/* Compute x * y + z as ternary operation.
  Copyright (C) 2010-2018 Free Software Foundation, Inc.
  This file is part of the GNU C Library.
  Contributed by Jakub Jelinek <[email protected]>, 2010.

  The GNU C Library is free software; you can redistribute it and/or
  modify it under the terms of the GNU Lesser General Public
  License as published by the Free Software Foundation; either
  version 2.1 of the License, or (at your option) any later version.

  The GNU C Library is distributed in the hope that it will be useful,
  but WITHOUT ANY WARRANTY; without even the implied warranty of
  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  Lesser General Public License for more details.

  You should have received a copy of the GNU Lesser General Public
  License along with the GNU C Library; if not, see
  <http://www.gnu.org/licenses/>.  */

#include "quadmath-imp.h"

/* This implementation uses rounding to odd to avoid problems with
  double rounding.  See a paper by Boldo and Melquiond:
  http://www.lri.fr/~melquion/doc/08-tc.pdf  */

__float128
fmaq (__float128 x, __float128 y, __float128 z)
{
 ieee854_float128 u, v, w;
 int adjust = 0;
 u.value = x;
 v.value = y;
 w.value = z;
 if (__builtin_expect (u.ieee.exponent + v.ieee.exponent
                       >= 0x7fff + IEEE854_FLOAT128_BIAS
                          - FLT128_MANT_DIG, 0)
     || __builtin_expect (u.ieee.exponent >= 0x7fff - FLT128_MANT_DIG, 0)
     || __builtin_expect (v.ieee.exponent >= 0x7fff - FLT128_MANT_DIG, 0)
     || __builtin_expect (w.ieee.exponent >= 0x7fff - FLT128_MANT_DIG, 0)
     || __builtin_expect (u.ieee.exponent + v.ieee.exponent
                          <= IEEE854_FLOAT128_BIAS + FLT128_MANT_DIG, 0))
   {
     /* If z is Inf, but x and y are finite, the result should be
        z rather than NaN.  */
     if (w.ieee.exponent == 0x7fff
         && u.ieee.exponent != 0x7fff
         && v.ieee.exponent != 0x7fff)
       return (z + x) + y;
     /* If z is zero and x are y are nonzero, compute the result
        as x * y to avoid the wrong sign of a zero result if x * y
        underflows to 0.  */
     if (z == 0 && x != 0 && y != 0)
       return x * y;
     /* If x or y or z is Inf/NaN, or if x * y is zero, compute as
        x * y + z.  */
     if (u.ieee.exponent == 0x7fff
         || v.ieee.exponent == 0x7fff
         || w.ieee.exponent == 0x7fff
         || x == 0
         || y == 0)
       return x * y + z;
     /* If fma will certainly overflow, compute as x * y.  */
     if (u.ieee.exponent + v.ieee.exponent
         > 0x7fff + IEEE854_FLOAT128_BIAS)
       return x * y;
     /* If x * y is less than 1/4 of FLT128_TRUE_MIN, neither the
        result nor whether there is underflow depends on its exact
        value, only on its sign.  */
     if (u.ieee.exponent + v.ieee.exponent
         < IEEE854_FLOAT128_BIAS - FLT128_MANT_DIG - 2)
       {
         int neg = u.ieee.negative ^ v.ieee.negative;
         __float128 tiny = neg ? -0x1p-16494Q : 0x1p-16494Q;
         if (w.ieee.exponent >= 3)
           return tiny + z;
         /* Scaling up, adding TINY and scaling down produces the
            correct result, because in round-to-nearest mode adding
            TINY has no effect and in other modes double rounding is
            harmless.  But it may not produce required underflow
            exceptions.  */
         v.value = z * 0x1p114Q + tiny;
         if (TININESS_AFTER_ROUNDING
             ? v.ieee.exponent < 115
             : (w.ieee.exponent == 0
                || (w.ieee.exponent == 1
                    && w.ieee.negative != neg
                    && w.ieee.mantissa3 == 0
