Home | History | Annotate | Download | only in Modules
      1 /* Math module -- standard C math library functions, pi and e */
      2 
      3 /* Here are some comments from Tim Peters, extracted from the
      4    discussion attached to http://bugs.python.org/issue1640.  They
      5    describe the general aims of the math module with respect to
      6    special values, IEEE-754 floating-point exceptions, and Python
      7    exceptions.
      8 
      9 These are the "spirit of 754" rules:
     10 
     11 1. If the mathematical result is a real number, but of magnitude too
     12 large to approximate by a machine float, overflow is signaled and the
     13 result is an infinity (with the appropriate sign).
     14 
     15 2. If the mathematical result is a real number, but of magnitude too
     16 small to approximate by a machine float, underflow is signaled and the
     17 result is a zero (with the appropriate sign).
     18 
     19 3. At a singularity (a value x such that the limit of f(y) as y
     20 approaches x exists and is an infinity), "divide by zero" is signaled
     21 and the result is an infinity (with the appropriate sign).  This is
     22 complicated a little by that the left-side and right-side limits may
     23 not be the same; e.g., 1/x approaches +inf or -inf as x approaches 0
     24 from the positive or negative directions.  In that specific case, the
     25 sign of the zero determines the result of 1/0.
     26 
     27 4. At a point where a function has no defined result in the extended
     28 reals (i.e., the reals plus an infinity or two), invalid operation is
     29 signaled and a NaN is returned.
     30 
     31 And these are what Python has historically /tried/ to do (but not
     32 always successfully, as platform libm behavior varies a lot):
     33 
     34 For #1, raise OverflowError.
     35 
     36 For #2, return a zero (with the appropriate sign if that happens by
     37 accident ;-)).
     38 
     39 For #3 and #4, raise ValueError.  It may have made sense to raise
     40 Python's ZeroDivisionError in #3, but historically that's only been
     41 raised for division by zero and mod by zero.
     42 
     43 */
     44 
     45 /*
     46    In general, on an IEEE-754 platform the aim is to follow the C99
     47    standard, including Annex 'F', whenever possible.  Where the
     48    standard recommends raising the 'divide-by-zero' or 'invalid'
     49    floating-point exceptions, Python should raise a ValueError.  Where
     50    the standard recommends raising 'overflow', Python should raise an
     51    OverflowError.  In all other circumstances a value should be
     52    returned.
     53  */
     54 
     55 #include "Python.h"
     56 #include "_math.h"
     57 
     58 #ifdef _OSF_SOURCE
     59 /* OSF1 5.1 doesn't make this available with XOPEN_SOURCE_EXTENDED defined */
     60 extern double copysign(double, double);
     61 #endif
     62 
     63 /*
     64    sin(pi*x), giving accurate results for all finite x (especially x
     65    integral or close to an integer).  This is here for use in the
     66    reflection formula for the gamma function.  It conforms to IEEE
     67    754-2008 for finite arguments, but not for infinities or nans.
     68 */
     69 
     70 static const double pi = 3.141592653589793238462643383279502884197;
     71 static const double sqrtpi = 1.772453850905516027298167483341145182798;
     72 
     73 static double
     74 sinpi(double x)
     75 {
     76     double y, r;
     77     int n;
     78     /* this function should only ever be called for finite arguments */
     79     assert(Py_IS_FINITE(x));
     80     y = fmod(fabs(x), 2.0);
     81     n = (int)round(2.0*y);
     82     assert(0 <= n && n <= 4);
     83     switch (n) {
     84     case 0:
     85         r = sin(pi*y);
     86         break;
     87     case 1:
     88         r = cos(pi*(y-0.5));
     89         break;
     90     case 2:
     91         /* N.B. -sin(pi*(y-1.0)) is *not* equivalent: it would give
     92            -0.0 instead of 0.0 when y == 1.0. */
     93         r = sin(pi*(1.0-y));
     94         break;
     95     case 3:
     96         r = -cos(pi*(y-1.5));
     97         break;
     98     case 4:
     99         r = sin(pi*(y-2.0));
    100         break;
    101     default:
    102         assert(0);  /* should never get here */
    103         r = -1.23e200; /* silence gcc warning */
    104     }
    105     return copysign(1.0, x)*r;
    106 }
    107 
    108 /* Implementation of the real gamma function.  In extensive but non-exhaustive
    109    random tests, this function proved accurate to within <= 10 ulps across the
    110    entire float domain.  Note that accuracy may depend on the quality of the
    111    system math functions, the pow function in particular.  Special cases
    112    follow C99 annex F.  The parameters and method are tailored to platforms
    113    whose double format is the IEEE 754 binary64 format.
    114 
    115    Method: for x > 0.0 we use the Lanczos approximation with parameters N=13
    116    and g=6.024680040776729583740234375; these parameters are amongst those
    117    used by the Boost library.  Following Boost (again), we re-express the
    118    Lanczos sum as a rational function, and compute it that way.  The
    119    coefficients below were computed independently using MPFR, and have been
    120    double-checked against the coefficients in the Boost source code.
    121 
    122    For x < 0.0 we use the reflection formula.
    123 
    124    There's one minor tweak that deserves explanation: Lanczos' formula for
    125    Gamma(x) involves computing pow(x+g-0.5, x-0.5) / exp(x+g-0.5).  For many x
    126    values, x+g-0.5 can be represented exactly.  However, in cases where it
    127    can't be represented exactly the small error in x+g-0.5 can be magnified
    128    significantly by the pow and exp calls, especially for large x.  A cheap
    129    correction is to multiply by (1 + e*g/(x+g-0.5)), where e is the error
    130    involved in the computation of x+g-0.5 (that is, e = computed value of
    131    x+g-0.5 - exact value of x+g-0.5).  Here's the proof:
    132 
    133    Correction factor
    134    -----------------
    135    Write x+g-0.5 = y-e, where y is exactly representable as an IEEE 754
    136    double, and e is tiny.  Then:
    137 
    138      pow(x+g-0.5,x-0.5)/exp(x+g-0.5) = pow(y-e, x-0.5)/exp(y-e)
    139      = pow(y, x-0.5)/exp(y) * C,
    140 
    141    where the correction_factor C is given by
    142 
    143      C = pow(1-e/y, x-0.5) * exp(e)
    144 
    145    Since e is tiny, pow(1-e/y, x-0.5) ~ 1-(x-0.5)*e/y, and exp(x) ~ 1+e, so:
    146 
    147      C ~ (1-(x-0.5)*e/y) * (1+e) ~ 1 + e*(y-(x-0.5))/y
    148 
    149    But y-(x-0.5) = g+e, and g+e ~ g.  So we get C ~ 1 + e*g/y, and
    150 
    151      pow(x+g-0.5,x-0.5)/exp(x+g-0.5) ~ pow(y, x-0.5)/exp(y) * (1 + e*g/y),
    152 
    153    Note that for accuracy, when computing r*C it's better to do
    154 
    155      r + e*g/y*r;
    156 
    157    than
    158 
    159      r * (1 + e*g/y);
    160 
    161    since the addition in the latter throws away most of the bits of
    162    information in e*g/y.
