On Thu, 3 Jul 2025, Michael Matz wrote: > Yes. And then the above is multiplied by PI, passed to cos/sin and that > one then tries to figure out the multiple of PI (i.e. the 'x' above) again > via range reduction (not a _terribly_ slow one anymore in a good > implementation, because of the limited input range, but still). That's > backwards.
People are encouraged to contribute format-specific implementations of these functions to glibc (for example, from CORE-MATH), after adding inputs to the glibc benchmarks so it can be verified if new implementations are a performance improvement. Indeed, that's already been done for various functions for binary32 format. The point of type-generic implementations is not to be optimal for performance or accuracy but to get the API supported across all glibc configurations within a reasonable time (including formats such as IBM long double that probably no-one cares enough about to write specific implementations for now), so that further improvements can be made incrementally. -- Joseph S. Myers josmy...@redhat.com
