path: root/src/rec_sqrt.c

/* mpfr_rec_sqrt -- inverse square root

Copyright 2008, 2009, 2010, 2011 Free Software Foundation, Inc.
Contributed by the Arenaire 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 <stdio.h>
#include <stdlib.h>

#define MPFR_NEED_LONGLONG_H /* for umul_ppmm */
#include "mpfr-impl.h"

#define LIMB_SIZE(x) ((((x)-1)>>MPFR_LOG2_GMP_NUMB_BITS) + 1)

#define MPFR_COM_N(x,y,n)                               \
  {                                                     \
    mp_size_t i;                                        \
    for (i = 0; i < n; i++)                             \
      *((x)+i) = ~*((y)+i);                             \
  }

/* Put in X a p-bit approximation of 1/sqrt(A),
   where X = {x, n}/B^n, n = ceil(p/GMP_NUMB_BITS),
   A = 2^(1+as)*{a, an}/B^an, as is 0 or 1, an = ceil(ap/GMP_NUMB_BITS),
   where B = 2^GMP_NUMB_BITS.

   We have 1 <= A < 4 and 1/2 <= X < 1.

   The error in the approximate result with respect to the true
   value 1/sqrt(A) is bounded by 1 ulp(X), i.e., 2^{-p} since 1/2 <= X < 1.

   Note: x and a are left-aligned, i.e., the most significant bit of
   a[an-1] is set, and so is the most significant bit of the output x[n-1].

   If p is not a multiple of GMP_NUMB_BITS, the extra low bits of the input
   A are taken into account to compute the approximation of 1/sqrt(A), but
   whether or not they are zero, the error between X and 1/sqrt(A) is bounded
   by 1 ulp(X) [in precision p].
   The extra low bits of the output X (if p is not a multiple of GMP_NUMB_BITS)
   are set to 0.

   Assumptions:
   (1) A should be normalized, i.e., the most significant bit of a[an-1]
       should be 1. If as=0, we have 1 <= A < 2; if as=1, we have 2 <= A < 4.
   (2) p >= 12
   (3) {a, an} and {x, n} should not overlap
   (4) GMP_NUMB_BITS >= 12 and is even

   Note: this routine is much more efficient when ap is small compared to p,
   including the case where ap <= GMP_NUMB_BITS, thus it can be used to
   implement an efficient mpfr_rec_sqrt_ui function.

   Reference: Modern Computer Algebra, Richard Brent and Paul Zimmermann,
   http://www.loria.fr/~zimmerma/mca/pub226.html
*/
static void
mpfr_mpn_rec_sqrt (mpfr_limb_ptr x, mpfr_prec_t p,
                   mpfr_limb_srcptr a, mpfr_prec_t ap, int as)

