1  /* Cray PVP/IEEE mpn_addmul_1 -- multiply a limb vector with a limb and add the
       2     result to a second limb vector.
       3  
       4  Copyright 2000-2002 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 either:
      10  
      11    * the GNU Lesser General Public License as published by the Free
      12      Software Foundation; either version 3 of the License, or (at your
      13      option) any later version.
      14  
      15  or
      16  
      17    * the GNU General Public License as published by the Free Software
      18      Foundation; either version 2 of the License, or (at your option) any
      19      later version.
      20  
      21  or both in parallel, as here.
      22  
      23  The GNU MP Library is distributed in the hope that it will be useful, but
      24  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
      25  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
      26  for more details.
      27  
      28  You should have received copies of the GNU General Public License and the
      29  GNU Lesser General Public License along with the GNU MP Library.  If not,
      30  see https://www.gnu.org/licenses/.  */
      31  
      32  /* This code runs at just under 9 cycles/limb on a T90.  That is not perfect,
      33     mainly due to vector register shortage in the main loop.  Assembly code
      34     should bring it down to perhaps 7 cycles/limb.  */
      35  
      36  #include <intrinsics.h>
      37  #include "gmp-impl.h"
      38  
      39  mp_limb_t
      40  mpn_addmul_1 (mp_ptr rp, mp_srcptr up, mp_size_t n, mp_limb_t vl)
      41  {
      42    mp_limb_t cy[n];
      43    mp_limb_t a, b, r, s0, s1, c0, c1;
      44    mp_size_t i;
      45    int more_carries;
      46  
      47    if (up == rp)
      48      {
      49        /* The algorithm used below cannot handle overlap.  Handle it here by
      50  	 making a temporary copy of the source vector, then call ourselves.  */
      51        mp_limb_t xp[n];
      52        MPN_COPY (xp, up, n);
      53        return mpn_addmul_1 (rp, xp, n, vl);
      54      }
      55  
      56    a = up[0] * vl;
      57    r = rp[0];
      58    s0 = a + r;
      59    rp[0] = s0;
      60    c0 = ((a & r) | ((a | r) & ~s0)) >> 63;
      61    cy[0] = c0;
      62  
      63    /* Main multiply loop.  Generate a raw accumulated output product in rp[]
      64       and a carry vector in cy[].  */
      65  #pragma _CRI ivdep
      66    for (i = 1; i < n; i++)
      67      {
      68        a = up[i] * vl;
      69        b = _int_mult_upper (up[i - 1], vl);
      70        s0 = a + b;
      71        c0 = ((a & b) | ((a | b) & ~s0)) >> 63;
      72        r = rp[i];
      73        s1 = s0 + r;
      74        rp[i] = s1;
      75        c1 = ((s0 & r) | ((s0 | r) & ~s1)) >> 63;
      76        cy[i] = c0 + c1;
      77      }
      78    /* Carry add loop.  Add the carry vector cy[] to the raw result rp[] and
      79       store the new result back to rp[].  */
      80    more_carries = 0;
      81  #pragma _CRI ivdep
      82    for (i = 1; i < n; i++)
      83      {
      84        r = rp[i];
      85        c0 = cy[i - 1];
      86        s0 = r + c0;
      87        rp[i] = s0;
      88        c0 = (r & ~s0) >> 63;
      89        more_carries += c0;
      90      }
      91    /* If that second loop generated carry, handle that in scalar loop.  */
      92    if (more_carries)
      93      {
      94        mp_limb_t cyrec = 0;
      95        /* Look for places where rp[k] == 0 and cy[k-1] == 1 or
      96  	 rp[k] == 1 and cy[k-1] == 2.
      97  	 These are where we got a recurrency carry.  */
      98        for (i = 1; i < n; i++)
      99  	{
     100  	  r = rp[i];
     101  	  c0 = r < cy[i - 1];
     102  	  s0 = r + cyrec;
     103  	  rp[i] = s0;
     104  	  c1 = (r & ~s0) >> 63;
     105  	  cyrec = c0 | c1;
     106  	}
     107        return _int_mult_upper (up[n - 1], vl) + cyrec + cy[n - 1];
     108      }
     109  
     110    return _int_mult_upper (up[n - 1], vl) + cy[n - 1];
     111  }