int
mpfr_mul (mpfr_ptr a, mpfr_srcptr b, mpfr_srcptr c, mpfr_rnd_t rnd_mode)
{
int sign, inexact;
mpfr_exp_t ax, ax2;
mp_limb_t *tmp;
mp_limb_t b1;
mpfr_prec_t bq, cq;
mp_size_t bn, cn, tn, k;
MPFR_TMP_DECL (marker);
MPFR_LOG_FUNC (("b[%#R]=%R c[%#R]=%R rnd=%d", b, b, c, c, rnd_mode),
("a[%#R]=%R inexact=%d", a, a, inexact));
/* deal with special cases */
if (MPFR_ARE_SINGULAR (b, c))
{
if (MPFR_IS_NAN (b) || MPFR_IS_NAN (c))
{
MPFR_SET_NAN (a);
MPFR_RET_NAN;
}
sign = MPFR_MULT_SIGN (MPFR_SIGN (b), MPFR_SIGN (c));
if (MPFR_IS_INF (b))
{
if (!MPFR_IS_ZERO (c))
{
MPFR_SET_SIGN (a, sign);
MPFR_SET_INF (a);
MPFR_RET (0);
}
else
{
MPFR_SET_NAN (a);
MPFR_RET_NAN;
}
}
else if (MPFR_IS_INF (c))
{
if (!MPFR_IS_ZERO (b))
{
MPFR_SET_SIGN (a, sign);
MPFR_SET_INF (a);
MPFR_RET(0);
}
else
{
MPFR_SET_NAN (a);
MPFR_RET_NAN;
}
}
else
{
MPFR_ASSERTD (MPFR_IS_ZERO(b) || MPFR_IS_ZERO(c));
MPFR_SET_SIGN (a, sign);
MPFR_SET_ZERO (a);
MPFR_RET (0);
}
}
sign = MPFR_MULT_SIGN (MPFR_SIGN (b), MPFR_SIGN (c));
ax = MPFR_GET_EXP (b) + MPFR_GET_EXP (c);
/* Note: the exponent of the exact result will be e = bx + cx + ec with
ec in {-1,0,1} and the following assumes that e is representable. */
/* FIXME: Useful since we do an exponent check after ?
* It is useful iff the precision is big, there is an overflow
* and we are doing further mults...*/
#ifdef HUGE
if (MPFR_UNLIKELY (ax > __gmpfr_emax + 1))
return mpfr_overflow (a, rnd_mode, sign);
if (MPFR_UNLIKELY (ax < __gmpfr_emin - 2))
return mpfr_underflow (a, rnd_mode == MPFR_RNDN ? MPFR_RNDZ : rnd_mode,
sign);
#endif
bq = MPFR_PREC (b);
cq = MPFR_PREC (c);
MPFR_ASSERTD (bq+cq > bq); /* PREC_MAX is /2 so no integer overflow */
bn = (bq+GMP_NUMB_BITS-1)/GMP_NUMB_BITS; /* number of limbs of b */
cn = (cq+GMP_NUMB_BITS-1)/GMP_NUMB_BITS; /* number of limbs of c */
k = bn + cn; /* effective nb of limbs used by b*c (= tn or tn+1) below */
tn = (bq + cq + GMP_NUMB_BITS - 1) / GMP_NUMB_BITS;
MPFR_ASSERTD (tn <= k); /* tn <= k, thus no int overflow */
/* Check for no size_t overflow*/
MPFR_ASSERTD ((size_t) k <= ((size_t) -1) / BYTES_PER_MP_LIMB);
MPFR_TMP_MARK (marker);
