Skip to main content

Include/pymath.h

Source:

cpython 3.14 @ ab2d84fe1023/Include/pymath.h

pymath.h defines portable macros for IEEE 754 float classification and special values. It also declares _Py_dg_strtod and _Py_dg_dtoa, the David Gay float-to-string and string-to-float routines used for round-trip float formatting.

Map​

LinesSymbolRole
1-30Py_IS_NAN, Py_IS_INFINITY, Py_IS_FINITEIEEE 754 classification
31-50Py_HUGE_VAL, Py_NANSpecial float constants
51-80_Py_dg_strtod, _Py_dg_dtoaRound-trip float I/O

Reading​

Py_IS_NAN, Py_IS_INFINITY, Py_IS_FINITE​

// CPython: Include/pymath.h:28
#define Py_IS_NAN(X)      isnan(X)
#define Py_IS_INFINITY(X) isinf(X)
#define Py_IS_FINITE(X)   isfinite(X)

Wrappers around the C99 macros from <math.h>. Prior to C99 these required compiler-specific tricks; the wrappers hide that history.

_Py_dg_strtod and _Py_dg_dtoa​

David Gay's strtod/dtoa library (Python/dtoa.c) provides correctly-rounded float parsing and formatting. These functions are used by float.__repr__ and float(string) to ensure round-trip fidelity: float(repr(x)) == x for all finite floats.

// CPython: Include/pymath.h:62
extern double _Py_dg_strtod(const char *s, char **spe);
extern char *_Py_dg_dtoa(double dd, int mode, int ndigits,
                         int *decpt, int *sign, char **rve);

_Py_dg_dtoa with mode=0 returns the shortest decimal string that round-trips back to the same double; this is what repr(1.1) uses to produce '1.1' rather than '1.1000000000000001'.

Py_HUGE_VAL and Py_NAN​

// CPython: Include/pymath.h:40
#define Py_HUGE_VAL HUGE_VAL
#ifdef __has_builtin
#  if __has_builtin(__builtin_nan)
#    define Py_NAN __builtin_nan("")
#  endif
#endif

Py_NAN is the standard IEEE 754 quiet NaN. Used to initialize math.nan and float('nan').

gopy notes​

Py_IS_NAN / Py_IS_INFINITY / Py_IS_FINITE map to math.IsNaN, math.IsInf, and !math.IsNaN(x) && !math.IsInf(x, 0) in Go. Round-trip float formatting uses strconv.FormatFloat(x, 'r', -1, 64) in Go 1.15+, which produces the shortest round-trip representation equivalent to David Gay's mode=0.