div64.h (461a5e51060c93f5844113f4be9dba513cc92830) div64.h (f682b27c57aec2f0ca8927f9bb7c267c6165ad5a)
1#ifndef _ASM_GENERIC_DIV64_H
2#define _ASM_GENERIC_DIV64_H
3/*
4 * Copyright (C) 2003 Bernardo Innocenti <bernie@develer.com>
5 * Based on former asm-ppc/div64.h and asm-m68knommu/div64.h
6 *
7 * Optimization for constant divisors on 32-bit machines:
8 * Copyright (C) 2006-2015 Nicolas Pitre

--- 49 unchanged lines hidden (view full) ---

58 * \
59 * We rely on the fact that most of this code gets optimized \
60 * away at compile time due to constant propagation and only \
61 * a few multiplication instructions should remain. \
62 * Hence this monstrous macro (static inline doesn't always \
63 * do the trick here). \
64 */ \
65 uint64_t ___res, ___x, ___t, ___m, ___n = (n); \
1#ifndef _ASM_GENERIC_DIV64_H
2#define _ASM_GENERIC_DIV64_H
3/*
4 * Copyright (C) 2003 Bernardo Innocenti <bernie@develer.com>
5 * Based on former asm-ppc/div64.h and asm-m68knommu/div64.h
6 *
7 * Optimization for constant divisors on 32-bit machines:
8 * Copyright (C) 2006-2015 Nicolas Pitre

--- 49 unchanged lines hidden (view full) ---

58 * \
59 * We rely on the fact that most of this code gets optimized \
60 * away at compile time due to constant propagation and only \
61 * a few multiplication instructions should remain. \
62 * Hence this monstrous macro (static inline doesn't always \
63 * do the trick here). \
64 */ \
65 uint64_t ___res, ___x, ___t, ___m, ___n = (n); \
66 uint32_t ___p, ___bias, ___m_lo, ___m_hi, ___n_lo, ___n_hi; \
66 uint32_t ___p, ___bias; \
67 \
68 /* determine MSB of b */ \
69 ___p = 1 << ilog2(___b); \
70 \
71 /* compute m = ((p << 64) + b - 1) / b */ \
72 ___m = (~0ULL / ___b) * ___p; \
73 ___m += (((~0ULL % ___b + 1) * ___p) + ___b - 1) / ___b; \
74 \

--- 58 unchanged lines hidden (view full) ---

133 /* \
134 * Now we have a combination of 2 conditions: \
135 * \
136 * 1) whether or not we need to apply a bias, and \
137 * \
138 * 2) whether or not there might be an overflow in the cross \
139 * product determined by (___m & ((1 << 63) | (1 << 31))). \
140 * \
67 \
68 /* determine MSB of b */ \
69 ___p = 1 << ilog2(___b); \
70 \
71 /* compute m = ((p << 64) + b - 1) / b */ \
72 ___m = (~0ULL / ___b) * ___p; \
73 ___m += (((~0ULL % ___b + 1) * ___p) + ___b - 1) / ___b; \
74 \