tmp = (mp_limb_t *) MPFR_TMP_ALLOC ((size_t) k * BYTES_PER_MP_LIMB);
/* multiplies two mantissa in temporary allocated space */
if (MPFR_UNLIKELY (bn < cn))
{
mpfr_srcptr z = b;
mp_size_t zn = bn;
b = c;
bn = cn;
c = z;
cn = zn;
}
MPFR_ASSERTD (bn >= cn);
if (MPFR_LIKELY (bn <= 2))
{
if (bn == 1)
{
/* 1 limb * 1 limb */
umul_ppmm (tmp[1], tmp[0], MPFR_MANT (b)[0], MPFR_MANT (c)[0]);
b1 = tmp[1];
}
else if (MPFR_UNLIKELY (cn == 1))
{
/* 2 limbs * 1 limb */
mp_limb_t t;
umul_ppmm (tmp[1], tmp[0], MPFR_MANT (b)[0], MPFR_MANT (c)[0]);
umul_ppmm (tmp[2], t, MPFR_MANT (b)[1], MPFR_MANT (c)[0]);
add_ssaaaa (tmp[2], tmp[1], tmp[2], tmp[1], 0, t);
b1 = tmp[2];
}
else
{
/* 2 limbs * 2 limbs */
mp_limb_t t1, t2, t3;
/* First 2 limbs * 1 limb */
umul_ppmm (tmp[1], tmp[0], MPFR_MANT (b)[0], MPFR_MANT (c)[0]);
umul_ppmm (tmp[2], t1, MPFR_MANT (b)[1], MPFR_MANT (c)[0]);
add_ssaaaa (tmp[2], tmp[1], tmp[2], tmp[1], 0, t1);
/* Second, the other 2 limbs * 1 limb product */
umul_ppmm (t1, t2, MPFR_MANT (b)[0], MPFR_MANT (c)[1]);
umul_ppmm (tmp[3], t3, MPFR_MANT (b)[1], MPFR_MANT (c)[1]);
add_ssaaaa (tmp[3], t1, tmp[3], t1, 0, t3);
/* Sum those two partial products */
add_ssaaaa (tmp[2], tmp[1], tmp[2], tmp[1], t1, t2);
tmp[3] += (tmp[2] < t1);
b1 = tmp[3];
}
b1 >>= (GMP_NUMB_BITS - 1);
tmp += k - tn;
if (MPFR_UNLIKELY (b1 == 0))
mpn_lshift (tmp, tmp, tn, 1); /* tn <= k, so no stack corruption */
}
else
/* Mulders' mulhigh. Disable if squaring, since it is not tuned for
such a case */
if (MPFR_UNLIKELY (bn > MPFR_MUL_THRESHOLD && b != c))
MPFR_HOT_FUNCTION_ATTR int
mpfr_cmp3 (mpfr_srcptr b, mpfr_srcptr c, int s)
{
mpfr_exp_t be, ce;
mp_size_t bn, cn;
mp_limb_t *bp, *cp;
s = MPFR_MULT_SIGN( s , MPFR_SIGN(c) );
if (MPFR_ARE_SINGULAR(b, c))
{
if (MPFR_IS_NAN (b) || MPFR_IS_NAN (c))
{
MPFR_SET_ERANGEFLAG ();
return 0;
}
else if (MPFR_IS_INF(b))
{
if (MPFR_IS_INF(c) && s == MPFR_SIGN(b) )
return 0;
else
return MPFR_SIGN(b);
}
else if (MPFR_IS_INF(c))
return -s;
else if (MPFR_IS_ZERO(b))
return MPFR_IS_ZERO(c) ? 0 : -s;
else /* necessarily c=0 */
return MPFR_SIGN(b);
}
/* b and c are real numbers */
if (s != MPFR_SIGN(b))
