1 /* Target-dependent costs for expmed.cc.
2 Copyright (C) 1987-2024 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
23 #include "insn-codes.h"
38 /* Indicates the type of fixup needed after a constant multiplication.
39 BASIC_VARIANT means no fixup is needed, NEGATE_VARIANT means that
40 the result should be negated, and ADD_VARIANT means that the
41 multiplicand should be added to the result. */
42 enum mult_variant
{basic_variant
, negate_variant
, add_variant
};
44 bool choose_mult_variant (machine_mode
, HOST_WIDE_INT
,
45 struct algorithm
*, enum mult_variant
*, int);
47 /* This structure holds the "cost" of a multiply sequence. The
48 "cost" field holds the total rtx_cost of every operator in the
49 synthetic multiplication sequence, hence cost(a op b) is defined
50 as rtx_cost(op) + cost(a) + cost(b), where cost(leaf) is zero.
51 The "latency" field holds the minimum possible latency of the
52 synthetic multiply, on a hypothetical infinitely parallel CPU.
53 This is the critical path, or the maximum height, of the expression
54 tree which is the sum of rtx_costs on the most expensive path from
55 any leaf to the root. Hence latency(a op b) is defined as zero for
56 leaves and rtx_cost(op) + max(latency(a), latency(b)) otherwise. */
59 short cost
; /* Total rtx_cost of the multiplication sequence. */
60 short latency
; /* The latency of the multiplication sequence. */
63 /* This macro is used to compare a pointer to a mult_cost against an
64 single integer "rtx_cost" value. This is equivalent to the macro
65 CHEAPER_MULT_COST(X,Z) where Z = {Y,Y}. */
66 #define MULT_COST_LESS(X,Y) ((X)->cost < (Y) \
67 || ((X)->cost == (Y) && (X)->latency < (Y)))
69 /* This macro is used to compare two pointers to mult_costs against
70 each other. The macro returns true if X is cheaper than Y.
71 Currently, the cheaper of two mult_costs is the one with the
72 lower "cost". If "cost"s are tied, the lower latency is cheaper. */
73 #define CHEAPER_MULT_COST(X,Y) ((X)->cost < (Y)->cost \
74 || ((X)->cost == (Y)->cost \
75 && (X)->latency < (Y)->latency))
77 /* This structure records a sequence of operations.
78 `ops' is the number of operations recorded.
79 `cost' is their total cost.
80 The operations are stored in `op' and the corresponding
81 logarithms of the integer coefficients in `log'.
83 These are the operations:
85 alg_m total := multiplicand;
86 alg_shift total := total * coeff
87 alg_add_t_m2 total := total + multiplicand * coeff;
88 alg_sub_t_m2 total := total - multiplicand * coeff;
89 alg_add_factor total := total * coeff + total;
90 alg_sub_factor total := total * coeff - total;
91 alg_add_t2_m total := total * coeff + multiplicand;
92 alg_sub_t2_m total := total * coeff - multiplicand;
94 The first operand must be either alg_zero or alg_m. */
98 struct mult_cost cost
;
100 /* The size of the OP and LOG fields are not directly related to the
101 word size, but the worst-case algorithms will be if we have few
102 consecutive ones or zeros, i.e., a multiplicand like 10101010101...
103 In that case we will generate shift-by-2, add, shift-by-2, add,...,
104 in total wordsize operations. */
105 enum alg_code op
[MAX_BITS_PER_WORD
];
106 char log
[MAX_BITS_PER_WORD
];
109 /* The entry for our multiplication cache/hash table. */
110 struct alg_hash_entry
{
111 /* The number we are multiplying by. */
112 unsigned HOST_WIDE_INT t
;
114 /* The mode in which we are multiplying something by T. */
117 /* The best multiplication algorithm for t. */
120 /* The cost of multiplication if ALG_CODE is not alg_impossible.
