int
mpfr_max (mpfr_ptr z, mpfr_srcptr x, mpfr_srcptr y, mpfr_rnd_t rnd_mode)
{
if (MPFR_ARE_SINGULAR(x,y))
{
if (MPFR_IS_NAN(x) && MPFR_IS_NAN(y) )
{
MPFR_SET_NAN(z);
MPFR_RET_NAN;
}
else if (MPFR_IS_NAN(x))
return mpfr_set(z, y, rnd_mode);
else if (MPFR_IS_NAN(y))
return mpfr_set(z, x, rnd_mode);
else if (MPFR_IS_ZERO(x) && MPFR_IS_ZERO(y))
{
if (MPFR_IS_NEG(x))
return mpfr_set(z, y, rnd_mode);
else
return mpfr_set(z, x, rnd_mode);
}
}
if (mpfr_cmp(x,y) <= 0)
return mpfr_set(z, y, rnd_mode);
else
return mpfr_set(z, x, rnd_mode);
}
static int
test_add (mpfr_ptr a, mpfr_srcptr b, mpfr_srcptr c, mpfr_rnd_t rnd_mode)
{
int res;
#ifdef CHECK_EXTERNAL
int ok = rnd_mode == MPFR_RNDN && mpfr_number_p (b) && mpfr_number_p (c);
if (ok)
{
mpfr_print_raw (b);
printf (" ");
mpfr_print_raw (c);
}
#endif
if (usesp || MPFR_ARE_SINGULAR(b,c) || MPFR_SIGN(b) != MPFR_SIGN(c))
res = mpfr_add (a, b, c, rnd_mode);
else
{
if (MPFR_GET_EXP(b) < MPFR_GET_EXP(c))
res = mpfr_add1(a, c, b, rnd_mode);
else
res = mpfr_add1(a, b, c, rnd_mode);
}
#ifdef CHECK_EXTERNAL
if (ok)
{
printf (" ");
mpfr_print_raw (a);
printf ("\n");
}
#endif
return res;
}
int
mpfr_cmpabs (mpfr_srcptr b, mpfr_srcptr c)
{
mp_exp_t be, ce;
mp_size_t bn, cn;
mp_limb_t *bp, *cp;
if (MPFR_ARE_SINGULAR (b, c))
{
if (MPFR_IS_NAN (b) || MPFR_IS_NAN (c))
{
MPFR_SET_ERANGE ();
return 0;
}
else if (MPFR_IS_INF (b))
return ! MPFR_IS_INF (c);
else if (MPFR_IS_INF (c))
return -1;
else if (MPFR_IS_ZERO (c))
return ! MPFR_IS_ZERO (b);
else /* b == 0 */
return -1;
}
be = MPFR_GET_EXP (b);
ce = MPFR_GET_EXP (c);
if (be > ce)
return 1;
if (be < ce)
return -1;
/* exponents are equal */
bn = MPFR_LIMB_SIZE(b)-1;
cn = MPFR_LIMB_SIZE(c)-1;
bp = MPFR_MANT(b);
cp = MPFR_MANT(c);
for ( ; bn >= 0 && cn >= 0; bn--, cn--)
{
if (bp[bn] > cp[cn])
return 1;
if (bp[bn] < cp[cn])
return -1;
}
for ( ; bn >= 0; bn--)
if (bp[bn])
return 1;
for ( ; cn >= 0; cn--)
if (cp[cn])
return -1;
return 0;
}
/* reldiff(b, c) = abs(b-c)/b */
void
mpfr_reldiff (mpfr_ptr a, mpfr_srcptr b, mpfr_srcptr c, mpfr_rnd_t rnd_mode)
{
mpfr_t b_copy;
if (MPFR_ARE_SINGULAR (b, c))
{
if (MPFR_IS_NAN(b) || MPFR_IS_NAN(c))
{
MPFR_SET_NAN(a);
return;
}
else if (MPFR_IS_INF(b))
{
if (MPFR_IS_INF (c) && (MPFR_SIGN (c) == MPFR_SIGN (b)))
MPFR_SET_ZERO(a);
else
MPFR_SET_NAN(a);
return;
}
else if (MPFR_IS_INF(c))
{
MPFR_SET_SAME_SIGN (a, b);
MPFR_SET_INF (a);
return;
}
else if (MPFR_IS_ZERO(b)) /* reldiff = abs(c)/c = sign(c) */
{
mpfr_set_si (a, MPFR_INT_SIGN (c), rnd_mode);
return;
}
/* Fall through */
}
if (a == b)
{
mpfr_init2 (b_copy, MPFR_PREC(b));
mpfr_set (b_copy, b, MPFR_RNDN);
}
mpfr_sub (a, b, c, rnd_mode);
mpfr_abs (a, a, rnd_mode); /* for compatibility with MPF */
mpfr_div (a, a, (a == b) ? b_copy : b, rnd_mode);
if (a == b)
mpfr_clear (b_copy);
}
/* agm(x,y) is between x and y, so we don't need to save exponent range */
int
mpfr_agm (mpfr_ptr r, mpfr_srcptr op2, mpfr_srcptr op1, mp_rnd_t rnd_mode)
{
int compare, inexact;
mp_size_t s;
mp_prec_t p, q;
mp_limb_t *up, *vp, *tmpp;
mpfr_t u, v, tmp;
unsigned long n; /* number of iterations */
unsigned long err = 0;
MPFR_ZIV_DECL (loop);
MPFR_TMP_DECL(marker);
MPFR_LOG_FUNC (("op2[%#R]=%R op1[%#R]=%R rnd=%d", op2,op2,op1,op1,rnd_mode),
("r[%#R]=%R inexact=%d", r, r, inexact));
/* Deal with special values */
if (MPFR_ARE_SINGULAR (op1, op2))
{
/* If a or b is NaN, the result is NaN */
if (MPFR_IS_NAN(op1) || MPFR_IS_NAN(op2))
{
MPFR_SET_NAN(r);
MPFR_RET_NAN;
}
/* now one of a or b is Inf or 0 */
/* If a and b is +Inf, the result is +Inf.
