/* mpfr_sqrt -- square root of a floating-point number

Copyright 1999-2015 Free Software Foundation, Inc.
Contributed by the AriC and Caramel projects, INRIA.

This file is part of the GNU MPFR Library.

The GNU MPFR 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 3 of the License, or (at your
option) any later version.

The GNU MPFR 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 MPFR Library; see the file COPYING.LESSER.  If not, see
http://www.gnu.org/licenses/ or write to the Free Software Foundation, Inc.,
51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. */

#include "mpfr-impl.h"

int
mpfr_sqrt (mpfr_ptr r, mpfr_srcptr u, mpfr_rnd_t rnd_mode)
{
  mp_size_t rsize; /* number of limbs of r (plus 1 if exact limb multiple) */
  mp_size_t rrsize;
  mp_size_t usize; /* number of limbs of u */
  mp_size_t tsize; /* number of limbs of the sqrtrem remainder */
  mp_size_t k;
  mp_size_t l;
  mpfr_limb_ptr rp, rp0;
  mpfr_limb_ptr up;
  mpfr_limb_ptr sp;
  mp_limb_t sticky0; /* truncated part of input */
  mp_limb_t sticky1; /* truncated part of rp[0] */
  mp_limb_t sticky;
  int odd_exp;
  int sh; /* number of extra bits in rp[0] */
  int inexact; /* return ternary flag */
  mpfr_exp_t expr;
  MPFR_TMP_DECL(marker);

  MPFR_LOG_FUNC
    (("x[%Pu]=%.*Rg rnd=%d", mpfr_get_prec (u), mpfr_log_prec, u, rnd_mode),
     ("y[%Pu]=%.*Rg inexact=%d",
      mpfr_get_prec (r), mpfr_log_prec, r, inexact));

  if (MPFR_UNLIKELY(MPFR_IS_SINGULAR(u)))
    {
      if (MPFR_IS_NAN(u))
        {
          MPFR_SET_NAN(r);
          MPFR_RET_NAN;
        }
      else if (MPFR_IS_ZERO(u))
        {
          /* 0+ or 0- */
          MPFR_SET_SAME_SIGN(r, u);
          MPFR_SET_ZERO(r);
          MPFR_RET(0); /* zero is exact */
        }
      else
        {
          MPFR_ASSERTD(MPFR_IS_INF(u));
          /* sqrt(-Inf) = NAN */
          if (MPFR_IS_NEG(u))
            {
              MPFR_SET_NAN(r);
              MPFR_RET_NAN;
            }
          MPFR_SET_POS(r);
          MPFR_SET_INF(r);
          MPFR_RET(0);
        }
    }
  if (MPFR_UNLIKELY(MPFR_IS_NEG(u)))
    {
      MPFR_SET_NAN(r);
      MPFR_RET_NAN;
    }
  MPFR_SET_POS(r);

  MPFR_TMP_MARK (marker);
  MPFR_UNSIGNED_MINUS_MODULO(sh,MPFR_PREC(r));
  if (sh == 0 && rnd_mode == MPFR_RNDN)
    sh = GMP_NUMB_BITS; /* ugly case */
  rsize = MPFR_LIMB_SIZE(r) + (sh == GMP_NUMB_BITS);
  /* rsize is the number of limbs of r + 1 if exact limb multiple and rounding
     to nearest, this is the number of wanted limbs for the square root */
  rrsize = rsize + rsize;
  usize = MPFR_LIMB_SIZE(u); /* number of limbs of u */
  rp0 = MPFR_MANT(r);
  rp = (sh < GMP_NUMB_BITS) ? rp0 : MPFR_TMP_LIMBS_ALLOC (rsize);
  up = MPFR_MANT(u);
  sticky0 = MPFR_LIMB_ZERO; /* truncated part of input */
  sticky1 = MPFR_LIMB_ZERO; /* truncated part of rp[0] */
  odd_exp = (unsigned int) MPFR_GET_EXP (u) & 1;
  inexact = -1; /* return ternary flag */

  sp = MPFR_TMP_LIMBS_ALLOC (rrsize);

