1 /* Target-dependent costs for expmed.c.
2 Copyright (C) 1987-2021 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_INT \
137 (MAX_MODE_INT - MIN_MODE_INT + 1)
138 #define NUM_MODE_PARTIAL_INT \
139 (MIN_MODE_PARTIAL_INT == E_VOIDmode ? 0 \
140 : MAX_MODE_PARTIAL_INT - MIN_MODE_PARTIAL_INT + 1)
141 #define NUM_MODE_VECTOR_INT \
142 (MIN_MODE_VECTOR_INT == E_VOIDmode ? 0 \
143 : MAX_MODE_VECTOR_INT - MIN_MODE_VECTOR_INT + 1)
145 #define NUM_MODE_IP_INT (NUM_MODE_INT + NUM_MODE_PARTIAL_INT)
146 #define NUM_MODE_IPV_INT (NUM_MODE_IP_INT + NUM_MODE_VECTOR_INT)
148 struct expmed_op_cheap
{
149 bool cheap
[2][NUM_MODE_IPV_INT
];
152 struct expmed_op_costs
{
153 int cost
[2][NUM_MODE_IPV_INT
];
156 /* Target-dependent globals. */
157 struct target_expmed
{
158 /* Each entry of ALG_HASH caches alg_code for some integer. This is
159 actually a hash table. If we have a collision, that the older
160 entry is kicked out. */
161 struct alg_hash_entry x_alg_hash
[NUM_ALG_HASH_ENTRIES
];
163 /* True if x_alg_hash might already have been used. */
164 bool x_alg_hash_used_p
;
166 /* Nonzero means divides or modulus operations are relatively cheap for
167 powers of two, so don't use branches; emit the operation instead.
168 Usually, this will mean that the MD file will emit non-branch
170 struct expmed_op_cheap x_sdiv_pow2_cheap
;
171 struct expmed_op_cheap x_smod_pow2_cheap
;
173 /* Cost of various pieces of RTL. Note that some of these are indexed by
174 shift count and some by mode. */
176 struct expmed_op_costs x_add_cost
;
177 struct expmed_op_costs x_neg_cost
;
178 struct expmed_op_costs x_shift_cost
[MAX_BITS_PER_WORD
];
179 struct expmed_op_costs x_shiftadd_cost
[MAX_BITS_PER_WORD
];
180 struct expmed_op_costs x_shiftsub0_cost
[MAX_BITS_PER_WORD
];
181 struct expmed_op_costs x_shiftsub1_cost
[MAX_BITS_PER_WORD
];
182 struct expmed_op_costs x_mul_cost
;
183 struct expmed_op_costs x_sdiv_cost
;
184 struct expmed_op_costs x_udiv_cost
;
185 int x_mul_widen_cost
[2][NUM_MODE_INT
];
186 int x_mul_highpart_cost
[2][NUM_MODE_INT
];
188 /* Conversion costs are only defined between two scalar integer modes
189 of different sizes. The first machine mode is the destination mode,
190 and the second is the source mode. */
191 int x_convert_cost
[2][NUM_MODE_IP_INT
][NUM_MODE_IP_INT
];
194 extern struct target_expmed default_target_expmed
;
195 #if SWITCHABLE_TARGET
196 extern struct target_expmed
*this_target_expmed
;
198 #define this_target_expmed (&default_target_expmed)
201 /* Return a pointer to the alg_hash_entry at IDX. */
203 static inline struct alg_hash_entry
*
204 alg_hash_entry_ptr (int idx
)
206 return &this_target_expmed
->x_alg_hash
[idx
];
209 /* Return true if the x_alg_hash field might have been used. */
212 alg_hash_used_p (void)
214 return this_target_expmed
->x_alg_hash_used_p
;
217 /* Set whether the x_alg_hash field might have been used. */
220 set_alg_hash_used_p (bool usedp
)
222 this_target_expmed
->x_alg_hash_used_p
= usedp
;
225 /* Compute an index into the cost arrays by mode class. */
228 expmed_mode_index (machine_mode mode
)
230 switch (GET_MODE_CLASS (mode
))
233 return mode
- MIN_MODE_INT
;
234 case MODE_PARTIAL_INT
