1 /* Target-dependent costs for expmed.cc.
2 Copyright (C) 1987-2022 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. Note that some of these are indexed by
165 shift count and some by mode. */
167 struct expmed_op_costs x_add_cost;
168 struct expmed_op_costs x_neg_cost;
169 struct expmed_op_costs x_shift_cost[MAX_BITS_PER_WORD];
170 struct expmed_op_costs x_shiftadd_cost[MAX_BITS_PER_WORD];
171 struct expmed_op_costs x_shiftsub0_cost[MAX_BITS_PER_WORD];
172 struct expmed_op_costs x_shiftsub1_cost[MAX_BITS_PER_WORD];
173 struct expmed_op_costs x_mul_cost;
174 struct expmed_op_costs x_sdiv_cost;
175 struct expmed_op_costs x_udiv_cost;
176 int x_mul_widen_cost[2][NUM_MODE_INT];
177 int x_mul_highpart_cost[2][NUM_MODE_INT];
179 /* Conversion costs are only defined between two scalar integer modes
180 of different sizes. The first machine mode is the destination mode,
181 and the second is the source mode. */
182 int x_convert_cost[2][NUM_MODE_IP_INT][NUM_MODE_IP_INT];
185 extern struct target_expmed default_target_expmed;
186 #if SWITCHABLE_TARGET
187 extern struct target_expmed *this_target_expmed;
189 #define this_target_expmed (&default_target_expmed)
192 /* Return a pointer to the alg_hash_entry at IDX. */
194 static inline struct alg_hash_entry *
195 alg_hash_entry_ptr (int idx)
197 return &this_target_expmed->x_alg_hash[idx];
200 /* Return true if the x_alg_hash field might have been used. */
203 alg_hash_used_p (void)
205 return this_target_expmed->x_alg_hash_used_p;
208 /* Set whether the x_alg_hash field might have been used. */
211 set_alg_hash_used_p (bool usedp)
213 this_target_expmed->x_alg_hash_used_p = usedp;
216 /* Compute an index into the cost arrays by mode class. */
219 expmed_mode_index (machine_mode mode)
221 switch (GET_MODE_CLASS (mode))
224 return mode - MIN_MODE_INT;
225 case MODE_PARTIAL_INT:
226 /* If there are no partial integer modes, help the compiler
227 to figure out this will never happen. See PR59934. */
228 if (MIN_MODE_PARTIAL_INT != VOIDmode)
229 return mode - MIN_MODE_PARTIAL_INT + NUM_MODE_INT;
231 case MODE_VECTOR_INT:
232 /* If there are no vector integer modes, help the compiler
233 to figure out this will never happen. See PR59934. */
234 if (MIN_MODE_VECTOR_INT != VOIDmode)
235 return mode - MIN_MODE_VECTOR_INT + NUM_MODE_IP_INT;
243 /* Return a pointer to a boolean contained in EOC indicating whether
244 a particular operation performed in MODE is cheap when optimizing
248 expmed_op_cheap_ptr (struct expmed_op_cheap *eoc, bool speed,
251 int idx = expmed_mode_index (mode);
252 return &eoc->cheap[speed][idx];
255 /* Return a pointer to a cost contained in COSTS when a particular
256 operation is performed in MODE when optimizing for SPEED. */
259 expmed_op_cost_ptr (struct expmed_op_costs *costs, bool speed,
262 int idx = expmed_mode_index (mode);
263 return &costs->cost[speed][idx];
266 /* Subroutine of {set_,}sdiv_pow2_cheap. Not to be used otherwise. */
269 sdiv_pow2_cheap_ptr (bool speed, machine_mode mode)
271 return expmed_op_cheap_ptr (&this_target_expmed->x_sdiv_pow2_cheap,
275 /* Set whether a signed division by a power of 2 is cheap in MODE
276 when optimizing for SPEED. */
279 set_sdiv_pow2_cheap (bool speed, machine_mode mode, bool cheap_p)
281 *sdiv_pow2_cheap_ptr (speed, mode) = cheap_p;
284 /* Return whether a signed division by a power of 2 is cheap in MODE
285 when optimizing for SPEED. */
288 sdiv_pow2_cheap (bool speed, machine_mode mode)
290 return *sdiv_pow2_cheap_ptr (speed, mode);
293 /* Subroutine of {set_,}smod_pow2_cheap. Not to be used otherwise. */
