https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117151
--- Comment #2 from vincenzo Innocente <vincenzo.innocente at cern dot ch> --- fine. I was trying to understand (reverse-engineering) what cuda and hip were using in device and host code... In case anybody is interested I think that cuda/nvcc uses their own implementation (in their own cuda::std namespace) on host and device (of course if one on host uses std namespace gcc libstd will be used) while hipcc seems to use gcc libstd on both host and device (on device the inline code is used) I used acosh on the negative axis to black-box test it as gcc libstd is the only one which implements the "kahan" formula (for which I haven't found any other reference)
