xref: /linux/include/linux/reciprocal_div.h (revision 3eb66e91a25497065c5322b1268cbc3953642227)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _LINUX_RECIPROCAL_DIV_H
3 #define _LINUX_RECIPROCAL_DIV_H
4 
5 #include <linux/types.h>
6 
7 /*
8  * This algorithm is based on the paper "Division by Invariant
9  * Integers Using Multiplication" by Torbjörn Granlund and Peter
10  * L. Montgomery.
11  *
12  * The assembler implementation from Agner Fog, which this code is
13  * based on, can be found here:
14  * http://www.agner.org/optimize/asmlib.zip
15  *
16  * This optimization for A/B is helpful if the divisor B is mostly
17  * runtime invariant. The reciprocal of B is calculated in the
18  * slow-path with reciprocal_value(). The fast-path can then just use
19  * a much faster multiplication operation with a variable dividend A
20  * to calculate the division A/B.
21  */
22 
23 struct reciprocal_value {
24 	u32 m;
25 	u8 sh1, sh2;
26 };
27 
28 /* "reciprocal_value" and "reciprocal_divide" together implement the basic
29  * version of the algorithm described in Figure 4.1 of the paper.
30  */
31 struct reciprocal_value reciprocal_value(u32 d);
32 
reciprocal_divide(u32 a,struct reciprocal_value R)33 static inline u32 reciprocal_divide(u32 a, struct reciprocal_value R)
34 {
35 	u32 t = (u32)(((u64)a * R.m) >> 32);
36 	return (t + ((a - t) >> R.sh1)) >> R.sh2;
37 }
38 
39 struct reciprocal_value_adv {
40 	u32 m;
41 	u8 sh, exp;
42 	bool is_wide_m;
43 };
44 
45 /* "reciprocal_value_adv" implements the advanced version of the algorithm
46  * described in Figure 4.2 of the paper except when "divisor > (1U << 31)" whose
47  * ceil(log2(d)) result will be 32 which then requires u128 divide on host. The
48  * exception case could be easily handled before calling "reciprocal_value_adv".
49  *
50  * The advanced version requires more complex calculation to get the reciprocal
51  * multiplier and other control variables, but then could reduce the required
52  * emulation operations.
53  *
54  * It makes no sense to use this advanced version for host divide emulation,
55  * those extra complexities for calculating multiplier etc could completely
56  * waive our saving on emulation operations.
57  *
58  * However, it makes sense to use it for JIT divide code generation for which
59  * we are willing to trade performance of JITed code with that of host. As shown
60  * by the following pseudo code, the required emulation operations could go down
61  * from 6 (the basic version) to 3 or 4.
62  *
63  * To use the result of "reciprocal_value_adv", suppose we want to calculate
64  * n/d, the pseudo C code will be:
65  *
66  *   struct reciprocal_value_adv rvalue;
67  *   u8 pre_shift, exp;
68  *
69  *   // handle exception case.
70  *   if (d >= (1U << 31)) {
71  *     result = n >= d;
72  *     return;
73  *   }
74  *
75  *   rvalue = reciprocal_value_adv(d, 32)
76  *   exp = rvalue.exp;
77  *   if (rvalue.is_wide_m && !(d & 1)) {
78  *     // floor(log2(d & (2^32 -d)))
79  *     pre_shift = fls(d & -d) - 1;
80  *     rvalue = reciprocal_value_adv(d >> pre_shift, 32 - pre_shift);
81  *   } else {
82  *     pre_shift = 0;
83  *   }
84  *
85  *   // code generation starts.
86  *   if (imm == 1U << exp) {
87  *     result = n >> exp;
88  *   } else if (rvalue.is_wide_m) {
89  *     // pre_shift must be zero when reached here.
90  *     t = (n * rvalue.m) >> 32;
91  *     result = n - t;
92  *     result >>= 1;
93  *     result += t;
94  *     result >>= rvalue.sh - 1;
95  *   } else {
96  *     if (pre_shift)
97  *       result = n >> pre_shift;
98  *     result = ((u64)result * rvalue.m) >> 32;
99  *     result >>= rvalue.sh;
100  *   }
101  */
102 struct reciprocal_value_adv reciprocal_value_adv(u32 d, u8 prec);
103 
104 #endif /* _LINUX_RECIPROCAL_DIV_H */
105