The 128-bit floating point division routines need to update the floating point hardware state when performing operations (e.g. division by NaN). This needs to account for the various architectures in the target agnostic code path (e.g. targeting Cortex M3 could not do this).
It seems like none of the fp128 routines set FE_INVALID, etc. This causes failures in the math tests of FreeBSD test suite for non-x86 architectures since it means that functions such as cosl/sinl/tanl/ etc. don't set any floating point exception flags.
https://reviews.llvm.org/D98332