 conditionally executed as part of a conditional candidate replacement,
2452 IS_PHI_ADJUST is true, otherwise false. If an initializer
2453 T_0 = stride * I is provided by a candidate that dominates all
2454 candidates with the same increment, also record T_0 for subsequent use. */
2457 record_increment (slsr_cand_t c
, widest_int increment
, bool is_phi_adjust
)
2462 /* Treat increments that differ only in sign as identical so as to
2463 share initializers, unless we are generating pointer arithmetic. */
2464 if (!address_arithmetic_p
&& wi::neg_p (increment
))
2465 increment
= -increment
;
2467 for (i
= 0; i
< incr_vec_len
; i
++)
2469 if (incr_vec
[i
].incr
== increment
)
2471 incr_vec
[i
].count
++;
2474 /* If we previously recorded an initializer that doesn't
2475 dominate this candidate, it's not going to be useful to
2477 if (incr_vec
[i
].initializer
2478 && !dominated_by_p (CDI_DOMINATORS
,
2479 gimple_bb (c
->cand_stmt
),
2480 incr_vec
[i
].init_bb
))
2482 incr_vec
[i
].initializer
= NULL_TREE
;
2483 incr_vec
[i
].init_bb
= NULL
;
2490 if (!found
&& incr_vec_len
< MAX_INCR_VEC_LEN
- 1)
2492 /* The first time we see an increment, create the entry for it.
2493 If this is the root candidate which doesn't have a basis, set
2494 the count to zero. We're only processing it so it can possibly
2495 provide an initializer for other candidates. */
2496 incr_vec
[incr_vec_len
].incr
= increment
;
2497 incr_vec
[incr_vec_len
].count
= c
->basis
|| is_phi_adjust
? 1 : 0;
2498 incr_vec
[incr_vec_len
].cost
= COST_INFINITE
;
2500 /* Optimistically record the first occurrence of this increment
2501 as providing an initializer (if it does); we will revise this
2502 opinion later if it doesn't dominate all other occurrences.
2503 Exception: increments of -1, 0, 1 never need initializers;
2504 and phi adjustments don't ever provide initializers. */
2505 if (c
->kind
== CAND_ADD
2507 && c
->index
== increment
2508 && (increment
> 1 || increment
< -1)
2509 && (gimple_assign_rhs_code (c
->cand_stmt
) == PLUS_EXPR
2510 || gimple_assign_rhs_code (c
->cand_stmt
) == POINTER_PLUS_EXPR
))
2512 tree t0
= NULL_TREE
;
2513 tree rhs1
= gimple_assign_rhs1 (c
->cand_stmt
);
2514 tree rhs2
= gimple_assign_rhs2 (c
->cand_stmt
);
2515 if (operand_equal_p (rhs1
, c
->base_expr
, 0))
2517 else if (operand_equal_p (rhs2
, c
->base_expr
, 0))
2520 && SSA_NAME_DEF_STMT (t0
)
2521 && gimple_bb (SSA_NAME_DEF_STMT (t0
)))
2523 incr_vec
[incr_vec_len
].initializer
= t0
;
2524 incr_vec
[incr_vec_len
++].init_bb
2525 = gimple_bb (SSA_NAME_DEF_STMT (t0
));
2529 incr_vec
[incr_vec_len
].initializer
= NULL_TREE
;
2530 incr_vec
[incr_vec_len
++].init_bb
= NULL
;
2535 incr_vec
[incr_vec_len
].initializer
= NULL_TREE
;
2536 incr_vec
[incr_vec_len
++].init_bb
= NULL
;
2541 /* Given phi statement PHI that hides a candidate from its BASIS, find
2542 the increments along each incoming arc (recursively handling additional
2543 phis that may be present) and record them. These increments are the
2544 difference in index between the index-adjusting statements and the
2545 index of the basis. */
2548 record_phi_increments (slsr_cand_t basis
, gimple
*phi
)
2551 slsr_cand_t phi_cand
= *stmt_cand_map
->get (phi
);
2553 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2555 tree arg
= gimple_phi_arg_def (phi
, i
);
2557 if (!operand_equal_p (arg
, phi_cand
->base_expr
, 0))
2559 gimple
*arg_def
= SSA_NAME_DEF_STMT (arg
);
2561 if (gimple_code (arg_def
) == GIMPLE_PHI
)
2562 record_phi_increments (basis
, arg_def
);
2565 slsr_cand_t arg_cand
= base_cand_from_table (arg
);
2566 widest_int diff
= arg_cand
->index
- basis
->index
;
2567 record_increment (arg_cand
, diff
, PHI_ADJUST
);
2573 /* Determine how many times each unique increment occurs in the set
2574 of candidates rooted at C's parent, recording the data in the
2575 increment vector. For each unique increment I, if an initializer
2576 T_0 = stride * I is provided by a candidate that dominates all
2577 candidates with the same increment, also record T_0 for subsequent
2581 record_increments (slsr_cand_t c
)
2583 if (!cand_already_replaced (c
))
2585 if (!phi_dependent_cand_p (c
))
2586 record_increment (c