Otherwise if a or b is -Inf or 0, the result is NaN */
else if (MPFR_IS_INF(op1) || MPFR_IS_INF(op2))
{
if (MPFR_IS_STRICTPOS(op1) && MPFR_IS_STRICTPOS(op2))
{
MPFR_SET_INF(r);
MPFR_SET_SAME_SIGN(r, op1);
MPFR_RET(0); /* exact */
}
else
{
MPFR_SET_NAN(r);
MPFR_RET_NAN;
}
}
else /* a and b are neither NaN nor Inf, and one is zero */
{ /* If a or b is 0, the result is +0 since a sqrt is positive */
MPFR_ASSERTD (MPFR_IS_ZERO (op1) || MPFR_IS_ZERO (op2));
MPFR_SET_POS (r);
MPFR_SET_ZERO (r);
MPFR_RET (0); /* exact */
}
}
MPFR_CLEAR_FLAGS (r);
/* If a or b is negative (excluding -Infinity), the result is NaN */
if (MPFR_UNLIKELY(MPFR_IS_NEG(op1) || MPFR_IS_NEG(op2)))
{
MPFR_SET_NAN(r);
MPFR_RET_NAN;
}
/* Precision of the following calculus */
q = MPFR_PREC(r);
p = q + MPFR_INT_CEIL_LOG2(q) + 15;
MPFR_ASSERTD (p >= 7); /* see algorithms.tex */
s = (p - 1) / BITS_PER_MP_LIMB + 1;
/* b (op2) and a (op1) are the 2 operands but we want b >= a */
compare = mpfr_cmp (op1, op2);
if (MPFR_UNLIKELY( compare == 0 ))
{
mpfr_set (r, op1, rnd_mode);
MPFR_RET (0); /* exact */
}
else if (compare > 0)
{
mpfr_srcptr t = op1;
op1 = op2;
op2 = t;
}
/* Now b(=op2) >= a (=op1) */
MPFR_TMP_MARK(marker);
/* Main loop */
MPFR_ZIV_INIT (loop, p);
for (;;)
{
mp_prec_t eq;
/* Init temporary vars */
MPFR_TMP_INIT (up, u, p, s);
MPFR_TMP_INIT (vp, v, p, s);
MPFR_TMP_INIT (tmpp, tmp, p, s);
/* Calculus of un and vn */
mpfr_mul (u, op1, op2, GMP_RNDN); /* Faster since PREC(op) < PREC(u) */
mpfr_sqrt (u, u, GMP_RNDN);
mpfr_add (v, op1, op2, GMP_RNDN); /* add with !=prec is still good*/
mpfr_div_2ui (v, v, 1, GMP_RNDN);
n = 1;
while (mpfr_cmp2 (u, v, &eq) != 0 && eq <= p - 2)
{
mpfr_add (tmp, u, v, GMP_RNDN);
mpfr_div_2ui (tmp, tmp, 1, GMP_RNDN);
/* See proof in algorithms.tex */
if (4*eq > p)
{
mpfr_t w;
/* tmp = U(k) */
mpfr_init2 (w, (p + 1) / 2);
mpfr_sub (w, v, u, GMP_RNDN); /* e = V(k-1)-U(k-1) */
mpfr_sqr (w, w, GMP_RNDN); /* e = e^2 */
mpfr_div_2ui (w, w, 4, GMP_RNDN); /* e*= (1/2)^2*1/4 */
mpfr_div (w, w, tmp, GMP_RNDN); /* 1/4*e^2/U(k) */
mpfr_sub (v, tmp, w, GMP_RNDN);
err = MPFR_GET_EXP (tmp) - MPFR_GET_EXP (v); /* 0 or 1 */
mpfr_clear (w);
break;
}
mpfr_mul (u, u, v, GMP_RNDN);
mpfr_sqrt (u, u, GMP_RNDN);
mpfr_swap (v, tmp);
n ++;
}
/* the error on v is bounded by (18n+51) ulps, or twice if there
was an exponent loss in the final subtraction */
err += MPFR_INT_CEIL_LOG2(18 * n + 51); /* 18n+51 should not overflow
since n is about log(p) */
/* we should have n+2 <= 2^(p/4) [see algorithms.tex] */
if (MPFR_LIKELY (MPFR_INT_CEIL_LOG2(n + 2) <= p / 4 &&
MPFR_CAN_ROUND (v, p - err, q, rnd_mode)))
break; /* Stop the loop */
/* Next iteration */
MPFR_ZIV_NEXT (loop, p);
s = (p - 1) / BITS_PER_MP_LIMB + 1;
}
MPFR_ZIV_FREE (loop);
/* Setting of the result */
inexact = mpfr_set (r, v, rnd_mode);
/* Let's clean */
MPFR_TMP_FREE(marker);
return inexact; /* agm(u,v) can be exact for u, v rational only for u=v.
Proof (due to Nicolas Brisebarre): it suffices to consider
u=1 and v<1. Then 1/AGM(1,v) = 2F1(1/2,1/2,1;1-v^2),
and a theorem due to G.V. Chudnovsky states that for x a
non-zero algebraic number with |x|<1, then
2F1(1/2,1/2,1;x) and 2F1(-1/2,1/2,1;x) are algebraically
independent over Q. */
}
/* Exact product. The number a is assumed to have enough allocated memory,
where the trailing bits are regarded as being part of the input numbers
(no reallocation is attempted and no check is performed as MPFR_TMP_INIT
could have been used). The arguments b and c may actually be UBF numbers
(mpfr_srcptr can be seen a bit like void *, but is stronger).
This function does not change the flags, except in case of NaN. */
void
mpfr_ubf_mul_exact (mpfr_ubf_ptr a, mpfr_srcptr b, mpfr_srcptr c)
{
MPFR_LOG_FUNC
(("b[%Pu]=%.*Rg c[%Pu]=%.*Rg",
mpfr_get_prec (b), mpfr_log_prec, b,
mpfr_get_prec (c), mpfr_log_prec, c),
("a[%Pu]=%.*Rg",
mpfr_get_prec (a), mpfr_log_prec, a));
MPFR_ASSERTD ((mpfr_ptr) a != b);
MPFR_ASSERTD ((mpfr_ptr) a != c);
MPFR_SIGN (a) = MPFR_MULT_SIGN (MPFR_SIGN (b), MPFR_SIGN (c));
if (MPFR_ARE_SINGULAR (b, c))
{
if (MPFR_IS_NAN (b) || MPFR_IS_NAN (c))
MPFR_SET_NAN (a);
else if (MPFR_IS_INF (b))
{
if (MPFR_NOTZERO (c))
MPFR_SET_INF (a);
else
MPFR_SET_NAN (a);
}
else if (MPFR_IS_INF (c))
{
if (!MPFR_IS_ZERO (b))
MPFR_SET_INF (a);
else
MPFR_SET_NAN (a);
}
else
{
MPFR_ASSERTD (MPFR_IS_ZERO(b) || MPFR_IS_ZERO(c));
MPFR_SET_ZERO (a);
}
}
else
{
mpfr_exp_t e;
mp_size_t bn, cn;
mpfr_limb_ptr ap;
mp_limb_t u, v;
int m;
/* Note about the code below: For the choice of the precision of
* the result a, one could choose PREC(b) + PREC(c), instead of
* taking whole limbs into account, but in most cases where one
* would gain one limb, one would need to copy the significand
* instead of a no-op (see the mul.c code).