121 Otherwise, the cost within which multiplication by T is
123 struct mult_cost cost
;
125 /* Optimized for speed? */
129 /* The number of cache/hash entries. */
130 #if HOST_BITS_PER_WIDE_INT == 64
131 #define NUM_ALG_HASH_ENTRIES 1031
133 #define NUM_ALG_HASH_ENTRIES 307
136 #define NUM_MODE_IP_INT (NUM_MODE_INT + NUM_MODE_PARTIAL_INT)
137 #define NUM_MODE_IPV_INT (NUM_MODE_IP_INT + NUM_MODE_VECTOR_INT)
139 struct expmed_op_cheap
{
140 bool cheap
[2][NUM_MODE_IPV_INT
];
143 struct expmed_op_costs
{
144 int cost
[2][NUM_MODE_IPV_INT
];
147 /* Target-dependent globals. */
148 struct target_expmed
{
149 /* Each entry of ALG_HASH caches alg_code for some integer. This is
150 actually a hash table. If we have a collision, that the older
151 entry is kicked out. */
152 struct alg_hash_entry x_alg_hash
[NUM_ALG_HASH_ENTRIES
];
154 /* True if x_alg_hash might already have been used. */
155 bool x_alg_hash_used_p
;
157 /* Nonzero means divides or modulus operations are relatively cheap for
158 powers of two, so don't use branches; emit the operation instead.
159 Usually, this will mean that the MD file will emit non-branch
161 struct expmed_op_cheap x_sdiv_pow2_cheap
;
162 struct expmed_op_cheap x_smod_pow2_cheap
;
164 /* Cost of various pieces of RTL. */
166 struct expmed_op_costs x_add_cost
;
167 struct expmed_op_costs x_neg_cost
;
168 int x_shift_cost
[2][NUM_MODE_IPV_INT
][MAX_BITS_PER_WORD
];
169 int x_shiftadd_cost
[2][NUM_MODE_IPV_INT
][MAX_BITS_PER_WORD
];
170 int x_shiftsub0_cost
[2][NUM_MODE_IPV_INT
][MAX_BITS_PER_WORD
];
171 int x_shiftsub1_cost
[2][NUM_MODE_IPV_INT
][MAX_BITS_PER_WORD
];
172 struct expmed_op_costs x_mul_cost
;
173 struct expmed_op_costs x_sdiv_cost
;
174 struct expmed_op_costs x_udiv_cost
;
175 int x_mul_widen_cost
[2][NUM_MODE_INT
];
176 int x_mul_highpart_cost
[2][NUM_MODE_INT
];
178 /* Conversion costs are only defined between two scalar integer modes
179 of different sizes. The first machine mode is the destination mode,
180 and the second is the source mode. */
181 int x_convert_cost
[2][NUM_MODE_IP_INT
][NUM_MODE_IP_INT
];
184 extern struct target_expmed default_target_expmed
;
185 #if SWITCHABLE_TARGET
186 extern struct target_expmed
*this_target_expmed
;
188 #define this_target_expmed (&default_target_expmed)
191 /* Return a pointer to the alg_hash_entry at IDX. */
193 inline struct alg_hash_entry
*
194 alg_hash_entry_ptr (int idx
)
196 return &this_target_expmed
->x_alg_hash
[idx
];
199 /* Return true if the x_alg_hash field might have been used. */
202 alg_hash_used_p (void)
204 return this_target_expmed
->x_alg_hash_used_p
;
207 /* Set whether the x_alg_hash field might have been used. */
210 set_alg_hash_used_p (bool usedp
)
212 this_target_expmed
->x_alg_hash_used_p
= usedp
;
215 /* Compute an index into the cost arrays by mode class. */
218 expmed_mode_index (machine_mode mode
)
220 switch (GET_MODE_CLASS (mode
))
223 return mode
- MIN_MODE_INT
;
224 case MODE_PARTIAL_INT
:
225 /* If there are no partial integer modes, help the compiler
226 to figure out this will never happen. See PR59934. */
227 if (MIN_MODE_PARTIAL_INT
!= VOIDmode
)
228 return mode
- MIN_MODE_PARTIAL_INT
+ NUM_MODE_INT
