xref: /openbmc/u-boot/include/div64.h (revision b8448051)
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
9  *
10  * The semantics of do_div() are:
11  *
12  * uint32_t do_div(uint64_t *n, uint32_t base)
13  * {
14  * 	uint32_t remainder = *n % base;
15  * 	*n = *n / base;
16  * 	return remainder;
17  * }
18  *
19  * NOTE: macro parameter n is evaluated multiple times,
20  *       beware of side effects!
21  */
22 
23 #include <linux/types.h>
24 #include <linux/compiler.h>
25 
26 #if BITS_PER_LONG == 64
27 
28 # define do_div(n,base) ({					\
29 	uint32_t __base = (base);				\
30 	uint32_t __rem;						\
31 	__rem = ((uint64_t)(n)) % __base;			\
32 	(n) = ((uint64_t)(n)) / __base;				\
33 	__rem;							\
34  })
35 
36 #elif BITS_PER_LONG == 32
37 
38 #include <linux/log2.h>
39 
40 /*
41  * If the divisor happens to be constant, we determine the appropriate
42  * inverse at compile time to turn the division into a few inline
43  * multiplications which ought to be much faster. And yet only if compiling
44  * with a sufficiently recent gcc version to perform proper 64-bit constant
45  * propagation.
46  *
47  * (It is unfortunate that gcc doesn't perform all this internally.)
48  */
49 
50 #ifndef __div64_const32_is_OK
51 #define __div64_const32_is_OK (__GNUC__ >= 4)
52 #endif
53 
54 #define __div64_const32(n, ___b)					\
55 ({									\
56 	/*								\
57 	 * Multiplication by reciprocal of b: n / b = n * (p / b) / p	\
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;						\
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 									\
75 	/* one less than the dividend with highest result */		\
76 	___x = ~0ULL / ___b * ___b - 1;					\
77 									\
78 	/* test our ___m with res = m * x / (p << 64) */		\
79 	___res = ((___m & 0xffffffff) * (___x & 0xffffffff)) >> 32;	\
80 	___t = ___res += (___m & 0xffffffff) * (___x >> 32);		\
81 	___res += (___x & 0xffffffff) * (___m >> 32);			\
82 	___t = (___res < ___t) ? (1ULL << 32) : 0;			\
83 	___res = (___res >> 32) + ___t;					\
84 	___res += (___m >> 32) * (___x >> 32);				\
85 	___res /= ___p;							\
86 									\
87 	/* Now sanitize and optimize what we've got. */			\
88 	if (~0ULL % (___b / (___b & -___b)) == 0) {			\
89 		/* special case, can be simplified to ... */		\
90 		___n /= (___b & -___b);					\
91 		___m = ~0ULL / (___b / (___b & -___b));			\
92 		___p = 1;						\
93 		___bias = 1;						\
94 	} else if (___res != ___x / ___b) {				\
95 		/*							\
96 		 * We can't get away without a bias to compensate	\
97 		 * for bit truncation errors.  To avoid it we'd need an	\
98 		 * additional bit to represent m which would overflow	\
99 		 * a 64-bit variable.					\
100 		 *							\
101 		 * Instead we do m = p / b and n / b = (n * m + m) / p.	\
102 		 */							\
103 		___bias = 1;						\
104 		/* Compute m = (p << 64) / b */				\
105 		___m = (~0ULL / ___b) * ___p;				\
106 		___m += ((~0ULL % ___b + 1) * ___p) / ___b;		\
107 	} else {							\
108 		/*							\
109 		 * Reduce m / p, and try to clear bit 31 of m when	\
110 		 * possible, otherwise that'll need extra overflow	\
111 		 * handling later.					\
112 		 */							\
113 		uint32_t ___bits = -(___m & -___m);			\
114 		___bits |= ___m >> 32;					\
115 		___bits = (~___bits) << 1;				\
116 		/*							\
117 		 * If ___bits == 0 then setting bit 31 is  unavoidable.	\
118 		 * Simply apply the maximum possible reduction in that	\
119 		 * case. Otherwise the MSB of ___bits indicates the	\
120 		 * best reduction we should apply.			\
121 		 */							\
122 		if (!___bits) {						\
123 			___p /= (___m & -___m);				\
124 			___m /= (___m & -___m);				\
125 		} else {						\
126 			___p >>= ilog2(___bits);			\
127 			___m >>= ilog2(___bits);			\
128 		}							\
129 		/* No bias needed. */					\
130 		___bias = 0;						\
131 	}								\
132 									\
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) / (1 << 64).	\
142 	 * From now on there will be actual runtime code generated.	\
143 	 */								\
144 	___res = __arch_xprod_64(___m, ___n, ___bias);			\
145 									\
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  */
160 static 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 
197 #ifndef __div64_32
198 extern uint32_t __div64_32(uint64_t *dividend, uint32_t divisor);
199 #endif
200 
201 /* The unnecessary pointer compare is there
202  * to check for type safety (n must be 64bit)
203  */
204 # define do_div(n,base) ({				\
205 	uint32_t __base = (base);			\
206 	uint32_t __rem;					\
207 	(void)(((typeof((n)) *)0) == ((uint64_t *)0));	\
208 	if (__builtin_constant_p(__base) &&		\
209 	    is_power_of_2(__base)) {			\
210 		__rem = (n) & (__base - 1);		\
211 		(n) >>= ilog2(__base);			\
212 	} else if (__div64_const32_is_OK &&		\
213 		   __builtin_constant_p(__base) &&	\
214 		   __base != 0) {			\
215 		uint32_t __res_lo, __n_lo = (n);	\
216 		(n) = __div64_const32(n, __base);	\
217 		/* the remainder can be computed with 32-bit regs */ \
218 		__res_lo = (n);				\
219 		__rem = __n_lo - __res_lo * __base;	\
220 	} else if (likely(((n) >> 32) == 0)) {		\
221 		__rem = (uint32_t)(n) % __base;		\
222 		(n) = (uint32_t)(n) / __base;		\
223 	} else 						\
224 		__rem = __div64_32(&(n), __base);	\
225 	__rem;						\
226  })
227 
228 #else /* BITS_PER_LONG == ?? */
229 
230 # error do_div() does not yet support the C64
231 
232 #endif /* BITS_PER_LONG */
233 
234 /* Wrapper for do_div(). Doesn't modify dividend and returns
235  * the result, not remainder.
236  */
237 static inline uint64_t lldiv(uint64_t dividend, uint32_t divisor)
238 {
239 	uint64_t __res = dividend;
240 	do_div(__res, divisor);
241 	return(__res);
242 }
243 
244 #endif /* _ASM_GENERIC_DIV64_H */
245