* But in the case MPFR_LIMB_MSB (u) == 0, if the result fits in
* an-1 limbs, one could actually do
* mpn_rshift (ap, ap, k, GMP_NUMB_BITS - 1)
* instead of
* mpn_lshift (ap, ap, k, 1)
* to gain one limb (and reduce the precision), replacing a shift
* by another one. Would this be interesting?
*/
bn = MPFR_LIMB_SIZE (b);
cn = MPFR_LIMB_SIZE (c);
ap = MPFR_MANT (a);
u = (bn >= cn) ?
mpn_mul (ap, MPFR_MANT (b), bn, MPFR_MANT (c), cn) :
mpn_mul (ap, MPFR_MANT (c), cn, MPFR_MANT (b), bn);
if (MPFR_UNLIKELY (MPFR_LIMB_MSB (u) == 0))
{
m = 1;
MPFR_DBGRES (v = mpn_lshift (ap, ap, bn + cn, 1));
MPFR_ASSERTD (v == 0);
}
else
m = 0;
if (! MPFR_IS_UBF (b) && ! MPFR_IS_UBF (c) &&
(e = MPFR_GET_EXP (b) + MPFR_GET_EXP (c) - m,
MPFR_EXP_IN_RANGE (e)))
{
MPFR_SET_EXP (a, e);
}
else
{
mpz_t be, ce;
mpz_init (MPFR_ZEXP (a));
/* This may involve copies of mpz_t, but exponents should not be
very large integers anyway. */
mpfr_get_zexp (be, b);
mpfr_get_zexp (ce, c);
mpz_add (MPFR_ZEXP (a), be, ce);
mpz_clear (be);
mpz_clear (ce);
mpz_sub_ui (MPFR_ZEXP (a), MPFR_ZEXP (a), m);
MPFR_SET_UBF (a);
}
}
}
int
mpfr_sub (mpfr_ptr a, mpfr_srcptr b, mpfr_srcptr c, mpfr_rnd_t rnd_mode)
{
MPFR_LOG_FUNC (("b[%#R]=%R c[%#R]=%R rnd=%d", b, b, c, c, rnd_mode),
("a[%#R]=%R", a, a));
if (MPFR_ARE_SINGULAR (b,c))
{
if (MPFR_IS_NAN (b) || MPFR_IS_NAN (c))
{
MPFR_SET_NAN (a);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF (b))
{
if (!MPFR_IS_INF (c) || MPFR_SIGN (b) != MPFR_SIGN(c))
{
MPFR_SET_INF (a);
MPFR_SET_SAME_SIGN (a, b);
MPFR_RET (0); /* exact */
}
else
{
MPFR_SET_NAN (a); /* Inf - Inf */
MPFR_RET_NAN;
}
}
else if (MPFR_IS_INF (c))
{
MPFR_SET_INF (a);
MPFR_SET_OPPOSITE_SIGN (a, c);
MPFR_RET (0); /* exact */
}
else if (MPFR_IS_ZERO (b))
{
if (MPFR_IS_ZERO (c))
{
int sign = rnd_mode != MPFR_RNDD
? ((MPFR_IS_NEG(b) && MPFR_IS_POS(c)) ? -1 : 1)
: ((MPFR_IS_POS(b) && MPFR_IS_NEG(c)) ? 1 : -1);
MPFR_SET_SIGN (a, sign);
MPFR_SET_ZERO (a);
MPFR_RET(0); /* 0 - 0 is exact */
}
else
return mpfr_neg (a, c, rnd_mode);
}
else
{
MPFR_ASSERTD (MPFR_IS_ZERO (c));
return mpfr_set (a, b, rnd_mode);
}
}
MPFR_ASSERTD (MPFR_IS_PURE_FP (b) && MPFR_IS_PURE_FP (c));
if (MPFR_LIKELY (MPFR_SIGN (b) == MPFR_SIGN (c)))
{ /* signs are equal, it's a real subtraction */
if (MPFR_LIKELY (MPFR_PREC (a) == MPFR_PREC (b)
&& MPFR_PREC (b) == MPFR_PREC (c)))
return mpfr_sub1sp (a, b, c, rnd_mode);
else
return mpfr_sub1 (a, b, c, rnd_mode);
}
else
{ /* signs differ, it's an addition */
if (MPFR_GET_EXP (b) < MPFR_GET_EXP (c))
{ /* exchange rounding modes toward +/- infinity */
int inexact;
rnd_mode = MPFR_INVERT_RND (rnd_mode);
if (MPFR_LIKELY (MPFR_PREC (a) == MPFR_PREC (b)
&& MPFR_PREC (b) == MPFR_PREC (c)))
inexact = mpfr_add1sp (a, c, b, rnd_mode);
else
inexact = mpfr_add1 (a, c, b, rnd_mode);
MPFR_CHANGE_SIGN (a);
return -inexact;
}
else
{
if (MPFR_LIKELY (MPFR_PREC (a) == MPFR_PREC (b)
&& MPFR_PREC (b) == MPFR_PREC (c)))
return mpfr_add1sp (a, b, c, rnd_mode);
else
return mpfr_add1 (a, b, c, rnd_mode);
}
}
}
int
mpfr_cmp3 (mpfr_srcptr b, mpfr_srcptr c, int s)
{
mp_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_ERANGE ();
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_PREC(b)-1)/BITS_PER_MP_LIMB;
cn = (MPFR_PREC(c)-1)/BITS_PER_MP_LIMB;
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;
}
int
mpfr_div (mpfr_ptr q, mpfr_srcptr u, mpfr_srcptr v, mp_rnd_t rnd_mode)
{
mp_srcptr up, vp, bp;
mp_size_t usize, vsize;
mp_ptr ap, qp, rp;
mp_size_t asize, bsize, qsize, rsize;
mp_exp_t qexp;
mp_size_t err, k;
mp_limb_t tonearest;
int inex, sh, can_round = 0, sign_quotient;
unsigned int cc = 0, rw;
TMP_DECL (marker);
/**************************************************************************
* *
* This part of the code deals with special cases *
* *
**************************************************************************/
if (MPFR_ARE_SINGULAR(u,v))
{
if (MPFR_IS_NAN(u) || MPFR_IS_NAN(v))
{
MPFR_SET_NAN(q);
MPFR_RET_NAN;
}
sign_quotient = MPFR_MULT_SIGN( MPFR_SIGN(u) , MPFR_SIGN(v) );
MPFR_SET_SIGN(q, sign_quotient);
if (MPFR_IS_INF(u))
{
if (MPFR_IS_INF(v))
{
MPFR_SET_NAN(q);
MPFR_RET_NAN;
}
else
{
MPFR_SET_INF(q);
MPFR_RET(0);
}
}
else if (MPFR_IS_INF(v))
{
MPFR_SET_ZERO(q);
MPFR_RET(0);
}
else if (MPFR_IS_ZERO(v))
{
if (MPFR_IS_ZERO(u))
{
MPFR_SET_NAN(q);
MPFR_RET_NAN;
}
else
{
MPFR_SET_INF(q);
MPFR_RET(0);
}
}
else
{
MPFR_ASSERTD(MPFR_IS_ZERO(u));
MPFR_SET_ZERO(q);
MPFR_RET(0);
}
}
MPFR_CLEAR_FLAGS(q);
/**************************************************************************
* *
* End of the part concerning special values. *
* *
**************************************************************************/
sign_quotient = MPFR_MULT_SIGN( MPFR_SIGN(u) , MPFR_SIGN(v) );
up = MPFR_MANT(u);
vp = MPFR_MANT(v);
MPFR_SET_SIGN(q, sign_quotient);
