Loading...
1/* SPDX-License-Identifier: GPL-2.0 */
2#ifndef _ASM_GENERIC_DIV64_H
3#define _ASM_GENERIC_DIV64_H
4/*
5 * Copyright (C) 2003 Bernardo Innocenti <bernie@develer.com>
6 * Based on former asm-ppc/div64.h and asm-m68knommu/div64.h
7 *
8 * Optimization for constant divisors on 32-bit machines:
9 * Copyright (C) 2006-2015 Nicolas Pitre
10 *
11 * The semantics of do_div() are:
12 *
13 * uint32_t do_div(uint64_t *n, uint32_t base)
14 * {
15 * uint32_t remainder = *n % base;
16 * *n = *n / base;
17 * return remainder;
18 * }
19 *
20 * NOTE: macro parameter n is evaluated multiple times,
21 * beware of side effects!
22 */
23
24#include <linux/types.h>
25#include <linux/compiler.h>
26
27#if BITS_PER_LONG == 64
28
29/**
30 * do_div - returns 2 values: calculate remainder and update new dividend
31 * @n: uint64_t dividend (will be updated)
32 * @base: uint32_t divisor
33 *
34 * Summary:
35 * ``uint32_t remainder = n % base;``
36 * ``n = n / base;``
37 *
38 * Return: (uint32_t)remainder
39 *
40 * NOTE: macro parameter @n is evaluated multiple times,
41 * beware of side effects!
42 */
43# define do_div(n,base) ({ \
44 uint32_t __base = (base); \
45 uint32_t __rem; \
46 __rem = ((uint64_t)(n)) % __base; \
47 (n) = ((uint64_t)(n)) / __base; \
48 __rem; \
49 })
50
51#elif BITS_PER_LONG == 32
52
53#include <linux/log2.h>
54
55/*
56 * If the divisor happens to be constant, we determine the appropriate
57 * inverse at compile time to turn the division into a few inline
58 * multiplications which ought to be much faster. And yet only if compiling
59 * with a sufficiently recent gcc version to perform proper 64-bit constant
60 * propagation.
61 *
62 * (It is unfortunate that gcc doesn't perform all this internally.)
63 */
64
65#ifndef __div64_const32_is_OK
66#define __div64_const32_is_OK (__GNUC__ >= 4)
67#endif
68
69#define __div64_const32(n, ___b) \
70({ \
71 /* \
72 * Multiplication by reciprocal of b: n / b = n * (p / b) / p \
73 * \
74 * We rely on the fact that most of this code gets optimized \
75 * away at compile time due to constant propagation and only \
76 * a few multiplication instructions should remain. \
77 * Hence this monstrous macro (static inline doesn't always \
78 * do the trick here). \
79 */ \
80 uint64_t ___res, ___x, ___t, ___m, ___n = (n); \
81 uint32_t ___p, ___bias; \
82 \
83 /* determine MSB of b */ \
84 ___p = 1 << ilog2(___b); \
85 \
86 /* compute m = ((p << 64) + b - 1) / b */ \
87 ___m = (~0ULL / ___b) * ___p; \
88 ___m += (((~0ULL % ___b + 1) * ___p) + ___b - 1) / ___b; \
89 \
90 /* one less than the dividend with highest result */ \
91 ___x = ~0ULL / ___b * ___b - 1; \
92 \
93 /* test our ___m with res = m * x / (p << 64) */ \
