xref: /linux/arch/microblaze/include/asm/hash.h (revision 8c994eff8fcfe8ecb1f1dbebed25b4d7bb75be12)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _ASM_HASH_H
3 #define _ASM_HASH_H
4 
5 /*
6  * Fortunately, most people who want to run Linux on Microblaze enable
7  * both multiplier and barrel shifter, but omitting them is technically
8  * a supported configuration.
9  *
10  * With just a barrel shifter, we can implement an efficient constant
11  * multiply using shifts and adds.  GCC can find a 9-step solution, but
12  * this 6-step solution was found by Yevgen Voronenko's implementation
13  * of the Hcub algorithm at http://spiral.ece.cmu.edu/mcm/gen.html.
14  *
15  * That software is really not designed for a single multiplier this large,
16  * but if you run it enough times with different seeds, it'll find several
17  * 6-shift, 6-add sequences for computing x * 0x61C88647.  They are all
18  *	c = (x << 19) + x;
19  *	a = (x <<  9) + c;
20  *	b = (x << 23) + a;
21  *	return (a<<11) + (b<<6) + (c<<3) - b;
22  * with variations on the order of the final add.
23  *
24  * Without even a shifter, it's hopless; any hash function will suck.
25  */
26 
27 #if CONFIG_XILINX_MICROBLAZE0_USE_HW_MUL == 0
28 
29 #define HAVE_ARCH__HASH_32 1
30 
31 /* Multiply by GOLDEN_RATIO_32 = 0x61C88647 */
32 static inline u32 __attribute_const__ __hash_32(u32 a)
33 {
34 #if CONFIG_XILINX_MICROBLAZE0_USE_BARREL
35 	unsigned int b, c;
36 
37 	/* Phase 1: Compute three intermediate values */
38 	b =  a << 23;
39 	c = (a << 19) + a;
40 	a = (a <<  9) + c;
41 	b += a;
42 
43 	/* Phase 2: Compute (a << 11) + (b << 6) + (c << 3) - b */
44 	a <<= 5;
45 	a += b;		/* (a << 5) + b */
46 	a <<= 3;
47 	a += c;		/* (a << 8) + (b << 3) + c */
48 	a <<= 3;
49 	return a - b;	/* (a << 11) + (b << 6) + (c << 3) - b */
50 #else
51 	/*
52 	 * "This is really going to hurt."
53 	 *
54 	 * Without a barrel shifter, left shifts are implemented as
55 	 * repeated additions, and the best we can do is an optimal
56 	 * addition-subtraction chain.  This one is not known to be
57 	 * optimal, but at 37 steps, it's decent for a 31-bit multiplier.
58 	 *
59 	 * Question: given its size (37*4 = 148 bytes per instance),
60 	 * and slowness, is this worth having inline?
61 	 */
62 	unsigned int b, c, d;
63 
64 	b = a << 4;	/* 4    */
65 	c = b << 1;	/* 1  5 */
66 	b += a;		/* 1  6 */
67 	c += b;		/* 1  7 */
68 	c <<= 3;	/* 3 10 */
69 	c -= a;		/* 1 11 */
70 	d = c << 7;	/* 7 18 */
71 	d += b;		/* 1 19 */
72 	d <<= 8;	/* 8 27 */
73 	d += a;		/* 1 28 */
74 	d <<= 1;	/* 1 29 */
75 	d += b;		/* 1 30 */
76 	d <<= 6;	/* 6 36 */
77 	return d + c;	/* 1 37 total instructions*/
78 #endif
79 }
80 
81 #endif /* !CONFIG_XILINX_MICROBLAZE0_USE_HW_MUL */
82 #endif /* _ASM_HASH_H */
83