TMP_MARK (marker);
usize = MPFR_LIMB_SIZE(u);
vsize = MPFR_LIMB_SIZE(v);
/**************************************************************************
* *
* First try to use only part of u, v. If this is not sufficient, *
* use the full u and v, to avoid long computations eg. in the case *
* u = v. *
* *
**************************************************************************/
/* The dividend is a, length asize. The divisor is b, length bsize. */
qsize = (MPFR_PREC(q) + 3) / BITS_PER_MP_LIMB + 1;
/* in case PREC(q)=PREC(v), then vsize=qsize with probability 1-4/b
where b is the number of bits per limb */
if (MPFR_LIKELY(vsize <= qsize))
{
bsize = vsize;
bp = vp;
}
else /* qsize < vsize: take only the qsize high limbs of the divisor */
{
bsize = qsize;
bp = (mp_srcptr) vp + (vsize - qsize);
}
/* we have {bp, bsize} * (1 + errb) = (true divisor)
with 0 <= errb < 2^(-qsize*BITS_PER_MP_LIMB+1) */
asize = bsize + qsize;
ap = (mp_ptr) TMP_ALLOC (asize * BYTES_PER_MP_LIMB);
/* if all arguments have same precision, then asize will be about 2*usize */
if (MPFR_LIKELY(asize > usize))
{
/* copy u into the high limbs of {ap, asize}, and pad with zeroes */
/* FIXME: could we copy only the qsize high limbs of the dividend? */
MPN_COPY (ap + asize - usize, up, usize);
MPN_ZERO (ap, asize - usize);
}
else /* truncate the high asize limbs of u into {ap, asize} */
MPN_COPY (ap, up + usize - asize, asize);
/* we have {ap, asize} = (true dividend) * (1 - erra)
with 0 <= erra < 2^(-asize*BITS_PER_MP_LIMB).
This {ap, asize} / {bp, bsize} =
(true dividend) / (true divisor) * (1 - erra) (1 + errb) */
/* Allocate limbs for quotient and remainder. */
qp = (mp_ptr) TMP_ALLOC ((qsize + 1) * BYTES_PER_MP_LIMB);
rp = (mp_ptr) TMP_ALLOC (bsize * BYTES_PER_MP_LIMB);
rsize = bsize;
mpn_tdiv_qr (qp, rp, 0, ap, asize, bp, bsize);
sh = - (int) qp[qsize];
/* since u and v are normalized, sh is 0 or -1 */
/* we have {qp, qsize + 1} = {ap, asize} / {bp, bsize} (1 - errq)
with 0 <= errq < 2^(-qsize*BITS_PER_MP_LIMB+1+sh)
thus {qp, qsize + 1} =
(true dividend) / (true divisor) * (1 - erra) (1 + errb) (1 - errq).
In fact, since the truncated dividend and {rp, bsize} do not overlap,
we have: {qp, qsize + 1} =
(true dividend) / (true divisor) * (1 - erra') (1 + errb)
where 0 <= erra' < 2^(-qsize*BITS_PER_MP_LIMB+sh) */
/* Estimate number of correct bits. */
err = qsize * BITS_PER_MP_LIMB;
/* We want to check if rounding is possible, but without normalizing
because we might have to divide again if rounding is impossible, or
if the result might be exact. We have however to mimic normalization */
/*
To detect asap if the result is inexact, so as to avoid doing the
division completely, we perform the following check :
- if rnd_mode != GMP_RNDN, and the result is exact, we are unable
to round simultaneously to zero and to infinity ;
- if rnd_mode == GMP_RNDN, and if we can round to zero with one extra
bit of precision, we can decide rounding. Hence in that case, check
as in the case of GMP_RNDN, with one extra bit. Note that in the case
of close to even rounding we shall do the division completely, but
this is necessary anyway : we need to know whether this is really
even rounding or not.
*/
if (MPFR_UNLIKELY(asize < usize || bsize < vsize))
{
{
mp_rnd_t rnd_mode1, rnd_mode2;
mp_exp_t tmp_exp;
mp_prec_t tmp_prec;
if (bsize < vsize)
err -= 2; /* divisor is truncated */
#if 0 /* commented this out since the truncation of the dividend is already
taken into account in {rp, bsize}, which does not overlap with the
neglected part of the dividend */
else if (asize < usize)
err --; /* dividend is truncated */
#endif
if (MPFR_LIKELY(rnd_mode == GMP_RNDN))
{
rnd_mode1 = GMP_RNDZ;
rnd_mode2 = MPFR_IS_POS_SIGN(sign_quotient) ? GMP_RNDU : GMP_RNDD;
sh++;
}
else
{
rnd_mode1 = rnd_mode;
switch (rnd_mode)
{
case GMP_RNDU:
rnd_mode2 = GMP_RNDD; break;
case GMP_RNDD:
rnd_mode2 = GMP_RNDU; break;
default:
rnd_mode2 = MPFR_IS_POS_SIGN(sign_quotient) ?
GMP_RNDU : GMP_RNDD;
break;
}
}
tmp_exp = err + sh + BITS_PER_MP_LIMB;
tmp_prec = MPFR_PREC(q) + sh + BITS_PER_MP_LIMB;
can_round =
mpfr_can_round_raw (qp, qsize + 1, sign_quotient, tmp_exp,
GMP_RNDN, rnd_mode1, tmp_prec)
& mpfr_can_round_raw (qp, qsize + 1, sign_quotient, tmp_exp,
GMP_RNDN, rnd_mode2, tmp_prec);
/* restore original value of sh, i.e. sh = - qp[qsize] */
sh -= (rnd_mode == GMP_RNDN);
}
/* The computation of z = pow(x,y) is done by
z = exp(y * log(x)) = x^y
For the special cases, see Section F.9.4.4 of the C standard:
_ pow(±0, y) = ±inf for y an odd integer < 0.
_ pow(±0, y) = +inf for y < 0 and not an odd integer.
_ pow(±0, y) = ±0 for y an odd integer > 0.
_ pow(±0, y) = +0 for y > 0 and not an odd integer.
_ pow(-1, ±inf) = 1.
_ pow(+1, y) = 1 for any y, even a NaN.
_ pow(x, ±0) = 1 for any x, even a NaN.