    163 */
    164 
    165 #define LANCZOS_N 13
    166 static const double lanczos_g = 6.024680040776729583740234375;
    167 static const double lanczos_g_minus_half = 5.524680040776729583740234375;
    168 static const double lanczos_num_coeffs[LANCZOS_N] = {
    169     23531376880.410759688572007674451636754734846804940,
    170     42919803642.649098768957899047001988850926355848959,
    171     35711959237.355668049440185451547166705960488635843,
    172     17921034426.037209699919755754458931112671403265390,
    173     6039542586.3520280050642916443072979210699388420708,
    174     1439720407.3117216736632230727949123939715485786772,
    175     248874557.86205415651146038641322942321632125127801,
    176     31426415.585400194380614231628318205362874684987640,
    177     2876370.6289353724412254090516208496135991145378768,
    178     186056.26539522349504029498971604569928220784236328,
    179     8071.6720023658162106380029022722506138218516325024,
    180     210.82427775157934587250973392071336271166969580291,
    181     2.5066282746310002701649081771338373386264310793408
    182 };
    183 
    184 /* denominator is x*(x+1)*...*(x+LANCZOS_N-2) */
    185 static const double lanczos_den_coeffs[LANCZOS_N] = {
    186     0.0, 39916800.0, 120543840.0, 150917976.0, 105258076.0, 45995730.0,
    187     13339535.0, 2637558.0, 357423.0, 32670.0, 1925.0, 66.0, 1.0};
    188 
    189 /* gamma values for small positive integers, 1 though NGAMMA_INTEGRAL */
    190 #define NGAMMA_INTEGRAL 23
    191 static const double gamma_integral[NGAMMA_INTEGRAL] = {
    192     1.0, 1.0, 2.0, 6.0, 24.0, 120.0, 720.0, 5040.0, 40320.0, 362880.0,
    193     3628800.0, 39916800.0, 479001600.0, 6227020800.0, 87178291200.0,
    194     1307674368000.0, 20922789888000.0, 355687428096000.0,
    195     6402373705728000.0, 121645100408832000.0, 2432902008176640000.0,
    196     51090942171709440000.0, 1124000727777607680000.0,
    197 };
    198 
    199 /* Lanczos' sum L_g(x), for positive x */
    200 
    201 static double
    202 lanczos_sum(double x)
    203 {
    204     double num = 0.0, den = 0.0;
    205     int i;
    206     assert(x > 0.0);
    207     /* evaluate the rational function lanczos_sum(x).  For large
    208        x, the obvious algorithm risks overflow, so we instead
    209        rescale the denominator and numerator of the rational
    210        function by x**(1-LANCZOS_N) and treat this as a
    211        rational function in 1/x.  This also reduces the error for
    212        larger x values.  The choice of cutoff point (5.0 below) is
    213        somewhat arbitrary; in tests, smaller cutoff values than
    214        this resulted in lower accuracy. */
    215     if (x < 5.0) {
    216         for (i = LANCZOS_N; --i >= 0; ) {
    217             num = num * x + lanczos_num_coeffs[i];
    218             den = den * x + lanczos_den_coeffs[i];
    219         }
    220     }
    221     else {
    222         for (i = 0; i < LANCZOS_N; i++) {
    223             num = num / x + lanczos_num_coeffs[i];
    224             den = den / x + lanczos_den_coeffs[i];
    225         }
    226     }
    227     return num/den;
    228 }
    229 
    230 static double
    231 m_tgamma(double x)
    232 {
    233     double absx, r, y, z, sqrtpow;
    234 
    235     /* special cases */
    236     if (!Py_IS_FINITE(x)) {
    237         if (Py_IS_NAN(x) || x > 0.0)
    238             return x;  /* tgamma(nan) = nan, tgamma(inf) = inf */
    239         else {
    240             errno = EDOM;
    241             return Py_NAN;  /* tgamma(-inf) = nan, invalid */
    242         }
    243     }
    244     if (x == 0.0) {
    245         errno = EDOM;
    246         return 1.0/x; /* tgamma(+-0.0) = +-inf, divide-by-zero */
    247     }
    248 
    249     /* integer arguments */
    250     if (x == floor(x)) {
    251         if (x < 0.0) {
    252             errno = EDOM;  /* tgamma(n) = nan, invalid for */
    253             return Py_NAN; /* negative integers n */
    254         }
    255         if (x <= NGAMMA_INTEGRAL)
    256             return gamma_integral[(int)x - 1];
    257     }
    258     absx = fabs(x);
    259 
    260     /* tiny arguments:  tgamma(x) ~ 1/x for x near 0 */
    261     if (absx < 1e-20) {
    262         r = 1.0/x;
    263         if (Py_IS_INFINITY(r))
    264             errno = ERANGE;
    265         return r;
    266     }
    267 
    268     /* large arguments: assuming IEEE 754 doubles, tgamma(x) overflows for
    269        x > 200, and underflows to +-0.0 for x < -200, not a negative
    270        integer. */
    271     if (absx > 200.0) {
    272         if (x < 0.0) {
    273             return 0.0/sinpi(x);
    274         }
    275         else {
    276             errno = ERANGE;
    277             return Py_HUGE_VAL;
    278         }
    279     }
    280 
    281     y = absx + lanczos_g_minus_half;
    282     /* compute error in sum */
    283     if (absx > lanczos_g_minus_half) {
    284         /* note: the correction can be foiled by an optimizing
    285            compiler that (incorrectly) thinks that an expression like
    286            a + b - a - b can be optimized to 0.0.  This shouldn't
    287            happen in a standards-conforming compiler. */
    288         double q = y - absx;
    289         z = q - lanczos_g_minus_half;
    290     }
    291     else {
    292         double q = y - lanczos_g_minus_half;
    293         z = q - absx;
    294     }
    295     z = z * lanczos_g / y;
    296     if (x < 0.0) {
    297         r = -pi / sinpi(absx) / absx * exp(y) / lanczos_sum(absx);
    298         r -= z * r;
    299         if (absx < 140.0) {
    300             r /= pow(y, absx - 0.5);
    301         }
    302         else {
    303             sqrtpow = pow(y, absx / 2.0 - 0.25);
    304             r /= sqrtpow;
    305             r /= sqrtpow;
    306         }
    307     }
    308     else {
    309         r = lanczos_sum(absx) / exp(y);
    310         r += z * r;
    311         if (absx < 140.0) {
    312             r *= pow(y, absx - 0.5);
    313         }
    314         else {
    315             sqrtpow = pow(y, absx / 2.0 - 0.25);
    316             r *= sqrtpow;
    317             r *= sqrtpow;
    318         }
    319     }
    320     if (Py_IS_INFINITY(r))
    321         errno = ERANGE;
    322     return r;
    323 }
    324 
    325 /*
    326    lgamma:  natural log of the absolute value of the Gamma function.
    327    For large arguments, Lanczos' formula works extremely well here.
    328 */
    329 
    330 static double
    331 m_lgamma(double x)
    332 {
    333     double r, absx;
    334 
    335     /* special cases */
    336     if (!Py_IS_FINITE(x)) {
    337         if (Py_IS_NAN(x))
    338             return x;  /* lgamma(nan) = nan */
    339         else
    340             return Py_HUGE_VAL; /* lgamma(+-inf) = +inf */
    341     }
    342 
    343     /* integer arguments */
    344     if (x == floor(x) && x <= 2.0) {
    345         if (x <= 0.0) {
    346             errno = EDOM;  /* lgamma(n) = inf, divide-by-zero for */
    347             return Py_HUGE_VAL; /* integers n <= 0 */
    348         }
    349         else {
    350             return 0.0; /* lgamma(1) = lgamma(2) = 0.0 */
    351         }
    352     }
    353 
    354     absx = fabs(x);
    355     /* tiny arguments: lgamma(x) ~ -log(fabs(x)) for small x */
    356     if (absx < 1e-20)
    357         return -log(absx);
    358 
    359     /* Lanczos' formula */
    360     if (x > 0.0) {
    361         /* we could save a fraction of a ulp in accuracy by having a
    362            second set of numerator coefficients for lanczos_sum that
    363            absorbed the exp(-lanczos_g) term, and throwing out the
    364            lanczos_g subtraction below; it's probably not worth it. */
    365         r = log(lanczos_sum(x)) - lanczos_g +
    366             (x-0.5)*(log(x+lanczos_g-0.5)-1);
    367     }
    368     else {
    369         r = log(pi) - log(fabs(sinpi(absx))) - log(absx) -
    370             (log(lanczos_sum(absx)) - lanczos_g +
    371              (absx-0.5)*(log(absx+lanczos_g-0.5)-1));
    372     }
    373     if (Py_IS_INFINITY(r))
    374         errno = ERANGE;
    375     return r;
    376 }
    377 
    378 /*
    379    Implementations of the error function erf(x) and the complementary error
    380    function erfc(x).
    381 
    382    Method: we use a series approximation for erf for small x, and a continued
    383    fraction approximation for erfc(x) for larger x;
    384    combined with the relations erf(-x) = -erf(x) and erfc(x) = 1.0 - erf(x),
    385    this gives us erf(x) and erfc(x) for all x.
    386 
    387    The series expansion used is:
    388 
    389       erf(x) = x*exp(-x*x)/sqrt(pi) * [
    390                      2/1 + 4/3 x**2 + 8/15 x**4 + 16/105 x**6 + ...]
    391 
    392    The coefficient of x**(2k-2) here is 4**k*factorial(k)/factorial(2*k).
    393    This series converges well for smallish x, but slowly for larger x.
    394 
    395    The continued fraction expansion used is:
    396 
    397       erfc(x) = x*exp(-x*x)/sqrt(pi) * [1/(0.5 + x**2 -) 0.5/(2.5 + x**2 - )
    398                               3.0/(4.5 + x**2 - ) 7.5/(6.5 + x**2 - ) ...]
    399 
    400    after the first term, the general term has the form:
    401 
    402       k*(k-0.5)/(2*k+0.5 + x**2 - ...).
    403 
    404    This expansion converges fast for larger x, but convergence becomes
    405    infinitely slow as x approaches 0.0.  The (somewhat naive) continued
    406    fraction evaluation algorithm used below also risks overflow for large x;
    407    but for large x, erfc(x) == 0.0 to within machine precision.  (For
    408    example, erfc(30.0) is approximately 2.56e-393).