{
  /* the following T1 and T2 are bipartite tables giving initial
     approximation for the inverse square root, with 13-bit input split in
     5+4+4, and 11-bit output. More precisely, if 2048 <= i < 8192,
     with i = a*2^8 + b*2^4 + c, we use for approximation of
     2048/sqrt(i/2048) the value x = T1[16*(a-8)+b] + T2[16*(a-8)+c].
     The largest error is obtained for i = 2054, where x = 2044,
     and 2048/sqrt(i/2048) = 2045.006576...
  */
  static short int T1[384] = {
2040, 2033, 2025, 2017, 2009, 2002, 1994, 1987, 1980, 1972, 1965, 1958, 1951,
1944, 1938, 1931, /* a=8 */
1925, 1918, 1912, 1905, 1899, 1892, 1886, 1880, 1874, 1867, 1861, 1855, 1849,
1844, 1838, 1832, /* a=9 */
1827, 1821, 1815, 1810, 1804, 1799, 1793, 1788, 1783, 1777, 1772, 1767, 1762,
1757, 1752, 1747, /* a=10 */
1742, 1737, 1733, 1728, 1723, 1718, 1713, 1709, 1704, 1699, 1695, 1690, 1686,
1681, 1677, 1673, /* a=11 */
1669, 1664, 1660, 1656, 1652, 1647, 1643, 1639, 1635, 1631, 1627, 1623, 1619,
1615, 1611, 1607, /* a=12 */
1603, 1600, 1596, 1592, 1588, 1585, 1581, 1577, 1574, 1570, 1566, 1563, 1559,
1556, 1552, 1549, /* a=13 */
1545, 1542, 1538, 1535, 1532, 1528, 1525, 1522, 1518, 1515, 1512, 1509, 1505,
1502, 1499, 1496, /* a=14 */
1493, 1490, 1487, 1484, 1481, 1478, 1475, 1472, 1469, 1466, 1463, 1460, 1457,
1454, 1451, 1449, /* a=15 */
1446, 1443, 1440, 1438, 1435, 1432, 1429, 1427, 1424, 1421, 1419, 1416, 1413,
1411, 1408, 1405, /* a=16 */
1403, 1400, 1398, 1395, 1393, 1390, 1388, 1385, 1383, 1380, 1378, 1375, 1373,
1371, 1368, 1366, /* a=17 */
1363, 1360, 1358, 1356, 1353, 1351, 1349, 1346, 1344, 1342, 1340, 1337, 1335,
1333, 1331, 1329, /* a=18 */
1327, 1325, 1323, 1321, 1319, 1316, 1314, 1312, 1310, 1308, 1306, 1304, 1302,
1300, 1298, 1296, /* a=19 */
1294, 1292, 1290, 1288, 1286, 1284, 1282, 1280, 1278, 1276, 1274, 1272, 1270,
1268, 1266, 1265, /* a=20 */
1263, 1261, 1259, 1257, 1255, 1253, 1251, 1250, 1248, 1246, 1244, 1242, 1241,
1239, 1237, 1235, /* a=21 */
1234, 1232, 1230, 1229, 1227, 1225, 1223, 1222, 1220, 1218, 1217, 1215, 1213,
1212, 1210, 1208, /* a=22 */
1206, 1204, 1203, 1201, 1199, 1198, 1196, 1195, 1193, 1191, 1190, 1188, 1187,
1185, 1184, 1182, /* a=23 */
1181, 1180, 1178, 1177, 1175, 1174, 1172, 1171, 1169, 1168, 1166, 1165, 1163,
1162, 1160, 1159, /* a=24 */
1157, 1156, 1154, 1153, 1151, 1150, 1149, 1147, 1146, 1144, 1143, 1142, 1140,
1139, 1137, 1136, /* a=25 */
1135, 1133, 1132, 1131, 1129, 1128, 1127, 1125, 1124, 1123, 1121, 1120, 1119,
1117, 1116, 1115, /* a=26 */
1114, 1113, 1111, 1110, 1109, 1108, 1106, 1105, 1104, 1103, 1101, 1100, 1099,
1098, 1096, 1095, /* a=27 */
1093, 1092, 1091, 1090, 1089, 1087, 1086, 1085, 1084, 1083, 1081, 1080, 1079,
1078, 1077, 1076, /* a=28 */
1075, 1073, 1072, 1071, 1070, 1069, 1068, 1067, 1065, 1064, 1063, 1062, 1061,
1060, 1059, 1058, /* a=29 */
1057, 1056, 1055, 1054, 1052, 1051, 1050, 1049, 1048, 1047, 1046, 1045, 1044,
1043, 1042, 1041, /* a=30 */
1040, 1039, 1038, 1037, 1036, 1035, 1034, 1033, 1032, 1031, 1030, 1029, 1028,
1027, 1026, 1025 /* a=31 */
};
  static unsigned char T2[384] = {
    7, 7, 6, 6, 5, 5, 4, 4, 4, 3, 3, 2, 2, 1, 1, 0, /* a=8 */
    6, 5, 5, 5, 4, 4, 3, 3, 3, 2, 2, 2, 1, 1, 0, 0, /* a=9 */
    5, 5, 4, 4, 4, 3, 3, 3, 2, 2, 2, 1, 1, 1, 0, 0, /* a=10 */
    4, 4, 3, 3, 3, 3, 2, 2, 2, 1, 1, 1, 1, 0, 0, 0, /* a=11 */
    3, 3, 3, 3, 2, 2, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, /* a=12 */
    3, 3, 3, 2, 2, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, /* a=13 */
    3, 3, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 0, 0, 0, 0, /* a=14 */
    2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, /* a=15 */
    2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, /* a=16 */
    2, 2, 2, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, /* a=17 */
    3, 2, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, 0, /* a=18 */
    2, 2, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, /* a=19 */
    1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, /* a=20 */
    1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, /* a=21 */
    1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, /* a=22 */
    2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, /* a=23 */
    1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, /* a=24 */
    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, /* a=25 */
    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, /* a=26 */
    1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, /* a=27 */
    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, /* a=28 */
    1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, /* a=29 */
    1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, /* a=30 */
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0  /* a=31 */
};
  mp_size_t n = LIMB_SIZE(p);   /* number of limbs of X */
  mp_size_t an = LIMB_SIZE(ap); /* number of limbs of A */