, cand_increment (c
), NOT_PHI_ADJUST
);
2589 /* A candidate with a basis hidden by a phi will have one
2590 increment for its relationship to the index represented by
2591 the phi, and potentially additional increments along each
2592 incoming edge. For the root of the dependency tree (which
2593 has no basis), process just the initial index in case it has
2594 an initializer that can be used by subsequent candidates. */
2595 record_increment (c
, c
->index
, NOT_PHI_ADJUST
);
2598 record_phi_increments (lookup_cand (c
->basis
),
2599 lookup_cand (c
->def_phi
)->cand_stmt
);
2604 record_increments (lookup_cand (c
->sibling
));
2607 record_increments (lookup_cand (c
->dependent
));
2610 /* Add up and return the costs of introducing add statements that
2611 require the increment INCR on behalf of candidate C and phi
2612 statement PHI. Accumulate into *SAVINGS the potential savings
2613 from removing existing statements that feed PHI and have no other
2617 phi_incr_cost (slsr_cand_t c
, const widest_int
&incr
, gimple
*phi
,
2622 slsr_cand_t basis
= lookup_cand (c
->basis
);
2623 slsr_cand_t phi_cand
= *stmt_cand_map
->get (phi
);
2625 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2627 tree arg
= gimple_phi_arg_def (phi
, i
);
2629 if (!operand_equal_p (arg
, phi_cand
->base_expr
, 0))
2631 gimple
*arg_def
= SSA_NAME_DEF_STMT (arg
);
2633 if (gimple_code (arg_def
) == GIMPLE_PHI
)
2635 int feeding_savings
= 0;
2636 cost
+= phi_incr_cost (c
, incr
, arg_def
, &feeding_savings
);
2637 if (has_single_use (gimple_phi_result (arg_def
)))
2638 *savings
+= feeding_savings
;
2642 slsr_cand_t arg_cand
= base_cand_from_table (arg
);
2643 widest_int diff
= arg_cand
->index
- basis
->index
;
2647 tree basis_lhs
= gimple_assign_lhs (basis
->cand_stmt
);
2648 tree lhs
= gimple_assign_lhs (arg_cand
->cand_stmt
);
2649 cost
+= add_cost (true, TYPE_MODE (TREE_TYPE (basis_lhs
)));
2650 if (has_single_use (lhs
))
2651 *savings
+= stmt_cost (arg_cand
->cand_stmt
, true);
2660 /* Return the first candidate in the tree rooted at C that has not
2661 already been replaced, favoring siblings over dependents. */
2664 unreplaced_cand_in_tree (slsr_cand_t c
)
2666 if (!cand_already_replaced (c
))
2671 slsr_cand_t sib
= unreplaced_cand_in_tree (lookup_cand (c
->sibling
));
2678 slsr_cand_t dep
= unreplaced_cand_in_tree (lookup_cand (c
->dependent
));
2686 /* Return TRUE if the candidates in the tree rooted at C should be
2687 optimized for speed, else FALSE. We estimate this based on the block
2688 containing the most dominant candidate in the tree that has not yet
2692 optimize_cands_for_speed_p (slsr_cand_t c
)
2694 slsr_cand_t c2
= unreplaced_cand_in_tree (c
);
2696 return optimize_bb_for_speed_p (gimple_bb (c2
->cand_stmt
));
2699 /* Add COST_IN to the lowest cost of any dependent path starting at
2700 candidate C or any of its siblings, counting only candidates along
2701 such paths with increment INCR. Assume that replacing a candidate
2702 reduces cost by REPL_SAVINGS. Also account for savings from any
2703 statements that would go dead. If COUNT_PHIS is true, include
2704 costs of introducing feeding statements for conditional candidates. */
2707 lowest_cost_path (int cost_in
, int repl_savings
, slsr_cand_t c
,
2708 const widest_int
&incr
, bool count_phis
)
2710 int local_cost
, sib_cost
, savings
= 0;
2711 widest_int cand_incr
= cand_abs_increment (c
);
2713 if (cand_already_replaced (c
))
2714 local_cost
= cost_in
;
2715 else if (incr
== cand_incr
)
2716 local_cost
= cost_in
- repl_savings
- c
->dead_savings
;
2718 local_cost
= cost_in
- c
->dead_savings
;
2721 && phi_dependent_cand_p (c
)
2722 && !cand_already_replaced (c
))
2724 gimple
*phi
= lookup_cand (c
->def_phi
)->cand_stmt
;
2725 local_cost
+= phi_incr_cost (c
, incr
, phi
, &savings
);
2727 if (has_single_use (gimple_phi_result (phi
)))
2728 local_cost
-= savings
;
2732 local_cost
= lowest_cost_path (local_cost
, repl_savings
,
2733 lookup_cand (c
->dependent
), incr
,
2738 sib_cost
= lowest_cost_path (cost_in
, repl_savings
,
2739 lookup_cand (c
->sibling
), incr
,
2741 local_cost
= MIN (local_cost
, sib_cost
);