                    && w.ieee.mantissa2 == 0
                    && w.ieee.mantissa1 == 0
                    && w.ieee.mantissa0 == 0)))
           {
             __float128 force_underflow = x * y;
             math_force_eval (force_underflow);
           }
         return v.value * 0x1p-114Q;
       }
     if (u.ieee.exponent + v.ieee.exponent
         >= 0x7fff + IEEE854_FLOAT128_BIAS - FLT128_MANT_DIG)
       {
         /* Compute 1p-113 times smaller result and multiply
            at the end.  */
         if (u.ieee.exponent > v.ieee.exponent)
           u.ieee.exponent -= FLT128_MANT_DIG;
         else
           v.ieee.exponent -= FLT128_MANT_DIG;
         /* If x + y exponent is very large and z exponent is very small,
            it doesn't matter if we don't adjust it.  */
         if (w.ieee.exponent > FLT128_MANT_DIG)
           w.ieee.exponent -= FLT128_MANT_DIG;
         adjust = 1;
       }
     else if (w.ieee.exponent >= 0x7fff - FLT128_MANT_DIG)
       {
         /* Similarly.
            If z exponent is very large and x and y exponents are
            very small, adjust them up to avoid spurious underflows,
            rather than down.  */
         if (u.ieee.exponent + v.ieee.exponent
             <= IEEE854_FLOAT128_BIAS + 2 * FLT128_MANT_DIG)
           {
             if (u.ieee.exponent > v.ieee.exponent)
               u.ieee.exponent += 2 * FLT128_MANT_DIG + 2;
             else
               v.ieee.exponent += 2 * FLT128_MANT_DIG + 2;
           }
         else if (u.ieee.exponent > v.ieee.exponent)
           {
             if (u.ieee.exponent > FLT128_MANT_DIG)
               u.ieee.exponent -= FLT128_MANT_DIG;
           }
         else if (v.ieee.exponent > FLT128_MANT_DIG)
           v.ieee.exponent -= FLT128_MANT_DIG;
         w.ieee.exponent -= FLT128_MANT_DIG;
         adjust = 1;
       }
     else if (u.ieee.exponent >= 0x7fff - FLT128_MANT_DIG)
       {
         u.ieee.exponent -= FLT128_MANT_DIG;
         if (v.ieee.exponent)
           v.ieee.exponent += FLT128_MANT_DIG;
         else
           v.value *= 0x1p113Q;
       }
     else if (v.ieee.exponent >= 0x7fff - FLT128_MANT_DIG)
       {
         v.ieee.exponent -= FLT128_MANT_DIG;
         if (u.ieee.exponent)
           u.ieee.exponent += FLT128_MANT_DIG;
         else
           u.value *= 0x1p113Q;
       }
     else /* if (u.ieee.exponent + v.ieee.exponent
                 <= IEEE854_FLOAT128_BIAS + FLT128_MANT_DIG) */
       {
         if (u.ieee.exponent > v.ieee.exponent)
           u.ieee.exponent += 2 * FLT128_MANT_DIG + 2;
         else
           v.ieee.exponent += 2 * FLT128_MANT_DIG + 2;
         if (w.ieee.exponent <= 4 * FLT128_MANT_DIG + 6)
           {
             if (w.ieee.exponent)
               w.ieee.exponent += 2 * FLT128_MANT_DIG + 2;
             else
               w.value *= 0x1p228Q;
             adjust = -1;
           }
         /* Otherwise x * y should just affect inexact
            and nothing else.  */
       }
     x = u.value;
     y = v.value;
     z = w.value;
   }

 /* Ensure correct sign of exact 0 + 0.  */
 if (__glibc_unlikely ((x == 0 || y == 0) && z == 0))
   {
     x = math_opt_barrier (x);
     return x * y + z;
   }

 fenv_t env;
 feholdexcept (&env);
 fesetround (FE_TONEAREST);

 /* Multiplication m1 + m2 = x * y using Dekker's algorithm.  */
#define C ((1LL << (FLT128_MANT_DIG + 1) / 2) + 1)
 __float128 x1 = x * C;
 __float128 y1 = y * C;
 __float128 m1 = x * y;
 x1 = (x - x1) + x1;
 y1 = (y - y1) + y1;
 __float128 x2 = x - x1;
 __float128 y2 = y - y1;
 __float128 m2 = (((x1 * y1 - m1) + x1 * y2) + x2 * y1) + x2 * y2;