  /* A should be normalized */
  MPFR_ASSERTD((a[an - 1] & MPFR_LIMB_HIGHBIT) != 0);
  /* We should have enough bits in one limb and GMP_NUMB_BITS should be even.
     Since that does not depend on MPFR, we always check this. */
  MPFR_ASSERTN((GMP_NUMB_BITS >= 12) && ((GMP_NUMB_BITS & 1) == 0));
  /* {a, an} and {x, n} should not overlap */
  MPFR_ASSERTD((a + an <= x) || (x + n <= a));
  MPFR_ASSERTD(p >= 11);

  if (MPFR_UNLIKELY(an > n)) /* we can cut the input to n limbs */
    {
      a += an - n;
      an = n;
    }

  if (p == 11) /* should happen only from recursive calls */
    {
      unsigned long i, ab, ac;
      mp_limb_t t;

      /* take the 12+as most significant bits of A */
      i = a[an - 1] >> (GMP_NUMB_BITS - (12 + as));
      /* if one wants faithful rounding for p=11, replace #if 0 by #if 1 */
      ab = i >> 4;
      ac = (ab & 0x3F0) | (i & 0x0F);
      t = (mp_limb_t) T1[ab - 0x80] + (mp_limb_t) T2[ac - 0x80];
      x[0] = t << (GMP_NUMB_BITS - p);
    }
  else /* p >= 12 */
    {
      mpfr_prec_t h, pl;
      mpfr_limb_ptr r, s, t, u;
      mp_size_t xn, rn, th, ln, tn, sn, ahn, un;
      mp_limb_t neg, cy, cu;
      MPFR_TMP_DECL(marker);

      /* h = max(11, ceil((p+3)/2)) is the bitsize of the recursive call */
      h = (p < 18) ? 11 : (p >> 1) + 2;

      xn = LIMB_SIZE(h);       /* limb size of the recursive Xh */
      rn = LIMB_SIZE(2 * h);   /* a priori limb size of Xh^2 */
      ln = n - xn;             /* remaining limbs to be computed */

      /* Since |Xh - A^{-1/2}| <= 2^{-h}, then by multiplying by Xh + A^{-1/2}
         we get |Xh^2 - 1/A| <= 2^{-h+1}, thus |A*Xh^2 - 1| <= 2^{-h+3},
         thus the h-3 most significant bits of t should be zero,
         which is in fact h+1+as-3 because of the normalization of A.
         This corresponds to th=floor((h+1+as-3)/GMP_NUMB_BITS) limbs. */
      th = (h + 1 + as - 3) >> MPFR_LOG2_GMP_NUMB_BITS;
      tn = LIMB_SIZE(2 * h + 1 + as);

      /* we need h+1+as bits of a */
      ahn = LIMB_SIZE(h + 1 + as); /* number of high limbs of A
                                      needed for the recursive call*/
      if (MPFR_UNLIKELY(ahn > an))
        ahn = an;
      mpfr_mpn_rec_sqrt (x + ln, h, a + an - ahn, ahn * GMP_NUMB_BITS, as);
      /* the most h significant bits of X are set, X has ceil(h/GMP_NUMB_BITS)
         limbs, the low (-h) % GMP_NUMB_BITS bits are zero */

      MPFR_TMP_MARK (marker);
      /* first step: square X in r, result is exact */
      un = xn + (tn - th);
      /* We use the same temporary buffer to store r and u: r needs 2*xn
         limbs where u needs xn+(tn-th) limbs. Since tn can store at least
         2h bits, and th at most h bits, then tn-th can store at least h bits,
         thus tn - th >= xn, and reserving the space for u is enough. */
      MPFR_ASSERTD(2 * xn <= un);
      u = r = MPFR_TMP_LIMBS_ALLOC (un);
      if (2 * h <= GMP_NUMB_BITS) /* xn=rn=1, and since p <= 2h-3, n=1,
                                     thus ln = 0 */
        {
          MPFR_ASSERTD(ln == 0);
          cy = x[0] >> (GMP_NUMB_BITS >> 1);
          r ++;
          r[0] = cy * cy;
        }
      else if (xn == 1) /* xn=1, rn=2 */
        umul_ppmm(r[1], r[0], x[ln], x[ln]);
      else
        {
          mpn_mul_n (r, x + ln, x + ln, xn);
          if (rn < 2 * xn)
            r ++;
        }
      /* now the 2h most significant bits of {r, rn} contains X^2, r has rn
         limbs, and the low (-2h) % GMP_NUMB_BITS bits are zero */