94 ___res = ((___m & 0xffffffff) * (___x & 0xffffffff)) >> 32; \
95 ___t = ___res += (___m & 0xffffffff) * (___x >> 32); \
96 ___res += (___x & 0xffffffff) * (___m >> 32); \
97 ___t = (___res < ___t) ? (1ULL << 32) : 0; \
98 ___res = (___res >> 32) + ___t; \
99 ___res += (___m >> 32) * (___x >> 32); \
100 ___res /= ___p; \
101 \
102 /* Now sanitize and optimize what we've got. */ \
103 if (~0ULL % (___b / (___b & -___b)) == 0) { \
104 /* special case, can be simplified to ... */ \
105 ___n /= (___b & -___b); \
106 ___m = ~0ULL / (___b / (___b & -___b)); \
107 ___p = 1; \
108 ___bias = 1; \
109 } else if (___res != ___x / ___b) { \
110 /* \
111 * We can't get away without a bias to compensate \
112 * for bit truncation errors. To avoid it we'd need an \
113 * additional bit to represent m which would overflow \
114 * a 64-bit variable. \
115 * \
116 * Instead we do m = p / b and n / b = (n * m + m) / p. \
117 */ \
118 ___bias = 1; \
119 /* Compute m = (p << 64) / b */ \
120 ___m = (~0ULL / ___b) * ___p; \
121 ___m += ((~0ULL % ___b + 1) * ___p) / ___b; \
122 } else { \
123 /* \
124 * Reduce m / p, and try to clear bit 31 of m when \
125 * possible, otherwise that'll need extra overflow \
126 * handling later. \
127 */ \
128 uint32_t ___bits = -(___m & -___m); \
129 ___bits |= ___m >> 32; \
130 ___bits = (~___bits) << 1; \
131 /* \
132 * If ___bits == 0 then setting bit 31 is unavoidable. \
133 * Simply apply the maximum possible reduction in that \
134 * case. Otherwise the MSB of ___bits indicates the \
135 * best reduction we should apply. \
136 */ \
137 if (!___bits) { \
138 ___p /= (___m & -___m); \
139 ___m /= (___m & -___m); \
140 } else { \
141 ___p >>= ilog2(___bits); \
142 ___m >>= ilog2(___bits); \
143 } \
144 /* No bias needed. */ \
145 ___bias = 0; \
146 } \
147 \
148 /* \
149 * Now we have a combination of 2 conditions: \
150 * \
151 * 1) whether or not we need to apply a bias, and \
152 * \
153 * 2) whether or not there might be an overflow in the cross \
154 * product determined by (___m & ((1 << 63) | (1 << 31))). \
155 * \
156 * Select the best way to do (m_bias + m * n) / (1 << 64). \
157 * From now on there will be actual runtime code generated. \
158 */ \
159 ___res = __arch_xprod_64(___m, ___n, ___bias); \
160 \
161 ___res /= ___p; \
162})
163
164#ifndef __arch_xprod_64
165/*
166 * Default C implementation for __arch_xprod_64()
167 *
168 * Prototype: uint64_t __arch_xprod_64(const uint64_t m, uint64_t n, bool bias)
169 * Semantic: retval = ((bias ? m : 0) + m * n) >> 64
170 *
171 * The product is a 128-bit value, scaled down to 64 bits.
172 * Assuming constant propagation to optimize away unused conditional code.
173 * Architectures may provide their own optimized assembly implementation.