2747 /* Compute the total savings that would accrue from all replacements
2748 in the candidate tree rooted at C, counting only candidates with
2749 increment INCR. Assume that replacing a candidate reduces cost
2750 by REPL_SAVINGS. Also account for savings from statements that
2754 total_savings (int repl_savings
, slsr_cand_t c
, const widest_int
&incr
,
2758 widest_int cand_incr
= cand_abs_increment (c
);
2760 if (incr
== cand_incr
&& !cand_already_replaced (c
))
2761 savings
+= repl_savings
+ c
->dead_savings
;
2764 && phi_dependent_cand_p (c
)
2765 && !cand_already_replaced (c
))
2767 int phi_savings
= 0;
2768 gimple
*phi
= lookup_cand (c
->def_phi
)->cand_stmt
;
2769 savings
-= phi_incr_cost (c
, incr
, phi
, &phi_savings
);
2771 if (has_single_use (gimple_phi_result (phi
)))
2772 savings
+= phi_savings
;
2776 savings
+= total_savings (repl_savings
, lookup_cand (c
->dependent
), incr
,
2780 savings
+= total_savings (repl_savings
, lookup_cand (c
->sibling
), incr
,
2786 /* Use target-specific costs to determine and record which increments
2787 in the current candidate tree are profitable to replace, assuming
2788 MODE and SPEED. FIRST_DEP is the first dependent of the root of
2791 One slight limitation here is that we don't account for the possible
2792 introduction of casts in some cases. See replace_one_candidate for
2793 the cases where these are introduced. This should probably be cleaned
2797 analyze_increments (slsr_cand_t first_dep
, machine_mode mode
, bool speed
)
2801 for (i
= 0; i
< incr_vec_len
; i
++)
2803 HOST_WIDE_INT incr
= incr_vec
[i
].incr
.to_shwi ();
2805 /* If somehow this increment is bigger than a HWI, we won't
2806 be optimizing candidates that use it. And if the increment
2807 has a count of zero, nothing will be done with it. */
2808 if (!wi::fits_shwi_p (incr_vec
[i
].incr
) || !incr_vec
[i
].count
)
2809 incr_vec
[i
].cost
= COST_INFINITE
;
2811 /* Increments of 0, 1, and -1 are always profitable to replace,
2812 because they always replace a multiply or add with an add or
2813 copy, and may cause one or more existing instructions to go
2814 dead. Exception: -1 can't be assumed to be profitable for
2815 pointer addition. */
2819 && !POINTER_TYPE_P (first_dep
->cand_type
)))
2820 incr_vec
[i
].cost
= COST_NEUTRAL
;
2822 /* FIXME: We don't handle pointers with a -1 increment yet.
2823 They are usually unprofitable anyway. */
2824 else if (incr
== -1 && POINTER_TYPE_P (first_dep
->cand_type
))
2825 incr_vec
[i
].cost
= COST_INFINITE
;
2827 /* FORNOW: If we need to add an initializer, give up if a cast from
2828 the candidate's type to its stride's type can lose precision.
2829 This could eventually be handled better by expressly retaining the
2830 result of a cast to a wider type in the stride. Example:
2835 _4 = x + _3; ADD: x + (10 * _1) : int
2837 _6 = x + _3; ADD: x + (15 * _1) : int
2839 Right now replacing _6 would cause insertion of an initializer
2840 of the form "short int T = _1 * 5;" followed by a cast to
2841 int, which could overflow incorrectly. Had we recorded _2 or
2842 (int)_1 as the stride, this wouldn't happen. However, doing
2843 this breaks other opportunities, so this will require some
2845 else if (!incr_vec
[i
].initializer
2846 && TREE_CODE (first_dep
->stride
) != INTEGER_CST
2847 && !legal_cast_p_1 (first_dep
->stride
,
2848 gimple_assign_lhs (first_dep
->cand_stmt
)))
2850 incr_vec
[i
].cost
= COST_INFINITE
;
2852 /* If we need to add an initializer, make sure we don't introduce
2853 a multiply by a pointer type, which can happen in certain cast
2854 scenarios. FIXME: When cleaning up these cast issues, we can
2855 afford to introduce the multiply provided we cast out to an
2856 unsigned int of appropriate size. */
2857 else if (!incr_vec
[i
].initializer
2858 && TREE_CODE (first_dep
->stride
) != INTEGER_CST
2859 && POINTER_TYPE_P (TREE_TYPE (first_dep
->stride
)))
2861 incr_vec
[i
].cost
= COST_INFINITE
;
2863 /* For any other increment, if this is a multiply candidate, we
2864 must introduce a temporary T and initialize it with
2865 T_0 = stride * increment. When optimizing for speed, walk the
2866 candidate tree to calculate the best cost reduction along any
