AFAIK, accuracy of transcendental functions in libm, like

sin
cos
atan
exp
log
pow
tan
atan2

is completly undocumented, and do not conform to IEEE 784-2008 chapter
9.2, table 9.1. These are recommended ("optional"), and not mandatory. It
is understood that adhering fully to this chapter even for some of the
recommended functions, would come at the performance cost without proper
hardware support, or hardware tricks to make library implementations fast
(i.e. like libultim, or CRlibm).

However, documenting properly accuracy and rounding behaviour of glibc
provided mathematical functions would serve people very well. At the
moment, it is a complete blackbox, and depends on a target hardware too.

Best regards,
Witold