174 */
175static inline uint64_t __arch_xprod_64(const uint64_t m, uint64_t n, bool bias)
176{
177 uint32_t m_lo = m;
178 uint32_t m_hi = m >> 32;
179 uint32_t n_lo = n;
180 uint32_t n_hi = n >> 32;
181 uint64_t res;
182 uint32_t res_lo, res_hi, tmp;
183
184 if (!bias) {
185 res = ((uint64_t)m_lo * n_lo) >> 32;
186 } else if (!(m & ((1ULL << 63) | (1ULL << 31)))) {
187 /* there can't be any overflow here */
188 res = (m + (uint64_t)m_lo * n_lo) >> 32;
189 } else {
190 res = m + (uint64_t)m_lo * n_lo;
191 res_lo = res >> 32;
192 res_hi = (res_lo < m_hi);
193 res = res_lo | ((uint64_t)res_hi << 32);
194 }
195
196 if (!(m & ((1ULL << 63) | (1ULL << 31)))) {
197 /* there can't be any overflow here */
198 res += (uint64_t)m_lo * n_hi;
199 res += (uint64_t)m_hi * n_lo;
200 res >>= 32;
201 } else {
202 res += (uint64_t)m_lo * n_hi;
203 tmp = res >> 32;
204 res += (uint64_t)m_hi * n_lo;
205 res_lo = res >> 32;
206 res_hi = (res_lo < tmp);
207 res = res_lo | ((uint64_t)res_hi << 32);
208 }
209
210 res += (uint64_t)m_hi * n_hi;
211
212 return res;
213}
214#endif
215
216#ifndef __div64_32
217extern uint32_t __div64_32(uint64_t *dividend, uint32_t divisor);
218#endif
219
220/* The unnecessary pointer compare is there
221 * to check for type safety (n must be 64bit)
222 */
223# define do_div(n,base) ({ \
224 uint32_t __base = (base); \
225 uint32_t __rem; \
226 (void)(((typeof((n)) *)0) == ((uint64_t *)0)); \
227 if (__builtin_constant_p(__base) && \
228 is_power_of_2(__base)) { \
229 __rem = (n) & (__base - 1); \
230 (n) >>= ilog2(__base); \
231 } else if (__div64_const32_is_OK && \
232 __builtin_constant_p(__base) && \
233 __base != 0) { \
234 uint32_t __res_lo, __n_lo = (n); \
235 (n) = __div64_const32(n, __base); \
236 /* the remainder can be computed with 32-bit regs */ \
237 __res_lo = (n); \
238 __rem = __n_lo - __res_lo * __base; \
239 } else if (likely(((n) >> 32) == 0)) { \
240 __rem = (uint32_t)(n) % __base; \
241 (n) = (uint32_t)(n) / __base; \
242 } else \
243 __rem = __div64_32(&(n), __base); \
244 __rem; \
245 })
246
247#else /* BITS_PER_LONG == ?? */
248
249# error do_div() does not yet support the C64
250
251#endif /* BITS_PER_LONG */
252
253#endif /* _ASM_GENERIC_DIV64_H */
1/* SPDX-License-Identifier: GPL-2.0 */
2#ifndef _ASM_GENERIC_DIV64_H
3#define _ASM_GENERIC_DIV64_H
4/*
5 * Copyright (C) 2003 Bernardo Innocenti <bernie@develer.com>
6 * Based on former asm-ppc/div64.h and asm-m68knommu/div64.h
7 *
8 * Optimization for constant divisors on 32-bit machines:
9 * Copyright (C) 2006-2015 Nicolas Pitre
10 *
11 * The semantics of do_div() are:
12 *
13 * uint32_t do_div(uint64_t *n, uint32_t base)
14 * {
15 * uint32_t remainder = *n % base;
16 * *n = *n / base;
17 * return remainder;
18 * }
19 *
20 * NOTE: macro parameter n is evaluated multiple times,
21 * beware of side effects!
22 */
23
24#include <linux/types.h>
25#include <linux/compiler.h>
26
27#if BITS_PER_LONG == 64
28
29/**
30 * do_div - returns 2 values: calculate remainder and update new dividend
31 * @n: uint64_t dividend (will be updated)
32 * @base: uint32_t divisor
33 *
34 * Summary:
35 * ``uint32_t remainder = n % base;``
36 * ``n = n / base;``
37 *
38 * Return: (uint32_t)remainder
39 *
40 * NOTE: macro parameter @n is evaluated multiple times,
41 * beware of side effects!
42 */
43# define do_div(n,base) ({ \
44 uint32_t __base = (base); \
45 uint32_t __rem; \
46 __rem = ((uint64_t)(n)) % __base; \
47 (n) = ((uint64_t)(n)) / __base; \
48 __rem; \
49 })
50
51#elif BITS_PER_LONG == 32
52
53#include <linux/log2.h>
54
55/*
56 * If the divisor happens to be constant, we determine the appropriate
57 * inverse at compile time to turn the division into a few inline
58 * multiplications which ought to be much faster. And yet only if compiling
59 * with a sufficiently recent gcc version to perform proper 64-bit constant
60 * propagation.
