MATH_ERRNO, MATH_ERREXCEPT, math_errhandling
<cmath>
The macro constant math_errhandling expands to an expression of type int that is either equal to MATH_ERRNO, or equal to MATH_ERREXCEPT, or equal to their bitwise OR (MATH_ERRNO | MATH_ERREXCEPT).
The value of math_errhandling indicates the type of error handling that is performed by the floating-point operators and functions:
MATH_ERREXCEPT
Indicates that floating-point exceptions are used: at least FE_DIVBYZERO , FE_INVALID , and FE_OVERFLOW are defined in <cfenv>.
MATH_ERRNO
Indicates that floating-point operations use the variable errno to report errors.
If the implementation supports IEEE floating-point arithmetic (IEC 60559), math_errhandling & MATH_ERREXCEPT is required to be non-zero.
The following floating-point error conditions are recognized:
| Condition | Explanation | errno | Floating-point exception | Example |
|---|---|---|---|---|
| Domain error | The argument is outside the range in which the operation is mathematically defined (the description of each function lists the required domain errors) | EDOM | FE_INVALID | std::acos (2) |
| Pole error | The mathematical result of the function is exactly infinite or undefined | ERANGE | FE_DIVBYZERO | std::log (0.0), 1.0 / 0.0 |
| Range error due to overflow | The mathematical result is finite, but becomes infinite after rounding, or becomes the largest representable finite value after rounding down | ERANGE | FE_OVERFLOW | std::pow (DBL_MAX, 2) |
| Range error due to underflow | The result is non-zero, but becomes zero after rounding, or becomes subnormal with a loss of precision | ERANGE or unchanged (implementation-defined) | FE_UNDERFLOW or nothing (implementation-defined) | DBL_TRUE_MIN / 2 |
| Inexact result | The result has to be rounded to fit in the destination type | Unchanged | FE_INEXACT or nothing (unspecified) | std::sqrt (2), 1.0 / 10.0 |
[edit] Notes
Whether FE_INEXACT is raised by the mathematical library functions is unspecified in general, but may be explicitly specified in the description of the function (e.g. std::rint vs std::nearbyint ).
Before C++11, floating-point exceptions were not specified, EDOM was required for any domain error, ERANGE was required for overflows and implementation-defined for underflows.
[edit] Example
#include <cerrno> #include <cfenv> #include <cmath> #include <cstring> #include <iostream> // #pragma STDC FENV_ACCESS ON int main() { std::cout << "MATH_ERRNO is " << (math_errhandling & MATH_ERRNO ? "set" : "not set") << '\n' << "MATH_ERREXCEPT is " << (math_errhandling & MATH_ERREXCEPT ? "set" : "not set") << '\n'; std::feclearexcept (FE_ALL_EXCEPT ); errno = 0; std::cout << "log(0) = " << std::log (0) << '\n'; if (errno == ERANGE ) std::cout << "errno = ERANGE (" << std::strerror (errno) << ")\n"; if (std::fetestexcept (FE_DIVBYZERO )) std::cout << "FE_DIVBYZERO (pole error) reported\n"; }
Possible output:
MATH_ERRNO is set MATH_ERREXCEPT is set log(0) = -inf errno = ERANGE (Numerical result out of range) FE_DIVBYZERO (pole error) reported