--- 58 unchanged lines hidden (view full) ---

133 /* \
134 * Now we have a combination of 2 conditions: \
135 * \
136 * 1) whether or not we need to apply a bias, and \
137 * \
138 * 2) whether or not there might be an overflow in the cross \
139 * product determined by (___m & ((1 << 63) | (1 << 31))). \
140 * \
141 * Select the best way to do (m_bias + m * n) / (p << 64). \
141 * Select the best way to do (m_bias + m * n) / (1 << 64). \
142 * From now on there will be actual runtime code generated. \
143 */ \
142 * From now on there will be actual runtime code generated. \
143 */ \
144 ___res = __arch_xprod_64(___m, ___n, ___bias); \
144 \
145 \
145 ___m_lo = ___m; \
146 ___m_hi = ___m >> 32; \
147 ___n_lo = ___n; \
148 ___n_hi = ___n >> 32; \
149 \
150 if (!___bias) { \
151 ___res = ((uint64_t)___m_lo * ___n_lo) >> 32; \
152 } else if (!(___m & ((1ULL << 63) | (1ULL << 31)))) { \
153 ___res = (___m + (uint64_t)___m_lo * ___n_lo) >> 32; \
154 } else { \
155 ___res = ___m + (uint64_t)___m_lo * ___n_lo; \
156 ___t = (___res < ___m) ? (1ULL << 32) : 0; \
157 ___res = (___res >> 32) + ___t; \
158 } \
159 \
160 if (!(___m & ((1ULL << 63) | (1ULL << 31)))) { \
161 ___res += (uint64_t)___m_lo * ___n_hi; \
162 ___res += (uint64_t)___m_hi * ___n_lo; \
163 ___res >>= 32; \
164 } else { \
165 ___t = ___res += (uint64_t)___m_lo * ___n_hi; \
166 ___res += (uint64_t)___m_hi * ___n_lo; \
167 ___t = (___res < ___t) ? (1ULL << 32) : 0; \
168 ___res = (___res >> 32) + ___t; \
169 } \
170 \
171 ___res += (uint64_t)___m_hi * ___n_hi; \
172 \
173 ___res /= ___p; \
174})
175
146 ___res /= ___p; \
147})
148
149#ifndef __arch_xprod_64
150/*
151 * Default C implementation for __arch_xprod_64()
152 *
153 * Prototype: uint64_t __arch_xprod_64(const uint64_t m, uint64_t n, bool bias)
154 * Semantic: retval = ((bias ? m : 0) + m * n) >> 64
155 *
156 * The product is a 128-bit value, scaled down to 64 bits.
157 * Assuming constant propagation to optimize away unused conditional code.
158 * Architectures may provide their own optimized assembly implementation.
159 */
160static inline uint64_t __arch_xprod_64(const uint64_t m, uint64_t n, bool bias)
161{
162 uint32_t m_lo = m;
163 uint32_t m_hi = m >> 32;
164 uint32_t n_lo = n;
165 uint32_t n_hi = n >> 32;
166 uint64_t res, tmp;
167
168 if (!bias) {
169 res = ((uint64_t)m_lo * n_lo) >> 32;
170 } else if (!(m & ((1ULL << 63) | (1ULL << 31)))) {
171 /* there can't be any overflow here */
172 res = (m + (uint64_t)m_lo * n_lo) >> 32;
173 } else {
174 res = m + (uint64_t)m_lo * n_lo;
175 tmp = (res < m) ? (1ULL << 32) : 0;
176 res = (res >> 32) + tmp;
177 }
178
179 if (!(m & ((1ULL << 63) | (1ULL << 31)))) {
180 /* there can't be any overflow here */
181 res += (uint64_t)m_lo * n_hi;
182 res += (uint64_t)m_hi * n_lo;
183 res >>= 32;
184 } else {
185 tmp = res += (uint64_t)m_lo * n_hi;
186 res += (uint64_t)m_hi * n_lo;
187 tmp = (res < tmp) ? (1ULL << 32) : 0;
188 res = (res >> 32) + tmp;
189 }
190
191 res += (uint64_t)m_hi * n_hi;
192
193 return res;
194}
195#endif
196
176extern uint32_t __div64_32(uint64_t *dividend, uint32_t divisor);
177
178/* The unnecessary pointer compare is there
179 * to check for type safety (n must be 64bit)
180 */
181# define do_div(n,base) ({ \
182 uint32_t __base = (base); \
183 uint32_t __rem; \

--- 28 unchanged lines hidden --- 		197extern uint32_t __div64_32(uint64_t *dividend, uint32_t divisor);
198
199/* The unnecessary pointer compare is there
200 * to check for type safety (n must be 64bit)
201 */
202# define do_div(n,base) ({ \
203 uint32_t __base = (base); \
204 uint32_t __rem; \

--- 28 unchanged lines hidden ---