source/libs/gmp/gmp-src/mpn/sparc64/ultrasparct3/bdiv_dbm1c.asm

   1 dnl  SPARC T3/T4/T5 mpn_bdiv_dbm1c.
   2
   3 dnl  Contributed to the GNU project by Torbjörn Granlund.
   4
   5 dnl  Copyright 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
  36 C UltraSPARC T3:        25
  37 C UltraSPARC T4/T5:      4
  38
  39 C INPUT PARAMETERS
  40 define(`qp',  `%i0')
  41 define(`ap',  `%i1')
  42 define(`n',   `%i2')
  43 define(`bd',  `%i3')
  44 define(`h',   `%i4')
  45
  46 define(`plo0',`%g4')  define(`plo1',`%g5')
  47 define(`phi0',`%l0')  define(`phi1',`%l1')
  48 define(`a0',  `%g1')  define(`a1',  `%g3')
  49
  50 ASM_START()
  51         REGISTER(%g2,#scratch)
  52         REGISTER(%g3,#scratch)
  53 PROLOGUE(mpn_bdiv_dbm1c)
  54         save    %sp, -176, %sp
  55
  56         and     n, 3, %g5
  57         ldx     [ap + 0], %g2
  58         add     n, -5, n
  59         brz     %g5, L(b0)
  60          cmp    %g5, 2
  61         bcs     %xcc, L(b1)
  62          nop
  63         be      %xcc, L(b2)
  64          nop
  65
  66 L(b3):  ldx     [ap + 8], a0
  67         mulx    bd, %g2, plo1
  68         umulxhi(bd, %g2, phi1)
  69         ldx     [ap + 16], a1
  70         add     qp, -24, qp
  71         b       L(lo3)
  72          add    ap, -8, ap
  73
  74 L(b2):  ldx     [ap + 8], a1
  75         mulx    bd, %g2, plo0
  76         umulxhi(bd, %g2, phi0)
  77         brlz,pt n, L(wd2)
  78          nop
  79 L(gt2): ldx     [ap + 16], a0
  80         add     ap, 16, ap
  81         b       L(lo2)
  82          add    n, -1, n
  83
  84 L(b1):  mulx    bd, %g2, plo1
  85          umulxhi(bd, %g2, phi1)
  86         brlz,pn n, L(wd1)
  87          add    qp, -8, qp
  88 L(gt1): ldx     [ap + 8], a0
  89         ldx     [ap + 16], a1
  90         b       L(lo1)
  91          add    ap, 8, ap
  92
  93 L(b0):  ldx     [ap + 8], a1
  94         mulx    bd, %g2, plo0
  95         umulxhi(bd, %g2, phi0)
  96         ldx     [ap + 16], a0
  97         b       L(lo0)
  98          add    qp, -16, qp
  99
 100 L(top): ldx     [ap + 0], a0
 101         sub     h, phi1, h
 102 L(lo2): mulx    bd, a1, plo1
 103         umulxhi(bd, a1, phi1)
 104         subcc   h, plo0, h
 105         addxc(  phi0, %g0, phi0)
 106         stx     h, [qp + 0]
 107         ldx     [ap + 8], a1
 108         sub     h, phi0, h
 109 L(lo1): mulx    bd, a0, plo0
 110         umulxhi(bd, a0, phi0)
 111         subcc   h, plo1, h
 112         addxc(  phi1, %g0, phi1)
 113         stx     h, [qp + 8]
 114         ldx     [ap + 16], a0
 115         sub     h, phi1, h
 116 L(lo0): mulx    bd, a1, plo1
 117         umulxhi(bd, a1, phi1)
 118         subcc   h, plo0, h
 119         addxc(  phi0, %g0, phi0)
 120         stx     h, [qp + 16]
 121         ldx     [ap + 24], a1
 122         sub     h, phi0, h
 123 L(lo3): mulx    bd, a0, plo0
 124         umulxhi(bd, a0, phi0)
 125         subcc   h, plo1, h
 126         addxc(  phi1, %g0, phi1)
 127         stx     h, [qp + 24]
 128         add     ap, 32, ap
 129         add     qp, 32, qp
 130         brgz,pt n, L(top)
 131          add    n, -4, n
 132
 133 L(end): sub     h, phi1, h
 134 L(wd2): mulx    bd, a1, plo1
 135         umulxhi(bd, a1, phi1)
 136         subcc   h, plo0, h
 137         addxc(  phi0, %g0, phi0)
 138         stx     h, [qp + 0]
 139         sub     h, phi0, h
 140 L(wd1): subcc   h, plo1, h
 141         addxc(  phi1, %g0, phi1)
 142         stx     h, [qp + 8]
 143         sub     h, phi1, %i0
 144
 145         ret
 146          restore
 147 EPILOGUE()