:
235 /* If there are no partial integer modes, help the compiler
236 to figure out this will never happen. See PR59934. */
237 if (MIN_MODE_PARTIAL_INT
!= VOIDmode
)
238 return mode
- MIN_MODE_PARTIAL_INT
+ NUM_MODE_INT
;
240 case MODE_VECTOR_INT
:
241 /* If there are no vector integer modes, help the compiler
242 to figure out this will never happen. See PR59934. */
243 if (MIN_MODE_VECTOR_INT
!= VOIDmode
)
244 return mode
- MIN_MODE_VECTOR_INT
+ NUM_MODE_IP_INT
;
252 /* Return a pointer to a boolean contained in EOC indicating whether
253 a particular operation performed in MODE is cheap when optimizing
257 expmed_op_cheap_ptr (struct expmed_op_cheap
*eoc
, bool speed
,
260 int idx
= expmed_mode_index (mode
);
261 return &eoc
->cheap
[speed
][idx
];
264 /* Return a pointer to a cost contained in COSTS when a particular
265 operation is performed in MODE when optimizing for SPEED. */
268 expmed_op_cost_ptr (struct expmed_op_costs
*costs
, bool speed
,
271 int idx
= expmed_mode_index (mode
);
272 return &costs
->cost
[speed
][idx
];
275 /* Subroutine of {set_,}sdiv_pow2_cheap. Not to be used otherwise. */
278 sdiv_pow2_cheap_ptr (bool speed
, machine_mode mode
)
280 return expmed_op_cheap_ptr (&this_target_expmed
->x_sdiv_pow2_cheap
,
284 /* Set whether a signed division by a power of 2 is cheap in MODE
285 when optimizing for SPEED. */
288 set_sdiv_pow2_cheap (bool speed
, machine_mode mode
, bool cheap_p
)
290 *sdiv_pow2_cheap_ptr (speed
, mode
) = cheap_p
;
293 /* Return whether a signed division by a power of 2 is cheap in MODE
294 when optimizing for SPEED. */
297 sdiv_pow2_cheap (bool speed
, machine_mode mode
)
299 return *sdiv_pow2_cheap_ptr (speed
, mode
);
302 /* Subroutine of {set_,}smod_pow2_cheap. Not to be used otherwise. */
305 smod_pow2_cheap_ptr (bool speed
, machine_mode mode
)
307 return expmed_op_cheap_ptr (&this_target_expmed
->x_smod_pow2_cheap
,
311 /* Set whether a signed modulo by a power of 2 is CHEAP in MODE when
312 optimizing for SPEED. */
315 set_smod_pow2_cheap (bool speed
, machine_mode mode
, bool cheap
)
317 *smod_pow2_cheap_ptr (speed
, mode
) = cheap
;
320 /* Return whether a signed modulo by a power of 2 is cheap in MODE
321 when optimizing for SPEED. */
324 smod_pow2_cheap (bool speed
, machine_mode mode
)
326 return *smod_pow2_cheap_ptr (speed
, mode
);
329 /* Subroutine of {set_,}zero_cost. Not to be used otherwise. */
332 zero_cost_ptr (bool speed
)
334 return &this_target_expmed
->x_zero_cost
[speed
];
337 /* Set the COST of loading zero when optimizing for SPEED. */
340 set_zero_cost (bool speed
, int cost
)
342 *zero_cost_ptr (speed
) = cost
;
345 /* Return the COST of loading zero when optimizing for SPEED. */
348 zero_cost (bool speed
)
350 return *zero_cost_ptr (speed
);
353 /* Subroutine of {set_,}add_cost. Not to be used otherwise. */
356 add_cost_ptr (bool speed
, machine_mode mode
)
358 return expmed_op_cost_ptr (&this_target_expmed
->x_add_cost
, speed
, mode
);
361 /* Set the COST of computing an add in MODE when optimizing for SPEED. */
364 set_add_cost (bool speed
, machine_mode mode
, int cost
)
366 *add_cost_ptr (speed
, mode
) = cost
;