    409 
    410    Parameters: use series expansion for abs(x) < ERF_SERIES_CUTOFF and
    411    continued fraction expansion for ERF_SERIES_CUTOFF <= abs(x) <
    412    ERFC_CONTFRAC_CUTOFF.  ERFC_SERIES_TERMS and ERFC_CONTFRAC_TERMS are the
    413    numbers of terms to use for the relevant expansions.  */
    414 
    415 #define ERF_SERIES_CUTOFF 1.5
    416 #define ERF_SERIES_TERMS 25
    417 #define ERFC_CONTFRAC_CUTOFF 30.0
    418 #define ERFC_CONTFRAC_TERMS 50
    419 
    420 /*
    421    Error function, via power series.
    422 
    423    Given a finite float x, return an approximation to erf(x).
    424    Converges reasonably fast for small x.
    425 */
    426 
    427 static double
    428 m_erf_series(double x)
    429 {
    430     double x2, acc, fk, result;
    431     int i, saved_errno;
    432 
    433     x2 = x * x;
    434     acc = 0.0;
    435     fk = (double)ERF_SERIES_TERMS + 0.5;
    436     for (i = 0; i < ERF_SERIES_TERMS; i++) {
    437         acc = 2.0 + x2 * acc / fk;
    438         fk -= 1.0;
    439     }
    440     /* Make sure the exp call doesn't affect errno;
    441        see m_erfc_contfrac for more. */
    442     saved_errno = errno;
    443     result = acc * x * exp(-x2) / sqrtpi;
    444     errno = saved_errno;
    445     return result;
    446 }
    447 
    448 /*
    449    Complementary error function, via continued fraction expansion.
    450 
    451    Given a positive float x, return an approximation to erfc(x).  Converges
    452    reasonably fast for x large (say, x > 2.0), and should be safe from
    453    overflow if x and nterms are not too large.  On an IEEE 754 machine, with x
    454    <= 30.0, we're safe up to nterms = 100.  For x >= 30.0, erfc(x) is smaller
    455    than the smallest representable nonzero float.  */
    456 
    457 static double
    458 m_erfc_contfrac(double x)
    459 {
    460     double x2, a, da, p, p_last, q, q_last, b, result;
    461     int i, saved_errno;
    462 
    463     if (x >= ERFC_CONTFRAC_CUTOFF)
    464         return 0.0;
    465 
    466     x2 = x*x;
    467     a = 0.0;
    468     da = 0.5;
    469     p = 1.0; p_last = 0.0;
    470     q = da + x2; q_last = 1.0;
    471     for (i = 0; i < ERFC_CONTFRAC_TERMS; i++) {
    472         double temp;
    473         a += da;
    474         da += 2.0;
    475         b = da + x2;
    476         temp = p; p = b*p - a*p_last; p_last = temp;
    477         temp = q; q = b*q - a*q_last; q_last = temp;
    478     }
    479     /* Issue #8986: On some platforms, exp sets errno on underflow to zero;
    480        save the current errno value so that we can restore it later. */
    481     saved_errno = errno;
    482     result = p / q * x * exp(-x2) / sqrtpi;
    483     errno = saved_errno;
    484     return result;
    485 }
    486 
    487 /* Error function erf(x), for general x */
    488 
    489 static double
    490 m_erf(double x)
    491 {
    492     double absx, cf;
    493 
    494     if (Py_IS_NAN(x))
    495         return x;
    496     absx = fabs(x);
    497     if (absx < ERF_SERIES_CUTOFF)
    498         return m_erf_series(x);
    499     else {
    500         cf = m_erfc_contfrac(absx);
    501         return x > 0.0 ? 1.0 - cf : cf - 1.0;
    502     }
    503 }
    504 
    505 /* Complementary error function erfc(x), for general x. */
    506 
    507 static double
    508 m_erfc(double x)
    509 {
    510     double absx, cf;
    511 
    512     if (Py_IS_NAN(x))
    513         return x;
    514     absx = fabs(x);
    515     if (absx < ERF_SERIES_CUTOFF)
    516         return 1.0 - m_erf_series(x);
    517     else {
    518         cf = m_erfc_contfrac(absx);
    519         return x > 0.0 ? cf : 2.0 - cf;
    520     }
    521 }
    522 
    523 /*
    524    wrapper for atan2 that deals directly with special cases before
    525    delegating to the platform libm for the remaining cases.  This
    526    is necessary to get consistent behaviour across platforms.
    527    Windows, FreeBSD and alpha Tru64 are amongst platforms that don't
    528    always follow C99.
    529 */
    530 
    531 static double
    532 m_atan2(double y, double x)
    533 {
    534     if (Py_IS_NAN(x) || Py_IS_NAN(y))
    535         return Py_NAN;
    536     if (Py_IS_INFINITY(y)) {
    537         if (Py_IS_INFINITY(x)) {
    538             if (copysign(1., x) == 1.)
    539                 /* atan2(+-inf, +inf) == +-pi/4 */
    540                 return copysign(0.25*Py_MATH_PI, y);
    541             else
    542                 /* atan2(+-inf, -inf) == +-pi*3/4 */
    543                 return copysign(0.75*Py_MATH_PI, y);
    544         }
    545         /* atan2(+-inf, x) == +-pi/2 for finite x */
    546         return copysign(0.5*Py_MATH_PI, y);
    547     }
    548     if (Py_IS_INFINITY(x) || y == 0.) {
    549         if (copysign(1., x) == 1.)
    550             /* atan2(+-y, +inf) = atan2(+-0, +x) = +-0. */
    551             return copysign(0., y);
    552         else
    553             /* atan2(+-y, -inf) = atan2(+-0., -x) = +-pi. */
    554             return copysign(Py_MATH_PI, y);
    555     }
    556     return atan2(y, x);
    557 }
    558 
    559 /*
    560     Various platforms (Solaris, OpenBSD) do nonstandard things for log(0),
    561     log(-ve), log(NaN).  Here are wrappers for log and log10 that deal with
    562     special values directly, passing positive non-special values through to
    563     the system log/log10.
    564  */
    565 
    566 static double
    567 m_log(double x)
    568 {
    569     if (Py_IS_FINITE(x)) {
    570         if (x > 0.0)
    571             return log(x);
    572         errno = EDOM;
    573         if (x == 0.0)
    574             return -Py_HUGE_VAL; /* log(0) = -inf */
    575         else
    576             return Py_NAN; /* log(-ve) = nan */
    577     }
    578     else if (Py_IS_NAN(x))
    579         return x; /* log(nan) = nan */
    580     else if (x > 0.0)
    581         return x; /* log(inf) = inf */
    582     else {
    583         errno = EDOM;
    584         return Py_NAN; /* log(-inf) = nan */
    585     }
    586 }
    587 
    588 static double
    589 m_log10(double x)
    590 {
    591     if (Py_IS_FINITE(x)) {
    592         if (x > 0.0)
    593             return log10(x);
    594         errno = EDOM;
    595         if (x == 0.0)
    596             return -Py_HUGE_VAL; /* log10(0) = -inf */
    597         else
    598             return Py_NAN; /* log10(-ve) = nan */
    599     }
    600     else if (Py_IS_NAN(x))
    601         return x; /* log10(nan) = nan */
    602     else if (x > 0.0)
    603         return x; /* log10(inf) = inf */
    604     else {
    605         errno = EDOM;
    606         return Py_NAN; /* log10(-inf) = nan */
    607     }
    608 }
    609 
    610 
    611 /* Call is_error when errno != 0, and where x is the result libm
    612  * returned.  is_error will usually set up an exception and return
    613  * true (1), but may return false (0) without setting up an exception.
    614  */
    615 static int
    616 is_error(double x)
    617 {
    618     int result = 1;     /* presumption of guilt */
    619     assert(errno);      /* non-zero errno is a precondition for calling */
    620     if (errno == EDOM)
    621         PyErr_SetString(PyExc_ValueError, "math domain error");
    622 
    623     else if (errno == ERANGE) {
    624         /* ANSI C generally requires libm functions to set ERANGE
    625          * on overflow, but also generally *allows* them to set
    626          * ERANGE on underflow too.  There's no consistency about
    627          * the latter across platforms.
    628          * Alas, C99 never requires that errno be set.
    629          * Here we suppress the underflow errors (libm functions
    630          * should return a zero on underflow, and +- HUGE_VAL on
    631          * overflow, so testing the result for zero suffices to
    632          * distinguish the cases).
    633          *
    634          * On some platforms (Ubuntu/ia64) it seems that errno can be
    635          * set to ERANGE for subnormal results that do *not* underflow
    636          * to zero.  So to be safe, we'll ignore ERANGE whenever the
    637          * function result is less than one in absolute value.