_ pow(x, y) = NaN for finite x < 0 and finite non-integer y.
_ pow(x, -inf) = +inf for |x| < 1.
_ pow(x, -inf) = +0 for |x| > 1.
_ pow(x, +inf) = +0 for |x| < 1.
_ pow(x, +inf) = +inf for |x| > 1.
_ pow(-inf, y) = -0 for y an odd integer < 0.
_ pow(-inf, y) = +0 for y < 0 and not an odd integer.
_ pow(-inf, y) = -inf for y an odd integer > 0.
_ pow(-inf, y) = +inf for y > 0 and not an odd integer.
_ pow(+inf, y) = +0 for y < 0.
_ pow(+inf, y) = +inf for y > 0. */
int
mpfr_pow (mpfr_ptr z, mpfr_srcptr x, mpfr_srcptr y, mpfr_rnd_t rnd_mode)
{
int inexact;
int cmp_x_1;
int y_is_integer;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg y[%Pu]=%.*Rg rnd=%d",
mpfr_get_prec (x), mpfr_log_prec, x,
mpfr_get_prec (y), mpfr_log_prec, y, rnd_mode),
("z[%Pu]=%.*Rg inexact=%d",
mpfr_get_prec (z), mpfr_log_prec, z, inexact));
if (MPFR_ARE_SINGULAR (x, y))
{
/* pow(x, 0) returns 1 for any x, even a NaN. */
if (MPFR_UNLIKELY (MPFR_IS_ZERO (y)))
return mpfr_set_ui (z, 1, rnd_mode);
else if (MPFR_IS_NAN (x))
{
MPFR_SET_NAN (z);
MPFR_RET_NAN;
}
else if (MPFR_IS_NAN (y))
{
/* pow(+1, NaN) returns 1. */
if (mpfr_cmp_ui (x, 1) == 0)
return mpfr_set_ui (z, 1, rnd_mode);
MPFR_SET_NAN (z);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF (y))
{
if (MPFR_IS_INF (x))
{
if (MPFR_IS_POS (y))
MPFR_SET_INF (z);
else
MPFR_SET_ZERO (z);
MPFR_SET_POS (z);
MPFR_RET (0);
}
else
{
int cmp;
cmp = mpfr_cmpabs (x, __gmpfr_one) * MPFR_INT_SIGN (y);
MPFR_SET_POS (z);
if (cmp > 0)
{
/* Return +inf. */
MPFR_SET_INF (z);
MPFR_RET (0);
}
else if (cmp < 0)
{
/* Return +0. */
MPFR_SET_ZERO (z);
MPFR_RET (0);
}
else
{
/* Return 1. */
return mpfr_set_ui (z, 1, rnd_mode);
}
}
}
else if (MPFR_IS_INF (x))
{
int negative;
/* Determine the sign now, in case y and z are the same object */
negative = MPFR_IS_NEG (x) && is_odd (y);
if (MPFR_IS_POS (y))
MPFR_SET_INF (z);
else
MPFR_SET_ZERO (z);
if (negative)
MPFR_SET_NEG (z);
else
MPFR_SET_POS (z);
MPFR_RET (0);
}
else
{
int negative;
MPFR_ASSERTD (MPFR_IS_ZERO (x));
/* Determine the sign now, in case y and z are the same object */
negative = MPFR_IS_NEG(x) && is_odd (y);
if (MPFR_IS_NEG (y))
{
MPFR_ASSERTD (! MPFR_IS_INF (y));
MPFR_SET_INF (z);
mpfr_set_divby0 ();
}
else
MPFR_SET_ZERO (z);
if (negative)
MPFR_SET_NEG (z);
else
MPFR_SET_POS (z);
MPFR_RET (0);
}
}
/* x^y for x < 0 and y not an integer is not defined */
y_is_integer = mpfr_integer_p (y);
if (MPFR_IS_NEG (x) && ! y_is_integer)
{
MPFR_SET_NAN (z);
MPFR_RET_NAN;
}
/* now the result cannot be NaN:
(1) either x > 0
(2) or x < 0 and y is an integer */
cmp_x_1 = mpfr_cmpabs (x, __gmpfr_one);
if (cmp_x_1 == 0)
return mpfr_set_si (z, MPFR_IS_NEG (x) && is_odd (y) ? -1 : 1, rnd_mode);
/* now we have:
(1) either x > 0
(2) or x < 0 and y is an integer
and in addition |x| <> 1.
*/
/* detect overflow: an overflow is possible if
(a) |x| > 1 and y > 0
(b) |x| < 1 and y < 0.
FIXME: this assumes 1 is always representable.
FIXME2: maybe we can test overflow and underflow simultaneously.
The idea is the following: first compute an approximation to
y * log2|x|, using rounding to nearest. If |x| is not too near from 1,
this approximation should be accurate enough, and in most cases enable
one to prove that there is no underflow nor overflow.
Otherwise, it should enable one to check only underflow or overflow,
instead of both cases as in the present case.
*/
if (cmp_x_1 * MPFR_SIGN (y) > 0)
{
mpfr_t t;
int negative, overflow;
MPFR_SAVE_EXPO_MARK (expo);
mpfr_init2 (t, 53);
/* we want a lower bound on y*log2|x|:
(i) if x > 0, it suffices to round log2(x) toward zero, and
to round y*o(log2(x)) toward zero too;
(ii) if x < 0, we first compute t = o(-x), with rounding toward 1,
and then follow as in case (1). */
if (MPFR_SIGN (x) > 0)
mpfr_log2 (t, x, MPFR_RNDZ);
else
{
mpfr_neg (t, x, (cmp_x_1 > 0) ? MPFR_RNDZ : MPFR_RNDU);
mpfr_log2 (t, t, MPFR_RNDZ);
}
mpfr_mul (t, t, y, MPFR_RNDZ);
overflow = mpfr_cmp_si (t, __gmpfr_emax) > 0;
mpfr_clear (t);
MPFR_SAVE_EXPO_FREE (expo);
if (overflow)
{
MPFR_LOG_MSG (("early overflow detection\n", 0));
negative = MPFR_SIGN(x) < 0 && is_odd (y);
return mpfr_overflow (z, rnd_mode, negative ? -1 : 1);
}
}
/* Basic underflow checking. One has:
* - if y > 0, |x^y| < 2^(EXP(x) * y);
* - if y < 0, |x^y| <= 2^((EXP(x) - 1) * y);
* so that one can compute a value ebound such that |x^y| < 2^ebound.
* If we have ebound <= emin - 2 (emin - 1 in directed rounding modes),
* then there is an underflow and we can decide the return value.