369 /* Return the cost of computing an add in MODE when optimizing for SPEED. */
372 add_cost (bool speed
, machine_mode mode
)
374 return *add_cost_ptr (speed
, mode
);
377 /* Subroutine of {set_,}neg_cost. Not to be used otherwise. */
380 neg_cost_ptr (bool speed
, machine_mode mode
)
382 return expmed_op_cost_ptr (&this_target_expmed
->x_neg_cost
, speed
, mode
);
385 /* Set the COST of computing a negation in MODE when optimizing for SPEED. */
388 set_neg_cost (bool speed
, machine_mode mode
, int cost
)
390 *neg_cost_ptr (speed
, mode
) = cost
;
393 /* Return the cost of computing a negation in MODE when optimizing for
397 neg_cost (bool speed
, machine_mode mode
)
399 return *neg_cost_ptr (speed
, mode
);
402 /* Subroutine of {set_,}shift_cost. Not to be used otherwise. */
405 shift_cost_ptr (bool speed
, machine_mode mode
, int bits
)
407 return expmed_op_cost_ptr (&this_target_expmed
->x_shift_cost
[bits
],
411 /* Set the COST of doing a shift in MODE by BITS when optimizing for SPEED. */
414 set_shift_cost (bool speed
, machine_mode mode
, int bits
, int cost
)
416 *shift_cost_ptr (speed
, mode
, bits
) = cost
;
419 /* Return the cost of doing a shift in MODE by BITS when optimizing for
423 shift_cost (bool speed
, machine_mode mode
, int bits
)
425 return *shift_cost_ptr (speed
, mode
, bits
);
428 /* Subroutine of {set_,}shiftadd_cost. Not to be used otherwise. */
431 shiftadd_cost_ptr (bool speed
, machine_mode mode
, int bits
)
433 return expmed_op_cost_ptr (&this_target_expmed
->x_shiftadd_cost
[bits
],
437 /* Set the COST of doing a shift in MODE by BITS followed by an add when
438 optimizing for SPEED. */
441 set_shiftadd_cost (bool speed
, machine_mode mode
, int bits
, int cost
)
443 *shiftadd_cost_ptr (speed
, mode
, bits
) = cost
;
446 /* Return the cost of doing a shift in MODE by BITS followed by an add
447 when optimizing for SPEED. */
450 shiftadd_cost (bool speed
, machine_mode mode
, int bits
)
452 return *shiftadd_cost_ptr (speed
, mode
, bits
);
455 /* Subroutine of {set_,}shiftsub0_cost. Not to be used otherwise. */
458 shiftsub0_cost_ptr (bool speed
, machine_mode mode
, int bits
)
460 return expmed_op_cost_ptr (&this_target_expmed
->x_shiftsub0_cost
[bits
],
464 /* Set the COST of doing a shift in MODE by BITS and then subtracting a
465 value when optimizing for SPEED. */
468 set_shiftsub0_cost (bool speed
, machine_mode mode
, int bits
, int cost
)
470 *shiftsub0_cost_ptr (speed
, mode
, bits
) = cost
;
473 /* Return the cost of doing a shift in MODE by BITS and then subtracting
474 a value when optimizing for SPEED. */
477 shiftsub0_cost (bool speed
, machine_mode mode
, int bits
)
479 return *shiftsub0_cost_ptr (speed
, mode
, bits
);
482 /* Subroutine of {set_,}shiftsub1_cost. Not to be used otherwise. */
485 shiftsub1_cost_ptr (bool speed
, machine_mode mode
, int bits
)
487 return expmed_op_cost_ptr (&this_target_expmed
->x_shiftsub1_cost
[bits
],
491 /* Set the COST of subtracting a shift in MODE by BITS from a value when
492 optimizing for SPEED. */
495 set_shiftsub1_cost (bool speed
, machine_mode mode
, int bits
, int cost
)
497 *shiftsub1_cost_ptr (speed
, mode
, bits
) = cost
;