;
230 case MODE_VECTOR_INT
:
231 /* If there are no vector integer modes, help the compiler
232 to figure out this will never happen. See PR59934. */
233 if (MIN_MODE_VECTOR_INT
!= VOIDmode
)
234 return mode
- MIN_MODE_VECTOR_INT
+ NUM_MODE_IP_INT
;
242 /* Return a pointer to a boolean contained in EOC indicating whether
243 a particular operation performed in MODE is cheap when optimizing
247 expmed_op_cheap_ptr (struct expmed_op_cheap
*eoc
, bool speed
,
250 int idx
= expmed_mode_index (mode
);
251 return &eoc
->cheap
[speed
][idx
];
254 /* Return a pointer to a cost contained in COSTS when a particular
255 operation is performed in MODE when optimizing for SPEED. */
258 expmed_op_cost_ptr (struct expmed_op_costs
*costs
, bool speed
,
261 int idx
= expmed_mode_index (mode
);
262 return &costs
->cost
[speed
][idx
];
265 /* Subroutine of {set_,}sdiv_pow2_cheap. Not to be used otherwise. */
268 sdiv_pow2_cheap_ptr (bool speed
, machine_mode mode
)
270 return expmed_op_cheap_ptr (&this_target_expmed
->x_sdiv_pow2_cheap
,
274 /* Set whether a signed division by a power of 2 is cheap in MODE
275 when optimizing for SPEED. */
278 set_sdiv_pow2_cheap (bool speed
, machine_mode mode
, bool cheap_p
)
280 *sdiv_pow2_cheap_ptr (speed
, mode
) = cheap_p
;
283 /* Return whether a signed division by a power of 2 is cheap in MODE
284 when optimizing for SPEED. */
287 sdiv_pow2_cheap (bool speed
, machine_mode mode
)
289 return *sdiv_pow2_cheap_ptr (speed
, mode
);
292 /* Subroutine of {set_,}smod_pow2_cheap. Not to be used otherwise. */
295 smod_pow2_cheap_ptr (bool speed
, machine_mode mode
)
297 return expmed_op_cheap_ptr (&this_target_expmed
->x_smod_pow2_cheap
,
301 /* Set whether a signed modulo by a power of 2 is CHEAP in MODE when
302 optimizing for SPEED. */
305 set_smod_pow2_cheap (bool speed
, machine_mode mode
, bool cheap
)
307 *smod_pow2_cheap_ptr (speed
, mode
) = cheap
;
310 /* Return whether a signed modulo by a power of 2 is cheap in MODE
311 when optimizing for SPEED. */
314 smod_pow2_cheap (bool speed
, machine_mode mode
)
316 return *smod_pow2_cheap_ptr (speed
, mode
);
319 /* Subroutine of {set_,}zero_cost. Not to be used otherwise. */
322 zero_cost_ptr (bool speed
)
324 return &this_target_expmed
->x_zero_cost
[speed
];
327 /* Set the COST of loading zero when optimizing for SPEED. */
330 set_zero_cost (bool speed
, int cost
)
332 *zero_cost_ptr (speed
) = cost
;
335 /* Return the COST of loading zero when optimizing for SPEED. */
338 zero_cost (bool speed
)
340 return *zero_cost_ptr (speed
);
343 /* Subroutine of {set_,}add_cost. Not to be used otherwise. */
346 add_cost_ptr (bool speed
, machine_mode mode
)
348 return expmed_op_cost_ptr (&this_target_expmed
->x_add_cost
, speed
, mode
);
351 /* Set the COST of computing an add in MODE when optimizing for SPEED. */
354 set_add_cost (bool speed
, machine_mode mode
, int cost
)
356 *add_cost_ptr (speed
, mode
) = cost
;
359 /* Return the cost of computing an add in MODE when optimizing for SPEED. */
362 add_cost (bool speed
, machine_mode mode
)
364 return *add_cost_ptr (speed
, mode
);
367 /* Subroutine of {set_,}neg_cost. Not to be used otherwise. */
370 neg_cost_ptr (bool speed