    638          */
    639         if (fabs(x) < 1.0)
    640             result = 0;
    641         else
    642             PyErr_SetString(PyExc_OverflowError,
    643                             "math range error");
    644     }
    645     else
    646         /* Unexpected math error */
    647         PyErr_SetFromErrno(PyExc_ValueError);
    648     return result;
    649 }
    650 
    651 /*
    652    math_1 is used to wrap a libm function f that takes a double
    653    arguments and returns a double.
    654 
    655    The error reporting follows these rules, which are designed to do
    656    the right thing on C89/C99 platforms and IEEE 754/non IEEE 754
    657    platforms.
    658 
    659    - a NaN result from non-NaN inputs causes ValueError to be raised
    660    - an infinite result from finite inputs causes OverflowError to be
    661      raised if can_overflow is 1, or raises ValueError if can_overflow
    662      is 0.
    663    - if the result is finite and errno == EDOM then ValueError is
    664      raised
    665    - if the result is finite and nonzero and errno == ERANGE then
    666      OverflowError is raised
    667 
    668    The last rule is used to catch overflow on platforms which follow
    669    C89 but for which HUGE_VAL is not an infinity.
    670 
    671    For the majority of one-argument functions these rules are enough
    672    to ensure that Python's functions behave as specified in 'Annex F'
    673    of the C99 standard, with the 'invalid' and 'divide-by-zero'
    674    floating-point exceptions mapping to Python's ValueError and the
    675    'overflow' floating-point exception mapping to OverflowError.
    676    math_1 only works for functions that don't have singularities *and*
    677    the possibility of overflow; fortunately, that covers everything we
    678    care about right now.
    679 */
    680 
    681 static PyObject *
    682 math_1(PyObject *arg, double (*func) (double), int can_overflow)
    683 {
    684     double x, r;
    685     x = PyFloat_AsDouble(arg);
    686     if (x == -1.0 && PyErr_Occurred())
    687         return NULL;
    688     errno = 0;
    689     PyFPE_START_PROTECT("in math_1", return 0);
    690     r = (*func)(x);
    691     PyFPE_END_PROTECT(r);
    692     if (Py_IS_NAN(r)) {
    693         if (!Py_IS_NAN(x))
    694             errno = EDOM;
    695         else
    696             errno = 0;
    697     }
    698     else if (Py_IS_INFINITY(r)) {
    699         if (Py_IS_FINITE(x))
    700             errno = can_overflow ? ERANGE : EDOM;
    701         else
    702             errno = 0;
    703     }
    704     if (errno && is_error(r))
    705         return NULL;
    706     else
    707         return PyFloat_FromDouble(r);
    708 }
    709 
    710 /* variant of math_1, to be used when the function being wrapped is known to
    711    set errno properly (that is, errno = EDOM for invalid or divide-by-zero,
    712    errno = ERANGE for overflow). */
    713 
    714 static PyObject *
    715 math_1a(PyObject *arg, double (*func) (double))
    716 {
    717     double x, r;
    718     x = PyFloat_AsDouble(arg);
    719     if (x == -1.0 && PyErr_Occurred())
    720         return NULL;
    721     errno = 0;
    722     PyFPE_START_PROTECT("in math_1a", return 0);
    723     r = (*func)(x);
    724     PyFPE_END_PROTECT(r);
    725     if (errno && is_error(r))
    726         return NULL;
    727     return PyFloat_FromDouble(r);
    728 }
    729 
    730 /*
    731    math_2 is used to wrap a libm function f that takes two double
    732    arguments and returns a double.
    733 
    734    The error reporting follows these rules, which are designed to do
    735    the right thing on C89/C99 platforms and IEEE 754/non IEEE 754
    736    platforms.
    737 
    738    - a NaN result from non-NaN inputs causes ValueError to be raised
    739    - an infinite result from finite inputs causes OverflowError to be
    740      raised.
    741    - if the result is finite and errno == EDOM then ValueError is
    742      raised
    743    - if the result is finite and nonzero and errno == ERANGE then
    744      OverflowError is raised
    745 
    746    The last rule is used to catch overflow on platforms which follow
    747    C89 but for which HUGE_VAL is not an infinity.
    748 
    749    For most two-argument functions (copysign, fmod, hypot, atan2)
    750    these rules are enough to ensure that Python's functions behave as
    751    specified in 'Annex F' of the C99 standard, with the 'invalid' and
    752    'divide-by-zero' floating-point exceptions mapping to Python's
    753    ValueError and the 'overflow' floating-point exception mapping to
    754    OverflowError.
    755 */
    756 
    757 static PyObject *
    758 math_2(PyObject *args, double (*func) (double, double), char *funcname)
    759 {
    760     PyObject *ox, *oy;
    761     double x, y, r;
    762     if (! PyArg_UnpackTuple(args, funcname, 2, 2, &ox, &oy))
    763         return NULL;
    764     x = PyFloat_AsDouble(ox);
    765     y = PyFloat_AsDouble(oy);
    766     if ((x == -1.0 || y == -1.0) && PyErr_Occurred())
    767         return NULL;
    768     errno = 0;
    769     PyFPE_START_PROTECT("in math_2", return 0);
    770     r = (*func)(x, y);
    771     PyFPE_END_PROTECT(r);
    772     if (Py_IS_NAN(r)) {
    773         if (!Py_IS_NAN(x) && !Py_IS_NAN(y))
    774             errno = EDOM;
    775         else
    776             errno = 0;
    777     }
    778     else if (Py_IS_INFINITY(r)) {
    779         if (Py_IS_FINITE(x) && Py_IS_FINITE(y))
    780             errno = ERANGE;
    781         else
    782             errno = 0;
    783     }
    784     if (errno && is_error(r))
    785         return NULL;
    786     else
    787         return PyFloat_FromDouble(r);
    788 }
    789 
    790 #define FUNC1(funcname, func, can_overflow, docstring)                  \
    791     static PyObject * math_##funcname(PyObject *self, PyObject *args) { \
    792         return math_1(args, func, can_overflow);                            \
    793     }\
    794     PyDoc_STRVAR(math_##funcname##_doc, docstring);
    795 
    796 #define FUNC1A(funcname, func, docstring)                               \
    797     static PyObject * math_##funcname(PyObject *self, PyObject *args) { \
    798         return math_1a(args, func);                                     \
    799     }\
    800     PyDoc_STRVAR(math_##funcname##_doc, docstring);
    801 
    802 #define FUNC2(funcname, func, docstring) \
    803     static PyObject * math_##funcname(PyObject *self, PyObject *args) { \
    804         return math_2(args, func, #funcname); \
    805     }\
    806     PyDoc_STRVAR(math_##funcname##_doc, docstring);
    807 
    808 FUNC1(acos, acos, 0,
    809       "acos(x)\n\nReturn the arc cosine (measured in radians) of x.")
    810 FUNC1(acosh, m_acosh, 0,
    811       "acosh(x)\n\nReturn the inverse hyperbolic cosine of x.")
    812 FUNC1(asin, asin, 0,
    813       "asin(x)\n\nReturn the arc sine (measured in radians) of x.")
    814 FUNC1(asinh, m_asinh, 0,
    815       "asinh(x)\n\nReturn the inverse hyperbolic sine of x.")
    816 FUNC1(atan, atan, 0,
    817       "atan(x)\n\nReturn the arc tangent (measured in radians) of x.")
    818 FUNC2(atan2, m_atan2,
    819       "atan2(y, x)\n\nReturn the arc tangent (measured in radians) of y/x.\n"
    820       "Unlike atan(y/x), the signs of both x and y are considered.")
    821 FUNC1(atanh, m_atanh, 0,
    822       "atanh(x)\n\nReturn the inverse hyperbolic tangent of x.")
    823 FUNC1(ceil, ceil, 0,
    824       "ceil(x)\n\nReturn the ceiling of x as a float.\n"
    825       "This is the smallest integral value >= x.")
    826 FUNC2(copysign, copysign,
    827       "copysign(x, y)\n\nReturn x with the sign of y.")
    828 FUNC1(cos, cos, 0,
    829       "cos(x)\n\nReturn the cosine of x (measured in radians).")
    830 FUNC1(cosh, cosh, 1,
    831       "cosh(x)\n\nReturn the hyperbolic cosine of x.")
    832 FUNC1A(erf, m_erf,
    833        "erf(x)\n\nError function at x.")
    834 FUNC1A(erfc, m_erfc,
    835        "erfc(x)\n\nComplementary error function at x.")
    836 FUNC1(exp, exp, 1,
    837       "exp(x)\n\nReturn e raised to the power of x.")
    838 FUNC1(expm1, m_expm1, 1,
    839       "expm1(x)\n\nReturn exp(x)-1.\n"
    840       "This function avoids the loss of precision involved in the direct "
    841       "evaluation of exp(x)-1 for small x.")