296 smod_pow2_cheap_ptr (bool speed, machine_mode mode)
298 return expmed_op_cheap_ptr (&this_target_expmed->x_smod_pow2_cheap,
302 /* Set whether a signed modulo by a power of 2 is CHEAP in MODE when
303 optimizing for SPEED. */
306 set_smod_pow2_cheap (bool speed, machine_mode mode, bool cheap)
308 *smod_pow2_cheap_ptr (speed, mode) = cheap;
311 /* Return whether a signed modulo by a power of 2 is cheap in MODE
312 when optimizing for SPEED. */
315 smod_pow2_cheap (bool speed, machine_mode mode)
317 return *smod_pow2_cheap_ptr (speed, mode);
320 /* Subroutine of {set_,}zero_cost. Not to be used otherwise. */
323 zero_cost_ptr (bool speed)
325 return &this_target_expmed->x_zero_cost[speed];
328 /* Set the COST of loading zero when optimizing for SPEED. */
331 set_zero_cost (bool speed, int cost)
333 *zero_cost_ptr (speed) = cost;
336 /* Return the COST of loading zero when optimizing for SPEED. */
339 zero_cost (bool speed)
341 return *zero_cost_ptr (speed);
344 /* Subroutine of {set_,}add_cost. Not to be used otherwise. */
347 add_cost_ptr (bool speed, machine_mode mode)
349 return expmed_op_cost_ptr (&this_target_expmed->x_add_cost, speed, mode);
352 /* Set the COST of computing an add in MODE when optimizing for SPEED. */
355 set_add_cost (bool speed, machine_mode mode, int cost)
357 *add_cost_ptr (speed, mode) = cost;
360 /* Return the cost of computing an add in MODE when optimizing for SPEED. */
363 add_cost (bool speed, machine_mode mode)
365 return *add_cost_ptr (speed, mode);
368 /* Subroutine of {set_,}neg_cost. Not to be used otherwise. */
371 neg_cost_ptr (bool speed, machine_mode mode)
373 return expmed_op_cost_ptr (&this_target_expmed->x_neg_cost, speed, mode);
376 /* Set the COST of computing a negation in MODE when optimizing for SPEED. */
379 set_neg_cost (bool speed, machine_mode mode, int cost)
381 *neg_cost_ptr (speed, mode) = cost;
384 /* Return the cost of computing a negation in MODE when optimizing for
388 neg_cost (bool speed, machine_mode mode)
390 return *neg_cost_ptr (speed, mode);
393 /* Subroutine of {set_,}shift_cost. Not to be used otherwise. */
396 shift_cost_ptr (bool speed, machine_mode mode, int bits)
398 return expmed_op_cost_ptr (&this_target_expmed->x_shift_cost[bits],
402 /* Set the COST of doing a shift in MODE by BITS when optimizing for SPEED. */
405 set_shift_cost (bool speed, machine_mode mode, int bits, int cost)
407 *shift_cost_ptr (speed, mode, bits) = cost;
410 /* Return the cost of doing a shift in MODE by BITS when optimizing for
414 shift_cost (bool speed, machine_mode mode, int bits)
416 return *shift_cost_ptr (speed, mode, bits);
419 /* Subroutine of {set_,}shiftadd_cost. Not to be used otherwise. */
422 shiftadd_cost_ptr (bool speed, machine_mode mode, int bits)
424 return expmed_op_cost_ptr (&this_target_expmed->x_shiftadd_cost[bits],
428 /* Set the COST of doing a shift in MODE by BITS followed by an add when
429 optimizing for SPEED. */
432 set_shiftadd_cost (bool speed, machine_mode mode, int bits, int cost)
434 *shiftadd_cost_ptr (speed, mode, bits) = cost;
437 /* Return the cost of doing a shift in MODE by BITS followed by an add
438 when optimizing for SPEED. */
441 shiftadd_cost (bool speed, machine_mode mode, int bits)
443 return *shiftadd_cost_ptr (speed, mode, bits);
446 /* Subroutine of {set_,}shiftsub0_cost. Not to be used otherwise. */
449 shiftsub0_cost_ptr (bool speed, machine_mode mode, int bits)
451 return expmed_op_cost_ptr (&this_target_expmed->x_shiftsub0_cost[bits],
455 /* Set the COST of doing a shift in MODE by BITS and then subtracting a