500 /* Return the cost of subtracting a shift in MODE by BITS from a value
501 when optimizing for SPEED. */
504 shiftsub1_cost (bool speed
, machine_mode mode
, int bits
)
506 return *shiftsub1_cost_ptr (speed
, mode
, bits
);
509 /* Subroutine of {set_,}mul_cost. Not to be used otherwise. */
512 mul_cost_ptr (bool speed
, machine_mode mode
)
514 return expmed_op_cost_ptr (&this_target_expmed
->x_mul_cost
, speed
, mode
);
517 /* Set the COST of doing a multiplication in MODE when optimizing for
521 set_mul_cost (bool speed
, machine_mode mode
, int cost
)
523 *mul_cost_ptr (speed
, mode
) = cost
;
526 /* Return the cost of doing a multiplication in MODE when optimizing
530 mul_cost (bool speed
, machine_mode mode
)
532 return *mul_cost_ptr (speed
, mode
);
535 /* Subroutine of {set_,}sdiv_cost. Not to be used otherwise. */
538 sdiv_cost_ptr (bool speed
, machine_mode mode
)
540 return expmed_op_cost_ptr (&this_target_expmed
->x_sdiv_cost
, speed
, mode
);
543 /* Set the COST of doing a signed division in MODE when optimizing
547 set_sdiv_cost (bool speed
, machine_mode mode
, int cost
)
549 *sdiv_cost_ptr (speed
, mode
) = cost
;
552 /* Return the cost of doing a signed division in MODE when optimizing
556 sdiv_cost (bool speed
, machine_mode mode
)
558 return *sdiv_cost_ptr (speed
, mode
);
561 /* Subroutine of {set_,}udiv_cost. Not to be used otherwise. */
564 udiv_cost_ptr (bool speed
, machine_mode mode
)
566 return expmed_op_cost_ptr (&this_target_expmed
->x_udiv_cost
, speed
, mode
);
569 /* Set the COST of doing an unsigned division in MODE when optimizing
573 set_udiv_cost (bool speed
, machine_mode mode
, int cost
)
575 *udiv_cost_ptr (speed
, mode
) = cost
;
578 /* Return the cost of doing an unsigned division in MODE when
579 optimizing for SPEED. */
582 udiv_cost (bool speed
, machine_mode mode
)
584 return *udiv_cost_ptr (speed
, mode
);
587 /* Subroutine of {set_,}mul_widen_cost. Not to be used otherwise. */
590 mul_widen_cost_ptr (bool speed
, machine_mode mode
)
592 gcc_assert (GET_MODE_CLASS (mode
) == MODE_INT
);
594 return &this_target_expmed
->x_mul_widen_cost
[speed
][mode
- MIN_MODE_INT
];
597 /* Set the COST for computing a widening multiplication in MODE when
598 optimizing for SPEED. */
601 set_mul_widen_cost (bool speed
, machine_mode mode
, int cost
)
603 *mul_widen_cost_ptr (speed
, mode
) = cost
;
606 /* Return the cost for computing a widening multiplication in MODE when
607 optimizing for SPEED. */
610 mul_widen_cost (bool speed
, machine_mode mode
)
612 return *mul_widen_cost_ptr (speed
, mode
);
615 /* Subroutine of {set_,}mul_highpart_cost. Not to be used otherwise. */
618 mul_highpart_cost_ptr (bool speed
, machine_mode mode
)
620 gcc_assert (GET_MODE_CLASS (mode
) == MODE_INT
);
621 int m
= mode
- MIN_MODE_INT
;
622 gcc_assert (m
< NUM_MODE_INT
);
624 return &this_target_expmed
->x_mul_highpart_cost
[speed
][m
];
627 /* Set the COST for computing the high part of a multiplication in MODE
628 when optimizing for SPEED. */
631 set_mul_highpart_cost (bool speed
, machine_mode mode
, int cost
)
633 *mul_highpart_cost_ptr (speed
, mode
) = cost
;
636 /* Return the cost for computing the high part of a multiplication in MODE