    842 FUNC1(fabs, fabs, 0,
    843       "fabs(x)\n\nReturn the absolute value of the float x.")
    844 FUNC1(floor, floor, 0,
    845       "floor(x)\n\nReturn the floor of x as a float.\n"
    846       "This is the largest integral value <= x.")
    847 FUNC1A(gamma, m_tgamma,
    848       "gamma(x)\n\nGamma function at x.")
    849 FUNC1A(lgamma, m_lgamma,
    850       "lgamma(x)\n\nNatural logarithm of absolute value of Gamma function at x.")
    851 FUNC1(log1p, m_log1p, 1,
    852       "log1p(x)\n\nReturn the natural logarithm of 1+x (base e).\n"
    853       "The result is computed in a way which is accurate for x near zero.")
    854 FUNC1(sin, sin, 0,
    855       "sin(x)\n\nReturn the sine of x (measured in radians).")
    856 FUNC1(sinh, sinh, 1,
    857       "sinh(x)\n\nReturn the hyperbolic sine of x.")
    858 FUNC1(sqrt, sqrt, 0,
    859       "sqrt(x)\n\nReturn the square root of x.")
    860 FUNC1(tan, tan, 0,
    861       "tan(x)\n\nReturn the tangent of x (measured in radians).")
    862 FUNC1(tanh, tanh, 0,
    863       "tanh(x)\n\nReturn the hyperbolic tangent of x.")
    864 
    865 /* Precision summation function as msum() by Raymond Hettinger in
    866    <http://aspn.activestate.com/ASPN/Cookbook/Python/Recipe/393090>,
    867    enhanced with the exact partials sum and roundoff from Mark
    868    Dickinson's post at <http://bugs.python.org/file10357/msum4.py>.
    869    See those links for more details, proofs and other references.
    870 
    871    Note 1: IEEE 754R floating point semantics are assumed,
    872    but the current implementation does not re-establish special
    873    value semantics across iterations (i.e. handling -Inf + Inf).
    874 
    875    Note 2:  No provision is made for intermediate overflow handling;
    876    therefore, sum([1e+308, 1e-308, 1e+308]) returns 1e+308 while
    877    sum([1e+308, 1e+308, 1e-308]) raises an OverflowError due to the
    878    overflow of the first partial sum.
    879 
    880    Note 3: The intermediate values lo, yr, and hi are declared volatile so
    881    aggressive compilers won't algebraically reduce lo to always be exactly 0.0.
    882    Also, the volatile declaration forces the values to be stored in memory as
    883    regular doubles instead of extended long precision (80-bit) values.  This
    884    prevents double rounding because any addition or subtraction of two doubles
    885    can be resolved exactly into double-sized hi and lo values.  As long as the
    886    hi value gets forced into a double before yr and lo are computed, the extra
    887    bits in downstream extended precision operations (x87 for example) will be
    888    exactly zero and therefore can be losslessly stored back into a double,
    889    thereby preventing double rounding.
    890 
    891    Note 4: A similar implementation is in Modules/cmathmodule.c.
    892    Be sure to update both when making changes.
    893 
    894    Note 5: The signature of math.fsum() differs from __builtin__.sum()
    895    because the start argument doesn't make sense in the context of
    896    accurate summation.  Since the partials table is collapsed before
    897    returning a result, sum(seq2, start=sum(seq1)) may not equal the
    898    accurate result returned by sum(itertools.chain(seq1, seq2)).
    899 */
    900 
    901 #define NUM_PARTIALS  32  /* initial partials array size, on stack */
    902 
    903 /* Extend the partials array p[] by doubling its size. */
    904 static int                          /* non-zero on error */
    905 _fsum_realloc(double **p_ptr, Py_ssize_t  n,
    906              double  *ps,    Py_ssize_t *m_ptr)
    907 {
    908     void *v = NULL;
    909     Py_ssize_t m = *m_ptr;
    910 
    911     m += m;  /* double */
    912     if (n < m && m < (PY_SSIZE_T_MAX / sizeof(double))) {
    913         double *p = *p_ptr;
    914         if (p == ps) {
    915             v = PyMem_Malloc(sizeof(double) * m);
    916             if (v != NULL)
    917                 memcpy(v, ps, sizeof(double) * n);
    918         }
    919         else
    920             v = PyMem_Realloc(p, sizeof(double) * m);
    921     }
    922     if (v == NULL) {        /* size overflow or no memory */
    923         PyErr_SetString(PyExc_MemoryError, "math.fsum partials");
    924         return 1;
    925     }
    926     *p_ptr = (double*) v;
    927     *m_ptr = m;
    928     return 0;
    929 }
    930 
    931 /* Full precision summation of a sequence of floats.
    932 
    933    def msum(iterable):
    934        partials = []  # sorted, non-overlapping partial sums
    935        for x in iterable:
    936            i = 0
    937            for y in partials:
    938                if abs(x) < abs(y):
    939                    x, y = y, x
    940                hi = x + y
    941                lo = y - (hi - x)
    942                if lo:
    943                    partials[i] = lo
    944                    i += 1
    945                x = hi
    946            partials[i:] = [x]
    947        return sum_exact(partials)
    948 
    949    Rounded x+y stored in hi with the roundoff stored in lo.  Together hi+lo
    950    are exactly equal to x+y.  The inner loop applies hi/lo summation to each
    951    partial so that the list of partial sums remains exact.
    952 
    953    Sum_exact() adds the partial sums exactly and correctly rounds the final
    954    result (using the round-half-to-even rule).  The items in partials remain
    955    non-zero, non-special, non-overlapping and strictly increasing in
    956    magnitude, but possibly not all having the same sign.
    957 
    958    Depends on IEEE 754 arithmetic guarantees and half-even rounding.
    959 */
    960 
    961 static PyObject*
    962 math_fsum(PyObject *self, PyObject *seq)
    963 {
    964     PyObject *item, *iter, *sum = NULL;
    965     Py_ssize_t i, j, n = 0, m = NUM_PARTIALS;
    966     double x, y, t, ps[NUM_PARTIALS], *p = ps;
    967     double xsave, special_sum = 0.0, inf_sum = 0.0;
    968     volatile double hi, yr, lo;
    969 
    970     iter = PyObject_GetIter(seq);
    971     if (iter == NULL)
    972         return NULL;
    973 
    974     PyFPE_START_PROTECT("fsum", Py_DECREF(iter); return NULL)
    975 
    976     for(;;) {           /* for x in iterable */
    977         assert(0 <= n && n <= m);
    978         assert((m == NUM_PARTIALS && p == ps) ||
    979                (m >  NUM_PARTIALS && p != NULL));
    980 
    981         item = PyIter_Next(iter);
    982         if (item == NULL) {
    983             if (PyErr_Occurred())
    984                 goto _fsum_error;
    985             break;
    986         }
    987         x = PyFloat_AsDouble(item);
    988         Py_DECREF(item);
    989         if (PyErr_Occurred())
    990             goto _fsum_error;
    991 
    992         xsave = x;
    993         for (i = j = 0; j < n; j++) {       /* for y in partials */
    994             y = p[j];
    995             if (fabs(x) < fabs(y)) {
    996                 t = x; x = y; y = t;
    997             }
    998             hi = x + y;
    999             yr = hi - x;
   1000             lo = y - yr;
   1001             if (lo != 0.0)
   1002                 p[i++] = lo;
   1003             x = hi;
   1004         }
   1005 
   1006         n = i;                              /* ps[i:] = [x] */
   1007         if (x != 0.0) {
   1008             if (! Py_IS_FINITE(x)) {
   1009                 /* a nonfinite x could arise either as
   1010                    a result of intermediate overflow, or
   1011                    as a result of a nan or inf in the
   1012                    summands */
   1013                 if (Py_IS_FINITE(xsave)) {
   1014                     PyErr_SetString(PyExc_OverflowError,
   1015                           "intermediate overflow in fsum");
   1016                     goto _fsum_error;
   1017                 }
   1018                 if (Py_IS_INFINITY(xsave))
   1019                     inf_sum += xsave;
   1020                 special_sum += xsave;
   1021                 /* reset partials */
   1022                 n = 0;
   1023             }
   1024             else if (n >= m && _fsum_realloc(&p, n, ps, &m))
   1025                 goto _fsum_error;
   1026             else
   1027                 p[n++] = x;
   1028         }
   1029     }
   1030 
   1031     if (special_sum != 0.0) {
   1032         if (Py_IS_NAN(inf_sum))
   1033             PyErr_SetString(PyExc_ValueError,
   1034                             "-inf + inf in fsum");
   1035         else
   1036             sum = PyFloat_FromDouble(special_sum);
   1037         goto _fsum_error;
   1038     }
   1039 
   1040     hi = 0.0;
   1041     if (n > 0) {
   1042         hi = p[--n];
   1043         /* sum_exact(ps, hi) from the top, stop when the sum becomes
   1044            inexact. */
   1045         while (n > 0) {
   1046             x = hi;
   1047             y = p[--n];
   1048             assert(fabs(y) < fabs(x));
   1049             hi = x + y;
   1050             yr = hi - x;
   1051             lo = y - yr;
   1052             if (lo != 0.0)
   1053                 break;
   1054         }
   1055         /* Make half-even rounding work across multiple partials.