*/
if (MPFR_IS_NEG (y) ? (MPFR_GET_EXP (x) > 1) : (MPFR_GET_EXP (x) < 0))
{
mpfr_t tmp;
mpfr_eexp_t ebound;
int inex2;
/* We must restore the flags. */
MPFR_SAVE_EXPO_MARK (expo);
mpfr_init2 (tmp, sizeof (mpfr_exp_t) * CHAR_BIT);
inex2 = mpfr_set_exp_t (tmp, MPFR_GET_EXP (x), MPFR_RNDN);
MPFR_ASSERTN (inex2 == 0);
if (MPFR_IS_NEG (y))
{
inex2 = mpfr_sub_ui (tmp, tmp, 1, MPFR_RNDN);
MPFR_ASSERTN (inex2 == 0);
}
mpfr_mul (tmp, tmp, y, MPFR_RNDU);
if (MPFR_IS_NEG (y))
mpfr_nextabove (tmp);
/* tmp doesn't necessarily fit in ebound, but that doesn't matter
since we get the minimum value in such a case. */
ebound = mpfr_get_exp_t (tmp, MPFR_RNDU);
mpfr_clear (tmp);
MPFR_SAVE_EXPO_FREE (expo);
if (MPFR_UNLIKELY (ebound <=
__gmpfr_emin - (rnd_mode == MPFR_RNDN ? 2 : 1)))
{
/* warning: mpfr_underflow rounds away from 0 for MPFR_RNDN */
MPFR_LOG_MSG (("early underflow detection\n", 0));
return mpfr_underflow (z,
rnd_mode == MPFR_RNDN ? MPFR_RNDZ : rnd_mode,
MPFR_SIGN (x) < 0 && is_odd (y) ? -1 : 1);
}
}
/* If y is an integer, we can use mpfr_pow_z (based on multiplications),
but if y is very large (I'm not sure about the best threshold -- VL),
we shouldn't use it, as it can be very slow and take a lot of memory
(and even crash or make other programs crash, as several hundred of
MBs may be necessary). Note that in such a case, either x = +/-2^b
(this case is handled below) or x^y cannot be represented exactly in
any precision supported by MPFR (the general case uses this property).
*/
if (y_is_integer && (MPFR_GET_EXP (y) <= 256))
{
mpz_t zi;
MPFR_LOG_MSG (("special code for y not too large integer\n", 0));
mpz_init (zi);
mpfr_get_z (zi, y, MPFR_RNDN);
inexact = mpfr_pow_z (z, x, zi, rnd_mode);
mpz_clear (zi);
return inexact;
}
/* Special case (+/-2^b)^Y which could be exact. If x is negative, then
necessarily y is a large integer. */
{
mpfr_exp_t b = MPFR_GET_EXP (x) - 1;
MPFR_ASSERTN (b >= LONG_MIN && b <= LONG_MAX); /* FIXME... */
if (mpfr_cmp_si_2exp (x, MPFR_SIGN(x), b) == 0)
{
mpfr_t tmp;
int sgnx = MPFR_SIGN (x);
MPFR_LOG_MSG (("special case (+/-2^b)^Y\n", 0));
/* now x = +/-2^b, so x^y = (+/-1)^y*2^(b*y) is exact whenever b*y is
an integer */
MPFR_SAVE_EXPO_MARK (expo);
mpfr_init2 (tmp, MPFR_PREC (y) + sizeof (long) * CHAR_BIT);
inexact = mpfr_mul_si (tmp, y, b, MPFR_RNDN); /* exact */
MPFR_ASSERTN (inexact == 0);
/* Note: as the exponent range has been extended, an overflow is not
possible (due to basic overflow and underflow checking above, as
the result is ~ 2^tmp), and an underflow is not possible either
because b is an integer (thus either 0 or >= 1). */
MPFR_CLEAR_FLAGS ();
inexact = mpfr_exp2 (z, tmp, rnd_mode);
mpfr_clear (tmp);
if (sgnx < 0 && is_odd (y))
{
mpfr_neg (z, z, rnd_mode);
inexact = -inexact;
}
/* Without the following, the overflows3 test in tpow.c fails. */
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, __gmpfr_flags);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (z, inexact, rnd_mode);
}
}
MPFR_SAVE_EXPO_MARK (expo);
/* Case where |y * log(x)| is very small. Warning: x can be negative, in
that case y is a large integer. */
{
mpfr_t t;
mpfr_exp_t err;
/* We need an upper bound on the exponent of y * log(x). */
mpfr_init2 (t, 16);
if (MPFR_IS_POS(x))
mpfr_log (t, x, cmp_x_1 < 0 ? MPFR_RNDD : MPFR_RNDU); /* away from 0 */
else
{
/* if x < -1, round to +Inf, else round to zero */
mpfr_neg (t, x, (mpfr_cmp_si (x, -1) < 0) ? MPFR_RNDU : MPFR_RNDD);
mpfr_log (t, t, (mpfr_cmp_ui (t, 1) < 0) ? MPFR_RNDD : MPFR_RNDU);
}
MPFR_ASSERTN (MPFR_IS_PURE_FP (t));
err = MPFR_GET_EXP (y) + MPFR_GET_EXP (t);
mpfr_clear (t);
MPFR_CLEAR_FLAGS ();
MPFR_SMALL_INPUT_AFTER_SAVE_EXPO (z, __gmpfr_one, - err, 0,
(MPFR_SIGN (y) > 0) ^ (cmp_x_1 < 0),
rnd_mode, expo, {});
}
/* General case */
inexact = mpfr_pow_general (z, x, y, rnd_mode, y_is_integer, &expo);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (z, inexact, rnd_mode);
}
int
mpfr_add (mpfr_ptr a, mpfr_srcptr b, mpfr_srcptr c, mpfr_rnd_t rnd_mode)
{
MPFR_LOG_FUNC (("b[%#R]=%R c[%#R]=%R rnd=%d", b, b, c, c, rnd_mode),
("a[%#R]=%R", a, a));
if (MPFR_ARE_SINGULAR(b,c))
{
if (MPFR_IS_NAN(b) || MPFR_IS_NAN(c))
{
MPFR_SET_NAN(a);
MPFR_RET_NAN;
}
/* neither b nor c is NaN here */
else if (MPFR_IS_INF(b))
{
if (!MPFR_IS_INF(c) || MPFR_SIGN(b) == MPFR_SIGN(c))
{
MPFR_SET_INF(a);
MPFR_SET_SAME_SIGN(a, b);
MPFR_RET(0); /* exact */
}
else
{
MPFR_SET_NAN(a);
MPFR_RET_NAN;
}
}
else if (MPFR_IS_INF(c))
{
MPFR_SET_INF(a);
MPFR_SET_SAME_SIGN(a, c);
MPFR_RET(0); /* exact */
}
/* now either b or c is zero */
else if (MPFR_IS_ZERO(b))
{
if (MPFR_IS_ZERO(c))
{
/* for round away, we take the same convention for 0 + 0
as for round to zero or to nearest: it always gives +0,
except (-0) + (-0) = -0. */
MPFR_SET_SIGN(a,
(rnd_mode != MPFR_RNDD ?