return MPFR_SIGN(b);
/* now signs are equal */
be = MPFR_GET_EXP (b);
ce = MPFR_GET_EXP (c);
if (be > ce)
return s;
if (be < ce)
return -s;
/* both signs and exponents are equal */
bn = MPFR_LAST_LIMB (b);
cn = MPFR_LAST_LIMB (c);
bp = MPFR_MANT(b);
cp = MPFR_MANT(c);
for ( ; bn >= 0 && cn >= 0; bn--, cn--)
{
if (bp[bn] > cp[cn])
return s;
if (bp[bn] < cp[cn])
return -s;
}
for ( ; bn >= 0; bn--)
if (bp[bn])
return s;
for ( ; cn >= 0; cn--)
if (cp[cn])
return -s;
return 0;
}
static int
mpfr_mul3 (mpfr_ptr a, mpfr_srcptr b, mpfr_srcptr c, mpfr_rnd_t rnd_mode)
{
/* Old implementation */
int sign_product, cc, inexact;
mpfr_exp_t ax;
mp_limb_t *tmp;
mp_limb_t b1;
mpfr_prec_t bq, cq;
mp_size_t bn, cn, tn, k;
MPFR_TMP_DECL(marker);
/* deal with special cases */
if (MPFR_ARE_SINGULAR(b,c))
{
if (MPFR_IS_NAN(b) || MPFR_IS_NAN(c))
{
MPFR_SET_NAN(a);
MPFR_RET_NAN;
}
sign_product = MPFR_MULT_SIGN( MPFR_SIGN(b) , MPFR_SIGN(c) );
if (MPFR_IS_INF(b))
{
if (MPFR_IS_INF(c) || MPFR_NOTZERO(c))
{
MPFR_SET_SIGN(a,sign_product);
MPFR_SET_INF(a);
MPFR_RET(0); /* exact */
}
else
{
MPFR_SET_NAN(a);
MPFR_RET_NAN;
}
}
else if (MPFR_IS_INF(c))
{
if (MPFR_NOTZERO(b))
{
MPFR_SET_SIGN(a, sign_product);
MPFR_SET_INF(a);
MPFR_RET(0); /* exact */
}
else
{
MPFR_SET_NAN(a);
MPFR_RET_NAN;
}
}
else
{
MPFR_ASSERTD(MPFR_IS_ZERO(b) || MPFR_IS_ZERO(c));
MPFR_SET_SIGN(a, sign_product);
MPFR_SET_ZERO(a);
MPFR_RET(0); /* 0 * 0 is exact */
}
}
sign_product = MPFR_MULT_SIGN( MPFR_SIGN(b) , MPFR_SIGN(c) );
ax = MPFR_GET_EXP (b) + MPFR_GET_EXP (c);
bq = MPFR_PREC (b);
cq = MPFR_PREC (c);
MPFR_ASSERTN ((mpfr_uprec_t) bq + cq <= MPFR_PREC_MAX);
bn = MPFR_PREC2LIMBS (bq); /* number of limbs of b */
cn = MPFR_PREC2LIMBS (cq); /* number of limbs of c */
k = bn + cn; /* effective nb of limbs used by b*c (= tn or tn+1) below */
tn = MPFR_PREC2LIMBS (bq + cq);
/* <= k, thus no int overflow */
MPFR_ASSERTD(tn <= k);
/* Check for no size_t overflow*/
MPFR_ASSERTD((size_t) k <= ((size_t) -1) / MPFR_BYTES_PER_MP_LIMB);
MPFR_TMP_MARK(marker);
tmp = MPFR_TMP_LIMBS_ALLOC (k);
/* multiplies two mantissa in temporary allocated space */
b1 = (MPFR_LIKELY(bn >= cn)) ?