2867 path; if it offsets the fixed cost of inserting the initializer,
2868 replacing the increment is profitable. When optimizing for
2869 size, instead calculate the total cost reduction from replacing
2870 all candidates with this increment. */
2871 else if (first_dep
->kind
== CAND_MULT
)
2873 int cost
= mult_by_coeff_cost (incr
, mode
, speed
);
2874 int repl_savings
= mul_cost (speed
, mode
) - add_cost (speed
, mode
);
2876 cost
= lowest_cost_path (cost
, repl_savings
, first_dep
,
2877 incr_vec
[i
].incr
, COUNT_PHIS
);
2879 cost
-= total_savings (repl_savings
, first_dep
, incr_vec
[i
].incr
,
2882 incr_vec
[i
].cost
= cost
;
2885 /* If this is an add candidate, the initializer may already
2886 exist, so only calculate the cost of the initializer if it
2887 doesn't. We are replacing one add with another here, so the
2888 known replacement savings is zero. We will account for removal
2889 of dead instructions in lowest_cost_path or total_savings. */
2893 if (!incr_vec
[i
].initializer
)
2894 cost
= mult_by_coeff_cost (incr
, mode
, speed
);
2897 cost
= lowest_cost_path (cost
, 0, first_dep
, incr_vec
[i
].incr
,
2900 cost
-= total_savings (0, first_dep
, incr_vec
[i
].incr
,
2903 incr_vec
[i
].cost
= cost
;
2908 /* Return the nearest common dominator of BB1 and BB2. If the blocks
2909 are identical, return the earlier of C1 and C2 in *WHERE. Otherwise,
2910 if the NCD matches BB1, return C1 in *WHERE; if the NCD matches BB2,
2911 return C2 in *WHERE; and if the NCD matches neither, return NULL in
2912 *WHERE. Note: It is possible for one of C1 and C2 to be NULL. */
2915 ncd_for_two_cands (basic_block bb1
, basic_block bb2
,
2916 slsr_cand_t c1
, slsr_cand_t c2
, slsr_cand_t
*where
)
2932 ncd
= nearest_common_dominator (CDI_DOMINATORS
, bb1
, bb2
);
2934 /* If both candidates are in the same block, the earlier
2936 if (bb1
== ncd
&& bb2
== ncd
)
2938 if (!c1
|| (c2
&& c2
->cand_num
< c1
->cand_num
))
2944 /* Otherwise, if one of them produced a candidate in the
2945 dominator, that one wins. */
2946 else if (bb1
== ncd
)
2949 else if (bb2
== ncd
)
2952 /* If neither matches the dominator, neither wins. */
2959 /* Consider all candidates that feed PHI. Find the nearest common
2960 dominator of those candidates requiring the given increment INCR.
2961 Further find and return the nearest common dominator of this result
2962 with block NCD. If the returned block contains one or more of the
2963 candidates, return the earliest candidate in the block in *WHERE. */
2966 ncd_with_phi (slsr_cand_t c
, const widest_int
&incr
, gphi
*phi
,
2967 basic_block ncd
, slsr_cand_t
*where
)
2970 slsr_cand_t basis
= lookup_cand (c
->basis
);
2971 slsr_cand_t phi_cand
= *stmt_cand_map
->get (phi
);
2973 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
2975 tree arg
= gimple_phi_arg_def (phi
, i
);
2977 if (!operand_equal_p (arg
, phi_cand
->base_expr
, 0))
2979 gimple
*arg_def
= SSA_NAME_DEF_STMT (arg
);
2981 if (gimple_code (arg_def
) == GIMPLE_PHI
)
2982 ncd
= ncd_with_phi (c
, incr
, as_a
<gphi
*> (arg_def
), ncd
,
2986 slsr_cand_t arg_cand
= base_cand_from_table (arg
);
2987 widest_int diff
= arg_cand
->index
- basis
->index
;
2988 basic_block pred
= gimple_phi_arg_edge (phi
, i
)->src
;
2990 if ((incr
== diff
) || (!address_arithmetic_p
&& incr
== -diff
))
2991 ncd
= ncd_for_two_cands (ncd
, pred
, *where
, NULL
, where
);
2999 /* Consider the candidate C together with any candidates that feed
3000 C's phi dependence (if any). Find and return the nearest common
3001 dominator of those candidates requiring the given increment INCR.
3002 If the returned block contains one or more of the candidates,
3003 return the earliest candidate in the block in *WHERE. */
3006 ncd_of_cand_and_phis (slsr_cand_t c
, const widest_int
&incr
, slsr_cand_t
*where
)
3008 basic_block ncd
= NULL
;
3010 if (cand_abs_increment (c
) == incr
)
3012 ncd
= gimple_bb (c
->cand_stmt
);
3016 if (phi_dependent_cand_p (c
))
3017 ncd
= ncd_with_phi (c
, incr
,
3018 as_a
<gphi
*> (lookup_cand (c
->def_phi
)->cand_stmt
),
3024 /* Consider all candidates in the tree rooted at C for which INCR
3025 represents the required increment of C relative to its basis.