  /* copy the most significant limbs of u to {sp, rrsize} */
  if (MPFR_LIKELY(usize <= rrsize)) /* in case r and u have the same precision,
                                       we have indeed rrsize = 2 * usize */
    {
      k = rrsize - usize;
      if (MPFR_LIKELY(k))
        MPN_ZERO (sp, k);
      if (odd_exp)
        {
          if (MPFR_LIKELY(k))
            sp[k - 1] = mpn_rshift (sp + k, up, usize, 1);
          else
            sticky0 = mpn_rshift (sp, up, usize, 1);
        }
      else
        MPN_COPY (sp + rrsize - usize, up, usize);
    }
  else /* usize > rrsize: truncate the input */
    {
      k = usize - rrsize;
      if (odd_exp)
        sticky0 = mpn_rshift (sp, up + k, rrsize, 1);
      else
        MPN_COPY (sp, up + k, rrsize);
      l = k;
      while (sticky0 == MPFR_LIMB_ZERO && l != 0)
        sticky0 = up[--l];
    }

  /* sticky0 is non-zero iff the truncated part of the input is non-zero */

  /* mpn_rootrem with NULL 2nd argument is faster than mpn_sqrtrem, thus use
     it if available and if the user asked to use GMP internal functions */
#if defined(WANT_GMP_INTERNALS) && defined(HAVE___GMPN_ROOTREM)
  tsize = __gmpn_rootrem (rp, NULL, sp, rrsize, 2);
#else
  tsize = mpn_sqrtrem (rp, NULL, sp, rrsize);
#endif

  /* a return value of zero in mpn_sqrtrem indicates a perfect square */
  sticky = sticky0 || tsize != 0;

  /* truncate low bits of rp[0] */
  sticky1 = rp[0] & ((sh < GMP_NUMB_BITS) ? MPFR_LIMB_MASK(sh)
                     : ~MPFR_LIMB_ZERO);
  rp[0] -= sticky1;

  sticky = sticky || sticky1;

  expr = (MPFR_GET_EXP(u) + odd_exp) / 2;  /* exact */

  if (rnd_mode == MPFR_RNDZ || rnd_mode == MPFR_RNDD || sticky == MPFR_LIMB_ZERO)
    {
      inexact = (sticky == MPFR_LIMB_ZERO) ? 0 : -1;
      goto truncate;
    }
  else if (rnd_mode == MPFR_RNDN)
    {
      /* if sh < GMP_NUMB_BITS, the round bit is bit (sh-1) of sticky1
                  and the sticky bit is formed by the low sh-1 bits from
                  sticky1, together with the sqrtrem remainder and sticky0. */
      if (sh < GMP_NUMB_BITS)
        {
          if (sticky1 & (MPFR_LIMB_ONE << (sh - 1)))
            { /* round bit is set */
              if (sticky1 == (MPFR_LIMB_ONE << (sh - 1)) && tsize == 0
                  && sticky0 == 0)
                goto even_rule;
              else
                goto add_one_ulp;
            }
          else /* round bit is zero */
            goto truncate; /* with the default inexact=-1 */
        }
      else /* sh = GMP_NUMB_BITS: the round bit is the most significant bit
              of rp[0], and the remaining GMP_NUMB_BITS-1 bits contribute to
              the sticky bit */
        {
          if (sticky1 & MPFR_LIMB_HIGHBIT)
            { /* round bit is set */
              if (sticky1 == MPFR_LIMB_HIGHBIT && tsize == 0 && sticky0 == 0)
                goto even_rule;
              else
                goto add_one_ulp;
            }
          else /* round bit is zero */
            goto truncate; /* with the default inexact=-1 */
        }
    }
  else /* rnd_mode=GMP_RDNU, necessarily sticky <> 0, thus add 1 ulp */
    goto add_one_ulp;

 even_rule: /* has to set inexact */
  if (sh < GMP_NUMB_BITS)
    inexact = (rp[0] & (MPFR_LIMB_ONE << sh)) ? 1 : -1;
  else
    inexact = (rp[1] & MPFR_LIMB_ONE) ? 1 : -1;
  if (inexact == -1)
    goto truncate;
  /* else go through add_one_ulp */

 add_one_ulp:
  inexact = 1; /* always here */
  if (sh == GMP_NUMB_BITS)
    {
      rp ++;
      rsize --;
      sh = 0;
    }
  if (mpn_add_1 (rp0, rp, rsize, MPFR_LIMB_ONE << sh))
    {
      expr ++;
      rp[rsize - 1] = MPFR_LIMB_HIGHBIT;
    }
  goto end;

 truncate: /* inexact = 0 or -1 */
  if (sh == GMP_NUMB_BITS)
    MPN_COPY (rp0, rp + 1, rsize - 1);

 end:
  MPFR_ASSERTN (expr >= MPFR_EMIN_MIN && expr <= MPFR_EMAX_MAX);
  MPFR_EXP (r) = expr;
  MPFR_TMP_FREE(marker);

  return mpfr_check_range (r, inexact, rnd_mode);
}