mpn_mul (tmp, MPFR_MANT(b), bn, MPFR_MANT(c), cn)
: mpn_mul (tmp, MPFR_MANT(c), cn, MPFR_MANT(b), bn);
/* now tmp[0]..tmp[k-1] contains the product of both mantissa,
with tmp[k-1]>=2^(GMP_NUMB_BITS-2) */
b1 >>= GMP_NUMB_BITS - 1; /* msb from the product */
/* if the mantissas of b and c are uniformly distributed in ]1/2, 1],
then their product is in ]1/4, 1/2] with probability 2*ln(2)-1 ~ 0.386
and in [1/2, 1] with probability 2-2*ln(2) ~ 0.614 */
tmp += k - tn;
if (MPFR_UNLIKELY(b1 == 0))
mpn_lshift (tmp, tmp, tn, 1); /* tn <= k, so no stack corruption */
cc = mpfr_round_raw (MPFR_MANT (a), tmp, bq + cq,
MPFR_IS_NEG_SIGN(sign_product),
MPFR_PREC (a), rnd_mode, &inexact);
/* cc = 1 ==> result is a power of two */
if (MPFR_UNLIKELY(cc))
MPFR_MANT(a)[MPFR_LIMB_SIZE(a)-1] = MPFR_LIMB_HIGHBIT;
MPFR_TMP_FREE(marker);
{
mpfr_exp_t ax2 = ax + (mpfr_exp_t) (b1 - 1 + cc);
if (MPFR_UNLIKELY( ax2 > __gmpfr_emax))
return mpfr_overflow (a, rnd_mode, sign_product);
if (MPFR_UNLIKELY( ax2 < __gmpfr_emin))
{
/* In the rounding to the nearest mode, if the exponent of the exact
result (i.e. before rounding, i.e. without taking cc into account)
is < __gmpfr_emin - 1 or the exact result is a power of 2 (i.e. if
both arguments are powers of 2) in absolute value, then round to
zero. */
if (rnd_mode == MPFR_RNDN &&
(ax + (mpfr_exp_t) b1 < __gmpfr_emin ||
(mpfr_powerof2_raw (b) && mpfr_powerof2_raw (c))))
rnd_mode = MPFR_RNDZ;
return mpfr_underflow (a, rnd_mode, sign_product);
}
MPFR_SET_EXP (a, ax2);
MPFR_SET_SIGN(a, sign_product);
}
MPFR_RET (inexact);
}
static void
underflowed_cothinf (void)
{
mpfr_t x, y;
int i, inex, rnd, err = 0;
mpfr_exp_t old_emin;
old_emin = mpfr_get_emin ();
mpfr_init2 (x, 8);
mpfr_init2 (y, 8);
for (i = -1; i <= 1; i += 2)
RND_LOOP (rnd)
{
mpfr_set_inf (x, i);
mpfr_clear_flags ();
set_emin (2); /* 1 is not representable. */
inex = mpfr_coth (x, x, (mpfr_rnd_t) rnd);
set_emin (old_emin);
if (! mpfr_underflow_p ())
{
printf ("Error in underflowed_cothinf (i = %d, rnd = %s):\n"
" The underflow flag is not set.\n",
i, mpfr_print_rnd_mode ((mpfr_rnd_t) rnd));
err = 1;
}
mpfr_set_si (y, (i < 0 && (rnd == MPFR_RNDD || rnd == MPFR_RNDA)) ||
(i > 0 && (rnd == MPFR_RNDU || rnd == MPFR_RNDA))
? 2 : 0, MPFR_RNDN);
if (i < 0)
mpfr_neg (y, y, MPFR_RNDN);
if (! (mpfr_equal_p (x, y) &&
MPFR_MULT_SIGN (MPFR_SIGN (x), MPFR_SIGN (y)) > 0))
{
printf ("Error in underflowed_cothinf (i = %d, rnd = %s):\n"
" Got ", i, mpfr_print_rnd_mode ((mpfr_rnd_t) rnd));
mpfr_print_binary (x);
printf (" instead of ");
mpfr_print_binary (y);
printf (".\n");
err = 1;
}
if ((rnd == MPFR_RNDD ||
(i > 0 && (rnd == MPFR_RNDN || rnd == MPFR_RNDZ))) && inex >= 0)
{
printf ("Error in underflowed_cothinf (i = %d, rnd = %s):\n"
" The inexact value must be negative.\n",
i, mpfr_print_rnd_mode ((mpfr_rnd_t) rnd));
err = 1;
}
if ((rnd == MPFR_RNDU ||
(i < 0 && (rnd == MPFR_RNDN || rnd == MPFR_RNDZ))) && inex <= 0)
{
printf ("Error in underflowed_cothinf (i = %d, rnd = %s):\n"
" The inexact value must be positive.\n",
i, mpfr_print_rnd_mode ((mpfr_rnd_t) rnd));
err = 1;
}
}
if (err)
exit (1);
mpfr_clear (x);
mpfr_clear (y);
}