3026 Find and return the basic block that most nearly dominates all
3027 such candidates. If the returned block contains one or more of
3028 the candidates, return the earliest candidate in the block in
3032 nearest_common_dominator_for_cands (slsr_cand_t c
, const widest_int
&incr
,
3035 basic_block sib_ncd
= NULL
, dep_ncd
= NULL
, this_ncd
= NULL
, ncd
;
3036 slsr_cand_t sib_where
= NULL
, dep_where
= NULL
, this_where
= NULL
, new_where
;
3038 /* First find the NCD of all siblings and dependents. */
3040 sib_ncd
= nearest_common_dominator_for_cands (lookup_cand (c
->sibling
),
3043 dep_ncd
= nearest_common_dominator_for_cands (lookup_cand (c
->dependent
),
3045 if (!sib_ncd
&& !dep_ncd
)
3050 else if (sib_ncd
&& !dep_ncd
)
3052 new_where
= sib_where
;
3055 else if (dep_ncd
&& !sib_ncd
)
3057 new_where
= dep_where
;
3061 ncd
= ncd_for_two_cands (sib_ncd
, dep_ncd
, sib_where
,
3062 dep_where
, &new_where
);
3064 /* If the candidate's increment doesn't match the one we're interested
3065 in (and nor do any increments for feeding defs of a phi-dependence),
3066 then the result depends only on siblings and dependents. */
3067 this_ncd
= ncd_of_cand_and_phis (c
, incr
, &this_where
);
3069 if (!this_ncd
|| cand_already_replaced (c
))
3075 /* Otherwise, compare this candidate with the result from all siblings
3077 ncd
= ncd_for_two_cands (ncd
, this_ncd
, new_where
, this_where
, where
);
3082 /* Return TRUE if the increment indexed by INDEX is profitable to replace. */
3085 profitable_increment_p (unsigned index
)
3087 return (incr_vec
[index
].cost
<= COST_NEUTRAL
);
3090 /* For each profitable increment in the increment vector not equal to
3091 0 or 1 (or -1, for non-pointer arithmetic), find the nearest common
3092 dominator of all statements in the candidate chain rooted at C
3093 that require that increment, and insert an initializer
3094 T_0 = stride * increment at that location. Record T_0 with the
3095 increment record. */
3098 insert_initializers (slsr_cand_t c
)
3102 for (i
= 0; i
< incr_vec_len
; i
++)
3105 slsr_cand_t where
= NULL
;
3107 tree stride_type
, new_name
, incr_tree
;
3108 widest_int incr
= incr_vec
[i
].incr
;
3110 if (!profitable_increment_p (i
)
3113 && gimple_assign_rhs_code (c
->cand_stmt
) != POINTER_PLUS_EXPR
)
3117 /* We may have already identified an existing initializer that
3119 if (incr_vec
[i
].initializer
)
3121 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3123 fputs ("Using existing initializer: ", dump_file
);
3124 print_gimple_stmt (dump_file
,
3125 SSA_NAME_DEF_STMT (incr_vec
[i
].initializer
),
3131 /* Find the block that most closely dominates all candidates
3132 with this increment. If there is at least one candidate in
3133 that block, the earliest one will be returned in WHERE. */
3134 bb
= nearest_common_dominator_for_cands (c
, incr
, &where
);
3136 /* Create a new SSA name to hold the initializer's value. */
3137 stride_type
= TREE_TYPE (c
->stride
);
3138 new_name
= make_temp_ssa_name (stride_type
, NULL
, "slsr");
3139 incr_vec
[i
].initializer
= new_name
;
3141 /* Create the initializer and insert it in the latest possible
3142 dominating position. */
3143 incr_tree
= wide_int_to_tree (stride_type
, incr
);
3144 init_stmt
= gimple_build_assign (new_name
, MULT_EXPR
,
3145 c
->stride
, incr_tree
);
3148 gimple_stmt_iterator gsi
= gsi_for_stmt (where
->cand_stmt
);
3149 gsi_insert_before (&gsi
, init_stmt
, GSI_SAME_STMT
);
3150 gimple_set_location (init_stmt
, gimple_location (where
->cand_stmt
));
3154 gimple_stmt_iterator gsi
= gsi_last_bb (bb
);
3155 gimple
*basis_stmt
= lookup_cand (c
->basis
)->cand_stmt
;
3157 if (!gsi_end_p (gsi
) && is_ctrl_stmt (gsi_stmt (gsi
)))
3158 gsi_insert_before (&gsi
, init_stmt
, GSI_SAME_STMT
);
3160 gsi_insert_after (&gsi
, init_stmt
, GSI_SAME_STMT
);
3162 gimple_set_location (init_stmt
, gimple_location (basis_stmt
));
3165 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3167 fputs ("Inserting initializer: ", dump_file
);
3168 print_gimple_stmt (dump_file
, init_stmt
, 0, 0);
3173 /* Return TRUE iff all required increments for candidates feeding PHI
3174 are profitable to replace on behalf of candidate C. */
3177 all_phi_incrs_profitable (slsr_cand_t c
, gimple
*phi
)
3180 slsr_cand_t basis
= lookup_cand (c
->basis
);
3181 slsr_cand_t phi_cand
= *stmt_cand_map
->get (phi
);
3183 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
3185 tree arg
= gimple_phi_arg_def (phi
, i
);
3187 if (!operand_equal_p (arg
, phi_cand
->base_expr
, 0))
3189 gimple
*arg_def
= SSA_NAME_DEF_STMT (arg
);
3191 if (gimple_code (arg_def
) == GIMPLE_PHI
)
3193 if (!all_phi_incrs_profitable (c
, arg_def
))
3199 slsr_cand_t arg_cand
= base_cand_from_table (arg
);
3200 widest_int increment
= arg_cand
->index
- basis
->index
;
3202 if (!address_arithmetic_p
&& wi::neg_p (increment
))
3203 increment
= -increment
;
3205 j
= incr_vec_index (increment
);
3207 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3209 fprintf (dump_file
, " Conditional candidate %d, phi: ",
3211 print_gimple_stmt (dump_file
, phi
, 0, 0);
3212 fputs (" increment: ", dump_file
);
3213 print_decs (increment
, dump_file
);
3216 "\n Not replaced; incr_vec overflow.\n");
3218 fprintf (dump_file
, "\n cost: %d\n", incr_vec
[j
].cost
);
3219 if (profitable_increment_p (j
))
3220 fputs (" Replacing...\n", dump_file
);
3222 fputs (" Not replaced.\n", dump_file
);
3226 if (j
< 0 || !profitable_increment_p (j
))
3235 /* Create a NOP_EXPR that copies FROM_EXPR into a new SSA name of
3236 type TO_TYPE, and insert it in front of the statement represented
3237 by candidate C. Use *NEW_VAR to create the new SSA name. Return
3238 the new SSA name. */
3241 introduce_cast_before_cand (slsr_cand_t c
, tree to_type
, tree from_expr
)
3245 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
3247 cast_lhs
= make_temp_ssa_name (to_type
, NULL
, "slsr");
3248 cast_stmt
= gimple_build_assign (cast_lhs
, NOP_EXPR
, from_expr
);
3249 gimple_set_location (cast_stmt
, gimple_location (c
->cand_stmt
));
3250 gsi_insert_before (&gsi
, cast_stmt
, GSI_SAME_STMT
);
3252 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3254 fputs (" Inserting: ", dump_file
);
3255 print_gimple_stmt (dump_file
, cast_stmt
, 0, 0);
3261 /* Replace the RHS of the statement represented by candidate C with
3262 NEW_CODE, NEW_RHS1, and NEW_RHS2, provided that to do so doesn't
3263 leave C unchanged or just interchange its operands. The original
3264 operation and operands are in OLD_CODE, OLD_RHS1, and OLD_RHS2.
