When erf() is given a subnormal argument, it raises an underflow exception. However errno is not set. It should be set to ERANGE. Background: On error, many glibc math functions both set errno and raise an exception (fetestexcept(3)). For example, the following function all do this: acos(), asin(), cosh(), sinh(), acosh(), asinh(), exp(), exp2(), ldexp(), log(), log10(), log2(). However, there is much inconsistency. Some functions raise an exception, but don't set errno. Some functions set errno for some errors, but not others. A few set errno, but don't raise an exception. This series of bug reports documents deviations from what I consider the ideal: all functions should BOTH set errno AND raise an exception for all errors. All of these reports relate to tests on glibc 2.8 (as provided by SUSE 11.0).
Created attachment 2850 [details] test program Sample run showing problem: $ /tmp/mt_erf subnormal:20 errno == 0 fetestexcept() says: FE_UNDERFLOW FE_INEXACT erf(2.12199579096527232e-314)=2.39441584310907145e-314 0 FE_UNDERFLOW subnormal
Confirmed with current sources on both x86 and x86_64.
It's now documented that errno is only intended to be set to ERANGE for underflow to 0, not for underflow to subnormals, and erf can't underflow to 0 (the result for small x is about x * 2/sqrt(pi)).
IMHO there is a related bug in gcc. See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64101