 /* Addition a1 + a2 = z + m1 using Knuth's algorithm.  */
 __float128 a1 = z + m1;
 __float128 t1 = a1 - z;
 __float128 t2 = a1 - t1;
 t1 = m1 - t1;
 t2 = z - t2;
 __float128 a2 = t1 + t2;
 /* Ensure the arithmetic is not scheduled after feclearexcept call.  */
 math_force_eval (m2);
 math_force_eval (a2);
 feclearexcept (FE_INEXACT);

 /* If the result is an exact zero, ensure it has the correct sign.  */
 if (a1 == 0 && m2 == 0)
   {
     feupdateenv (&env);
     /* Ensure that round-to-nearest value of z + m1 is not reused.  */
     z = math_opt_barrier (z);
     return z + m1;
   }

 fesetround (FE_TOWARDZERO);
 /* Perform m2 + a2 addition with round to odd.  */
 u.value = a2 + m2;

 if (__glibc_likely (adjust == 0))
   {
     if ((u.ieee.mantissa3 & 1) == 0 && u.ieee.exponent != 0x7fff)
       u.ieee.mantissa3 |= fetestexcept (FE_INEXACT) != 0;
     feupdateenv (&env);
     /* Result is a1 + u.value.  */
     return a1 + u.value;
   }
 else if (__glibc_likely (adjust > 0))
   {
     if ((u.ieee.mantissa3 & 1) == 0 && u.ieee.exponent != 0x7fff)
       u.ieee.mantissa3 |= fetestexcept (FE_INEXACT) != 0;
     feupdateenv (&env);
     /* Result is a1 + u.value, scaled up.  */
     return (a1 + u.value) * 0x1p113Q;
   }
 else
   {
     if ((u.ieee.mantissa3 & 1) == 0)
       u.ieee.mantissa3 |= fetestexcept (FE_INEXACT) != 0;
     v.value = a1 + u.value;
     /* Ensure the addition is not scheduled after fetestexcept call.  */
     math_force_eval (v.value);
     int j = fetestexcept (FE_INEXACT) != 0;
     feupdateenv (&env);
     /* Ensure the following computations are performed in default rounding
        mode instead of just reusing the round to zero computation.  */
     asm volatile ("" : "=m" (u) : "m" (u));
     /* If a1 + u.value is exact, the only rounding happens during
        scaling down.  */
     if (j == 0)
       return v.value * 0x1p-228Q;
     /* If result rounded to zero is not subnormal, no double
        rounding will occur.  */
     if (v.ieee.exponent > 228)
       return (a1 + u.value) * 0x1p-228Q;
     /* If v.value * 0x1p-228L with round to zero is a subnormal above
        or equal to FLT128_MIN / 2, then v.value * 0x1p-228L shifts mantissa
        down just by 1 bit, which means v.ieee.mantissa3 |= j would
        change the round bit, not sticky or guard bit.
        v.value * 0x1p-228L never normalizes by shifting up,
        so round bit plus sticky bit should be already enough
        for proper rounding.  */
     if (v.ieee.exponent == 228)
       {
         /* If the exponent would be in the normal range when
            rounding to normal precision with unbounded exponent
            range, the exact result is known and spurious underflows
            must be avoided on systems detecting tininess after
            rounding.  */
         if (TININESS_AFTER_ROUNDING)
           {
             w.value = a1 + u.value;
             if (w.ieee.exponent == 229)
               return w.value * 0x1p-228Q;
           }
         /* v.ieee.mantissa3 & 2 is LSB bit of the result before rounding,
            v.ieee.mantissa3 & 1 is the round bit and j is our sticky
            bit.  */
         w.value = 0;
         w.ieee.mantissa3 = ((v.ieee.mantissa3 & 3) << 1) | j;
         w.ieee.negative = v.ieee.negative;
         v.ieee.mantissa3 &= ~3U;
         v.value *= 0x1p-228Q;
         w.value *= 0x1p-2Q;
         return v.value + w.value;
       }
     v.ieee.mantissa3 |= j;
     return v.value * 0x1p-228Q;
   }
}

You are viewing proxied material from ftp.icm.edu.pl. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.