((MPFR_IS_NEG(b) && MPFR_IS_NEG(c)) ? -1 : 1) :
((MPFR_IS_POS(b) && MPFR_IS_POS(c)) ? 1 : -1)));
MPFR_SET_ZERO(a);
MPFR_RET(0); /* 0 + 0 is exact */
}
return mpfr_set (a, c, rnd_mode);
}
else
{
MPFR_ASSERTD(MPFR_IS_ZERO(c));
return mpfr_set (a, b, rnd_mode);
}
}
MPFR_ASSERTD(MPFR_IS_PURE_FP(b) && MPFR_IS_PURE_FP(c));
if (MPFR_UNLIKELY(MPFR_SIGN(b) != MPFR_SIGN(c)))
{ /* signs differ, it's a subtraction */
if (MPFR_LIKELY(MPFR_PREC(a) == MPFR_PREC(b)
&& MPFR_PREC(b) == MPFR_PREC(c)))
return mpfr_sub1sp(a,b,c,rnd_mode);
else
return mpfr_sub1(a, b, c, rnd_mode);
}
else
{ /* signs are equal, it's an addition */
if (MPFR_LIKELY(MPFR_PREC(a) == MPFR_PREC(b)
&& MPFR_PREC(b) == MPFR_PREC(c)))
if (MPFR_GET_EXP(b) < MPFR_GET_EXP(c))
return mpfr_add1sp(a, c, b, rnd_mode);
else
return mpfr_add1sp(a, b, c, rnd_mode);
else
if (MPFR_GET_EXP(b) < MPFR_GET_EXP(c))
return mpfr_add1(a, c, b, rnd_mode);
else
return mpfr_add1(a, b, c, rnd_mode);
}
}
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_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;
/* <= k, thus no int overflow */
MPFR_ASSERTD(tn <= k);
/* 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 */
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), 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);
}
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))
int
mpfr_cmpabs (mpfr_srcptr b, mpfr_srcptr c)
{
mpfr_exp_t be, ce;
mp_size_t bn, cn;
mp_limb_t *bp, *cp;
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))
return ! MPFR_IS_INF (c);
else if (MPFR_IS_INF (c))
return -1;
else if (MPFR_IS_ZERO (c))
return ! MPFR_IS_ZERO (b);
else /* b == 0 */
return -1;
}
MPFR_ASSERTD (MPFR_IS_PURE_FP (b));
MPFR_ASSERTD (MPFR_IS_PURE_FP (c));
/* Now that we know that b and c are pure FP numbers (i.e. they have
a meaningful exponent), we use MPFR_EXP instead of MPFR_GET_EXP to
allow exponents outside the current exponent range. For instance,
this is useful for mpfr_pow, which compares values to __gmpfr_one.
This is for internal use only! For compatibility with other MPFR
versions, the user must still provide values that are representable
in the current exponent range. */
be = MPFR_EXP (b);
ce = MPFR_EXP (c);
if (be > ce)
return 1;
if (be < ce)
return -1;
/* exponents are equal */
bn = MPFR_LIMB_SIZE(b)-1;
cn = MPFR_LIMB_SIZE(c)-1;
bp = MPFR_MANT(b);
cp = MPFR_MANT(c);
for ( ; bn >= 0 && cn >= 0; bn--, cn--)
{
if (bp[bn] > cp[cn])
return 1;
if (bp[bn] < cp[cn])
return -1;
}
for ( ; bn >= 0; bn--)
if (bp[bn])
return 1;
for ( ; cn >= 0; cn--)
if (cp[cn])
return -1;
return 0;
}
int
mpfr_atan2 (mpfr_ptr dest, mpfr_srcptr y, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
mpfr_t tmp, pi;
int inexact;
mpfr_prec_t prec;
mpfr_exp_t e;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_ZIV_DECL (loop);
MPFR_LOG_FUNC
(("y[%Pu]=%.*Rg x[%Pu]=%.*Rg rnd=%d",
mpfr_get_prec (y), mpfr_log_prec, y,
mpfr_get_prec (x), mpfr_log_prec, x, rnd_mode),
("atan[%Pu]=%.*Rg inexact=%d",
mpfr_get_prec (dest), mpfr_log_prec, dest, inexact));
/* Special cases */
if (MPFR_ARE_SINGULAR (x, y))
{
/* atan2(0, 0) does not raise the "invalid" floating-point
exception, nor does atan2(y, 0) raise the "divide-by-zero"
floating-point exception.
-- atan2(±0, -0) returns ±pi.313)
-- atan2(±0, +0) returns ±0.
-- atan2(±0, x) returns ±pi, for x < 0.
-- atan2(±0, x) returns ±0, for x > 0.
-- atan2(y, ±0) returns -pi/2 for y < 0.
-- atan2(y, ±0) returns pi/2 for y > 0.
-- atan2(±oo, -oo) returns ±3pi/4.
-- atan2(±oo, +oo) returns ±pi/4.
-- atan2(±oo, x) returns ±pi/2, for finite x.
-- atan2(±y, -oo) returns ±pi, for finite y > 0.
-- atan2(±y, +oo) returns ±0, for finite y > 0.