3265 If the replacement was made and we are doing a details dump,
3266 return the revised statement, else NULL. */
3269 replace_rhs_if_not_dup (enum tree_code new_code
, tree new_rhs1
, tree new_rhs2
,
3270 enum tree_code old_code
, tree old_rhs1
, tree old_rhs2
,
3273 if (new_code
!= old_code
3274 || ((!operand_equal_p (new_rhs1
, old_rhs1
, 0)
3275 || !operand_equal_p (new_rhs2
, old_rhs2
, 0))
3276 && (!operand_equal_p (new_rhs1
, old_rhs2
, 0)
3277 || !operand_equal_p (new_rhs2
, old_rhs1
, 0))))
3279 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
3280 gimple_assign_set_rhs_with_ops (&gsi
, new_code
, new_rhs1
, new_rhs2
);
3281 update_stmt (gsi_stmt (gsi
));
3282 c
->cand_stmt
= gsi_stmt (gsi
);
3284 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3285 return gsi_stmt (gsi
);
3288 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3289 fputs (" (duplicate, not actually replacing)\n", dump_file
);
3294 /* Strength-reduce the statement represented by candidate C by replacing
3295 it with an equivalent addition or subtraction. I is the index into
3296 the increment vector identifying C's increment. NEW_VAR is used to
3297 create a new SSA name if a cast needs to be introduced. BASIS_NAME
3298 is the rhs1 to use in creating the add/subtract. */
3301 replace_one_candidate (slsr_cand_t c
, unsigned i
, tree basis_name
)
3303 gimple
*stmt_to_print
= NULL
;
3304 tree orig_rhs1
, orig_rhs2
;
3306 enum tree_code orig_code
, repl_code
;
3307 widest_int cand_incr
;
3309 orig_code
= gimple_assign_rhs_code (c
->cand_stmt
);
3310 orig_rhs1
= gimple_assign_rhs1 (c
->cand_stmt
);
3311 orig_rhs2
= gimple_assign_rhs2 (c
->cand_stmt
);
3312 cand_incr
= cand_increment (c
);
3314 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3316 fputs ("Replacing: ", dump_file
);
3317 print_gimple_stmt (dump_file
, c
->cand_stmt
, 0, 0);
3318 stmt_to_print
= c
->cand_stmt
;
3321 if (address_arithmetic_p
)
3322 repl_code
= POINTER_PLUS_EXPR
;
3324 repl_code
= PLUS_EXPR
;
3326 /* If the increment has an initializer T_0, replace the candidate
3327 statement with an add of the basis name and the initializer. */
3328 if (incr_vec
[i
].initializer
)
3330 tree init_type
= TREE_TYPE (incr_vec
[i
].initializer
);
3331 tree orig_type
= TREE_TYPE (orig_rhs2
);
3333 if (types_compatible_p (orig_type
, init_type
))
3334 rhs2
= incr_vec
[i
].initializer
;
3336 rhs2
= introduce_cast_before_cand (c
, orig_type
,
3337 incr_vec
[i
].initializer
);
3339 if (incr_vec
[i
].incr
!= cand_incr
)
3341 gcc_assert (repl_code
== PLUS_EXPR
);
3342 repl_code
= MINUS_EXPR
;
3345 stmt_to_print
= replace_rhs_if_not_dup (repl_code
, basis_name
, rhs2
,
3346 orig_code
, orig_rhs1
, orig_rhs2
,
3350 /* Otherwise, the increment is one of -1, 0, and 1. Replace
3351 with a subtract of the stride from the basis name, a copy
3352 from the basis name, or an add of the stride to the basis
3353 name, respectively. It may be necessary to introduce a
3354 cast (or reuse an existing cast). */
3355 else if (cand_incr
== 1)
3357 tree stride_type
= TREE_TYPE (c
->stride
);
3358 tree orig_type
= TREE_TYPE (orig_rhs2
);
3360 if (types_compatible_p (orig_type
, stride_type
))
3363 rhs2
= introduce_cast_before_cand (c
, orig_type
, c
->stride
);
3365 stmt_to_print
= replace_rhs_if_not_dup (repl_code
, basis_name
, rhs2
,
3366 orig_code
, orig_rhs1
, orig_rhs2
,
3370 else if (cand_incr
== -1)
3372 tree stride_type
= TREE_TYPE (c
->stride
);
3373 tree orig_type
= TREE_TYPE (orig_rhs2
);
3374 gcc_assert (repl_code
!= POINTER_PLUS_EXPR
);
3376 if (types_compatible_p (orig_type
, stride_type
))
3379 rhs2
= introduce_cast_before_cand (c
, orig_type
, c