637 when optimizing for SPEED. */
640 mul_highpart_cost (bool speed
, machine_mode mode
)
642 return *mul_highpart_cost_ptr (speed
, mode
);
645 /* Subroutine of {set_,}convert_cost. Not to be used otherwise. */
648 convert_cost_ptr (machine_mode to_mode
, machine_mode from_mode
,
651 int to_idx
= expmed_mode_index (to_mode
);
652 int from_idx
= expmed_mode_index (from_mode
);
654 gcc_assert (IN_RANGE (to_idx
, 0, NUM_MODE_IP_INT
- 1));
655 gcc_assert (IN_RANGE (from_idx
, 0, NUM_MODE_IP_INT
- 1));
657 return &this_target_expmed
->x_convert_cost
[speed
][to_idx
][from_idx
];
660 /* Set the COST for converting from FROM_MODE to TO_MODE when optimizing
664 set_convert_cost (machine_mode to_mode
, machine_mode from_mode
,
665 bool speed
, int cost
)
667 *convert_cost_ptr (to_mode
, from_mode
, speed
) = cost
;
670 /* Return the cost for converting from FROM_MODE to TO_MODE when optimizing
674 convert_cost (machine_mode to_mode
, machine_mode from_mode
,
677 return *convert_cost_ptr (to_mode
, from_mode
, speed
);
680 extern int mult_by_coeff_cost (HOST_WIDE_INT
, machine_mode
, bool);
681 extern rtx
emit_cstore (rtx target
, enum insn_code icode
, enum rtx_code code
,
682 machine_mode mode
, machine_mode compare_mode
,
683 int unsignedp
, rtx x
, rtx y
, int normalizep
,
684 machine_mode target_mode
);
686 /* Arguments MODE, RTX: return an rtx for the negation of that value.
688 extern rtx
negate_rtx (machine_mode
, rtx
);
690 /* Arguments MODE, RTX: return an rtx for the flipping of that value.
692 extern rtx
flip_storage_order (machine_mode
, rtx
);
694 /* Expand a logical AND operation. */
695 extern rtx
expand_and (machine_mode
, rtx
, rtx
, rtx
);
697 /* Emit a store-flag operation. */
698 extern rtx
emit_store_flag (rtx
, enum rtx_code
, rtx
, rtx
, machine_mode
,
701 /* Like emit_store_flag, but always succeeds. */
702 extern rtx
emit_store_flag_force (rtx
, enum rtx_code
, rtx
, rtx
,
703 machine_mode
, int, int);
705 extern void canonicalize_comparison (machine_mode
, enum rtx_code
*, rtx
*);
707 /* Choose a minimal N + 1 bit approximation to 1/D that can be used to
708 replace division by D, and put the least significant N bits of the result
709 in *MULTIPLIER_PTR and return the most significant bit. */
710 extern unsigned HOST_WIDE_INT
choose_multiplier (unsigned HOST_WIDE_INT
, int,
711 int, unsigned HOST_WIDE_INT
*,
715 extern rtx
expand_variable_shift (enum tree_code
, machine_mode
,
716 rtx
, tree
, rtx
, int);
717 extern rtx
expand_shift (enum tree_code
, machine_mode
, rtx
, poly_int64
, rtx
,
720 extern rtx
expand_divmod (int, enum tree_code
, machine_mode
, rtx
, rtx
,
721 rtx
, int, enum optab_methods
= OPTAB_LIB_WIDEN
);
725 extern void store_bit_field (rtx
, poly_uint64
, poly_uint64
,
726 poly_uint64
, poly_uint64
,
727 machine_mode
, rtx
, bool);
728 extern rtx
extract_bit_field (rtx
, poly_uint64
, poly_uint64
, int, rtx
,
729 machine_mode
, machine_mode
, bool, rtx
*);
730 extern rtx
extract_low_bits (machine_mode
, machine_mode
, rtx
);
731 extern rtx
expand_mult (machine_mode
, rtx
, rtx
, rtx
, int, bool = false);
732 extern rtx
expand_mult_highpart_adjust (scalar_int_mode
, rtx
, rtx
, rtx
,