456 value when optimizing for SPEED. */
459 set_shiftsub0_cost (bool speed, machine_mode mode, int bits, int cost)
461 *shiftsub0_cost_ptr (speed, mode, bits) = cost;
464 /* Return the cost of doing a shift in MODE by BITS and then subtracting
465 a value when optimizing for SPEED. */
468 shiftsub0_cost (bool speed, machine_mode mode, int bits)
470 return *shiftsub0_cost_ptr (speed, mode, bits);
473 /* Subroutine of {set_,}shiftsub1_cost. Not to be used otherwise. */
476 shiftsub1_cost_ptr (bool speed, machine_mode mode, int bits)
478 return expmed_op_cost_ptr (&this_target_expmed->x_shiftsub1_cost[bits],
482 /* Set the COST of subtracting a shift in MODE by BITS from a value when
483 optimizing for SPEED. */
486 set_shiftsub1_cost (bool speed, machine_mode mode, int bits, int cost)
488 *shiftsub1_cost_ptr (speed, mode, bits) = cost;
491 /* Return the cost of subtracting a shift in MODE by BITS from a value
492 when optimizing for SPEED. */
495 shiftsub1_cost (bool speed, machine_mode mode, int bits)
497 return *shiftsub1_cost_ptr (speed, mode, bits);
500 /* Subroutine of {set_,}mul_cost. Not to be used otherwise. */
503 mul_cost_ptr (bool speed, machine_mode mode)
505 return expmed_op_cost_ptr (&this_target_expmed->x_mul_cost, speed, mode);
508 /* Set the COST of doing a multiplication in MODE when optimizing for
512 set_mul_cost (bool speed, machine_mode mode, int cost)
514 *mul_cost_ptr (speed, mode) = cost;
517 /* Return the cost of doing a multiplication in MODE when optimizing
521 mul_cost (bool speed, machine_mode mode)
523 return *mul_cost_ptr (speed, mode);
526 /* Subroutine of {set_,}sdiv_cost. Not to be used otherwise. */
529 sdiv_cost_ptr (bool speed, machine_mode mode)
531 return expmed_op_cost_ptr (&this_target_expmed->x_sdiv_cost, speed, mode);
534 /* Set the COST of doing a signed division in MODE when optimizing
538 set_sdiv_cost (bool speed, machine_mode mode, int cost)
540 *sdiv_cost_ptr (speed, mode) = cost;
543 /* Return the cost of doing a signed division in MODE when optimizing
547 sdiv_cost (bool speed, machine_mode mode)
549 return *sdiv_cost_ptr (speed, mode);
552 /* Subroutine of {set_,}udiv_cost. Not to be used otherwise. */
555 udiv_cost_ptr (bool speed, machine_mode mode)
557 return expmed_op_cost_ptr (&this_target_expmed->x_udiv_cost, speed, mode);
560 /* Set the COST of doing an unsigned division in MODE when optimizing
564 set_udiv_cost (bool speed, machine_mode mode, int cost)
566 *udiv_cost_ptr (speed, mode) = cost;
569 /* Return the cost of doing an unsigned division in MODE when
570 optimizing for SPEED. */
573 udiv_cost (bool speed, machine_mode mode)
575 return *udiv_cost_ptr (speed, mode);
578 /* Subroutine of {set_,}mul_widen_cost. Not to be used otherwise. */
581 mul_widen_cost_ptr (bool speed, machine_mode mode)
583 gcc_assert (GET_MODE_CLASS (mode) == MODE_INT);
585 return &this_target_expmed->x_mul_widen_cost[speed][mode - MIN_MODE_INT];
588 /* Set the COST for computing a widening multiplication in MODE when
589 optimizing for SPEED. */
592 set_mul_widen_cost (bool speed, machine_mode mode, int cost)
594 *mul_widen_cost_ptr (speed, mode) = cost;
597 /* Return the cost for computing a widening multiplication in MODE when
598 optimizing for SPEED. */
601 mul_widen_cost (bool speed, machine_mode mode)
603 return *mul_widen_cost_ptr (speed, mode);
606 /* Subroutine of {set_,}mul_highpart_cost. Not to be used otherwise. */
609 mul_highpart_cost_ptr (bool speed, machine_mode mode)
611 gcc_assert (GET_MODE_CLASS (mode) == MODE_INT);