->stride
);
3381 if (orig_code
!= MINUS_EXPR
3382 || !operand_equal_p (basis_name
, orig_rhs1
, 0)
3383 || !operand_equal_p (rhs2
, orig_rhs2
, 0))
3385 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
3386 gimple_assign_set_rhs_with_ops (&gsi
, MINUS_EXPR
, basis_name
, rhs2
);
3387 update_stmt (gsi_stmt (gsi
));
3388 c
->cand_stmt
= gsi_stmt (gsi
);
3390 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3391 stmt_to_print
= gsi_stmt (gsi
);
3393 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3394 fputs (" (duplicate, not actually replacing)\n", dump_file
);
3397 else if (cand_incr
== 0)
3399 tree lhs
= gimple_assign_lhs (c
->cand_stmt
);
3400 tree lhs_type
= TREE_TYPE (lhs
);
3401 tree basis_type
= TREE_TYPE (basis_name
);
3403 if (types_compatible_p (lhs_type
, basis_type
))
3405 gassign
*copy_stmt
= gimple_build_assign (lhs
, basis_name
);
3406 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
3407 gimple_set_location (copy_stmt
, gimple_location (c
->cand_stmt
));
3408 gsi_replace (&gsi
, copy_stmt
, false);
3409 c
->cand_stmt
= copy_stmt
;
3411 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3412 stmt_to_print
= copy_stmt
;
3416 gimple_stmt_iterator gsi
= gsi_for_stmt (c
->cand_stmt
);
3417 gassign
*cast_stmt
= gimple_build_assign (lhs
, NOP_EXPR
, basis_name
);
3418 gimple_set_location (cast_stmt
, gimple_location (c
->cand_stmt
));
3419 gsi_replace (&gsi
, cast_stmt
, false);
3420 c
->cand_stmt
= cast_stmt
;
3422 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3423 stmt_to_print
= cast_stmt
;
3429 if (dump_file
&& (dump_flags
& TDF_DETAILS
) && stmt_to_print
)
3431 fputs ("With: ", dump_file
);
3432 print_gimple_stmt (dump_file
, stmt_to_print
, 0, 0);
3433 fputs ("\n", dump_file
);
3437 /* For each candidate in the tree rooted at C, replace it with
3438 an increment if such has been shown to be profitable. */
3441 replace_profitable_candidates (slsr_cand_t c
)
3443 if (!cand_already_replaced (c
))
3445 widest_int increment
= cand_abs_increment (c
);
3446 enum tree_code orig_code
= gimple_assign_rhs_code (c
->cand_stmt
);
3449 i
= incr_vec_index (increment
);
3451 /* Only process profitable increments. Nothing useful can be done
3452 to a cast or copy. */
3454 && profitable_increment_p (i
)
3455 && orig_code
!= SSA_NAME
3456 && !CONVERT_EXPR_CODE_P (orig_code
))
3458 if (phi_dependent_cand_p (c
))
3460 gimple
*phi
= lookup_cand (c
->def_phi
)->cand_stmt
;
3462 if (all_phi_incrs_profitable (c
, phi
))
3464 /* Look up the LHS SSA name from C's basis. This will be
3465 the RHS1 of the adds we will introduce to create new
3467 slsr_cand_t basis
= lookup_cand (c
->basis
);
3468 tree basis_name
= gimple_assign_lhs (basis
->cand_stmt
);
3470 /* Create a new phi statement that will represent C's true
3471 basis after the transformation is complete. */
3472 location_t loc
= gimple_location (c
->cand_stmt
);
3473 tree name
= create_phi_basis (c
, phi
, basis_name
,
3474 loc
, UNKNOWN_STRIDE
);
3476 /* Replace C with an add of the new basis phi and the
3478 replace_one_candidate (c
, i
, name
);
3483 slsr_cand_t basis
= lookup_cand (c
->basis
);
3484 tree basis_name
= gimple_assign_lhs (basis
->cand_stmt
);
3485 replace_one_candidate (c
, i
, basis_name
);
3491 replace_profitable_candidates (lookup_cand (c
->sibling
));
3494 replace_profitable_candidates (lookup_cand (c
->dependent
));
3497 /* Analyze costs of related candidates in the candidate vector,
3498 and make beneficial replacements. */
3501 analyze_candidates_and_replace (void)
3506 /* Each candidate that has a null basis and a non-null
3507 dependent is the root of a tree of related statements.
