rts/gmp/mpn/hppa/udiv_qrnnd.s

   1 ; HP-PA  __udiv_qrnnd division support, used from longlong.h.
   2 ; This version runs fast on pre-PA7000 CPUs.
   3
   4 ; Copyright (C) 1993, 1994, 2000 Free Software Foundation, Inc.
   5
   6 ; This file is part of the GNU MP Library.
   7
   8 ; The GNU MP Library is free software; you can redistribute it and/or modify
   9 ; it under the terms of the GNU Lesser General Public License as published by
  10 ; the Free Software Foundation; either version 2.1 of the License, or (at your
  11 ; option) any later version.
  12
  13 ; The GNU MP Library is distributed in the hope that it will be useful, but
  14 ; WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  15 ; or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
  16 ; License for more details.
  17
  18 ; You should have received a copy of the GNU Lesser General Public License
  19 ; along with the GNU MP Library; see the file COPYING.LIB.  If not, write to
  20 ; the Free Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
  21 ; MA 02111-1307, USA.
  22
  23
  24 ; INPUT PARAMETERS
  25 ; rem_ptr       gr26
  26 ; n1            gr25
  27 ; n0            gr24
  28 ; d             gr23
  29
  30 ; The code size is a bit excessive.  We could merge the last two ds;addc
  31 ; sequences by simply moving the "bb,< Odd" instruction down.  The only
  32 ; trouble is the FFFFFFFF code that would need some hacking.
  33
  34         .code
  35         .export         __gmpn_udiv_qrnnd
  36 __gmpn_udiv_qrnnd
  37         .proc
  38         .callinfo       frame=0,no_calls
  39         .entry
  40
  41         comb,<          %r23,0,L$largedivisor
  42          sub            %r0,%r23,%r1            ; clear cy as side-effect
  43         ds              %r0,%r1,%r0
  44         addc            %r24,%r24,%r24
  45         ds              %r25,%r23,%r25
  46         addc            %r24,%r24,%r24
  47         ds              %r25,%r23,%r25
  48         addc            %r24,%r24,%r24
  49         ds              %r25,%r23,%r25
  50         addc            %r24,%r24,%r24
  51         ds              %r25,%r23,%r25
  52         addc            %r24,%r24,%r24
  53         ds              %r25,%r23,%r25
  54         addc            %r24,%r24,%r24
  55         ds              %r25,%r23,%r25
  56         addc            %r24,%r24,%r24
  57         ds              %r25,%r23,%r25
  58         addc            %r24,%r24,%r24
  59         ds              %r25,%r23,%r25
  60         addc            %r24,%r24,%r24
  61         ds              %r25,%r23,%r25
  62         addc            %r24,%r24,%r24
  63         ds              %r25,%r23,%r25
  64         addc            %r24,%r24,%r24
  65         ds              %r25,%r23,%r25
  66         addc            %r24,%r24,%r24
  67         ds              %r25,%r23,%r25
  68         addc            %r24,%r24,%r24
  69         ds              %r25,%r23,%r25
  70         addc            %r24,%r24,%r24
  71         ds              %r25,%r23,%r25
  72         addc            %r24,%r24,%r24
  73         ds              %r25,%r23,%r25
  74         addc            %r24,%r24,%r24
  75         ds              %r25,%r23,%r25
  76         addc            %r24,%r24,%r24
  77         ds              %r25,%r23,%r25
  78         addc            %r24,%r24,%r24
  79         ds              %r25,%r23,%r25
  80         addc            %r24,%r24,%r24
  81         ds              %r25,%r23,%r25
  82         addc            %r24,%r24,%r24
  83         ds              %r25,%r23,%r25
  84         addc            %r24,%r24,%r24
  85         ds              %r25,%r23,%r25
  86         addc            %r24,%r24,%r24
  87         ds              %r25,%r23,%r25
  88         addc            %r24,%r24,%r24
  89         ds              %r25,%r23,%r25
  90         addc            %r24,%r24,%r24
  91         ds              %r25,%r23,%r25
  92         addc            %r24,%r24,%r24
  93         ds              %r25,%r23,%r25
  94         addc            %r24,%r24,%r24
  95         ds              %r25,%r23,%r25
  96         addc            %r24,%r24,%r24