, machine_mode mode
)
372 return expmed_op_cost_ptr (&this_target_expmed
->x_neg_cost
, speed
, mode
);
375 /* Set the COST of computing a negation in MODE when optimizing for SPEED. */
378 set_neg_cost (bool speed
, machine_mode mode
, int cost
)
380 *neg_cost_ptr (speed
, mode
) = cost
;
383 /* Return the cost of computing a negation in MODE when optimizing for
387 neg_cost (bool speed
, machine_mode mode
)
389 return *neg_cost_ptr (speed
, mode
);
392 /* Subroutine of {set_,}shift_cost. Not to be used otherwise. */
395 shift_cost_ptr (bool speed
, machine_mode mode
, int bits
)
397 int midx
= expmed_mode_index (mode
);
398 return &this_target_expmed
->x_shift_cost
[speed
][midx
][bits
];
401 /* Set the COST of doing a shift in MODE by BITS when optimizing for SPEED. */
404 set_shift_cost (bool speed
, machine_mode mode
, int bits
, int cost
)
406 *shift_cost_ptr (speed
, mode
, bits
) = cost
;
409 /* Return the cost of doing a shift in MODE by BITS when optimizing for
413 shift_cost (bool speed
, machine_mode mode
, int bits
)
415 return *shift_cost_ptr (speed
, mode
, bits
);
418 /* Subroutine of {set_,}shiftadd_cost. Not to be used otherwise. */
421 shiftadd_cost_ptr (bool speed
, machine_mode mode
, int bits
)
423 int midx
= expmed_mode_index (mode
);
424 return &this_target_expmed
->x_shiftadd_cost
[speed
][midx
][bits
];
427 /* Set the COST of doing a shift in MODE by BITS followed by an add when
428 optimizing for SPEED. */
431 set_shiftadd_cost (bool speed
, machine_mode mode
, int bits
, int cost
)
433 *shiftadd_cost_ptr (speed
, mode
, bits
) = cost
;
436 /* Return the cost of doing a shift in MODE by BITS followed by an add
437 when optimizing for SPEED. */
440 shiftadd_cost (bool speed
, machine_mode mode
, int bits
)
442 return *shiftadd_cost_ptr (speed
, mode
, bits
);
445 /* Subroutine of {set_,}shiftsub0_cost. Not to be used otherwise. */
448 shiftsub0_cost_ptr (bool speed
, machine_mode mode
, int bits
)
450 int midx
= expmed_mode_index (mode
);
451 return &this_target_expmed
->x_shiftsub0_cost
[speed
][midx
][bits
];
454 /* Set the COST of doing a shift in MODE by BITS and then subtracting a
455 value when optimizing for SPEED. */
458 set_shiftsub0_cost (bool speed
, machine_mode mode
, int bits
, int cost
)
460 *shiftsub0_cost_ptr (speed
, mode
, bits
) = cost
;
463 /* Return the cost of doing a shift in MODE by BITS and then subtracting
464 a value when optimizing for SPEED. */
467 shiftsub0_cost (bool speed
, machine_mode mode
, int bits
)
469 return *shiftsub0_cost_ptr (speed
, mode
, bits
);
472 /* Subroutine of {set_,}shiftsub1_cost. Not to be used otherwise. */
475 shiftsub1_cost_ptr (bool speed
, machine_mode mode
, int bits
)
477 int midx
= expmed_mode_index (mode
);
478 return &this_target_expmed
->x_shiftsub1_cost
[speed
][midx
][bits
];
481 /* Set the COST of subtracting a shift in MODE by BITS from a value when
482 optimizing for SPEED. */
485 set_shiftsub1_cost (bool speed
, machine_mode mode
, int bits
, int cost
)
487 *shiftsub1_cost_ptr (speed
, mode
, bits
) = cost
;
490 /* Return the cost of subtracting a shift in MODE by BITS from a value
491 when optimizing for SPEED. */