3508 Analyze each tree to determine a subset of those
3509 statements that can be replaced with maximum benefit. */
3510 FOR_EACH_VEC_ELT (cand_vec
, i
, c
)
3512 slsr_cand_t first_dep
;
3514 if (c
->basis
!= 0 || c
->dependent
== 0)
3517 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3518 fprintf (dump_file
, "\nProcessing dependency tree rooted at %d.\n",
3521 first_dep
= lookup_cand (c
->dependent
);
3523 /* If this is a chain of CAND_REFs, unconditionally replace
3524 each of them with a strength-reduced data reference. */
3525 if (c
->kind
== CAND_REF
)
3528 /* If the common stride of all related candidates is a known
3529 constant, each candidate without a phi-dependence can be
3530 profitably replaced. Each replaces a multiply by a single
3531 add, with the possibility that a feeding add also goes dead.
3532 A candidate with a phi-dependence is replaced only if the
3533 compensation code it requires is offset by the strength
3534 reduction savings. */
3535 else if (TREE_CODE (c
->stride
) == INTEGER_CST
)
3536 replace_uncond_cands_and_profitable_phis (first_dep
);
3538 /* When the stride is an SSA name, it may still be profitable
3539 to replace some or all of the dependent candidates, depending
3540 on whether the introduced increments can be reused, or are
3541 less expensive to calculate than the replaced statements. */
3547 /* Determine whether we'll be generating pointer arithmetic
3548 when replacing candidates. */
3549 address_arithmetic_p
= (c
->kind
== CAND_ADD
3550 && POINTER_TYPE_P (c
->cand_type
));
3552 /* If all candidates have already been replaced under other
3553 interpretations, nothing remains to be done. */
3554 if (!count_candidates (c
))
3557 /* Construct an array of increments for this candidate chain. */
3558 incr_vec
= XNEWVEC (incr_info
, MAX_INCR_VEC_LEN
);
3560 record_increments (c
);
3562 /* Determine which increments are profitable to replace. */
3563 mode
= TYPE_MODE (TREE_TYPE (gimple_assign_lhs (c
->cand_stmt
)));
3564 speed
= optimize_cands_for_speed_p (c
);
3565 analyze_increments (first_dep
, mode
, speed
);
3567 /* Insert initializers of the form T_0 = stride * increment
3568 for use in profitable replacements. */
3569 insert_initializers (first_dep
);
3572 /* Perform the replacements. */
3573 replace_profitable_candidates (first_dep
);
3581 const pass_data pass_data_strength_reduction
=
3583 GIMPLE_PASS
, /* type */
3585 OPTGROUP_NONE
, /* optinfo_flags */
3586 TV_GIMPLE_SLSR
, /* tv_id */
3587 ( PROP_cfg
| PROP_ssa
), /* properties_required */
3588 0, /* properties_provided */
3589 0, /* properties_destroyed */
3590 0, /* todo_flags_start */
3591 0, /* todo_flags_finish */
3594 class pass_strength_reduction
: public gimple_opt_pass
3597 pass_strength_reduction (gcc::context
*ctxt
)
3598 : gimple_opt_pass (pass_data_strength_reduction
, ctxt
)
3601 /* opt_pass methods: */
3602 virtual bool gate (function
*) { return flag_tree_slsr
; }
3603 virtual unsigned int execute (function
*);
3605 }; // class pass_strength_reduction
3608 pass_strength_reduction::execute (function
*fun
)
3610 /* Create the obstack where candidates will reside. */
3611 gcc_obstack_init (&cand_obstack
);
3613 /* Allocate the candidate vector. */
3614 cand_vec
.create (128);
3616 /* Allocate the mapping from statements to candidate indices. */
3617 stmt_cand_map
= new hash_map
<gimple
*, slsr_cand_t
>;
3619 /* Create the obstack where candidate chains will reside. */
3620 gcc_obstack_init (&chain_obstack
);
3622 /* Allocate the mapping from base expressions to candidate chains. */
3623 base_cand_map
= new hash_table
<cand_chain_hasher
> (500);
3625 /* Allocate the mapping from bases to alternative bases. */
3626 alt_base_map
= new hash_map
<tree
, tree
>;
3628 /* Initialize the loop optimizer. We need to detect flow across
3629 back edges, and this gives us dominator information as well. */
3630 loop_optimizer_init (AVOID_CFG_MODIFICATIONS
);
3632 /* Walk the CFG in predominator order looking for strength reduction
3634 find_candidates_dom_walker (CDI_DOMINATORS
)
3635 .walk (fun
->cfg
->x_entry_block_ptr
);
3637 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3640 dump_cand_chains ();
3643 delete alt_base_map
;
3644 free_affine_expand_cache (&name_expansions
);
3646 /* Analyze costs and make appropriate replacements. */
3647 analyze_candidates_and_replace ();
3649 loop_optimizer_finalize ();
3650 delete base_cand_map
;
3651 base_cand_map
= NULL
;
3652 obstack_free (&chain_obstack
, NULL
);
3653 delete stmt_cand_map
;
3654 cand_vec
.release ();
3655 obstack_free (&cand_obstack
, NULL
);
3663 make_pass_strength_reduction (gcc::context
*ctxt
)
3665 return new pass_strength_reduction (ctxt
);