  97         ds              %r25,%r23,%r25
  98         addc            %r24,%r24,%r24
  99         ds              %r25,%r23,%r25
 100         addc            %r24,%r24,%r24
 101         ds              %r25,%r23,%r25
 102         addc            %r24,%r24,%r24
 103         ds              %r25,%r23,%r25
 104         addc            %r24,%r24,%r24
 105         ds              %r25,%r23,%r25
 106         addc            %r24,%r24,%r28
 107         ds              %r25,%r23,%r25
 108         comclr,>=       %r25,%r0,%r0
 109         addl            %r25,%r23,%r25
 110         stws            %r25,0(0,%r26)
 111         bv              0(%r2)
 112          addc           %r28,%r28,%r28
 113
 114 L$largedivisor
 115         extru           %r24,31,1,%r19          ; r19 = n0 & 1
 116         bb,<            %r23,31,L$odd
 117          extru          %r23,30,31,%r22         ; r22 = d >> 1
 118         shd             %r25,%r24,1,%r24        ; r24 = new n0
 119         extru           %r25,30,31,%r25         ; r25 = new n1
 120         sub             %r0,%r22,%r21
 121         ds              %r0,%r21,%r0
 122         addc            %r24,%r24,%r24
 123         ds              %r25,%r22,%r25
 124         addc            %r24,%r24,%r24
 125         ds              %r25,%r22,%r25
 126         addc            %r24,%r24,%r24
 127         ds              %r25,%r22,%r25
 128         addc            %r24,%r24,%r24
 129         ds              %r25,%r22,%r25
 130         addc            %r24,%r24,%r24
 131         ds              %r25,%r22,%r25
 132         addc            %r24,%r24,%r24
 133         ds              %r25,%r22,%r25
 134         addc            %r24,%r24,%r24
 135         ds              %r25,%r22,%r25
 136         addc            %r24,%r24,%r24
 137         ds              %r25,%r22,%r25
 138         addc            %r24,%r24,%r24
 139         ds              %r25,%r22,%r25
 140         addc            %r24,%r24,%r24
 141         ds              %r25,%r22,%r25
 142         addc            %r24,%r24,%r24
 143         ds              %r25,%r22,%r25
 144         addc            %r24,%r24,%r24
 145         ds              %r25,%r22,%r25
 146         addc            %r24,%r24,%r24
 147         ds              %r25,%r22,%r25
 148         addc            %r24,%r24,%r24
 149         ds              %r25,%r22,%r25
 150         addc            %r24,%r24,%r24
 151         ds              %r25,%r22,%r25
 152         addc            %r24,%r24,%r24
 153         ds              %r25,%r22,%r25
 154         addc            %r24,%r24,%r24
 155         ds              %r25,%r22,%r25
 156         addc            %r24,%r24,%r24
 157         ds              %r25,%r22,%r25
 158         addc            %r24,%r24,%r24
 159         ds              %r25,%r22,%r25
 160         addc            %r24,%r24,%r24
 161         ds              %r25,%r22,%r25
 162         addc            %r24,%r24,%r24
 163         ds              %r25,%r22,%r25
 164         addc            %r24,%r24,%r24
 165         ds              %r25,%r22,%r25
 166         addc            %r24,%r24,%r24
 167         ds              %r25,%r22,%r25
 168         addc            %r24,%r24,%r24
 169         ds              %r25,%r22,%r25
 170         addc            %r24,%r24,%r24
 171         ds              %r25,%r22,%r25
 172         addc            %r24,%r24,%r24
 173         ds              %r25,%r22,%r25
 174         addc            %r24,%r24,%r24
 175         ds              %r25,%r22,%r25
 176         addc            %r24,%r24,%r24
 177         ds              %r25,%r22,%r25
 178         addc            %r24,%r24,%r24
 179         ds              %r25,%r22,%r25
 180         addc            %r24,%r24,%r24
 181         ds              %r25,%r22,%r25
 182         addc            %r24,%r24,%r24
 183         ds              %r25,%r22,%r25
 184         addc            %r24,%r24,%r24
 185         ds              %r25,%r22,%r25
 186         comclr,>=       %r25,%r0,%r0
 187         addl            %r25,%r22,%r25
 188         sh1addl         %r25,%r19,%r25
 189         stws            %r25,0(0,%r26)
 190         bv              0(%r2)