61 *
62 * (It is unfortunate that gcc doesn't perform all this internally.)
63 */
64
65#ifndef __div64_const32_is_OK
66#define __div64_const32_is_OK (__GNUC__ >= 4)
67#endif
68
69#define __div64_const32(n, ___b) \
70({ \
71 /* \
72 * Multiplication by reciprocal of b: n / b = n * (p / b) / p \
73 * \
74 * We rely on the fact that most of this code gets optimized \
75 * away at compile time due to constant propagation and only \
76 * a few multiplication instructions should remain. \
77 * Hence this monstrous macro (static inline doesn't always \
78 * do the trick here). \
79 */ \
80 uint64_t ___res, ___x, ___t, ___m, ___n = (n); \
81 uint32_t ___p, ___bias; \
82 \
83 /* determine MSB of b */ \
84 ___p = 1 << ilog2(___b); \
85 \
86 /* compute m = ((p << 64) + b - 1) / b */ \
87 ___m = (~0ULL / ___b) * ___p; \
88 ___m += (((~0ULL % ___b + 1) * ___p) + ___b - 1) / ___b; \
89 \
90 /* one less than the dividend with highest result */ \
91 ___x = ~0ULL / ___b * ___b - 1; \
92 \
93 /* test our ___m with res = m * x / (p << 64) */ \
94 ___res = ((___m & 0xffffffff) * (___x & 0xffffffff)) >> 32; \
95 ___t = ___res += (___m & 0xffffffff) * (___x >> 32); \
96 ___res += (___x & 0xffffffff) * (___m >> 32); \
97 ___t = (___res < ___t) ? (1ULL << 32) : 0; \
98 ___res = (___res >> 32) + ___t; \
99 ___res += (___m >> 32) * (___x >> 32); \
100 ___res /= ___p; \
101 \
102 /* Now sanitize and optimize what we've got. */ \
103 if (~0ULL % (___b / (___b & -___b)) == 0) { \
104 /* special case, can be simplified to ... */ \
105 ___n /= (___b & -___b); \
106 ___m = ~0ULL / (___b / (___b & -___b)); \
107 ___p = 1; \
108 ___bias = 1; \
109 } else if (___res != ___x / ___b) { \
110 /* \
111 * We can't get away without a bias to compensate \
112 * for bit truncation errors. To avoid it we'd need an \
113 * additional bit to represent m which would overflow \
114 * a 64-bit variable. \
115 * \
116 * Instead we do m = p / b and n / b = (n * m + m) / p. \
117 */ \
118 ___bias = 1; \
119 /* Compute m = (p << 64) / b */ \
120 ___m = (~0ULL / ___b) * ___p; \
121 ___m += ((~0ULL % ___b + 1) * ___p) / ___b; \
122 } else { \
123 /* \
124 * Reduce m / p, and try to clear bit 31 of m when \
125 * possible, otherwise that'll need extra overflow \
126 * handling later. \
127 */ \
128 uint32_t ___bits = -(___m & -___m); \
129 ___bits |= ___m >> 32; \
130 ___bits = (~___bits) << 1; \
131 /* \
132 * If ___bits == 0 then setting bit 31 is unavoidable. \
133 * Simply apply the maximum possible reduction in that \
134 * case. Otherwise the MSB of ___bits indicates the \
135 * best reduction we should apply. \
136 */ \
137 if (!___bits) { \
138 ___p /= (___m & -___m); \
139 ___m /= (___m & -___m); \
140 } else { \
141 ___p >>= ilog2(___bits); \
142 ___m >>= ilog2(___bits); \
143 } \
144 /* No bias needed. */ \
145 ___bias = 0; \
146 } \
147 \
148 /* \
149 * Now we have a combination of 2 conditions: \
150 * \
151 * 1) whether or not we need to apply a bias, and \
152 * \
153 * 2) whether or not there might be an overflow in the cross \
154 * product determined by (___m & ((1 << 63) | (1 << 31))). \
155 * \
156 * Select the best way to do (m_bias + m * n) / (1 << 64). \
157 * From now on there will be actual runtime code generated. \
158 */ \
159 ___res = __arch_xprod_64(___m, ___n, ___bias); \
160 \
161 ___res /= ___p; \
162})
163
164#ifndef __arch_xprod_64
165/*
166 * Default C implementation for __arch_xprod_64()
167 *
168 * Prototype: uint64_t __arch_xprod_64(const uint64_t m, uint64_t n, bool bias)
169 * Semantic: retval = ((bias ? m : 0) + m * n) >> 64
170 *
171 * The product is a 128-bit value, scaled down to 64 bits.
