source/libs/gmp/gmp-src/mpn/arm/v6/mul_2.asm

   1 dnl  ARM mpn_mul_2.
   2
   3 dnl  Contributed to the GNU project by Torbjörn Granlund.
   4
   5 dnl  Copyright 2012 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
  36 C StrongARM:     -
  37 C XScale         -
  38 C ARM11          5.25
  39 C Cortex-A7      3.13
  40 C Cortex-A8      5
  41 C Cortex-A9      2.25
  42 C Cortex-A15     2.5
  43
  44 C TODO
  45 C  * This is a trivial edit of the addmul_2 code.  Check for simplifications,
  46 C    and possible speedups to 2.0 c/l.
  47
  48 define(`rp',`r0')
  49 define(`up',`r1')
  50 define(`n', `r2')
  51 define(`vp',`r3')
  52
  53 define(`v0',`r6')
  54 define(`v1',`r7')
  55 define(`u0',`r3')
  56 define(`u1',`r9')
  57
  58 define(`cya',`r8')
  59 define(`cyb',`r12')
  60
  61
  62 ASM_START()
  63 PROLOGUE(mpn_mul_2)
  64         push    { r4, r5, r6, r7, r8, r9 }
  65
  66         ldm     vp, { v0, v1 }
  67         mov     cya, #0
  68         mov     cyb, #0
  69
  70         tst     n, #1
  71         beq     L(evn)
  72 L(odd): mov     r5, #0
  73         ldr     u0, [up, #0]
  74         mov     r4, #0
  75         tst     n, #2
  76         beq     L(fi1)
  77 L(fi3): sub     up, up, #12
  78         sub     rp, rp, #16
  79         b       L(lo3)
  80 L(fi1): sub     n, n, #1
  81         sub     up, up, #4
  82         sub     rp, rp, #8
  83         b       L(lo1)
  84 L(evn): mov     r4, #0
  85         ldr     u1, [up, #0]
  86         mov     r5, #0
  87         tst     n, #2
  88         bne     L(fi2)
  89 L(fi0): sub     up, up, #8
  90         sub     rp, rp, #12
  91         b       L(lo0)
  92 L(fi2): subs    n, n, #2
  93         sub     rp, rp, #4
  94         bls     L(end)
  95
  96         ALIGN(16)
  97 L(top): ldr     u0, [up, #4]
  98         umaal   r4, cya, u1, v0
  99         str     r4, [rp, #4]
 100         mov     r4, #0
 101         umaal   r5, cyb, u1, v1
 102 L(lo1): ldr     u1, [up, #8]
 103         umaal   r5, cya, u0, v0
 104         str     r5, [rp, #8]
 105         mov     r5, #0
 106         umaal   r4, cyb, u0, v1
 107 L(lo0): ldr     u0, [up, #12]
 108         umaal   r4, cya, u1, v0
 109         str     r4, [rp, #12]
 110         mov     r4, #0
 111         umaal   r5, cyb, u1, v1
 112 L(lo3): ldr     u1, [up, #16]!
 113         umaal   r5, cya, u0, v0
 114         str     r5, [rp, #16]!
 115         mov     r5, #0
 116         umaal   r4, cyb, u0, v1
 117         subs    n, n, #4
 118         bhi     L(top)
 119
 120 L(end): umaal   r4, cya, u1, v0
 121         ldr     u0, [up, #4]
 122         umaal   r5, cyb, u1, v1
 123         str     r4, [rp, #4]
 124         umaal   r5, cya, u0, v0
 125         umaal   cya, cyb, u0, v1
 126         str     r5, [rp, #8]
 127         str     cya, [rp, #12]
 128         mov     r0, cyb
 129
 130         pop     { r4, r5, r6, r7, r8, r9 }
 131         bx      r14
 132 EPILOGUE()