 191          addc           %r24,%r24,%r28
 192
 193 L$odd   addib,sv,n      1,%r22,L$FF..           ; r22 = (d / 2 + 1)
 194         shd             %r25,%r24,1,%r24        ; r24 = new n0
 195         extru           %r25,30,31,%r25         ; r25 = new n1
 196         sub             %r0,%r22,%r21
 197         ds              %r0,%r21,%r0
 198         addc            %r24,%r24,%r24
 199         ds              %r25,%r22,%r25
 200         addc            %r24,%r24,%r24
 201         ds              %r25,%r22,%r25
 202         addc            %r24,%r24,%r24
 203         ds              %r25,%r22,%r25
 204         addc            %r24,%r24,%r24
 205         ds              %r25,%r22,%r25
 206         addc            %r24,%r24,%r24
 207         ds              %r25,%r22,%r25
 208         addc            %r24,%r24,%r24
 209         ds              %r25,%r22,%r25
 210         addc            %r24,%r24,%r24
 211         ds              %r25,%r22,%r25
 212         addc            %r24,%r24,%r24
 213         ds              %r25,%r22,%r25
 214         addc            %r24,%r24,%r24
 215         ds              %r25,%r22,%r25
 216         addc            %r24,%r24,%r24
 217         ds              %r25,%r22,%r25
 218         addc            %r24,%r24,%r24
 219         ds              %r25,%r22,%r25
 220         addc            %r24,%r24,%r24
 221         ds              %r25,%r22,%r25
 222         addc            %r24,%r24,%r24
 223         ds              %r25,%r22,%r25
 224         addc            %r24,%r24,%r24
 225         ds              %r25,%r22,%r25
 226         addc            %r24,%r24,%r24
 227         ds              %r25,%r22,%r25
 228         addc            %r24,%r24,%r24
 229         ds              %r25,%r22,%r25
 230         addc            %r24,%r24,%r24
 231         ds              %r25,%r22,%r25
 232         addc            %r24,%r24,%r24
 233         ds              %r25,%r22,%r25
 234         addc            %r24,%r24,%r24
 235         ds              %r25,%r22,%r25
 236         addc            %r24,%r24,%r24
 237         ds              %r25,%r22,%r25
 238         addc            %r24,%r24,%r24
 239         ds              %r25,%r22,%r25
 240         addc            %r24,%r24,%r24
 241         ds              %r25,%r22,%r25
 242         addc            %r24,%r24,%r24
 243         ds              %r25,%r22,%r25
 244         addc            %r24,%r24,%r24
 245         ds              %r25,%r22,%r25
 246         addc            %r24,%r24,%r24
 247         ds              %r25,%r22,%r25
 248         addc            %r24,%r24,%r24
 249         ds              %r25,%r22,%r25
 250         addc            %r24,%r24,%r24
 251         ds              %r25,%r22,%r25
 252         addc            %r24,%r24,%r24
 253         ds              %r25,%r22,%r25
 254         addc            %r24,%r24,%r24
 255         ds              %r25,%r22,%r25
 256         addc            %r24,%r24,%r24
 257         ds              %r25,%r22,%r25
 258         addc            %r24,%r24,%r24
 259         ds              %r25,%r22,%r25
 260         addc            %r24,%r24,%r24
 261         ds              %r25,%r22,%r25
 262         addc            %r24,%r24,%r28
 263         comclr,>=       %r25,%r0,%r0
 264         addl            %r25,%r22,%r25
 265         sh1addl         %r25,%r19,%r25
 266 ; We have computed (n1,,n0) / (d + 1), q' = r28, r' = r25
 267         add,nuv         %r28,%r25,%r25
 268         addl            %r25,%r1,%r25
 269         addc            %r0,%r28,%r28
 270         sub,<<          %r25,%r23,%r0
 271         addl            %r25,%r1,%r25
 272         stws            %r25,0(0,%r26)
 273         bv              0(%r2)
 274          addc           %r0,%r28,%r28
 275
 276 ; This is just a special case of the code above.
 277 ; We come here when d == 0xFFFFFFFF
 278 L$FF..  add,uv          %r25,%r24,%r24
 279         sub,<<          %r24,%r23,%r0
 280         ldo             1(%r24),%r24
 281         stws            %r24,0(0,%r26)
 282         bv              0(%r2)
 283          addc           %r0,%r25,%r28
 284
 285         .exit
 286         .procend