Daniel James wrote:
Hello,
From the test results at:
It appears that std::numeric_limits<long double> might be defined incorrectly on OpenBSD on x86 (for example, the 1 + epsilon != 1 test is failing for long double). Is there someone with access to this operating system who can look into this?
On FreeBSD long doubles are defined as having the same accuracy as doubles, so maybe that's the case for OpenBSD as well? Another possibility is that OpenBSD doesn't have floating point functions defined for long doubles - but that wouldn't explain why the epsilon test is failing.
There are a couple of problems: and you are correct on both counts, there are no long double versions of the math functions (there are no *l versions, the std::func versions do have long double overloads but they call the double versions internally). And numeric_limits<long double> is wrong as well, see http://tinyurl.com/mof7j In fact numeric_limits<long double> has some really strange properties: it's min/max values correspond to an 80-bit long double with an extended exponent range, but the digits and epsilon values are those of a 64-bit double. Oh but LDBL_EPSILON and LDBL_MANT_DIG are correct! So go figure :-( BTW the math_info program results were designed to diagnose precisely these kinds of issues, HTH, John.