   1056            Needed so that sum([1e-16, 1, 1e16]) will round-up the last
   1057            digit to two instead of down to zero (the 1e-16 makes the 1
   1058            slightly closer to two).  With a potential 1 ULP rounding
   1059            error fixed-up, math.fsum() can guarantee commutativity. */
   1060         if (n > 0 && ((lo < 0.0 && p[n-1] < 0.0) ||
   1061                       (lo > 0.0 && p[n-1] > 0.0))) {
   1062             y = lo * 2.0;
   1063             x = hi + y;
   1064             yr = x - hi;
   1065             if (y == yr)
   1066                 hi = x;
   1067         }
   1068     }
   1069     sum = PyFloat_FromDouble(hi);
   1070 
   1071 _fsum_error:
   1072     PyFPE_END_PROTECT(hi)
   1073     Py_DECREF(iter);
   1074     if (p != ps)
   1075         PyMem_Free(p);
   1076     return sum;
   1077 }
   1078 
   1079 #undef NUM_PARTIALS
   1080 
   1081 PyDoc_STRVAR(math_fsum_doc,
   1082 "fsum(iterable)\n\n\
   1083 Return an accurate floating point sum of values in the iterable.\n\
   1084 Assumes IEEE-754 floating point arithmetic.");
   1085 
   1086 static PyObject *
   1087 math_factorial(PyObject *self, PyObject *arg)
   1088 {
   1089     long i, x;
   1090     PyObject *result, *iobj, *newresult;
   1091 
   1092     if (PyFloat_Check(arg)) {
   1093         PyObject *lx;
   1094         double dx = PyFloat_AS_DOUBLE((PyFloatObject *)arg);
   1095         if (!(Py_IS_FINITE(dx) && dx == floor(dx))) {
   1096             PyErr_SetString(PyExc_ValueError,
   1097                 "factorial() only accepts integral values");
   1098             return NULL;
   1099         }
   1100         lx = PyLong_FromDouble(dx);
   1101         if (lx == NULL)
   1102             return NULL;
   1103         x = PyLong_AsLong(lx);
   1104         Py_DECREF(lx);
   1105     }
   1106     else
   1107         x = PyInt_AsLong(arg);
   1108 
   1109     if (x == -1 && PyErr_Occurred())
   1110         return NULL;
   1111     if (x < 0) {
   1112         PyErr_SetString(PyExc_ValueError,
   1113             "factorial() not defined for negative values");
   1114         return NULL;
   1115     }
   1116 
   1117     result = (PyObject *)PyInt_FromLong(1);
   1118     if (result == NULL)
   1119         return NULL;
   1120     for (i=1 ; i<=x ; i++) {
   1121         iobj = (PyObject *)PyInt_FromLong(i);
   1122         if (iobj == NULL)
   1123             goto error;
   1124         newresult = PyNumber_Multiply(result, iobj);
   1125         Py_DECREF(iobj);
   1126         if (newresult == NULL)
   1127             goto error;
   1128         Py_DECREF(result);
   1129         result = newresult;
   1130     }
   1131     return result;
   1132 
   1133 error:
   1134     Py_DECREF(result);
   1135     return NULL;
   1136 }
   1137 
   1138 PyDoc_STRVAR(math_factorial_doc,
   1139 "factorial(x) -> Integral\n"
   1140 "\n"
   1141 "Find x!. Raise a ValueError if x is negative or non-integral.");
   1142 
   1143 static PyObject *
   1144 math_trunc(PyObject *self, PyObject *number)
   1145 {
   1146     return PyObject_CallMethod(number, "__trunc__", NULL);
   1147 }
   1148 
   1149 PyDoc_STRVAR(math_trunc_doc,
   1150 "trunc(x:Real) -> Integral\n"
   1151 "\n"
   1152 "Truncates x to the nearest Integral toward 0. Uses the __trunc__ magic method.");
   1153 
   1154 static PyObject *
   1155 math_frexp(PyObject *self, PyObject *arg)
   1156 {
   1157     int i;
   1158     double x = PyFloat_AsDouble(arg);
   1159     if (x == -1.0 && PyErr_Occurred())
   1160         return NULL;
   1161     /* deal with special cases directly, to sidestep platform
   1162        differences */
   1163     if (Py_IS_NAN(x) || Py_IS_INFINITY(x) || !x) {
   1164         i = 0;
   1165     }
   1166     else {
   1167         PyFPE_START_PROTECT("in math_frexp", return 0);
   1168         x = frexp(x, &i);
   1169         PyFPE_END_PROTECT(x);
   1170     }
   1171     return Py_BuildValue("(di)", x, i);
   1172 }
   1173 
   1174 PyDoc_STRVAR(math_frexp_doc,
   1175 "frexp(x)\n"
   1176 "\n"
   1177 "Return the mantissa and exponent of x, as pair (m, e).\n"
   1178 "m is a float and e is an int, such that x = m * 2.**e.\n"
   1179 "If x is 0, m and e are both 0.  Else 0.5 <= abs(m) < 1.0.");
   1180 
   1181 static PyObject *
   1182 math_ldexp(PyObject *self, PyObject *args)
   1183 {
   1184     double x, r;
   1185     PyObject *oexp;
   1186     long exp;
   1187     int overflow;
   1188     if (! PyArg_ParseTuple(args, "dO:ldexp", &x, &oexp))
   1189         return NULL;
   1190 
   1191     if (PyLong_Check(oexp) || PyInt_Check(oexp)) {
   1192         /* on overflow, replace exponent with either LONG_MAX
   1193            or LONG_MIN, depending on the sign. */
   1194         exp = PyLong_AsLongAndOverflow(oexp, &overflow);
   1195         if (exp == -1 && PyErr_Occurred())
   1196             return NULL;
   1197         if (overflow)
   1198             exp = overflow < 0 ? LONG_MIN : LONG_MAX;
   1199     }
   1200     else {
   1201         PyErr_SetString(PyExc_TypeError,
   1202                         "Expected an int or long as second argument "
   1203                         "to ldexp.");
   1204         return NULL;
   1205     }
   1206 
   1207     if (x == 0. || !Py_IS_FINITE(x)) {
   1208         /* NaNs, zeros and infinities are returned unchanged */
   1209         r = x;
   1210         errno = 0;
   1211     } else if (exp > INT_MAX) {
   1212         /* overflow */
   1213         r = copysign(Py_HUGE_VAL, x);
   1214         errno = ERANGE;
   1215     } else if (exp < INT_MIN) {
   1216         /* underflow to +-0 */
   1217         r = copysign(0., x);
   1218         errno = 0;
   1219     } else {
   1220         errno = 0;
   1221         PyFPE_START_PROTECT("in math_ldexp", return 0);
   1222         r = ldexp(x, (int)exp);
   1223         PyFPE_END_PROTECT(r);
   1224         if (Py_IS_INFINITY(r))
   1225             errno = ERANGE;
   1226     }
   1227 
   1228     if (errno && is_error(r))
   1229         return NULL;
   1230     return PyFloat_FromDouble(r);
   1231 }
   1232 
   1233 PyDoc_STRVAR(math_ldexp_doc,
   1234 "ldexp(x, i)\n\n\
   1235 Return x * (2**i).");
   1236 
   1237 static PyObject *
   1238 math_modf(PyObject *self, PyObject *arg)
   1239 {
   1240     double y, x = PyFloat_AsDouble(arg);
   1241     if (x == -1.0 && PyErr_Occurred())
   1242         return NULL;
   1243     /* some platforms don't do the right thing for NaNs and
   1244        infinities, so we take care of special cases directly. */
   1245     if (!Py_IS_FINITE(x)) {
   1246         if (Py_IS_INFINITY(x))
   1247             return Py_BuildValue("(dd)", copysign(0., x), x);
   1248         else if (Py_IS_NAN(x))
   1249             return Py_BuildValue("(dd)", x, x);
   1250     }
   1251 
   1252     errno = 0;
   1253     PyFPE_START_PROTECT("in math_modf", return 0);
   1254     x = modf(x, &y);
   1255     PyFPE_END_PROTECT(x);
   1256     return Py_BuildValue("(dd)", x, y);
   1257 }
   1258 
   1259 PyDoc_STRVAR(math_modf_doc,
   1260 "modf(x)\n"
   1261 "\n"
   1262 "Return the fractional and integer parts of x.  Both results carry the sign\n"
   1263 "of x and are floats.");
   1264 
   1265 /* A decent logarithm is easy to compute even for huge longs, but libm can't
   1266    do that by itself -- loghelper can.  func is log or log10, and name is
   1267    "log" or "log10".  Note that overflow of the result isn't possible: a long
   1268    can contain no more than INT_MAX * SHIFT bits, so has value certainly less
   1269    than 2**(2**64 * 2**16) == 2**2**80, and log2 of that is 2**80, which is
   1270    small enough to fit in an IEEE single.  log and log10 are even smaller.