      /* Second step: s <- A * (r^2), and truncate the low ap bits,
         i.e., at weight 2^{-2h} (s is aligned to the low significant bits)
       */
      sn = an + rn;
      s = MPFR_TMP_LIMBS_ALLOC (sn);
      if (rn == 1) /* rn=1 implies n=1, since rn*GMP_NUMB_BITS >= 2h,
                           and 2h >= p+3 */
        {
          /* necessarily p <= GMP_NUMB_BITS-3: we can ignore the two low
             bits from A */
          /* since n=1, and we ensured an <= n, we also have an=1 */
          MPFR_ASSERTD(an == 1);
          umul_ppmm (s[1], s[0], r[0], a[0]);
        }
      else
        {
          /* we have p <= n * GMP_NUMB_BITS
             2h <= rn * GMP_NUMB_BITS with p+3 <= 2h <= p+4
             thus n <= rn <= n + 1 */
          MPFR_ASSERTD(rn <= n + 1);
          /* since we ensured an <= n, we have an <= rn */
          MPFR_ASSERTD(an <= rn);
          mpn_mul (s, r, rn, a, an);
          /* s should be near B^sn/2^(1+as), thus s[sn-1] is either
             100000... or 011111... if as=0, or
             010000... or 001111... if as=1.
             We ignore the bits of s after the first 2h+1+as ones.
          */
        }

      /* We ignore the bits of s after the first 2h+1+as ones: s has rn + an
         limbs, where rn = LIMBS(2h), an=LIMBS(a), and tn = LIMBS(2h+1+as). */
      t = s + sn - tn; /* pointer to low limb of the high part of t */
      /* the upper h-3 bits of 1-t should be zero,
         where 1 corresponds to the most significant bit of t[tn-1] if as=0,
         and to the 2nd most significant bit of t[tn-1] if as=1 */

      /* compute t <- 1 - t, which is B^tn - {t, tn+1},
         with rounding toward -Inf, i.e., rounding the input t toward +Inf.
         We could only modify the low tn - th limbs from t, but it gives only
         a small speedup, and would make the code more complex.
      */
      neg = t[tn - 1] & (MPFR_LIMB_HIGHBIT >> as);
      if (neg == 0) /* Ax^2 < 1: we have t = th + eps, where 0 <= eps < ulp(th)
                       is the part truncated above, thus 1 - t rounded to -Inf
                       is 1 - th - ulp(th) */
        {
          /* since the 1+as most significant bits of t are zero, set them
             to 1 before the one-complement */
          t[tn - 1] |= MPFR_LIMB_HIGHBIT | (MPFR_LIMB_HIGHBIT >> as);
          MPFR_COM_N (t, t, tn);
          /* we should add 1 here to get 1-th complement, and subtract 1 for
             -ulp(th), thus we do nothing */
        }
      else /* negative case: we want 1 - t rounded toward -Inf, i.e.,
              th + eps rounded toward +Inf, which is th + ulp(th):
              we discard the bit corresponding to 1,
              and we add 1 to the least significant bit of t */
        {
          t[tn - 1] ^= neg;
          mpn_add_1 (t, t, tn, 1);
        }
      tn -= th; /* we know at least th = floor((h+1+as-3)/GMP_NUMB_LIMBS) of
                   the high limbs of {t, tn} are zero */

      /* tn = rn - th, where rn * GMP_NUMB_BITS >= 2*h and
         th * GMP_NUMB_BITS <= h+1+as-3, thus tn > 0 */
      MPFR_ASSERTD(tn > 0);

      /* u <- x * t, where {t, tn} contains at least h+3 bits,
         and {x, xn} contains h bits, thus tn >= xn */
      MPFR_ASSERTD(tn >= xn);
      if (tn == 1) /* necessarily xn=1 */
        umul_ppmm (u[1], u[0], t[0], x[ln]);
      else
        mpn_mul (u, t, tn, x + ln, xn);