612 int m = mode - MIN_MODE_INT;
613 gcc_assert (m < NUM_MODE_INT);
615 return &this_target_expmed->x_mul_highpart_cost[speed][m];
618 /* Set the COST for computing the high part of a multiplication in MODE
619 when optimizing for SPEED. */
622 set_mul_highpart_cost (bool speed, machine_mode mode, int cost)
624 *mul_highpart_cost_ptr (speed, mode) = cost;
627 /* Return the cost for computing the high part of a multiplication in MODE
628 when optimizing for SPEED. */
631 mul_highpart_cost (bool speed, machine_mode mode)
633 return *mul_highpart_cost_ptr (speed, mode);
636 /* Subroutine of {set_,}convert_cost. Not to be used otherwise. */
639 convert_cost_ptr (machine_mode to_mode, machine_mode from_mode,
642 int to_idx = expmed_mode_index (to_mode);
643 int from_idx = expmed_mode_index (from_mode);
645 gcc_assert (IN_RANGE (to_idx, 0, NUM_MODE_IP_INT - 1));
646 gcc_assert (IN_RANGE (from_idx, 0, NUM_MODE_IP_INT - 1));
648 return &this_target_expmed->x_convert_cost[speed][to_idx][from_idx];
651 /* Set the COST for converting from FROM_MODE to TO_MODE when optimizing
655 set_convert_cost (machine_mode to_mode, machine_mode from_mode,
656 bool speed, int cost)
658 *convert_cost_ptr (to_mode, from_mode, speed) = cost;
661 /* Return the cost for converting from FROM_MODE to TO_MODE when optimizing
665 convert_cost (machine_mode to_mode, machine_mode from_mode,
668 return *convert_cost_ptr (to_mode, from_mode, speed);
671 extern int mult_by_coeff_cost (HOST_WIDE_INT, machine_mode, bool);
672 extern rtx emit_cstore (rtx target, enum insn_code icode, enum rtx_code code,
673 machine_mode mode, machine_mode compare_mode,
674 int unsignedp, rtx x, rtx y, int normalizep,
675 machine_mode target_mode);
677 /* Arguments MODE, RTX: return an rtx for the negation of that value.
679 extern rtx negate_rtx (machine_mode, rtx);
681 /* Arguments MODE, RTX: return an rtx for the flipping of that value.
683 extern rtx flip_storage_order (machine_mode, rtx);
685 /* Expand a logical AND operation. */
686 extern rtx expand_and (machine_mode, rtx, rtx, rtx);
688 /* Emit a store-flag operation. */
689 extern rtx emit_store_flag (rtx, enum rtx_code, rtx, rtx, machine_mode,
692 /* Like emit_store_flag, but always succeeds. */
693 extern rtx emit_store_flag_force (rtx, enum rtx_code, rtx, rtx,
694 machine_mode, int, int);
696 extern void canonicalize_comparison (machine_mode, enum rtx_code *, rtx *);
698 /* Choose a minimal N + 1 bit approximation to 1/D that can be used to
699 replace division by D, and put the least significant N bits of the result
700 in *MULTIPLIER_PTR and return the most significant bit. */
701 extern unsigned HOST_WIDE_INT choose_multiplier (unsigned HOST_WIDE_INT, int,
702 int, unsigned HOST_WIDE_INT *,
706 extern rtx expand_variable_shift (enum tree_code, machine_mode,
707 rtx, tree, rtx, int);
708 extern rtx expand_shift (enum tree_code, machine_mode, rtx, poly_int64, rtx,
710 extern rtx maybe_expand_shift (enum tree_code, machine_mode, rtx, int, rtx,
713 extern rtx expand_divmod (int, enum tree_code, machine_mode, tree, tree,
715 enum optab_methods = OPTAB_LIB_WIDEN);
719 extern void store_bit_field (rtx, poly_uint64, poly_uint64,
720 poly_uint64, poly_uint64,
721 machine_mode, rtx, bool, bool);
722 extern rtx extract_bit_field (rtx, poly_uint64, poly_uint64, int, rtx,
723 machine_mode, machine_mode, bool, rtx *);
724 extern rtx extract_low_bits (machine_mode, machine_mode, rtx);
725 extern rtx expand_mult (machine_mode, rtx, rtx, rtx, int, bool = false);
726 extern rtx expand_mult_highpart_adjust (scalar_int_mode, rtx, rtx, rtx,