494 shiftsub1_cost (bool speed
, machine_mode mode
, int bits
)
496 return *shiftsub1_cost_ptr (speed
, mode
, bits
);
499 /* Subroutine of {set_,}mul_cost. Not to be used otherwise. */
502 mul_cost_ptr (bool speed
, machine_mode mode
)
504 return expmed_op_cost_ptr (&this_target_expmed
->x_mul_cost
, speed
, mode
);
507 /* Set the COST of doing a multiplication in MODE when optimizing for
511 set_mul_cost (bool speed
, machine_mode mode
, int cost
)
513 *mul_cost_ptr (speed
, mode
) = cost
;
516 /* Return the cost of doing a multiplication in MODE when optimizing
520 mul_cost (bool speed
, machine_mode mode
)
522 return *mul_cost_ptr (speed
, mode
);
525 /* Subroutine of {set_,}sdiv_cost. Not to be used otherwise. */
528 sdiv_cost_ptr (bool speed
, machine_mode mode
)
530 return expmed_op_cost_ptr (&this_target_expmed
->x_sdiv_cost
, speed
, mode
);
533 /* Set the COST of doing a signed division in MODE when optimizing
537 set_sdiv_cost (bool speed
, machine_mode mode
, int cost
)
539 *sdiv_cost_ptr (speed
, mode
) = cost
;
542 /* Return the cost of doing a signed division in MODE when optimizing
546 sdiv_cost (bool speed
, machine_mode mode
)
548 return *sdiv_cost_ptr (speed
, mode
);
551 /* Subroutine of {set_,}udiv_cost. Not to be used otherwise. */
554 udiv_cost_ptr (bool speed
, machine_mode mode
)
556 return expmed_op_cost_ptr (&this_target_expmed
->x_udiv_cost
, speed
, mode
);
559 /* Set the COST of doing an unsigned division in MODE when optimizing
563 set_udiv_cost (bool speed
, machine_mode mode
, int cost
)
565 *udiv_cost_ptr (speed
, mode
) = cost
;
568 /* Return the cost of doing an unsigned division in MODE when
569 optimizing for SPEED. */
572 udiv_cost (bool speed
, machine_mode mode
)
574 return *udiv_cost_ptr (speed
, mode
);
577 /* Subroutine of {set_,}mul_widen_cost. Not to be used otherwise. */
580 mul_widen_cost_ptr (bool speed
, machine_mode mode
)
582 gcc_assert (GET_MODE_CLASS (mode
) == MODE_INT
);
584 return &this_target_expmed
->x_mul_widen_cost
[speed
][mode
- MIN_MODE_INT
];
587 /* Set the COST for computing a widening multiplication in MODE when
588 optimizing for SPEED. */
591 set_mul_widen_cost (bool speed
, machine_mode mode
, int cost
)
593 *mul_widen_cost_ptr (speed
, mode
) = cost
;
596 /* Return the cost for computing a widening multiplication in MODE when
597 optimizing for SPEED. */
600 mul_widen_cost (bool speed
, machine_mode mode
)
602 return *mul_widen_cost_ptr (speed
, mode
);
605 /* Subroutine of {set_,}mul_highpart_cost. Not to be used otherwise. */
608 mul_highpart_cost_ptr (bool speed
, machine_mode mode
)
610 gcc_assert (GET_MODE_CLASS (mode
) == MODE_INT
);
611 int m
= mode
- MIN_MODE_INT
;
612 gcc_assert (m
< NUM_MODE_INT
);
614 return &this_target_expmed
->x_mul_highpart_cost
[speed
][m
];
617 /* Set the COST for computing the high part of a multiplication in MODE
618 when optimizing for SPEED. */
621 set_mul_highpart_cost (bool speed
, machine_mode mode
, int cost
)
623 *mul_highpart_cost_ptr (speed
, mode
) = cost
;
626 /* Return the cost for computing the high part of a multiplication in MODE
627 when optimizing for SPEED. */
630 mul_highpart_cost (bool speed
, machine_mode mode