   1271    However, intermediate overflow is possible for a long if the number of bits
   1272    in that long is larger than PY_SSIZE_T_MAX. */
   1273 
   1274 static PyObject*
   1275 loghelper(PyObject* arg, double (*func)(double), char *funcname)
   1276 {
   1277     /* If it is long, do it ourselves. */
   1278     if (PyLong_Check(arg)) {
   1279         double x, result;
   1280         Py_ssize_t e;
   1281 
   1282         /* Negative or zero inputs give a ValueError. */
   1283         if (Py_SIZE(arg) <= 0) {
   1284             PyErr_SetString(PyExc_ValueError,
   1285                             "math domain error");
   1286             return NULL;
   1287         }
   1288 
   1289         x = PyLong_AsDouble(arg);
   1290         if (x == -1.0 && PyErr_Occurred()) {
   1291             if (!PyErr_ExceptionMatches(PyExc_OverflowError))
   1292                 return NULL;
   1293             /* Here the conversion to double overflowed, but it's possible
   1294                to compute the log anyway.  Clear the exception and continue. */
   1295             PyErr_Clear();
   1296             x = _PyLong_Frexp((PyLongObject *)arg, &e);
   1297             if (x == -1.0 && PyErr_Occurred())
   1298                 return NULL;
   1299             /* Value is ~= x * 2**e, so the log ~= log(x) + log(2) * e. */
   1300             result = func(x) + func(2.0) * e;
   1301         }
   1302         else
   1303             /* Successfully converted x to a double. */
   1304             result = func(x);
   1305         return PyFloat_FromDouble(result);
   1306     }
   1307 
   1308     /* Else let libm handle it by itself. */
   1309     return math_1(arg, func, 0);
   1310 }
   1311 
   1312 static PyObject *
   1313 math_log(PyObject *self, PyObject *args)
   1314 {
   1315     PyObject *arg;
   1316     PyObject *base = NULL;
   1317     PyObject *num, *den;
   1318     PyObject *ans;
   1319 
   1320     if (!PyArg_UnpackTuple(args, "log", 1, 2, &arg, &base))
   1321         return NULL;
   1322 
   1323     num = loghelper(arg, m_log, "log");
   1324     if (num == NULL || base == NULL)
   1325         return num;
   1326 
   1327     den = loghelper(base, m_log, "log");
   1328     if (den == NULL) {
   1329         Py_DECREF(num);
   1330         return NULL;
   1331     }
   1332 
   1333     ans = PyNumber_Divide(num, den);
   1334     Py_DECREF(num);
   1335     Py_DECREF(den);
   1336     return ans;
   1337 }
   1338 
   1339 PyDoc_STRVAR(math_log_doc,
   1340 "log(x[, base])\n\n\
   1341 Return the logarithm of x to the given base.\n\
   1342 If the base not specified, returns the natural logarithm (base e) of x.");
   1343 
   1344 static PyObject *
   1345 math_log10(PyObject *self, PyObject *arg)
   1346 {
   1347     return loghelper(arg, m_log10, "log10");
   1348 }
   1349 
   1350 PyDoc_STRVAR(math_log10_doc,
   1351 "log10(x)\n\nReturn the base 10 logarithm of x.");
   1352 
   1353 static PyObject *
   1354 math_fmod(PyObject *self, PyObject *args)
   1355 {
   1356     PyObject *ox, *oy;
   1357     double r, x, y;
   1358     if (! PyArg_UnpackTuple(args, "fmod", 2, 2, &ox, &oy))
   1359         return NULL;
   1360     x = PyFloat_AsDouble(ox);
   1361     y = PyFloat_AsDouble(oy);
   1362     if ((x == -1.0 || y == -1.0) && PyErr_Occurred())
   1363         return NULL;
   1364     /* fmod(x, +/-Inf) returns x for finite x. */
   1365     if (Py_IS_INFINITY(y) && Py_IS_FINITE(x))
   1366         return PyFloat_FromDouble(x);
   1367     errno = 0;
   1368     PyFPE_START_PROTECT("in math_fmod", return 0);
   1369     r = fmod(x, y);
   1370     PyFPE_END_PROTECT(r);
   1371     if (Py_IS_NAN(r)) {
   1372         if (!Py_IS_NAN(x) && !Py_IS_NAN(y))
   1373             errno = EDOM;
   1374         else
   1375             errno = 0;
   1376     }
   1377     if (errno && is_error(r))
   1378         return NULL;
   1379     else
   1380         return PyFloat_FromDouble(r);
   1381 }
   1382 
   1383 PyDoc_STRVAR(math_fmod_doc,
   1384 "fmod(x, y)\n\nReturn fmod(x, y), according to platform C."
   1385 "  x % y may differ.");
   1386 
   1387 static PyObject *
   1388 math_hypot(PyObject *self, PyObject *args)
   1389 {
   1390     PyObject *ox, *oy;
   1391     double r, x, y;
   1392     if (! PyArg_UnpackTuple(args, "hypot", 2, 2, &ox, &oy))
   1393         return NULL;
   1394     x = PyFloat_AsDouble(ox);
   1395     y = PyFloat_AsDouble(oy);
   1396     if ((x == -1.0 || y == -1.0) && PyErr_Occurred())
   1397         return NULL;
   1398     /* hypot(x, +/-Inf) returns Inf, even if x is a NaN. */
   1399     if (Py_IS_INFINITY(x))
   1400         return PyFloat_FromDouble(fabs(x));
   1401     if (Py_IS_INFINITY(y))
   1402         return PyFloat_FromDouble(fabs(y));
   1403     errno = 0;
   1404     PyFPE_START_PROTECT("in math_hypot", return 0);
   1405     r = hypot(x, y);
   1406     PyFPE_END_PROTECT(r);
   1407     if (Py_IS_NAN(r)) {
   1408         if (!Py_IS_NAN(x) && !Py_IS_NAN(y))
   1409             errno = EDOM;
   1410         else
   1411             errno = 0;
   1412     }
   1413     else if (Py_IS_INFINITY(r)) {
   1414         if (Py_IS_FINITE(x) && Py_IS_FINITE(y))
   1415             errno = ERANGE;
   1416         else
   1417             errno = 0;
   1418     }
   1419     if (errno && is_error(r))
   1420         return NULL;
   1421     else
   1422         return PyFloat_FromDouble(r);
   1423 }
   1424 
   1425 PyDoc_STRVAR(math_hypot_doc,
   1426 "hypot(x, y)\n\nReturn the Euclidean distance, sqrt(x*x + y*y).");
   1427 
   1428 /* pow can't use math_2, but needs its own wrapper: the problem is
   1429    that an infinite result can arise either as a result of overflow
   1430    (in which case OverflowError should be raised) or as a result of
   1431    e.g. 0.**-5. (for which ValueError needs to be raised.)
   1432 */
   1433 
   1434 static PyObject *
   1435 math_pow(PyObject *self, PyObject *args)
   1436 {
   1437     PyObject *ox, *oy;
   1438     double r, x, y;
   1439     int odd_y;
   1440 
   1441     if (! PyArg_UnpackTuple(args, "pow", 2, 2, &ox, &oy))
   1442         return NULL;
   1443     x = PyFloat_AsDouble(ox);
   1444     y = PyFloat_AsDouble(oy);
   1445     if ((x == -1.0 || y == -1.0) && PyErr_Occurred())
   1446         return NULL;
   1447 
   1448     /* deal directly with IEEE specials, to cope with problems on various
   1449        platforms whose semantics don't exactly match C99 */
   1450     r = 0.; /* silence compiler warning */
   1451     if (!Py_IS_FINITE(x) || !Py_IS_FINITE(y)) {
   1452         errno = 0;
   1453         if (Py_IS_NAN(x))
   1454             r = y == 0. ? 1. : x; /* NaN**0 = 1 */
   1455         else if (Py_IS_NAN(y))
   1456             r = x == 1. ? 1. : y; /* 1**NaN = 1 */
   1457         else if (Py_IS_INFINITY(x)) {
   1458             odd_y = Py_IS_FINITE(y) && fmod(fabs(y), 2.0) == 1.0;
   1459             if (y > 0.)