      /* we have already discarded the upper th high limbs of t, thus we only
         have to consider the upper n - th limbs of u */
      un = n - th; /* un cannot be zero, since p <= n*GMP_NUMB_BITS,
                      h = ceil((p+3)/2) <= (p+4)/2,
                      th*GMP_NUMB_BITS <= h-1 <= p/2+1,
                      thus (n-th)*GMP_NUMB_BITS >= p/2-1.
                   */
      MPFR_ASSERTD(un > 0);
      u += (tn + xn) - un; /* xn + tn - un = xn + (original_tn - th) - (n - th)
                                           = xn + original_tn - n
                              = LIMBS(h) + LIMBS(2h+1+as) - LIMBS(p) > 0
                              since 2h >= p+3 */
      MPFR_ASSERTD(tn + xn > un); /* will allow to access u[-1] below */

      /* In case as=0, u contains |x*(1-Ax^2)/2|, which is exactly what we
         need to add or subtract.
         In case as=1, u contains |x*(1-Ax^2)/4|, thus we need to multiply
         u by 2. */

      if (as == 1)
        /* shift on un+1 limbs to get most significant bit of u[-1] into
           least significant bit of u[0] */
        mpn_lshift (u - 1, u - 1, un + 1, 1);

      pl = n * GMP_NUMB_BITS - p;       /* low bits from x */
      /* We want that the low pl bits are zero after rounding to nearest,
         thus we round u to nearest at bit pl-1 of u[0] */
      if (pl > 0)
        {
          cu = mpn_add_1 (u, u, un, u[0] & (MPFR_LIMB_ONE << (pl - 1)));
          /* mask bits 0..pl-1 of u[0] */
          u[0] &= ~MPFR_LIMB_MASK(pl);
        }
      else /* round bit is in u[-1] */
        cu = mpn_add_1 (u, u, un, u[-1] >> (GMP_NUMB_BITS - 1));

      /* We already have filled {x + ln, xn = n - ln}, and we want to add or
         subtract cu*B^un + {u, un} at position x.
         un = n - th, where th contains <= h+1+as-3<=h-1 bits
         ln = n - xn, where xn contains >= h bits
         thus un > ln.
         Warning: ln might be zero.
      */
      MPFR_ASSERTD(un > ln);
      /* we can have un = ln + 2, for example with GMP_NUMB_BITS=32 and
         p=62, as=0, then h=33, n=2, th=0, xn=2, thus un=2 and ln=0. */
      MPFR_ASSERTD(un == ln + 1 || un == ln + 2);
      /* the high un-ln limbs of u will overlap the low part of {x+ln,xn},
         we need to add or subtract the overlapping part {u + ln, un - ln} */
      if (neg == 0)
        {
          if (ln > 0)
            MPN_COPY (x, u, ln);
          cy = mpn_add (x + ln, x + ln, xn, u + ln, un - ln);
          /* add cu at x+un */
          cy += mpn_add_1 (x + un, x + un, th, cu);
        }
      else /* negative case */
        {
          /* subtract {u+ln, un-ln} from {x+ln,un} */
          cy = mpn_sub (x + ln, x + ln, xn, u + ln, un - ln);
          /* carry cy is at x+un, like cu */
          cy = mpn_sub_1 (x + un, x + un, th, cy + cu); /* n - un = th */
          /* cy cannot be zero, since the most significant bit of Xh is 1,
             and the correction is bounded by 2^{-h+3} */
          MPFR_ASSERTD(cy == 0);
          if (ln > 0)
            {
              MPFR_COM_N (x, u, ln);
              /* we must add one for the 2-complement ... */
              cy = mpn_add_1 (x, x, n, MPFR_LIMB_ONE);
              /* ... and subtract 1 at x[ln], where n = ln + xn */
              cy -= mpn_sub_1 (x + ln, x + ln, xn, MPFR_LIMB_ONE);
            }
        }

      /* cy can be 1 when A=1, i.e., {a, n} = B^n. In that case we should
         have X = B^n, and setting X to 1-2^{-p} satisties the error bound
         of 1 ulp. */
      if (MPFR_UNLIKELY(cy != 0))
        {
          cy -= mpn_sub_1 (x, x, n, MPFR_LIMB_ONE << pl);
          MPFR_ASSERTD(cy == 0);
        }

      MPFR_TMP_FREE (marker);
    }
}

int
mpfr_rec_sqrt (mpfr_ptr r, mpfr_srcptr u, mpfr_rnd_t rnd_mode)
{
  mpfr_prec_t rp, up, wp;
  mp_size_t rn, wn;
  int s, cy, inex;
  mpfr_limb_ptr x;
  MPFR_TMP_DECL(marker);