*/
if (MPFR_IS_NAN (x) || MPFR_IS_NAN (y))
{
MPFR_SET_NAN (dest);
MPFR_RET_NAN;
}
if (MPFR_IS_ZERO (y))
{
if (MPFR_IS_NEG (x)) /* +/- PI */
{
set_pi:
if (MPFR_IS_NEG (y))
{
inexact = mpfr_const_pi (dest, MPFR_INVERT_RND (rnd_mode));
MPFR_CHANGE_SIGN (dest);
return -inexact;
}
else
return mpfr_const_pi (dest, rnd_mode);
}
else /* +/- 0 */
{
set_zero:
MPFR_SET_ZERO (dest);
MPFR_SET_SAME_SIGN (dest, y);
return 0;
}
}
if (MPFR_IS_ZERO (x))
{
return pi_div_2ui (dest, 1, MPFR_IS_NEG (y), rnd_mode);
}
if (MPFR_IS_INF (y))
{
if (!MPFR_IS_INF (x)) /* +/- PI/2 */
return pi_div_2ui (dest, 1, MPFR_IS_NEG (y), rnd_mode);
else if (MPFR_IS_POS (x)) /* +/- PI/4 */
return pi_div_2ui (dest, 2, MPFR_IS_NEG (y), rnd_mode);
else /* +/- 3*PI/4: Ugly since we have to round properly */
{
mpfr_t tmp2;
MPFR_ZIV_DECL (loop2);
mpfr_prec_t prec2 = MPFR_PREC (dest) + 10;
MPFR_SAVE_EXPO_MARK (expo);
mpfr_init2 (tmp2, prec2);
MPFR_ZIV_INIT (loop2, prec2);
for (;;)
{
mpfr_const_pi (tmp2, MPFR_RNDN);
mpfr_mul_ui (tmp2, tmp2, 3, MPFR_RNDN); /* Error <= 2 */
mpfr_div_2ui (tmp2, tmp2, 2, MPFR_RNDN);
if (mpfr_round_p (MPFR_MANT (tmp2), MPFR_LIMB_SIZE (tmp2),
MPFR_PREC (tmp2) - 2,
MPFR_PREC (dest) + (rnd_mode == MPFR_RNDN)))
break;
MPFR_ZIV_NEXT (loop2, prec2);
mpfr_set_prec (tmp2, prec2);
}
MPFR_ZIV_FREE (loop2);
if (MPFR_IS_NEG (y))
MPFR_CHANGE_SIGN (tmp2);
inexact = mpfr_set (dest, tmp2, rnd_mode);
mpfr_clear (tmp2);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (dest, inexact, rnd_mode);
}
}
MPFR_ASSERTD (MPFR_IS_INF (x));
if (MPFR_IS_NEG (x))
goto set_pi;
else
goto set_zero;
}
/* When x is a power of two, we call directly atan(y/x) since y/x is
exact. */
if (MPFR_UNLIKELY (MPFR_IS_POWER_OF_2 (x)))
{
int r;
mpfr_t yoverx;
unsigned int saved_flags = __gmpfr_flags;
mpfr_init2 (yoverx, MPFR_PREC (y));
if (MPFR_LIKELY (mpfr_div_2si (yoverx, y, MPFR_GET_EXP (x) - 1,
MPFR_RNDN) == 0))
{
/* Here the flags have not changed due to mpfr_div_2si. */
r = mpfr_atan (dest, yoverx, rnd_mode);
mpfr_clear (yoverx);
return r;
}
else
{
/* Division is inexact because of a small exponent range */
mpfr_clear (yoverx);
__gmpfr_flags = saved_flags;
}
}
MPFR_SAVE_EXPO_MARK (expo);
/* Set up initial prec */
prec = MPFR_PREC (dest) + 3 + MPFR_INT_CEIL_LOG2 (MPFR_PREC (dest));
mpfr_init2 (tmp, prec);
MPFR_ZIV_INIT (loop, prec);
if (MPFR_IS_POS (x))
/* use atan2(y,x) = atan(y/x) */
for (;;)
{
int div_inex;
MPFR_BLOCK_DECL (flags);
MPFR_BLOCK (flags, div_inex = mpfr_div (tmp, y, x, MPFR_RNDN));
if (div_inex == 0)
{
/* Result is exact. */
inexact = mpfr_atan (dest, tmp, rnd_mode);
goto end;
}
/* Error <= ulp (tmp) except in case of underflow or overflow. */
/* If the division underflowed, since |atan(z)/z| < 1, we have
an underflow. */
if (MPFR_UNDERFLOW (flags))
{
int sign;
/* In the case MPFR_RNDN with 2^(emin-2) < |y/x| < 2^(emin-1):
The smallest significand value S > 1 of |y/x| is:
* 1 / (1 - 2^(-px)) if py <= px,
* (1 - 2^(-px) + 2^(-py)) / (1 - 2^(-px)) if py >= px.
Therefore S - 1 > 2^(-pz), where pz = max(px,py). We have:
atan(|y/x|) > atan(z), where z = 2^(emin-2) * (1 + 2^(-pz)).
> z - z^3 / 3.
> 2^(emin-2) * (1 + 2^(-pz) - 2^(2 emin - 5))
Assuming pz <= -2 emin + 5, we can round away from zero
(this is what mpfr_underflow always does on MPFR_RNDN).
In the case MPFR_RNDN with |y/x| <= 2^(emin-2), we round
toward zero, as |atan(z)/z| < 1. */
MPFR_ASSERTN (MPFR_PREC_MAX <=
2 * (mpfr_uexp_t) - MPFR_EMIN_MIN + 5);
if (rnd_mode == MPFR_RNDN && MPFR_IS_ZERO (tmp))
rnd_mode = MPFR_RNDZ;
sign = MPFR_SIGN (tmp);
mpfr_clear (tmp);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_underflow (dest, rnd_mode, sign);
}
mpfr_atan (tmp, tmp, MPFR_RNDN); /* Error <= 2*ulp (tmp) since
abs(D(arctan)) <= 1 */
/* TODO: check that the error bound is correct in case of overflow. */
/* FIXME: Error <= ulp(tmp) ? */
if (MPFR_LIKELY (MPFR_CAN_ROUND (tmp, prec - 2, MPFR_PREC (dest),
rnd_mode)))
break;
MPFR_ZIV_NEXT (loop, prec);
mpfr_set_prec (tmp, prec);
}
else /* x < 0 */
/* Use sign(y)*(PI - atan (|y/x|)) */
{
mpfr_init2 (pi, prec);
for (;;)
{
mpfr_div (tmp, y, x, MPFR_RNDN); /* Error <= ulp (tmp) */
/* If tmp is 0, we have |y/x| <= 2^(-emin-2), thus
atan|y/x| < 2^(-emin-2). */
MPFR_SET_POS (tmp); /* no error */
mpfr_atan (tmp, tmp, MPFR_RNDN); /* Error <= 2*ulp (tmp) since
abs(D(arctan)) <= 1 */
mpfr_const_pi (pi, MPFR_RNDN); /* Error <= ulp(pi) /2 */
e = MPFR_NOTZERO(tmp) ? MPFR_GET_EXP (tmp) : __gmpfr_emin - 1;
mpfr_sub (tmp, pi, tmp, MPFR_RNDN); /* see above */
if (MPFR_IS_NEG (y))
MPFR_CHANGE_SIGN (tmp);
/* Error(tmp) <= (1/2+2^(EXP(pi)-EXP(tmp)-1)+2^(e-EXP(tmp)+1))*ulp
<= 2^(MAX (MAX (EXP(PI)-EXP(tmp)-1, e-EXP(tmp)+1),
-1)+2)*ulp(tmp) */
e = MAX (MAX (MPFR_GET_EXP (pi)-MPFR_GET_EXP (tmp) - 1,
e - MPFR_GET_EXP (tmp) + 1), -1) + 2;
if (MPFR_LIKELY (MPFR_CAN_ROUND (tmp, prec - e, MPFR_PREC (dest),
rnd_mode)))
break;
MPFR_ZIV_NEXT (loop, prec);
mpfr_set_prec (tmp, prec);
mpfr_set_prec (pi, prec);
}
mpfr_clear (pi);
}
inexact = mpfr_set (dest, tmp, rnd_mode);
end:
MPFR_ZIV_FREE (loop);
mpfr_clear (tmp);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (dest, inexact, rnd_mode);
}