   1460                 r = odd_y ? x : fabs(x);
   1461             else if (y == 0.)
   1462                 r = 1.;
   1463             else /* y < 0. */
   1464                 r = odd_y ? copysign(0., x) : 0.;
   1465         }
   1466         else if (Py_IS_INFINITY(y)) {
   1467             if (fabs(x) == 1.0)
   1468                 r = 1.;
   1469             else if (y > 0. && fabs(x) > 1.0)
   1470                 r = y;
   1471             else if (y < 0. && fabs(x) < 1.0) {
   1472                 r = -y; /* result is +inf */
   1473                 if (x == 0.) /* 0**-inf: divide-by-zero */
   1474                     errno = EDOM;
   1475             }
   1476             else
   1477                 r = 0.;
   1478         }
   1479     }
   1480     else {
   1481         /* let libm handle finite**finite */
   1482         errno = 0;
   1483         PyFPE_START_PROTECT("in math_pow", return 0);
   1484         r = pow(x, y);
   1485         PyFPE_END_PROTECT(r);
   1486         /* a NaN result should arise only from (-ve)**(finite
   1487            non-integer); in this case we want to raise ValueError. */
   1488         if (!Py_IS_FINITE(r)) {
   1489             if (Py_IS_NAN(r)) {
   1490                 errno = EDOM;
   1491             }
   1492             /*
   1493                an infinite result here arises either from:
   1494                (A) (+/-0.)**negative (-> divide-by-zero)
   1495                (B) overflow of x**y with x and y finite
   1496             */
   1497             else if (Py_IS_INFINITY(r)) {
   1498                 if (x == 0.)
   1499                     errno = EDOM;
   1500                 else
   1501                     errno = ERANGE;
   1502             }
   1503         }
   1504     }
   1505 
   1506     if (errno && is_error(r))
   1507         return NULL;
   1508     else
   1509         return PyFloat_FromDouble(r);
   1510 }
   1511 
   1512 PyDoc_STRVAR(math_pow_doc,
   1513 "pow(x, y)\n\nReturn x**y (x to the power of y).");
   1514 
   1515 static const double degToRad = Py_MATH_PI / 180.0;
   1516 static const double radToDeg = 180.0 / Py_MATH_PI;
   1517 
   1518 static PyObject *
   1519 math_degrees(PyObject *self, PyObject *arg)
   1520 {
   1521     double x = PyFloat_AsDouble(arg);
   1522     if (x == -1.0 && PyErr_Occurred())
   1523         return NULL;
   1524     return PyFloat_FromDouble(x * radToDeg);
   1525 }
   1526 
   1527 PyDoc_STRVAR(math_degrees_doc,
   1528 "degrees(x)\n\n\
   1529 Convert angle x from radians to degrees.");
   1530 
   1531 static PyObject *
   1532 math_radians(PyObject *self, PyObject *arg)
   1533 {
   1534     double x = PyFloat_AsDouble(arg);
   1535     if (x == -1.0 && PyErr_Occurred())
   1536         return NULL;
   1537     return PyFloat_FromDouble(x * degToRad);
   1538 }
   1539 
   1540 PyDoc_STRVAR(math_radians_doc,
   1541 "radians(x)\n\n\
   1542 Convert angle x from degrees to radians.");
   1543 
   1544 static PyObject *
   1545 math_isnan(PyObject *self, PyObject *arg)
   1546 {
   1547     double x = PyFloat_AsDouble(arg);
   1548     if (x == -1.0 && PyErr_Occurred())
   1549         return NULL;
   1550     return PyBool_FromLong((long)Py_IS_NAN(x));
   1551 }
   1552 
   1553 PyDoc_STRVAR(math_isnan_doc,
   1554 "isnan(x) -> bool\n\n\
   1555 Check if float x is not a number (NaN).");
   1556 
   1557 static PyObject *
   1558 math_isinf(PyObject *self, PyObject *arg)
   1559 {
   1560     double x = PyFloat_AsDouble(arg);
   1561     if (x == -1.0 && PyErr_Occurred())
   1562         return NULL;
   1563     return PyBool_FromLong((long)Py_IS_INFINITY(x));
   1564 }
   1565 
   1566 PyDoc_STRVAR(math_isinf_doc,
   1567 "isinf(x) -> bool\n\n\
   1568 Check if float x is infinite (positive or negative).");
   1569 
   1570 static PyMethodDef math_methods[] = {
   1571     {"acos",            math_acos,      METH_O,         math_acos_doc},
   1572     {"acosh",           math_acosh,     METH_O,         math_acosh_doc},
   1573     {"asin",            math_asin,      METH_O,         math_asin_doc},
   1574     {"asinh",           math_asinh,     METH_O,         math_asinh_doc},
   1575     {"atan",            math_atan,      METH_O,         math_atan_doc},
   1576     {"atan2",           math_atan2,     METH_VARARGS,   math_atan2_doc},
   1577     {"atanh",           math_atanh,     METH_O,         math_atanh_doc},
   1578     {"ceil",            math_ceil,      METH_O,         math_ceil_doc},
   1579     {"copysign",        math_copysign,  METH_VARARGS,   math_copysign_doc},
   1580     {"cos",             math_cos,       METH_O,         math_cos_doc},
   1581     {"cosh",            math_cosh,      METH_O,         math_cosh_doc},
   1582     {"degrees",         math_degrees,   METH_O,         math_degrees_doc},
   1583     {"erf",             math_erf,       METH_O,         math_erf_doc},
   1584     {"erfc",            math_erfc,      METH_O,         math_erfc_doc},
   1585     {"exp",             math_exp,       METH_O,         math_exp_doc},
   1586     {"expm1",           math_expm1,     METH_O,         math_expm1_doc},
   1587     {"fabs",            math_fabs,      METH_O,         math_fabs_doc},
   1588     {"factorial",       math_factorial, METH_O,         math_factorial_doc},
   1589     {"floor",           math_floor,     METH_O,         math_floor_doc},
   1590     {"fmod",            math_fmod,      METH_VARARGS,   math_fmod_doc},
   1591     {"frexp",           math_frexp,     METH_O,         math_frexp_doc},
   1592     {"fsum",            math_fsum,      METH_O,         math_fsum_doc},
   1593     {"gamma",           math_gamma,     METH_O,         math_gamma_doc},
   1594     {"hypot",           math_hypot,     METH_VARARGS,   math_hypot_doc},
   1595     {"isinf",           math_isinf,     METH_O,         math_isinf_doc},
   1596     {"isnan",           math_isnan,     METH_O,         math_isnan_doc},
   1597     {"ldexp",           math_ldexp,     METH_VARARGS,   math_ldexp_doc},
   1598     {"lgamma",          math_lgamma,    METH_O,         math_lgamma_doc},
   1599     {"log",             math_log,       METH_VARARGS,   math_log_doc},
   1600     {"log1p",           math_log1p,     METH_O,         math_log1p_doc},
   1601     {"log10",           math_log10,     METH_O,         math_log10_doc},
   1602     {"modf",            math_modf,      METH_O,         math_modf_doc},
   1603     {"pow",             math_pow,       METH_VARARGS,   math_pow_doc},
   1604     {"radians",         math_radians,   METH_O,         math_radians_doc},
   1605     {"sin",             math_sin,       METH_O,         math_sin_doc},
   1606     {"sinh",            math_sinh,      METH_O,         math_sinh_doc},
   1607     {"sqrt",            math_sqrt,      METH_O,         math_sqrt_doc},
   1608     {"tan",             math_tan,       METH_O,         math_tan_doc},
   1609     {"tanh",            math_tanh,      METH_O,         math_tanh_doc},
   1610     {"trunc",           math_trunc,     METH_O,         math_trunc_doc},
   1611     {NULL,              NULL}           /* sentinel */
   1612 };
   1613 
   1614 
   1615 PyDoc_STRVAR(module_doc,
   1616 "This module is always available.  It provides access to the\n"
   1617 "mathematical functions defined by the C standard.");
   1618 
   1619 PyMODINIT_FUNC
   1620 initmath(void)
   1621 {
   1622     PyObject *m;
   1623 
   1624     m = Py_InitModule3("math", math_methods, module_doc);
   1625     if (m == NULL)
   1626         goto finally;
   1627 
   1628     PyModule_AddObject(m, "pi", PyFloat_FromDouble(Py_MATH_PI));
   1629     PyModule_AddObject(m, "e", PyFloat_FromDouble(Py_MATH_E));
   1630 
   1631     finally:
   1632     return;
   1633 }
   1634