)
632 return *mul_highpart_cost_ptr (speed
, mode
);
635 /* Subroutine of {set_,}convert_cost. Not to be used otherwise. */
638 convert_cost_ptr (machine_mode to_mode
, machine_mode from_mode
,
641 int to_idx
= expmed_mode_index (to_mode
);
642 int from_idx
= expmed_mode_index (from_mode
);
644 gcc_assert (IN_RANGE (to_idx
, 0, NUM_MODE_IP_INT
- 1));
645 gcc_assert (IN_RANGE (from_idx
, 0, NUM_MODE_IP_INT
- 1));
647 return &this_target_expmed
->x_convert_cost
[speed
][to_idx
][from_idx
];
650 /* Set the COST for converting from FROM_MODE to TO_MODE when optimizing
654 set_convert_cost (machine_mode to_mode
, machine_mode from_mode
,
655 bool speed
, int cost
)
657 *convert_cost_ptr (to_mode
, from_mode
, speed
) = cost
;
660 /* Return the cost for converting from FROM_MODE to TO_MODE when optimizing
664 convert_cost (machine_mode to_mode
, machine_mode from_mode
,
667 return *convert_cost_ptr (to_mode
, from_mode
, speed
);
670 extern int mult_by_coeff_cost (HOST_WIDE_INT
, machine_mode
, bool);
671 extern rtx
emit_cstore (rtx target
, enum insn_code icode
, enum rtx_code code
,
672 machine_mode mode
, machine_mode compare_mode
,
673 int unsignedp
, rtx x
, rtx y
, int normalizep
,
674 machine_mode target_mode
);
676 /* Arguments MODE, RTX: return an rtx for the negation of that value.
678 extern rtx
negate_rtx (machine_mode
, rtx
);
680 /* Arguments MODE, RTX: return an rtx for the flipping of that value.
682 extern rtx
flip_storage_order (machine_mode
, rtx
);
684 /* Expand a logical AND operation. */
685 extern rtx
expand_and (machine_mode
, rtx
, rtx
, rtx
);
687 /* Emit a store-flag operation. */
688 extern rtx
emit_store_flag (rtx
, enum rtx_code
, rtx
, rtx
, machine_mode
,
691 /* Like emit_store_flag, but always succeeds. */
692 extern rtx
emit_store_flag_force (rtx
, enum rtx_code
, rtx
, rtx
,
693 machine_mode
, int, int);
695 extern void canonicalize_comparison (machine_mode
, enum rtx_code
*, rtx
*);
697 /* Choose a minimal N + 1 bit approximation to 1/D that can be used to
698 replace division by D, and put the least significant N bits of the result
699 in *MULTIPLIER_PTR and return the most significant bit. */
700 extern unsigned HOST_WIDE_INT
choose_multiplier (unsigned HOST_WIDE_INT
, int,
701 int, unsigned HOST_WIDE_INT
*,
705 extern rtx
expand_variable_shift (enum tree_code
, machine_mode
,
706 rtx
, tree
, rtx
, int);
707 extern rtx
expand_shift (enum tree_code
, machine_mode
, rtx
, poly_int64
, rtx
,
709 extern rtx
maybe_expand_shift (enum tree_code
, machine_mode
, rtx
, int, rtx
,
712 extern rtx
expand_divmod (int, enum tree_code
, machine_mode
, rtx
, rtx
,
713 rtx
, int, enum optab_methods
= OPTAB_LIB_WIDEN
);
717 extern void store_bit_field (rtx
, poly_uint64
, poly_uint64
,
718 poly_uint64
, poly_uint64
,
719 machine_mode
, rtx
, bool, bool);
720 extern rtx
extract_bit_field (rtx
, poly_uint64
, poly_uint64
, int, rtx
,
721 machine_mode
, machine_mode
, bool, rtx
*);
722 extern rtx
extract_low_bits (machine_mode
, machine_mode
, rtx
);
723 extern rtx
expand_mult (machine_mode
, rtx
, rtx
, rtx
, int, bool = false);
724 extern rtx
expand_mult_highpart_adjust (scalar_int_mode
, rtx
, rtx
, rtx
,