source/libs/gmp/gmp-src/mpn/arm/v6/dive_1.asm

   1 dnl  ARM v6 mpn_divexact_1
   2
   3 dnl  Contributed to the GNU project by Torbjörn Granlund.
   4
   5 dnl  Copyright 2012, 2013 Free Software Foundation, Inc.
   6
   7 dnl  This file is part of the GNU MP Library.
   8 dnl
   9 dnl  The GNU MP Library is free software; you can redistribute it and/or modify
  10 dnl  it under the terms of either:
  11 dnl
  12 dnl    * the GNU Lesser General Public License as published by the Free
  13 dnl      Software Foundation; either version 3 of the License, or (at your
  14 dnl      option) any later version.
  15 dnl
  16 dnl  or
  17 dnl
  18 dnl    * the GNU General Public License as published by the Free Software
  19 dnl      Foundation; either version 2 of the License, or (at your option) any
  20 dnl      later version.
  21 dnl
  22 dnl  or both in parallel, as here.
  23 dnl
  24 dnl  The GNU MP Library is distributed in the hope that it will be useful, but
  25 dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  26 dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  27 dnl  for more details.
  28 dnl
  29 dnl  You should have received copies of the GNU General Public License and the
  30 dnl  GNU Lesser General Public License along with the GNU MP Library.  If not,
  31 dnl  see https://www.gnu.org/licenses/.
  32
  33 include(`../config.m4')
  34
  35 C               cycles/limb       cycles/limb
  36 C               norm    unorm    modexact_1c_odd
  37 C StrongARM      -       -
  38 C XScale         -       -
  39 C Cortex-A7      ?       ?
  40 C Cortex-A8      ?       ?
  41 C Cortex-A9      9      10               9
  42 C Cortex-A15     7       7               7
  43
  44 C Architecture requirements:
  45 C v5    -
  46 C v5t   clz
  47 C v5te  -
  48 C v6    umaal
  49 C v6t2  -
  50 C v7a   -
  51
  52 define(`rp', `r0')
  53 define(`up', `r1')
  54 define(`n',  `r2')
  55 define(`d',  `r3')
  56
  57 define(`cy',  `r7')
  58 define(`cnt', `r6')
  59 define(`tnc', `r10')
  60
  61 ASM_START()
  62 PROLOGUE(mpn_divexact_1)
  63         push    {r4,r5,r6,r7,r8,r9}
  64
  65         tst     d, #1
  66
  67         rsb     r4, d, #0
  68         and     r4, r4, d
  69         clz     r4, r4
  70         rsb     cnt, r4, #31            C count_trailing_zeros
  71         mov     d, d, lsr cnt
  72
  73 C binvert limb
  74         LEA(    r4, binvert_limb_table)
  75         and     r12, d, #254
  76         ldrb    r4, [r4, r12, lsr #1]
  77         mul     r12, r4, r4
  78         mul     r12, d, r12
  79         rsb     r12, r12, r4, lsl #1
  80         mul     r4, r12, r12
  81         mul     r4, d, r4
  82         rsb     r4, r4, r12, lsl #1     C r4 = inverse
  83
  84         ldr     r5, [up], #4            C up[0]
  85         mov     cy, #0
  86         rsb     r8, r4, #0              C r8 = -inverse
  87         beq     L(unnorm)
  88
  89 L(norm):
  90         subs    n, n, #1
  91         mul     r5, r5, r4
  92         beq     L(end)
  93
  94         ALIGN(16)
  95 L(top): ldr     r9, [up], #4
  96         mov     r12, #0
  97         str     r5, [rp], #4
  98         umaal   r12, cy, r5, d
  99         mul     r5, r9, r4
 100         mla     r5, cy, r8, r5
 101         subs    n, n, #1
 102         bne     L(top)
 103
 104 L(end): str     r5, [rp]
 105         pop     {r4,r5,r6,r7,r8,r9}
 106         bx      r14
 107
 108 L(unnorm):
 109         push    {r10,r11}
 110         rsb     tnc, cnt, #32
 111         mov     r11, r5, lsr cnt
 112         subs    n, n, #1
 113         beq     L(edx)
 114
 115         ldr     r12, [up], #4
 116         orr     r9, r11, r12, lsl tnc
 117         mov     r11, r12, lsr cnt
 118         mul     r5, r9, r4
 119         subs    n, n, #1
 120         beq     L(edu)
 121
 122         ALIGN(16)
 123 L(tpu): ldr     r12, [up], #4
 124         orr     r9, r11, r12, lsl tnc
 125         mov     r11, r12, lsr cnt
 126         mov     r12, #0
 127         str     r5, [rp], #4
 128         umaal   r12, cy, r5, d
 129         mul     r5, r9, r4
 130         mla     r5, cy, r8, r5
 131         subs    n, n, #1
 132         bne     L(tpu)
 133
 134 L(edu): str     r5, [rp], #4
 135         mov     r12, #0
 136         umaal   r12, cy, r5, d
 137         mul     r5, r11, r4
 138         mla     r5, cy, r8, r5
 139         str     r5, [rp]
 140         pop     {r10,r11}
 141         pop     {r4,r5,r6,r7,r8,r9}
 142         bx      r14
 143
 144 L(edx): mul     r5, r11, r4
 145         str     r5, [rp]
 146         pop     {r10,r11}
 147         pop     {r4,r5,r6,r7,r8,r9}
 148         bx      r14
 149 EPILOGUE()