172 * Assuming constant propagation to optimize away unused conditional code.
173 * Architectures may provide their own optimized assembly implementation.
174 */
175static inline uint64_t __arch_xprod_64(const uint64_t m, uint64_t n, bool bias)
176{
177 uint32_t m_lo = m;
178 uint32_t m_hi = m >> 32;
179 uint32_t n_lo = n;
180 uint32_t n_hi = n >> 32;
181 uint64_t res;
182 uint32_t res_lo, res_hi, tmp;
183
184 if (!bias) {
185 res = ((uint64_t)m_lo * n_lo) >> 32;
186 } else if (!(m & ((1ULL << 63) | (1ULL << 31)))) {
187 /* there can't be any overflow here */
188 res = (m + (uint64_t)m_lo * n_lo) >> 32;
189 } else {
190 res = m + (uint64_t)m_lo * n_lo;
191 res_lo = res >> 32;
192 res_hi = (res_lo < m_hi);
193 res = res_lo | ((uint64_t)res_hi << 32);
194 }
195
196 if (!(m & ((1ULL << 63) | (1ULL << 31)))) {
197 /* there can't be any overflow here */
198 res += (uint64_t)m_lo * n_hi;
199 res += (uint64_t)m_hi * n_lo;
200 res >>= 32;
201 } else {
202 res += (uint64_t)m_lo * n_hi;
203 tmp = res >> 32;
204 res += (uint64_t)m_hi * n_lo;
205 res_lo = res >> 32;
206 res_hi = (res_lo < tmp);
207 res = res_lo | ((uint64_t)res_hi << 32);
208 }
209
210 res += (uint64_t)m_hi * n_hi;
211
212 return res;
213}
214#endif
215
216#ifndef __div64_32
217extern uint32_t __div64_32(uint64_t *dividend, uint32_t divisor);
218#endif
219
220/* The unnecessary pointer compare is there
221 * to check for type safety (n must be 64bit)
222 */
223# define do_div(n,base) ({ \
224 uint32_t __base = (base); \
225 uint32_t __rem; \
226 (void)(((typeof((n)) *)0) == ((uint64_t *)0)); \
227 if (__builtin_constant_p(__base) && \
228 is_power_of_2(__base)) { \
229 __rem = (n) & (__base - 1); \
230 (n) >>= ilog2(__base); \
231 } else if (__div64_const32_is_OK && \
232 __builtin_constant_p(__base) && \
233 __base != 0) { \
234 uint32_t __res_lo, __n_lo = (n); \
235 (n) = __div64_const32(n, __base); \
236 /* the remainder can be computed with 32-bit regs */ \
237 __res_lo = (n); \
238 __rem = __n_lo - __res_lo * __base; \
239 } else if (likely(((n) >> 32) == 0)) { \
240 __rem = (uint32_t)(n) % __base; \
241 (n) = (uint32_t)(n) / __base; \
242 } else \
243 __rem = __div64_32(&(n), __base); \
244 __rem; \
245 })
246
247#else /* BITS_PER_LONG == ?? */
248
249# error do_div() does not yet support the C64
250
251#endif /* BITS_PER_LONG */
252
253#endif /* _ASM_GENERIC_DIV64_H */