xref: /linux/arch/microblaze/include/asm/hash.h (revision e5c86679d5e864947a52fb31e45a425dea3e7fa9)
1 #ifndef _ASM_HASH_H
2 #define _ASM_HASH_H
3 
4 /*
5  * Fortunately, most people who want to run Linux on Microblaze enable
6  * both multiplier and barrel shifter, but omitting them is technically
7  * a supported configuration.
8  *
9  * With just a barrel shifter, we can implement an efficient constant
10  * multiply using shifts and adds.  GCC can find a 9-step solution, but
11  * this 6-step solution was found by Yevgen Voronenko's implementation
12  * of the Hcub algorithm at http://spiral.ece.cmu.edu/mcm/gen.html.
13  *
14  * That software is really not designed for a single multiplier this large,
15  * but if you run it enough times with different seeds, it'll find several
16  * 6-shift, 6-add sequences for computing x * 0x61C88647.  They are all
17  *	c = (x << 19) + x;
18  *	a = (x <<  9) + c;
19  *	b = (x << 23) + a;
20  *	return (a<<11) + (b<<6) + (c<<3) - b;
21  * with variations on the order of the final add.
22  *
23  * Without even a shifter, it's hopless; any hash function will suck.
24  */
25 
26 #if CONFIG_XILINX_MICROBLAZE0_USE_HW_MUL == 0
27 
28 #define HAVE_ARCH__HASH_32 1
29 
30 /* Multiply by GOLDEN_RATIO_32 = 0x61C88647 */
31 static inline u32 __attribute_const__ __hash_32(u32 a)
32 {
33 #if CONFIG_XILINX_MICROBLAZE0_USE_BARREL
34 	unsigned int b, c;
35 
36 	/* Phase 1: Compute three intermediate values */
37 	b =  a << 23;
38 	c = (a << 19) + a;
39 	a = (a <<  9) + c;
40 	b += a;
41 
42 	/* Phase 2: Compute (a << 11) + (b << 6) + (c << 3) - b */
43 	a <<= 5;
44 	a += b;		/* (a << 5) + b */
45 	a <<= 3;
46 	a += c;		/* (a << 8) + (b << 3) + c */
47 	a <<= 3;
48 	return a - b;	/* (a << 11) + (b << 6) + (c << 3) - b */
49 #else
50 	/*
51 	 * "This is really going to hurt."
52 	 *
53 	 * Without a barrel shifter, left shifts are implemented as
54 	 * repeated additions, and the best we can do is an optimal
55 	 * addition-subtraction chain.  This one is not known to be
56 	 * optimal, but at 37 steps, it's decent for a 31-bit multiplier.
57 	 *
58 	 * Question: given its size (37*4 = 148 bytes per instance),
59 	 * and slowness, is this worth having inline?
60 	 */
61 	unsigned int b, c, d;
62 
63 	b = a << 4;	/* 4    */
64 	c = b << 1;	/* 1  5 */
65 	b += a;		/* 1  6 */
66 	c += b;		/* 1  7 */
67 	c <<= 3;	/* 3 10 */
68 	c -= a;		/* 1 11 */
69 	d = c << 7;	/* 7 18 */
70 	d += b;		/* 1 19 */
71 	d <<= 8;	/* 8 27 */
72 	d += a;		/* 1 28 */
73 	d <<= 1;	/* 1 29 */
74 	d += b;		/* 1 30 */
75 	d <<= 6;	/* 6 36 */
76 	return d + c;	/* 1 37 total instructions*/
77 #endif
78 }
79 
80 #endif /* !CONFIG_XILINX_MICROBLAZE0_USE_HW_MUL */
81 #endif /* _ASM_HASH_H */
82