linux/arch/microblaze/include/asm/hash.h
<<
>>
Prefs
   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 */
  32static 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