  MPFR_LOG_FUNC (("x[%#R]=%R rnd=%d", u, u, rnd_mode),
                 ("y[%#R]=%R inexact=%d", r, r, inex));

  /* special values */
  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)) /* 1/sqrt(+0) = 1/sqrt(-0) = +Inf */
        {
          /* 0+ or 0- */
          MPFR_SET_INF(r);
          MPFR_SET_POS(r);
          mpfr_set_divby0 ();
          MPFR_RET(0); /* Inf is exact */
        }
      else
        {
          MPFR_ASSERTD(MPFR_IS_INF(u));
          /* 1/sqrt(-Inf) = NAN */
          if (MPFR_IS_NEG(u))
            {
              MPFR_SET_NAN(r);
              MPFR_RET_NAN;
            }
          /* 1/sqrt(+Inf) = +0 */
          MPFR_SET_POS(r);
          MPFR_SET_ZERO(r);
          MPFR_RET(0);
        }
    }

  /* if u < 0, 1/sqrt(u) is NaN */
  if (MPFR_UNLIKELY(MPFR_IS_NEG(u)))
    {
      MPFR_SET_NAN(r);
      MPFR_RET_NAN;
    }

  MPFR_SET_POS(r);

  rp = MPFR_PREC(r); /* output precision */
  up = MPFR_PREC(u); /* input precision */
  wp = rp + 11;      /* initial working precision */

  /* Let u = U*2^e, where e = EXP(u), and 1/2 <= U < 1.
     If e is even, we compute an approximation of X of (4U)^{-1/2},
     and the result is X*2^(-(e-2)/2) [case s=1].
     If e is odd, we compute an approximation of X of (2U)^{-1/2},
     and the result is X*2^(-(e-1)/2) [case s=0]. */

  /* parity of the exponent of u */
  s = 1 - ((mpfr_uexp_t) MPFR_GET_EXP (u) & 1);

  rn = LIMB_SIZE(rp);

  /* for the first iteration, if rp + 11 fits into rn limbs, we round up
     up to a full limb to maximize the chance of rounding, while avoiding
     to allocate extra space */
  wp = rp + 11;
  if (wp < rn * GMP_NUMB_BITS)
    wp = rn * GMP_NUMB_BITS;
  for (;;)
    {
      MPFR_TMP_MARK (marker);
      wn = LIMB_SIZE(wp);
      if (r == u || wn > rn) /* out of place, i.e., we cannot write to r */
        x = MPFR_TMP_LIMBS_ALLOC (wn);
      else
        x = MPFR_MANT(r);
      mpfr_mpn_rec_sqrt (x, wp, MPFR_MANT(u), up, s);
      /* If the input was not truncated, the error is at most one ulp;
         if the input was truncated, the error is at most two ulps
         (see algorithms.tex). */
      if (MPFR_LIKELY (mpfr_round_p (x, wn, wp - (wp < up),
                                     rp + (rnd_mode == MPFR_RNDN))))
        break;

      /* We detect only now the exact case where u=2^(2e), to avoid
         slowing down the average case. This can happen only when the
         mantissa is exactly 1/2 and the exponent is odd. */
      if (s == 0 && mpfr_cmp_ui_2exp (u, 1, MPFR_EXP(u) - 1) == 0)
        {
          mpfr_prec_t pl = wn * GMP_NUMB_BITS - wp;

          /* we should have x=111...111 */
          mpn_add_1 (x, x, wn, MPFR_LIMB_ONE << pl);
          x[wn - 1] = MPFR_LIMB_HIGHBIT;
          s += 2;
          break; /* go through */
        }
      MPFR_TMP_FREE(marker);

      wp += GMP_NUMB_BITS;
    }
  cy = mpfr_round_raw (MPFR_MANT(r), x, wp, 0, rp, rnd_mode, &inex);
  MPFR_EXP(r) = - (MPFR_EXP(u) - 1 - s) / 2;
  if (MPFR_UNLIKELY(cy != 0))
    {
      MPFR_EXP(r) ++;
      MPFR_MANT(r)[rn - 1] = MPFR_LIMB_HIGHBIT;
    }
  MPFR_TMP_FREE(marker);
  return mpfr_check_range (r, inex, rnd_mode);
}